YouTube Transcripts

Is this laptop the reason you won't need
to pay for AI subscriptions anymore? I
got my hands on a brand new laptop from
Microsoft and Nvidia that claims to be
the first laptop built for local AI.
With 128 GB of unified memory, this
laptop can handle some seriously
powerful local AI models that are just
as good as frontier AI models. In this
video, I'm going to put the Microsoft
Surface Ultra with RTX Spark to the
test. I'm going to load it with a bunch
of local AI models to see what this
thing is capable of. In this video,
you're going to learn why local AI is
important, why you should be loading
these local AI models no matter what
device or hardware you're using, even if
you don't have this laptop, and let you
know if you can replace all your
expensive AI subscriptions with just a
laptop. Let's lock in and get into it.
So, this is the Microsoft Surface Ultra
with the Nvidia RTX Spark chip. It's
their new flagship laptop competing
directly with the MacBook Pro. Microsoft
sent this to me for free. This is not a
sponsored video. They just sent it to me
to use. They didn't even ask for
content, but I find this tech so cool
that I wanted to make a video on it.
What's really interesting about this
laptop is the first laptop ever with the
RTX Spark chip. What's the RTX Spark
chip, you ask? Well, it's basically the
same chip that's in the DGX Spark, a
very powerful super AI computer. Tons of
people, professionals, regular people
are using it to load super powerful AI
models. Now, the same chip that's in
that super computer is in this laptop.
That chip, much like the DGX Spark,
supports up to 128
GB of unified memory. That means you can
load some of the best local AI models on
planet Earth on a laptop that you can
take anywhere. You can bring it on a
plane and have basically frontier
intelligence running on the computer
without any sort of internet required.
Now, I'm going to put it through its
paces. I'm going to tell you all about
local AI a little bit first, why it's
important, why you should be doing it,
then I'm going to load some local models
in here, show you some workflows, show
you some things you can do. There are
tons of reviews out there of this
laptop. If you're looking to find out
about the USB ports and how many pixels
per inch rate in the display, this isn't
the video you. This review is purely
from an AI enthusiast perspective,
someone who works with AI 15 hours a
day, loads local AI models, has a ton of
DGX sparks. I'm going to be showing you
this laptop from that perspective. If
you're into AI, if you're into like deep
tech, this is going to be the review for
you. And you don't even need to be
considering buying this laptop to watch
this video. I'm going to educate you a
ton on local AI and why it's important.
I'll also show you how to load local AI
models like even if you're on like a
$200 computer, I'll show you how to load
them up. No matter what computer you
have, this is going to be a good video
for you. Speaking of controlling AI, I
have a secret to tell you. For the first
time in my life, I have complete control
over my inbox. I get hundreds of emails
a week and they are impossible to
manage. Many of these are important cuz
they're businesses and sponsorships and
I'm constantly losing track of them.
Now, I use the Superhuman Mail and the
Superhuman Docs connector in Claude to
keep on top of all these important
emails and keep them organized.
Superhuman is sponsoring this video and
they put out these two amazing
connectors. They're a part of the
Superhuman Productivity Suite and they
help me keep all of my emails organized.
All you need to do is go to customize,
then search for the connectors for
Superhuman, then you install each and
every one. You just sign in and you're
good to go. It's the Superhuman Mail,
Superhuman Docs connectors. From there,
you can go and start interacting with
them in Claude. So, I'm going to say,
"Use the Superhuman connector to find me
all my unresponded business emails."
And boom, just like that, it connected
to Superhuman Mail and got me all the
emails I missed. Then I can have Claude
draft responses these emails and send
them directly from this chat. Then I
have it pull every invoice from my inbox
into a Superhuman doc so I can build
this spreadsheet directly out of this
information. Puts it in a table with
vendor, amount, due date, keeps it
organized for me. And
[laughter and clears throat] boom, just
like that, there is the doc it made for
me in the table with everything super
organized. This saves me literally hours
of work I used to have to do. Does it in
a couple minutes. You need to try the
Superhuman Connectors today. Sign up
using my link in the description and
thanks again to Superhuman for
sponsoring this video. Now, let's first
talk about this chip. As I mentioned
earlier, up to 128 GB of memory, this
one I have here is 64 gigs. We're still
going to be able to load some crazy
awesome models on it. So, as I said
before, up to 128 GB of unified memory
so you can load amazing models. But, the
difference here is this is the RTX Spark
rather than the DGX Spark. What does
that mean? The DGX Spark was built to be
a Linux computer that you load local
models on, which has advantages to it.
But, this is the RTX Spark, which is
built more for Windows. That brings a
ton of advantages, including one that's
near and dear to my heart. You can play
basically your entire gaming collection
on this computer. It's basically the
equivalent to an RTX 5070 from a gaming
perspective, meaning you can play many
of the games out there on high. I've
tried Cyberpunk. I've played my personal
addiction, Slay the Spire, too, a ton.
I'll be playing World of Warcraft
forever next week. They all work super
well. You can basically play everything
on high, which is amazing. So, you can
not only run local models like you do on
a DGX Spark, but now you can do it side
by side with all your video games, all
your video editing, all your creative
workflows since it is a Windows
computer. Now, over the last couple
years, Apple has made a move in the
personal computer space to try to take
over local AI. They were starting to
build their chips specifically for local
AI. This is Microsoft's and Nvidia's
kind of response to that.
From a pricing perspective, they're
pretty similar. This is a 15-in laptop,
so it is bigger than the MacBook I am
comparing here. That comes in at 3,600
for the 64 GB. This is 4,200, so you are
getting a bigger laptop and screen for
that price. You're also getting a chip
specifically optimized for local AI
model workflows, which is great. So, RTX
Spark, you can load local models, you
can play video games, you can do all
your creative work, you can edit videos,
all that, all in one chip, all in one
nice sleek package. The hardware is
great. I'd say it's top tier. I'd say
it's for sure the best hardware I've
used with a Microsoft computer. It feels
very professional, all aluminum, all
very nice. But, let's get into the AI.
Let's get into the fun stuff. Why would
you care about local AI? Well, there's
many reasons you should be getting into
local AI, no matter what computer you're
on. First, it's free and unlimited AI,
right? Right now, when you use ChatGPT,
Claude, whatever, you're paying a
monthly subscription. This is free and
unlimited. Yes, the computer costs
money, but the actual AI running itself
just costs the power of the electricity
going into your computer. It's
completely private, all your prompts
stay on your computer. I mean, you
literally don't even need the internet.
None of the prompts leave your computer
at all. It's all personal and private.
It's customizable, so you can fine-tune,
train LoRAs, customize the AI models any
way you want. So, if you want it to talk
to you in a very specific way, you can
do that, which is amazing. It's
educational. I mean, I personally
believe that AI is the most important
technology of our entire lifetimes. And
so, loading local models, playing with
them, fine-tuning them, it's super
educational. You'll learn about the most
important technology of our lifetimes. I
think there's no better way to learn
about it than going hands-on with it
like you do with local AI. And it's just
straight-up fun. It is fun using local
AI. It is fun looking at your computer,
looking at your laptop, looking at
whatever, and saying, "Hey, there's an
AI model running on that." That is sick.
That is fun. And if you haven't done
yet, you don't know what feeling's like,
and you really should get into it. And
no internet needed. So, if you're on an
airplane and you want to get some work
done and talk to an AI and do little
coding, whatever it is you do with AI,
you can do it. You do not need internet.
It can run all on your laptop, which is
amazing and it is a benefit of the
laptop form. Now, I'll be jumping around
a ton here, so feel free to look at the
chapters below if you just want to get
straight to me loading local AI models
up. But, let's first talk about as well
how local AI works. When you use AI, so
you talk to ChatGPT or whatever, your
prompts that you're sending from your
computer go through the internet and go
to a data center. In this data center,
there is a crap ton of GPUs. The GPUs
get your prompt, they run a bunch of
math, a bunch of calculations, a bunch
of cross matrix multiplication, and they
come up with the answer to your prompt.
They then take that answer and send it
back to your computer. That's why it
costs more money. That's why there's a
longer delay. That's why you need the
internet because you need to communicate
back and forth between your computer and
this massive data center. But, with
local AI, it's different. Everything
that happens in this data center happens
on your computer. So, when you send a
prompt, all the calculations, all the
multiplication happens in the chip on
your computer. It never leaves. That's
why you don't need internet. That's why
you're not paying any sort of extra
subscription prices. It's all happening
on your computer. Never leaves. It's
personal and private. So, the GPUs they
have in this massive data center now
lives on your own computer. That's what
they're putting inside of this. The RTX
Spark is one of those really powerful
GPUs. That's why this is very important
and very cool. So, let's talk about
loading local models onto this laptop.
Is it super technical? Is it difficult
to do? Uh the answer is no. You
literally need no technical experience
to be able to load local models on here.
Let me show you exactly how to do it
completely for free.
So, what I have loaded up here on my
screen is Hermes Agent. What is Hermes
Agent? It is a free open-source AI agent
that runs locally on your computer. You
can run it on anything, any computer you
have. Works on all computers. And what's
great is it is built with local AI in
mind. Yes, you can connect this to your
different subscriptions and have it run
through ChatGPT or Claude or Grok or
whatever, but what's great about it is
it makes it super easy to load local AI
models. So, I highly encourage you to
follow along side me here. Even if you
don't have a Microsoft Surface Ultra,
you should be doing this doesn't matter
what computer you have. You can do
exactly what I'm about to show you here
and load local models up. So, download
the Hermes desktop application. Again,
completely free. I'll put a link for it
down below. Not sponsored by them,
either. Once you have it loaded up, you
can at first, if you want, connect it to
one of your existing subscriptions just
to get to this screen faster. But, once
we have it loaded up, let's load local
models into it. Here's how you do that.
If you go into the settings in the top
right, you click that, then you go to
providers right there, and then we move
over here, and we click on run models
locally. What's going to happen is it is
going to show us a bunch of models that
will work on your computer. If this is
your first time doing it, what you
should see here is actually just one big
model that says, "Hey, this is your
recommended model." Because I've done
this before, it shows me a bunch of
other models I could load into it. For
me, it's recommending Qwen 38 27B, which
is a fantastic model. This is probably
the model most of you go with, depending
on which computer you have. Even if you
have like a dirt cheap computer, you can
load like a local model on it. So, from
there, you can just click download on
the model it recommends. For me, it's
recommending again Qwen 38. I click
download, it loads it into memory. This
might take a while, depending on the
size of the model. Some of these are 10
GB in size, some of these are hundreds
of GB in size. Depending on the size,
it'll download it, might take a little
bit. Once it's downloaded, loaded in,
come back here, let's get this going.
So, you want to create a new session.
Once you're in there, go to the bottom
right, make sure that model you
downloaded is listed. There's Qwen 38
for me. Send a message to your agent,
just say, "Hey, you there?" like I did
here. What that will do is trigger it to
load it into your memory. Once it's
loaded into your memory, it's good to
go. You can start talking to it. So, I
can say what's your name? I hit enter on
that. Let's see how long this takes. It
should be pretty quick. I'm Hermes, an
agent built by new research. As you can
see, that was super quick. That was 5
seconds total. On this laptop, you are
getting AI supercomputer level speed. I
was able to get that response in 5
seconds. Depending on your computer, the
memory, all that, might take longer,
might take shorter. For me, that's
pretty good. Now, let's take it up a
notch. Let's make it a little bit more
complicated. Build me a simple 2D
side-scrolling
hack-and-slash
game. And I'm going to hit enter on
that. I'm a gamer. A lot of the examples
you're going to see from me in my videos
are me building games cuz I absolutely
love games. If you're a gamer like me,
this is a good laptop for you because
you can basically play any games you
have on here. Works really, really well.
Let's see how well and how fast it can
build this game for us. And if you're
new to local AI 2, things like that, I
think building games is an awesome kind
of first project cuz it's super
creative. It gets your creative juices
flowing. It's a lot of fun. You don't
even need to come up with like ideas.
You can just like talk to the agent and
say, "Hey, what should we build? What
game can we build?" And it'll walk you
through it. So, it's super easy, super
fun, super simple to do. As this is
going, by the way, something I'm
surprised with is it's not warm at all.
It's like actually cool to the touch. I
can hear the fan spinning a little bit,
but it's not too bad at all. For a
laptop, this is rather impressive for
running local AI. I didn't think we
would ever have laptops running local
AI, but here we are, and it seems to be
working pretty well.
All right, so looks like it is done. It
took about 20 minutes for it to build
this app. So, not the quickest, but it
is a full game, so it's also not the
slowest. What's amazing, though, is it
was able to build the game out, test
itself, wrote all the code, tested all
the code, played it itself, and then
even put the game in the chat for us to
play as well. So, let's check it out.
So, as you can see, we can move around.
There's full graphics it built and
designed. There's like gems you can
collect, there's experience points. It's
actually a pretty playable game and it's
not bad for one shot for a local AI
model. Again, none of this left our
computer, went to internet, all the
calculations, all the math, all the AI
stuff was done locally on this laptop,
which is really amazing.
So, what are some amazing use cases you
can be doing local AI other than
building games and writing code and all
that cool stuff? Well, it can manage
your entire computer for you. One
awesome thing I did was I saw that Halo
3 had a new mod where you can play with
up to 128 players online. I went to my
local model. I said, "Hey, there's this
new mod I keep reading about with Halo
3. Can you figure out what's going on
and get it set up and installed for me?"
And all by itself, this local model went
in, downloaded Halo 3 onto my computer
from Steam, found the mod, which was
Project Reclaimer, installed it, set it
up, and boom, good to go. I did zero
work in this mod that makes Halo 3
really cool just worked. It set up
everything for me. It's my own personal
computer administrator. Note, usually
when you install mods, there's like a
lot of technical work and things like
that. I didn't have to do any of it. The
local agent did it all for me. So,
anything you want to do on your computer
now, having local intelligence is a
really amazing way to do it cuz it's not
going to cost you any money. The agent
lives on your computer and just does all
the work for you. It is always great to
kind of have that local assistant ready
to go whenever you want living on your
computer. Another use case you can do is
what I call ambient intelligence. What's
ambient intelligence? It's intelligence
that can just do things throughout the
day for you. Here's an example. Can you,
every hour, go on this computer and do a
security check to make sure everything
is safe? And I'm going to hit enter and
now it's going to set up a security
check where this local model is just
going to check everything over every
hour to make sure it is secure and
everything is safe and there's no
viruses and things like that. It's
ambient, so it can check every hour,
every half hour, and it costs us zero.
It just goes and checks for security
patches and vulnerabilities and things
like that. If you do this with a cloud
model, doing something like this every
hour is going to cost you money. You're
going to use your usage very quick. But
because this is local, you have
unlimited usage of it. So, ambient use
cases, things where your AI does things
over and over and over again, really
good use case for local AI, really good
use case for Hermes agent cuz Hermes
agent has built-in routines where it can
do things on a scheduled regular basis
for you. So, this is a really good one.
Another example could be having your
Hermes check your email every so often,
see if any important emails, and then
pinging you or messaging you when you
do. That's ambient intelligence. That's
something that's really, really good for
local AI.
Now, there's many other things this
computer can do. You can do editing of
videos, things like that, playing video
games. My channel, my content is all
around AI from an AI enthusiast
perspective, from a perspective of
someone who wants to maximize their
efficiency with AI and push technology
to the limit. If that's something that
you're interested in, make sure to hit
like and subscribe and turn on
notifications down below because all I
do is make amazing videos on AI. I don't
think there's many other videos
reviewing hardware from an AI enthusiast
perspective. So, I hope videos like this
are helpful. Now, can I recommend the
RTX laptop? It is a great laptop. It is
incredible hardware. It is very high
quality. All the materials are very
nice. But so, you are paying a premium
for the laptop format, right? Like you
pay extra money to get a laptop rather
than a desktop. Is it worth the extra
money to be able to bring AI around
wherever you go? Well, an alternative
would be buying a desktop version of
this. You can buy an RTX Spark desktop
or a DGX Spark desktop, run your local
intelligence there, and then connect it
from your other laptops and equipment.
That's an option if you're looking to
save money, and that is a great way to
do this as well. But, if you're looking
to bring the power everywhere you go,
bring local AI wherever you go, this is
a really, really good option. I also
really like this cuz I play video games
on the go. I'm going to play World of
Warcraft in bed next week, and there's
nothing you can do to tell me otherwise.
That is amazing. That is a really cool
use case. If you purely care about local
AI and a laptop version doesn't really
appeal to you, go with a DGX or go with
an RTX park. Both are great. Then you
can connect your laptop from there if
you want to use a local intelligence or
from any of your other computers, even
your phone you can connect to it and use
the local intelligence. It's up to you.
But, if you want power on the go, this
is a great option. Again, like and
subscribe if you found this helpful at
all. If you want more videos like this,
I don't do a ton of hardware reviews. If
you like this hardware review, I'm happy
to do way more of that. Let me know down
below in the comments as well. Happy to
review everything from AI enthusiast
perspective. Hope this is helpful. See
you in the next

Work mode is easily the most slept on
feature on Chad GBT right now because
instead of just answering your
questions, it goes and does the boring
tasks for you, like building
spreadsheets, making decks, and
researching your competitors. And in
this video, I'll show you how to set it
up from scratch and use it on real world
tasks you'd actually use every day. But
before we set it up, let's look at what
actually makes work mode different from
the normal chat. In the normal chat
window, you ask JGBT a question and it
gives you an answer back. It can write,
search the web, do some light coding,
and make images, but every reply ends
there, and nothing else happens until
you send the next message. Work is
different because it's built to go and
finish a task instead of just telling
you how. It's what's called an AI agent,
which just means you hand it a goal, and
it makes its own plan and keeps working
through each step until the job is done.
It runs on the same tech as codeex,
which is OpenAI's coding agent, just
made for everyday work. So, in the CHGBT
desktop app, I'll switch the toggle at
the top from chat to work. Next, I'll
click the model picker next to the
microphone and pick GBT 6.1 Soul because
it handles bigger tasks at a lower cost
than Astra. That's worth knowing because
a long work task uses up your limits
faster than a normal chat message.
There's also a small icon at the bottom
right of the box where you choose if it
runs on your computer or in the cloud.
I'll pick on your computer because I
want it to work with all the files on my
Mac. Next, I'll click the plus next to
projects in the sidebar. This is the
step that matters most because the
folder you add here is what work can
open and edit. So I'll give the project
a name. Click add under source folders.
Pick my folder and hit create project.
The last thing is the plugins icon in
the left sidebar which is where you
connect the apps you already use like
Gmail, Google Drive and Google Calendar.
A plug-in can connect a tool, add a
skill or both. And once it's installed,
you can use it in any task by typing the
add sign. So that's work set up with its
own project folder and the apps you've
connected. But right now, that project
is empty, so there's nothing in it for
work to use yet. And the messiest thing
I can put in there right now is a full
year of bank statement exports. So, I'll
click the plus button next to projects.
Name this one money and add the folder
with all the exports as the source.
Inside that folder are a bunch of files
with my expenses, ranging from
subscriptions to rent. And for the first
task, I'll write a prompt that asks it
to combine all of the data into one
spreadsheet. then sort every transaction
into a category and list all the
recurring charges one by one. I'm also
telling it to use formulas for every
total so I can check any number myself.
And if it can't tell what the charge is,
it has to mark it as unknown instead of
guessing. Oh, and before we continue, if
you want to go from casually using CHP
to actually getting work like this done
with it, that's exactly what my AI
fluency community is built for. There's
a new video every day for three months,
plus over a thousand ready-m made
prompts and agent workflows in the links
down in the description. Anyway, once I
hit enter, it works through the task for
a while and gives me back a spreadsheet
with everything sorted out. Next, I want
to be able to check all of this from
anywhere in a nice dashboard because
spreadsheet is a pain to read every
time. For that, I'll type the add
symbol, pick the sides plugin, and ask
it to turn the spreadsheet into a
private dashboard I can open on my phone
or in any browser. I'm also laying out
how I want it to look and the colors it
should go for. A few minutes later, it
builds the whole website in one go, and
it looks pretty good. At the top, I can
see how much money came into my account
and how much went out. Then there's a
chart of my spending by month. And if I
click on any bar, the total for that
month shows up right below it. Below
that are categories I've spent my money
on, and I can scroll from housing all
the way down to bank fees. And at the
end, there's a list of 16 services for
me to review and cancel. This is by far
one of the most useful things I've used
it for because everyone deals with
expenses like this. So, I want JGBT to
refresh the dashboard on the first of
every month. I'll just type in this
prompt that tells it to read any new
files I add to this folder and update
everything based on them. So, from now
on, I don't have to worry about sorting
out my bank statements because JGBT does
that for me every month. Next, I also
wanted to see how it handles creative
work like turning an ad script into a
full storyboard. So, to test that, I'll
make a new project and add a folder with
some product photos and a few ad scripts
for Helden, which is a cold brew coffee
brand I made up for this video. The
scripts are all saved in that folder, so
I'll just point it at the 3 p.m. slam
script and ask for a storyboard PDF with
the woman, the desk, and the coffee can
matching in every frame. After a few
minutes, it hands me back the whole
thing. So far, the storyboard looks
really promising, but I can see that
it's not perfect, and some images have a
few small issues. First, these two shots
are a little different, and I think the
continuity between them could be better.
In the first one, the Mac, the laptop,
and the can are in one spot on the desk,
and in the next one, they've moved
around. This is easy to miss in a steel
frame, but once it's a video, props
jumping around between cuts is something
people notice right away. I'll
screenshot them, paste them in, and
explain to JGBT what's wrong. After
that, it's fixed, and it looks a lot
better. Now, the Mac stays on the left,
the laptop is in the middle, and the can
is on the right in both shots. So her
hand reaches across the same desk. But
before moving on, there is one more
thing I want to change. After this
frame, I want another one where the
character holds the coffee can for a
little bit before placing it on the
table. Right now, she goes straight from
the fridge to opening it. And I want a
moment where she actually looks at the
can because that's when the viewer gets
a good look at it, too. I'll do the same
thing here. Screenshot it, paste it in,
and ask for exactly that. And shortly
after, I've got what I wanted. It adds
the new frame and renumbers everything
after it. So, the storyboard goes from
seven frames to eight. It also laid out
a voice over and how each shot should be
filmed based on the script, and I really
like that. So, after two quick rounds,
it went from a decent first try to a
storyboard I could actually hand to
someone to shoot. And for a brand like
this, I'd have to make storyboards all
the time. And writing the same prompt
for each one is a waste of time. So, to
fix that, I'll tell CHBT to turn
everything we just did into a skill
called storyboard maker, including both
of my correction notes. While it's
working on that, I'll quickly explain
what the skill is. It's basically a safe
set of instructions JGBT can follow.
Usually when you need to get a task like
this done, you end up explaining the
process over and over and that wastes a
lot of time. But for repetitive work, we
can turn it into a skill once and let
Chip follow it every single time without
starting from scratch. And to test this,
I'll open a new chat in the same folder.
Tag the skill and point it to the
rooftop script. It comes out looking
good, but it still needs a little bit of
work. This time, instead of pasting
screenshots, I'll click right on the
spot in each image and leave a comment
there. On this one, the hand holding the
can looks fake because the grip is
glitchy and it doesn't look natural. So,
I'll tell you to fix that. In the next
one, the people don't line up with the
shot before it. So, the empty space
doesn't make sense. And here, the watch
is facing the camera with the numbers
flipped when it should be shot over his
shoulder like we're checking the time
with him. I'll send all of these
comments together, and it fixes them in
one go and updates the whole storyboard.
So, now I have an automated process that
turns my ad scripts into storyboards
with just one line. For the next one, I
wanted to see if it could do actual
research for me, like checking what ads
my competitors are running right now.
So, I'll make a new project, add a
Helden folder with the brand sheet and
the product photos, and write a prompt
that sends it to the meta ad library.
I'm asking it to use the inapp browser
and run a separate agent for each of
three real coffee brands, so they all
work at the same time. Each agent
screenshots the current ads and writes
down the hook, the angle, and the format
of each one. Then it puts everything
into one table and tells me which angles
all three brands use and which ones
nobody is using. Once I hit enter, the
browser opens up on the right and the
agents start going through the ad
library on their own. About 10 minutes
later, I've got a folder of screenshots
for each brand and a table with all of
them in. Next, I'll ask you to read the
brand sheet and use what it found in the
competitor's ads to make five ad
concepts for Hayden with the can photos
as a reference so the can looks the same
everywhere. When it's done, it hands me
a deck with the competitor's table up
front and a slide for each concept with
image, the angle, and the copy. So far,
it looks very nice, but honestly, the
cans look really artificial. I like it
to make them look realistic with a
metallic aluminum fill and the
typography and design printed properly
on them. And while it's working, I also
noticed the text doesn't really wrap
around the can. So, I mentioned that
too, and it fixes it from there. In the
end, they are super nice with every can
finally looking like an actual product.
And all of these are ready to post on
social media right away. And through
this whole video, I didn't have to do
much of the work myself. I handed the
task to Chajib. It took care of
everything and all I did was review its
task a little bit. And reviewing an AI's
work is actually the hard part because
you need to know what to fix and how to
ask for it. Inside AI Fluency, you
rebuilt real projects like this yourself
so you get good at catching those fixes
on your own. So join through the link in
the description before the price goes up
again. Thanks for watching and I'll see
you in the next

Welcome to Tiny Software Factories
lesson one. Tiny Software Factories is a
crash course mini-series where we build
small useful software factories that
work for you around the clock so you can
move from being stuck in the loop to
being near the loop. In each lesson,
you'll see an empty factory and a
running factory. This lets you
understand the capabilities first and
then you can apply them to a real
problem. In lesson one, you'll dive into
the world of Tiny Software Factories and
learn agentic engineering done wrong.
Now, I know it's a little weird. Why
would you want to learn agentic
engineering incorrectly? Let me explain.
Tiny Software Factories is a crash
course on incrementally scaling your
agentic engineering to the closest
[music] useful threshold. In TSF, you're
going to learn many dark patterns of
agentic engineering that will not scale.
In exchange, you'll move fast, break
things, fix them, and ship. But, these
are anti-patterns. So, I want to make
this super clear. Tiny Software
Factories is a highly compressed lossy
version of the infinite agentic
engineering core lesson set. Here, we
focus on near-loop agentic engineering.
[music] IA3 focuses on the much
higher-leverage but more challenging
out-loop agentic engineering where we
build fully autonomous systems that
operate for you without you. What you'll
see here is not recommended as long-term
solutions. These anti-patterns are meant
to move you from zero to value at the
cost of basically everything except
getting the job done. This is
blitzscaling for agentic engineering.
[music] This is the quick fix. It's not
meditation, it's Xanax. It's not working
out, it's a GLP-1. This is a useful
local maximum for getting results at all
costs. Now that that's out of the way,
just because it isn't technically
correct doesn't mean your tiny software
factories can't be incredibly useful. In
TSF, [music] we stop worrying about
getting things right and we focus on
getting things shipped. If these ideas
[music] interest you and you want to get
things done at the speed of agents,
let's build some tiny software [music]
factories.
So, let's start by asking an obvious
question. What is a tiny [music]
software factory? A tiny software
factory is a software factory built to
solve one specific problem where a
proactive agent works around the clock
in a durable environment with you near
the loop instead of in it. There are key
pieces of this definition we need to
highlight. Built to solve one problem
with a proactive agent around the clock
in a durable environment with you near
the loop, not in the loop. You're going
to see these properties as repeat themes
as you master the tiny software factory.
Now that we're aligned on a definition,
let's ask the next important question.
Why is this valuable? What makes a tiny
software factory more valuable than
sitting in the terminal and prompting an
agent back and forth? In loop versus
near loop. The big advantage of your
tiny software factory is that you are
not required in the loop. Instead, you
move near the loop to give feedback,
make decisions, and approve work your
agent has already done. The feedback you
give your near loop agent creates a data
flywheel where the factory improves with
each signal you give it. Let's face it,
the facts are you and I are the
bottleneck, and 80% or more of the work
we're doing right now can run without us
with the right system in place. The tiny
software factory gets us there fast. In
loop, you prompt. Near loop, you
respond. The trick is having the right
system in place so you aren't stuck with
your bot in your dev chair. What else
makes a tiny software factory so
valuable? As you'll see, the tiny
software factory is rapidly deployable.
It's designed to be changed. With one
meta skill and one command, bam, you can
mount a new factory next to your code
base to solve a specific problem. Point
it at your repo, drop it in a sandbox,
and you go from zero to working agentic
system in minutes. Building a scalable
software factory takes a ton of time
even with powerful language models. The
result is always one more feature away,
and then you realize it doesn't do
everything for every scenario you're
trying to build for. The tiny software
factory gives you a massive advantage
because it's a highly compressed working
system with great defaults you can
rapidly deploy over and over. It's a
quick fix, and that's the point. What
else makes the tiny software factory
architecture you'll see so valuable?
It's extensible by design. It's designed
to be customized. Every prompt, tool,
hot key, extension, theme is all
accessible to you and your in-loop
agent. This means you can rapidly build
the system that builds the system,
speeding up your transition from in the
loop to near the loop. You can change it
so it bends to solve whatever problem
you need to solve. Initialize,
configure, specialize. Last but not
least, the tiny software factory
architecture is special because the
near-loop agents inside them don't stop
working for you. They contain a durable
agent plus a trigger inside a sandbox to
give you a proactive agent. You can
configure it so it starts working before
you wake up, and when you check in, the
work is already waiting for your review.
It doesn't wait for you to work, it
works for you, then waits. We have a
rapidly deployable agentic system you
can spin up at a moment's notice to
consistently solve one specific problem
around the clock with a proactive agent.
But, what [music] does a tiny software
factory actually look like?
Throughout tiny software factories,
we'll incrementally reveal features for
the tiny software factory using a
feature map. We're going to work through
four factory levels, and each factory
level contains unique features we'll
work through incrementally one by one.
Every feature sits on one of three
dimensions: effort required to unlock,
the value you get for putting in the
effort to build and use that feature.
And then every feature also has a
certain skill requirement that you must
meet. To keep this simple in the true
spirit of tiny software factory, we'll
[music] focus on value and effort.
The first and most important property of
tiny software factories is the empty
factory. The empty factory is the
template. It's fully built out. It has
every feature you need to specialize,
but no specialization yet. The empty
factory has no purpose until you give it
one. Let's take a look at a real empty
factory right now. In the browser, you
can see we have an empty factory here.
It's a minimal user interface with
[music] two sidebars we're going to work
through right now. Every feature we work
through is a capability you have [music]
that you can specialize to solve your
specific problems.
Let's look at what's inside the empty
factory with our next feature, the agent
on your right. Here, this is a durable
agent inside your factory that's always
on, always one message away. If we open
up the right sidebar here and right
ping, you can see we have a classic
agent ready to work for you. This is
something you're very familiar with.
Every application can be supercharged
with an agent so you move at the agentic
speed. Every software factory must
operate on something. What does this
empty factory operate on?
That brings us to our third feature. We
have lists on your left. Every product
is a list of something. What's in your
list? On the left here, we have pages.
So these are pages that you and your
agent operate on. As you'll see, this is
designed to be replaced. But with just
these three features, we can start
operating our tiny software factory.
Explain the basics of Python,
TypeScript, and Rust nested folders in
our pages. We'll kick that off. So you
can see Python, TypeScript, Rust in
folders in the tiny software factory.
It's creating these pages. It's
populating the tiny software factory.
There are pages we can go ahead and open
up every one of these and we can see
them. We have a start here and you can
see that our center view [music] has now
focused on the content.
That brings us to our next feature.
Focus is in the center. The primary
thing you and your agent work on is
located in the center of your tiny
software factory. Focus your attention
here. So for our empty software factory,
we have a simple generative UI based
application where we have pages, we have
artifacts, we have simple markdown
documents that our agent can operate,
build, tweak, and tune on our behalf.
You can see here we have the language
basics for Python, for TypeScript, and
for Rust generated by our agent and it's
all here for us to analyze, to look at,
and to tune and [music] tweak. But these
two calls our agent has just run hints
at a very important property of the tiny
software [music] factories.
The next key feature of the tiny
software factory is agent access parity.
Everything you can do, your agent can
do. Every control in the UI has a
matching agent tool. This has serious
implications that let you move much
faster than your competitors because
this allows you to move at the agentic
speed, the speed of agents. Let's focus
in on our agent here. List pages tools.
And our agent is listing all the tools
it has here for operating its pages. Our
agent truly has access to everything we
have inside [music] the tiny software
factory. Let's look at another core
feature and I'll show you exactly what I
mean.
You do not truly own your agent if you
do not have core four control. Context
model, prompt, tools. A true software
factory must give you full control over
your agents. In our user interface here,
we have that exactly. If we go to the
settings here, you can see we have a
comprehensive list of settings. We have
the essentials, we have the core four,
and then we have our advanced settings.
Let's talk about the core four, because
this is what makes up your agent. If you
master the core four, you'll master your
agent. And if you can configure and
customize it, you can build agents no
one else can. And in your tiny software
factory, that means everything. You can
see here, we have a context bar, we have
auto compaction at 200k, force
compaction at 500k. Once our agent
crosses the 200k threshold, it'll start
compacting in the background. We'll talk
about durable agents more in a moment,
but this is a key feature of them. You
can see we have model selection and
thinking level broken down into tiers.
This is very important. You don't want
to fix it on one model, you want to use
the right model for the job at the right
performance, cost, speed trade-off. We
have a system prompt that we have prompt
engineered. It does very specific things
inside the software factory. And then of
course, we have our tools. You can see
all the tools our agent has there. And
then we have some settings. So now we
have a couple key features stacking up
here. And we can showcase this. If we go
into the settings for the core four, you
can see context focus there. We have
auto compaction on. We can see our model
details. We can see all of our prompts
broken down, custom prompts, extension
prompts, which we'll get to, and custom
prompts that we can append to our agent
at any time. And of course, all the
tools our agents have access to, we can
see them in quick view mode here. We can
show the description of each, and then
we can show the full parameters in
whatever syntax we like to ingest our
tools in. But our features are starting
to stack up now. So let me show you what
it really means for your agent to have
agentic access parity. I'll close our
main view, focus on our agent, and then
I'll run disable all coding tools,
update thinking level to max. I'll kick
this off, and our agent can modify
itself. This is one of the key
advantages of having a durable agent
with the right tools. This is reflection
for agents. If you open our settings
now, you can see that our thinking is
now set to high. And if we go to our
tools, you can see that all of the
coding tools, read, write, edit, bash,
have now been disabled. Our agent has a
specific set of tools that we can tweak
and it can self-modify itself up to some
useful limit that again you can set.
Obviously, this capability can be very
dangerous if you don't control it
properly, but agentic access parity lets
your agent do everything you can do and
the most important customizations any
agentic engineer can make now is the
configuration of your agents and the
code surrounding your agents. And if you
can do that at the agentic speed, you'll
move a lot faster. We can [music]
customize the core four with agentic
access parity.
That leads us to another big feature
that the tiny software factory has that
you can tap into. You can build an
extensible harness. This tiny software
factory near loop agent is built on
slices of capability. What do I mean by
that? If we hop back into the settings
menu and we go down to our advanced
settings, inside of harness, you can see
that this agent is a composition of
several agent harnesses. We have the
factory harness. We have our coding
tools, which you can see here we've
fully disabled. We then have our page
harness. We have our image generation
harness, which is another layer of
capability, and then we have our factory
controls. Excluding the default coding
agent settings, we have four additional
layers we've stacked on top of this. And
this is fully customizable as you'll see
in subsequent lessons. We'll showcase
how to harness engineer new capabilities
inside your tiny software factory near
loop agent. Tools that feel great to
operate make for a better [music]
experience. And better experiences means
you're more willing to invest in the
tools you're building and using.
Which is why our next feature of our
tiny software factory is so important,
customizable DX UX. You can customize
your tiny software factory to the core.
We're talking themes, we're talking
updating the nav bars, we're talking hot
keys, animations. Shape your factory to
how you do your work. As you've likely
noticed, I've been opening the settings
with another hot key. If you want to
move fast, the tools you use must enable
you to do that. And the Tiny Software
Factory was built with this in mind.
What do I mean by that? Let's keep
extending on our theme here. So, I'll
run this. I'll say update to another
dark theme, then add context bar to the
agent nav bar, then add status in the
global nav bar on the left side. We'll
let our agent kick off, and it's going
to check the themes. It has full agent
access, and it's updated our settings.
So, couple things have changed. We have
our agent status in the nav bar. you saw
we updated our theme. So, once again, we
have agentic access to everything we can
do in the UI. This means we can move at
the speed of agents. And I mean full
agentic access parity, even simple
things like this. On select the page.
Everything you can do, your agent can do
as well. It's the small operations that
you take, that I take, that stack up big
over time. This context bar is a little
scrunched here. Move context bar into
full width mode. This is going to give
us more space. It's going to push it
down. There we go. You can take this.
You can customize it. You can really do
whatever you want with it. You can make
it your own. We're not going to use that
much context, so we're going to say drop
context bar. I don't quite like this
theme, so we're going to hop back into
our theme here in the settings, and I'm
just going to update this. We're going
to switch to a midnight. We can continue
to customize to our heart's content. Of
course, in the nav bar, you can see
exactly how our agent is organizing the
user interface. Fully customizable,
fully extendable. We can update the
animation so we can understand when our
agent is working versus when it's not.
And then we have hot keys. fully
configure, customize the application to
move very, very quickly. In a world
where software can change at light
speed, our core tooling, our core
applications must be built with that in
mind. This is liquid software. It's
meant to be changed. It's meant to be
customized. [music] That's a key
property of the Tiny Software Factory.
Switch back to default TSF theme. What
else makes the Tiny Software Factory
[music] so powerful?
If you want to move near the loop and
eventually out the loop, you're going to
need cron, and specifically cron
prompts. These are exactly what you
think they are. It's prompts on a
schedule. It's run by your agent while
you're away. Let's take a look at this
because this is a massive up level for
your agentic engineering because with
the tiny software factory, this idea is
rapidly deployable against any problem
you have. So, out of the box, the hot
key for cron is R. You can see here,
every cron has a simple set of items,
right? It has a name, it has a schedule,
and then it has a prompt. That's really
it. This lets you get up and running
with repeat schedules, repeat jobs
instantly with zero cost. Let's create a
cron and really start getting value out
of our near loop agent here. I'll say
add cron, every minute add a new file to
each of our programming languages
because remember we had it write out
some language basics here. And then I'll
say breakdown one new concept each.
We'll kick that off. Let's see what our
agent thinks about that. Two blockers,
has no file rights. That's fine. It
shouldn't need this. Every 3 minutes.
That's a lot per day. That's fine.
Confirm anyway. Use page tools. TSF
cron. There it goes. There it is. It's
kicking off that workflow every minute.
It's going to read the pages and it's
going to add a new fundamental into our
folders for every one of our languages.
Obviously, this is just a random example
to showcase what this really looks like,
what this really feels like inside your
tiny software factory. But you can see
here, we got context managers for
Python, classic width statements,
context manager decorator looks great.
TypeScript utility types looks great. We
got why it matters, common mistakes at
the bottom. Awesome. And then Rust
lifetimes, label of borrow, tell
compiler how long it's valid, why it
matters, common mistakes. Fantastic.
Here we are. We're getting to that point
where minute's almost complete and now
our agent is going [music] to get that
event once again. There it is. The cron
is running and it doesn't matter if
we're here. This is the important part.
I can close this. This is a durable
application. I can refresh as much as I
want. I can close this browser window if
I want and our agent is just going to
keep working, keep stacking up features.
Cron prompts are an incredibly
important, powerful, and fast, simple
way to add repeat jobs into your tiny
software factory. You're moving fast.
You're getting done, and now
[music] with this next feature, you're
turning your empty software factory into
a
>> [music]

Now, this might be the coolest thing you
see today. This panda right here is
climbing a stone pillar, and the whole
world is being made by AI typed [music]
one sentence at a time. This is H Highus
from Tencent. At least I think I'm
saying that right. And it's basically
their version of Google's Genie 4, which
I actually just showed you a couple days
ago, except this one is open source with
no weight list, and you can try it right
now. Here they typed, "Unfurl the bamboo
leaf paraglider and glide out." And then
the panda literally pulls out the giant
leaf and then floats off. You could
steer this with your keyboard for up to
one minute and theoretically it'll run
on just one consumer grade computer
graphics card. They even took Google's
own Genie 4 demos and then compared
theirs side by side claiming that theirs
followed the camera a lot better. And
with the Google version, I actually
struggled to get it to generate King
Kong. But in the 10-cent demo gallery,
they've already got videos with like
Charizard and Blastoise. So, they
apparently have no problem generating
other companies IP, at least for now,
until some of the lawyers get involved.
I didn't expect an open-source version
of GD4 to come out this quickly after
G4, but now I'm curious from you. Would
you rather use the version from Google
or the one that lets you play with
Charizard?

I created this whole video with Cling
4.0, which just came out, and I did it
all on art list. All you do is add a
reference image or video and type
exactly what you want. For this one, I
used a real wings suit flight and a
picture of me. I also built a claw skill
that does this for you automatically
with Art List MCP. It made each of these
ads from one prompt. Check out how in my
full video.

AI has truly leveled the playing field
for solo operators like myself. Small
businesses who can create a website that
would cost $10,000 plus hiring an
agency. All you have to do is open your
mind.
>> Compared to the lite that like burn down
the factory, they're going to take our
jobs. Growing up in that environment, if
that's your teacher or your parent, one
generation from now, you are already a
lost.
>> I don't understand AI, so therefore I'm
anti- AAI and I'm going to put a stance
on that and I'm never going to adapt.
Well, good luck with that. That's all
I'm going to say cuz it's not going to
end up for you really well.
>> The real superpower in 2026 and even
more going forward, it's curiosity.
That's what it is. Like even just ask
Claude Code, can you do this? 99% of
people will never do that. If you even
try, you're ahead of them.
Welcome back to another edition of AI
Rabbit Holes. This is a podcast where we
go down the endless amount of rabbit
holes in artificial intelligence. And
I'm happy to have back on once again
Aaron McCulkkey. And first of all, those
of you who are tuning in live on X, we
appreciate that. Uh those who are
listening on the Spotify, Apple, YouTube
podcast version, we appreciate your
listening and viewership as well. Uh
Aaron, we just chatted a few days ago
about the buy it or build it. Should we
vibe code it? Should we buy the
software? Anything in your world that
has changed in the past, what has it
been 48 hours when it comes to AI?
No, we got access to Claude Duet and we
mentioned that on the intro of the last
one. So, they said everybody on the wait
list, you had to go sign up. If you have
a max plan, they're going to give access
to this week. We'll see if that actually
happens. But another personal agent,
runs in the cloud, does everything for
you tool that everybody has. So,
>> I do like how they have channels though.
That's one thing it beats Chad GBT dots
on.
>> Here's the biggest one. It has opus.
That's what people don't realize. The
reason I like it is I want my Opus. And
yeah, no matter what features and tools
and interfaces you give, DOTs have some
cool new ones. It's in iMessage now,
which is amazing, but it doesn't run on
Opus. So,
>> yeah,
>> it's
Yeah, I think people are learning like,
man, I just got to pick one thing. I got
to stick with it. There's very little
that makes me want to move and and like
redo a subscription or a workflow unless
the model in it is that much better and
it's still Opus even though it's a month
old.
>> Yeah. And my thinking is like on my Mac
Mini, I have the Claw desktop app and I
have the Chat GPT desktop app running.
And so I'm going to have Cad GBT dots
and claw duet paired with those apps
where I can have them, you know, access
the Mac Mini to even go above and beyond
what I could do on the mobile app or in
the cloud. And so that's what I'm going
to be doing is running them on my Mac
Mini 24/7, but also having the mobile
apps on my phone so I can communicate
with them. And I saw they just had an
update, I think it was chat GPT, where
you can now text chat GPT dots into like
iMessage or whatever, which honestly I
don't think that's that big of a deal
personally because you can just do that
in the chat GPT app. Um, but that's kind
of the way that I'm looking at this
right now.
>> Yeah, I'm going to take the other
opinion. And I think it's a big deal to
a lot of people. And I always think of
who are your nine pinned chats in
iMessage cuz you can pin, you know,
group messages, your family, your wife,
whatever. And I bet there are some
people who their GPT dot is going to be
a top nine contact and you don't have to
open a different app. It's right there
where you're texting other people.
Biggest thing I was impressed with that
is that you can send them voice notes
just natively on an iMessage and they
don't ever tell you you ramble too much
like your friends you send voice notes
to. So
>> absolutely. So, the first thing I want
to talk about today is another tweet
that went semiviral for me, which as we
talked about before this, Aaron, my ex
or Twitter account seems to be getting
off uh a little more here as I started
to post a little more, I would say,
quote unquote controversial, not so
controversial post related to AI. Like,
this one's crazy. I don't know if we can
see this um cuz I have all the replies
zoomed in here, but this was just
literally like a one. And by the way,
when I'm posting things, like I'm just
like brain dumping what comes into my
head. If I'm sitting in the sauna, if
I'm laying, you know, on the couch with
my son and I think of something, I
literally either just voice dump it or
pull up my phone and just tweet it.
Like, I don't really think twice about a
lot of this stuff, which maybe I should.
I don't know. That's just how I am.
That's just like my authentic self, I
guess. And this was my latest one,
right? People can complain all they
want, but in reality, it will not end
well for the anti- AI crowd. Okay? And
you could take this in a million
different ways of what I mean by not end
well. I'm not talking about the end of
humanity or anything like that. I'm
talking about not end well in terms of
let's say someone's career or someone's
ability to take care of their family.
And that's where I'm going with this is
that I have a 2-month old son, Aaron.
You have kids, too. I don't like with
all of this coming into the abyss and
like all this AI stuff going on. I am
going to step up and make sure that I
know what I need to know and do what I
need to do regardless of all the changes
that are happening in the background.
Have an open mind. adapt where I can
adapt in the marketing industry and make
sure that I am on the top of my game so
I can support my two-month old son or
any other potential kids and any other
family that I have going on regardless
of all these updates instead of take a
victim mentality which is where I was
going with this tweet. This anti- AI
quote unquote mindset or movement where
these people just think I'm going to be
anti- AI and they're not going to stop
me and this and that. Well, news flash
here, Aaron, you and I both know this.
The economy is going to keep moving.
things are going to keep changing and if
you don't adapt to it, whatever field
you're in, I know there's nuance to this
whole thing. But whatever field that
you're in, AI is probably going to have
its hand in it in some regard. So, it's
really like back to the whole money ball
thing, which I always tell people when
Billy Bean went up to that old scout and
he said adapt or die and the scout
eventually got fired. And now look at
MLB today where it's basically all run
on analytics with of course some of the
other stuff like the fundamentals are
still important in the game of baseball.
But Moneyball is the best example that I
can give here of the adapter die
situation. And I could ramble about this
whole thing forever. Um, but before I
get into this too much, Aaron, what are
your thoughts on like the whole anti- AI
movement and this and that and you know,
last thing I will say though is like I'm
not a fanboy of these big tech oligarchs
like Zuckerberg and you know, Elon Musk.
Like I understand the downsides. I don't
I'm not a fanboy of any of these people.
I think they're very evil if I'm being
honest with you. But it doesn't matter
like there things are just going to keep
changing so I might as well use what's
at my disposal to help me and benefit my
friends and family. That's the way I
look at this whole thing. Yeah, I
actually had multiple people tag me on
the same topic, but it was with their
kids and their teachers, which is a
profession that is still near and dear
to my heart. My wife even showed me this
self-portrait my son did from parent
teacher conferences, and it's
symbolizing what's in his brain. And
he's got to be the only sixth grader in
[clears throat] America who has chat GPT
6.1 on his self-portrait in his brain.
And it's funny, he'll come home and tell
me what his teachers say about AI and
they just go on rants and they don't
know what they're talking about. Here's
my example I like to give people. You
have an advantage if you use these tools
right now. If you reinvent your
workflows and learn some new skills that
were very inaccessible to non-technical
people like GitHub today,
>> but imagine this, you teaching your son,
think of the advantage the next
generation has if they're learning from
people like you and I at home, just like
our own children, let alone on the
internet. Compared to the lite, the like
burn down the factory, they're going to
take our jobs. growing up in that
environment, if that's your teacher or
your parent, one generation from now,
you are already a lost cause. You better
hope there's some kind of universal
basic income or welfare program, which I
don't know too many people that are
confident enough in the government to do
that. Uh I certainly am not. I want my
kids to learn those skills and it's even
bigger advantage when you go through a
generation. So, what would you tell
farmers when the tractor came out?
>> Yeah, horses aren't bad, but learn
>> horse and buggies when the cars came
out. I mean, it's literally just like
you could go back in time and you're a
history teacher, so I mean, you could
probably give a more a ton more
examples than I could. I wanted to share
this reply quickly. This guy, uh, Mario,
he follows me. Like, we DM sometimes.
He's he's cool. And he was saying that
his son's kindergarten teacher is very
anti- AI considering homeschool now. So,
I gave the year you're out of here, like
get him out of the the class that he's
in right now. But, it's like it's this
whole case of you don't know what you
don't know. And I think there's just so
much fear that people have when they
hear AI because they freeze. They they
don't know how to use AI and so that
means AI bad. If I don't know how to use
it and I tune into Fox News and CNN and
pull up social media and there's all
this bad stuff about AI, AI bad. I don't
understand AI. So therefore, I'm anti-
AAI and I'm going to put a stance on
that and I'm never going to adapt. Well,
good luck with that. That's all I'm
going to say cuz it's not going to end
up for you really well.
>> Yeah. Here's the single best thing you
can do to adapt. If you don't have a
GitHub account, go make a free one. And
if you do, keep a list or give this
transcript to your agents and have them
star some of these that we're going to
show cuz
>> I actually see GitHub as my gateway drug
to all this AI coding, building um about
a year ago when I kind of got sucked in,
GitHub was the vehicle that took me
there more than chat GPT or Claude was.
>> Yes. And on that note, we're going to
transfer to the main topic of today's
episode, which are the best free GitHub
repositories that you can start using
today. Um, but before we actually get
into the repos, we had a whole episode
talking about, you know, how you should
use GitHub, what is GitHub, and I'll
leave that video link in the description
below. Um, but if you don't understand
this, what I would first recommend is
getting a free GitHub account, okay? and
then start to create free repositories,
private, public, you know, just just
dive in and figure this out like and you
can seamlessly integrate GitHub to
Claude or GitHub to chat GPT and you can
start having these AI agents create
GitHub repos for you. So, you know, you
might be looking at this and saying,
well, what is all this stuff? What are
all these repositories, right? And I'm
not going to get into all of them, but
basically every single website or
landing page that I have is stored on a
private GitHub repo. ryandozer.com. All
my website files, my codebase, that is
all stored on a private GitHub repo. All
these landing pages that I use for my
skill stack or digital products or
clients if I'm building landing pages
for paid ad campaigns, that's all stored
on a private GitHub repo. my chat GPT
dot. I'm not going to get into the whole
dots right now in the personal AI
assistance, but I have a whole GitHub
repo that backs up every single thing
that my chat GPT dot is doing behind the
scenes with SOPs with daily logs and and
all sorts voice transcripts. When I call
him on the phone, he's keeping those
transcripts stored in a folder on a
private GitHub repo, right? So literally
every single thing that I am doing with
AI personal assistance, websites,
landing pages, I am storing on a private
GitHub repo. Aaron, anything else you
want to comment on that before we
actually start getting into some public
re GitHub repos that anyone can start
using? And I'll leave all the links to
those in the description below as well,
just if anyone wants to try them out
themselves.
>> Yeah, the coolest thing about AI is your
microwave can't tell you how to use it,
but your AI tool can tell you how to use
itself and tools like GitHub. The
biggest question I get from people is I
don't know the vocabulary. It's saying
PR merged. What the heck is PR stand
for? What does merge mean?
>> Give your AI tool of choice the GitHub
connector. Log in like once and then
it's it's good. And tell it I'm not a
developer. Use common sense plain
English that I will know as a marketer,
as a teacher, as a nurse, whatever your
skill set is. And then anytime it slips
in, just say explain what a PR is.
explain what merge did and it'll tell
you and if you don't like it you can
undo it. So the whole point I didn't
actually know this about GitHub at first
that makes me feel dumb. It's version
control. Like a Google doc says Ryan
changed this sentence 5 minutes ago and
it's kind of highlighted. That's all Git
is. It's like if you broke something
it'd be easy to undo it or see who broke
it. And that's the whole point of GitHub
is like when you update your dot every
day if you somehow broke something you
could revert back to yesterday's version
just like a backup of your computer. Ask
your AI tool to explain this stuff and
you'll have it down in like two days.
>> Yeah, absolutely. By the way, every time
I do these slides, just for those tuning
in on X live or, you know, watching the
video version of this, this is all
one/comand cloud code, by the way, and I
teach how to do all of this stuff in my
community. If you want to join that,
it's free to join right now. I'll leave
a link to that in the description below.
But anyways, we talked about kind of a
repo just a little bit here, but some of
the repos that we're going to, and by
the way, repo short for repository. So,
if I keep saying repo, you don't know
what that is, just repository, short for
repository. And these are just some of
the GitHub repositories that we're going
to be diving into. And shameless plug,
of course, is I have one of my own
public free GitHub repositories that you
can literally and and if you want to
know how to use this, okay, so you can
literally come to a GitHub repository.
So, if I search like I don't even know
if I have it up here, but if I search
like hyperframes, that's another one um
I'm going to talk about here later on.
If you want to know how you can actually
use a GitHub repo, it's super simple.
All you have to do is take the URL of
the public GitHub repository, copy it,
paste it into your AI system of choice,
claude, chat, GPT, Claude Code, Codeex,
etc., and just say implement this. Now,
be very cautious of that because there's
a lot of potential prompt injections or
malicious things going on in public
GitHub repos. So, I would always tell it
to scan the GitHub repository first and
look for anything malicious with prompt
injections and malicious files. You can
create a skill for that if you really
want to. But that's how easy it is. You
take this URL, paste it into your AI
system, and just say implement this, and
it'll walk you through that step by
step. And let's go back to the slides
here. So, the first one that I had was
just my some of my free Claude skills.
Now, I do have a paid digital product of
all my Claude skills or skills, not just
for Claude. Um, but I do give away I
think about 10 or 15 of these. Yeah, 15
for free from SEO, keyword research,
CRO, which stands for conversion rate
optimization, landing page, anti-slop
stuff, like all the stuff that I've been
doing over my 10 plus year marketing
career packaged into some of these skill
markdown files, which if you don't know
what a skill is, it's essentially an SOP
or standard operating procedure on how
to do a specific task. So, um, before I
go on to the next one here of just the
Google Workspace CLI, Aaron, did you
have any GitHub repos that you wanted to
cover or did you just want me to keep
going through the slides here of what I
had?
>> Let's do yours and then we can put a
link to mine cuz I picked some that I uh
compliment yours. Another reason to get
a GitHub account, if you're a listener
who keeps hearing us talk about this and
you didn't make one,
>> being able to star it in GitHub is kind
of like a bookmark, but it doesn't have
to live in your browser. You don't have
to put it in Notion. So you can always
find starred repos. So if you see Ryan's
and go, I want to use that on GitHub,
you click the star, you'll always be
able to find it and it's easy to search
for
>> and then you can just tell your AI tool
like, hey, use GitHub to look at my
starred repos. Those are things we can
use for marketing purposes. So another
reason to get a GitHub repo or a GitHub
account if you don't have one already.
>> Yeah. And if my hunch is that stars are
the new Reddit up votes in terms of how
manipulated these things can be because
if you have a lot of stars, I mean that
gains a ton of traction from like the
vibe coding community and and everyone
else. Um, but the next one I wanted to
cover is the Google Workspace CLI. So
Google Workspace meaning Google Drive,
you know, Gmail, Calendar, Google
Sheets, Google Docs. Now this is
important because if you're using like
let's say VS code like I am and you have
claude code and codecs running in the
background there's no you know seamless
integration like there is on claude web
app or chat GPT web app where you can
just oneclick connect Google drive this
is a workaround if you're using an IDE
like VS Code like I am that still gets
you access to my Google Workspace and it
is the verified version this isn't some
third party so you got to be very
cautious here this is the verified
version and one of the things that I do
all the time with the Google Workspace
repo is I store a bunch of short form
videos in there in Google Drive folders.
And so I can literally take a Google
Drive folder URL, paste it in VS Code,
and have Claude Code scan through all
the videos, analyze the transcripts, and
then create captions and then use Blow,
which is my social media tool, to
automate all of the the videos from that
Google Drive folder. Like, it's just all
seamlessly connected, which is just that
little quick example here. Um, the next
one I wanted to quickly cover is
Enthropic, who owns Claude. They
actually have a free public repository
of skills. Everything from a front-end
web design skill, which is something
that I've used in the past, which
actually is very good in my opinion. Um,
they have an MCP skill, they have a
skill creator skill, which I've used
multiple times in the past to help
create skills with best practices. So,
for those who didn't know, Enthropic has
their own official GitHub repository of
skills, which is really cool to look at.
Um, and Aaron, feel free to chime in
anytime as I'm just I don't really want
to just go through in like a typical
school presentation, go through slides.
Like if you have thoughts or anything on
how you're using these, like feel free
to chime in as well.
>> Yes. Well, speaking of school, I do want
to shout out that Anthropic released
their own teacher skill repo. So, if
you're a professor, an educator, a
principal,
>> I can put a link in the description on
the video post, but it's things
specifically for teachers. And again,
the whole point of this, like Claude can
do pretty good front-end design, but
they're giving you a list of
instructions that has a more narrow
focus than if you just opened a chat and
said, "Hey, Claude, design a landing
page."
>> It's it's almost like steps or
procedures in plain text.
>> So, [clears throat] my other thing I
want to chime in is use people's skills,
but you can fork them cuz all these
things are free. They're public.
>> So, take your skill uh repo with 10
things in there. I might really like
something yours does with landing pages,
but add my own to it. Just tell your
agent, "Copy Ryan's thing, add this on
top of it, and put it in my GitHub." You
don't even have to share that with
people. You can make it private. I'm a
big fan of customizing cuz you're not
going to want everything somebody does,
but skills are amazing for inspiration
and ideas and like, "Oh, I didn't know
it could do this. Now make it your own."
>> Yeah. And that's a great point. You
always want to repurpose skills. You
never want to take Ryan Dozer's landing
page skill with my copywriting, my
branding, Ryan Dozer everywhere. You
want to repurpose that with your own
context and branding and and etc. Um,
the other one too, like I think I found
this one, I've used it before um with
the last 30 days where it can scrape the
last 30 days of your history of what
you've been doing inside Claude Code.
Um, so or any topic here is what this
says. Has 66K stars. Um, yeah, search
the last 30 days of sources you've set
up. So, it looks like it analyzes the
last 30 days of what you've been doing
and then it can find any insights or
update opportunities. And I know
Enthropic actually has I think it's
called insights. They have default slash
commands and insights is one of those. I
think it's very similar to this last 30
days.
>> The last 30 days is a good news one. So,
say there's a topic. Uh I actually
worked on a team at a company that we
use this as part of our daily things
like what's the internet saying about uh
claude AI data centers what's new in the
news and it can search Reddit, Twitter,
YouTube. You need API keys for some of
the sources but a lot of it's free. It's
like a pulse check on the internet.
>> Interesting. Yeah. And by the way, I'll
leave all these GitHub repos in the
description below where you guys can
literally copy and paste them and then
use them however you want to. Uh the
next one is firecrawl. So firecrawl I'
I've used this in the past is a I would
say an advanced web scraper, right? So
for those who have maybe seen the term
web fetch. So web fetch is the default
scraper of cloud models and sometimes
like the scraping world is such an
interesting battle between like websites
blocking certain scrapers and scrapers
finding workarounds and firecrawl from
my experience. I haven't used it
probably in the past few weeks or so,
but when I was using it, they have a
really generous free plan, by the way.
Um, they tend to have more advanced
scrapers, I would say, than what you get
with like a default web fetch. So,
that's something to consider if you're
doing a lot of web scraping.
Uh, this is one I use every day, the
Google Search Console MCP. So, Google
Search Console is a free tool offered by
Google that literally monitors all the
website traffic and views and clicks
that you're getting from Google Search.
They also have Google AI overviews now.
So you can look at your AI data inside
Google Search Console. And so from an
SEO perspective, think about the power
of integrating a Google Search Console
MCP with your existing AI setup versus
previously you'd have to manually go
into Search Console, look at all your
data the past 28 days, find impression
gaps or click gaps, try to find
different content update opportunities,
maybe there's some new ideas in there.
Well, guess what? You can literally in
one slash command or one prompt use the
Google search console mcp and ask things
like which pages are losing clicks, you
know, compare these things, compare
that, give me three update
opportunities. And if you're using an
Astro Cloudflare GitHub website stack,
you can then all implement that in the
same chat or a new chat. And it's
seamless. It's absolutely seamless. like
this is one of my favorite I would say
MCPs or GitHub repos that I'm using on a
daily basis for myself at ryondozer.com
but also the various SEO clients that
I'm working with.
>> And it's funny because look 167 stars
that's nothing compared to these big
popular ones and I feel like you're
calling me out a little bit.
>> I get the emails from Google Search
Console and it'll tell me there's a
problem with a page or a 404 thing.
>> I forward it. I forward it to my agent
and say fix it. But I should just go
straight to the MCP and let my agent see
these tools without me having to like
open it in a in a Chrome browser. So,
I'm I'm adding this one as soon as we're
done recording.
>> Yeah. Be careful though. This is like a
one that I found that doesn't have a lot
of like doesn't have any security issues
and it works flawlessly. Like there's
other ones out there that are probably
made by third parties like trying to
spoof a Google Search Console MCP and
like get access to your data for their
own benefits. So, just be very cautious
with that.
>> Give us So, give us a sample tip cuz I
I'll practice this. Let's say I open up
Claude and I say, "Hey, grab this repo,
but before you plug it in or before we
use it, what should I ask Claude to do?"
>> Uh, analyze, scrape the entire GitHub
repository and look for any malicious
files. Look for literally any signs of
malicious or mal intent of something
that it's going to, I don't know,
potentially take API keys or steal my
data. Like, I would just be broad and
just say like do a quick analyze or
check if it's going to do anything
malicious to my existing setup. I would
just start with that.
>> Gotcha. Yeah. Like prompt injecting.
I've seen that in GitHub repos where it
gives instructions in your AI agent. At
least older older models would follow it
and like
>> put in something that syncs all your
data with them. And
>> yeah,
>> I I'll try that with the the Google
search console.
>> So, on the same topic of SEO, um I
haven't actually I've tried this once,
but I don't use it daily. I use data for
SEO. It's kind of my go-to SEO tool now.
But those who have been in the SEO space
for a while, hrefs and SCM Rush, those
are the two big boys and they still are.
Um, but this guy, he basically cloned
SEM Rush or Hrefs, called it Open SEO,
and charges 10 bucks a month and
basically has the whole thing available
in a public GitHub repo. Now, he is
using data for SEO behind the scenes for
what he is doing. And so, that's why I
don't use it and I'll just go straight
to the source and pull data from SEO
from their MCP. But for those who are
looking for, you know, a $10 a month
solution that's similar to an SCM Rush
or an HFS type feel, this one might be a
good solution here. And here's the
thing, because the GitHub repo is
publicly available. Like if I type in
Open SEO GitHub, you could literally
come in here, copy the GitHub repo URL,
and say like clone my or make a fork of
my own version of OpenSEO and and here's
what I want in my a in my SEO tool, and
here's what I don't want. You can make
your own version of this. That's why
this is so powerful and you'll see it
has over 25,000 stars. So, it is gaining
popularity and all these giant SEO SAS
companies like HFS and SEM Rush, which
Sem Rush went public. I think they just
got bought by Adobe not too long ago.
Um, they they are coming after I mean
they are losing their moat one day at a
time. data is the only moat that those
tools have anymore just because they've
collected 10 15 years plus of highv
value website data from all the
different people that have used their
tools over the years. Um, but slowly but
surely we are going to have more open
SEOs come into the equation and I'm a
big fan of this. I'm all for it.
>> So let's say somebody goes and hits fork
on here since we talked about a prompt
to give your agent before you use or
clone a repo.
>> A fork is duplicate it but now it's my
own. And then what I would tell people
to do, same thing. Scan it for security
issues, malicious intent. Let's say it's
clean. Now tell it, I run these
websites. I have a blog. I do landing
pages to sell a digital product.
Whatever it is, what should I add or
take away from this to customize to my
use case? And if your AI agent has a
bunch of your work in it, like a cloud
project, it'll know exactly what to give
you or to take out.
>> Well, here's the other thing, too. Like
if I'm using VS Code and I have all my
client work and cloud code and codecs
all synced up to that, I could literally
just say scan through my entire VS Code
workspace and figure out all the work
that I'm doing with my current SEO
clients and figure out what's valuable
here of what I should include and what I
shouldn't include. Like again, you can
literally outsource a lot of this to
have AI help you out in building this.
If you want to provide more context and
more direction, obviously that's better.
But if you don't even know where to
start, just say, "Scan my entire VS Code
workspace or AI setup if you're not
using VS Code." And it will help provide
like actionable direction from there.
Um, well, cool. Let's move on to the
next one here. Uh, NewsJack. Uh,
NewsJack is an open-source basically
open-source skills that turn your AI
agent into a full PR team. Um, so this
one gained a lot of popularity on X. uh
shocker because I feel like a lot of
these news jack uh GitHub I feel like a
lot of these GitHub repos that's where
they start is they gain all this
popularity on X and then they blow up.
So yeah, 1.7K stars. So, this is
essentially where you could have an AI
agent uh train off of this and act as
your PR outreach person. meaning that,
and this is kind of what I'm starting to
do right now on rindozer.com, is I'm
putting together these like quote
unquote research articles where it's
actually taking my own data of things
that I'm working on behind the scenes,
turning these into articles, of course,
with not including any confidential data
by any means, but turning these into
actual things that would be worthy of
stories or something a journalist would
want to pick up and write an article
about. Um, and this is always this is
the hardest part of SEO is PR outreach,
getting high quality backlinks and brand
mentions. And this newsjack just helps
um, essentially automate that process. I
haven't actually I think I've used this
before um, but then I got caught up
doing something else. Um, so again, I
would like have your agent scan this
repo and just say in plain English, what
can this help me do? But is essentially
it's your PR outreach person where it'll
find like journalist emails. little
craft, you know, take your existing
content, craft emails, pitch
journalists. Basically trying to get you
featured in the media and get you back
links, uh, is what this whole thing is
trying to do here.
>> I haven't heard of that one either. So,
you're teaching me stuff.
>> Yeah, this is a good one. And again, you
can repurpose your own version. But one
quick point I will make is that when
people in SEO like they always say, "Oh,
I can't get back links and no one wants
to feature me and da da da." Well, it's
like usually the culprit is why don't
you actually create content that people
care enough to feature, right? Like
people just think that their respun
claude or chat GPT article is
good enough to be featured in the New
York Times or something that someone's
going to mention them or give them a
link. Like you have to create
interesting That's really what it
comes down to to start getting back
links and brand mentions and placements
like that. Um well, cool. Hyperframes.
Aaron, I think you have a lot to say on
this one because I know you use
hyperframes. Um, but this is a free
open- source framework from Hey Genen.
Hen being the popular AI avatar
generator uh released hyperframes which
basically I use hyperframes to actually
make this presentation like these slide
decks and I also use them in my videos.
So it's a free software you can you know
add to your video editing system and it
can create like motion graphics, visuals
uh completely free to use and really has
changed the game when it comes to AI
video editing in my opinion. So, uh, and
what are your what are your more
advanced thoughts on this than what I
have about Hyperframes? Cuz I think you
use this more than I do.
>> Yeah, I've used it three times already
today and it's only launch.
>> Here's what I'll say. I'm going to give
them credit. They're a video editing
company that did things the traditional
way. And then I want to say, I don't
know, February or so, early this year,
they said, "We're going to create a
whole new way that you can edit video
through HTML, which if you're not a
coder, I'm not either, don't be
intimidated. It's just a visual coding
language. is what every website is
fundamentally in is like HTML. It's a
way to display stuff. Well, guess what's
good at writing in HTML? Not Aaron.
Aaron's coding tool,
>> Cloud Code, uh, codecs, Grockbot,
whatever. And this idea is like now your
agent can do visual things with code.
And then it's easy to edit because one
of the problems with videos once you
render in captions, they're just burned
in the video. You can't get them out.
But HTML lets it say, "Here's a box with
captions behind it. Here's a box with
Aaron.
It's free. You don't even have to pay
per use. Like your AI agent is all you
need to run this. And then the other
thing that's cool is they actually have
a desktop app because a lot of people
said it opens in a local host in a
browser tab and it's like I still need a
visual interface,
some sort of timeline or like draw and
say put it over here. So they actually
have hyperframe studio which is a
desktop app. I know it's out on Mac. I
don't know about Windows and Linux. And
I put our Tuesday two days ago podcast
in there. And it has a transcriptbased
editor. I plugged in three feeds, your
camera, my camera, and your screen
share. It lined them up. It gave me a
transcript to edit by, and it's just
running through clawed code on my
computer, and I don't have to pay
anybody stupid AI credits to use it. So,
I'm a huge fan. Um,
>> yeah, it's free. That's the biggest
thing here.
>> Yeah, use a ski use a repo to come up
with a video editing skill. So maybe you
clip a podcast, maybe you do explainers,
maybe it's recorded lectures, whatever
your uses are, have it use hyperframes,
and then just update the skill every day
when you're done using it and say, "What
did we learn? What styles, colors,
captions, edits, uh, intros, outros,
music, whatever, and just keep that in a
GitHub repo of your own. And then when
you go to edit a video with Hyperframes
tomorrow, say, "Pull up my video editing
repo. Cut five clips from this to be a
Instagram reel." and it can just run on
your computer. So, I'm a huge fan. I use
it every day.
>> Yeah, Hyperframes is legit. Um, Postus
actually a competitor to Blot, which is
a client of mine. Um, but the reason I
included them is because they actually
have a full GitHub public GitHub repo.
Um, so if you do post GitHub and again
I'll leave all this in the description
below. But the main reason I I included
them is because basically you can take
this public GitHub repo 40K stars which
again this is actually how they got off
the ground and became popular is because
they open sourced and made a free public
GitHub repo. So that's an entire
marketing strategy in itself. Um, but
what you can do here is again clone this
or fork it. And if you want to make your
own social media scheduling app for AI
agents to use in Claude or ChatgPT or
whatever, you can make your own version
of this and create your own social media
scheduler. And that's why I wanted to
include it. Um, because I think that's
something a lot of people might be
interested in if they like tinkering
around with their own social media and
they don't want to pay, you know, like
Buffer or, you know, Hootsweet
>> 40 50 bucks a month for some of those.
>> Exactly. And it's not free. you still
have to pay some API stuff right on the
back end to post through social media
and whatnot. Um, but just something
>> that's my question for you though, Ryan.
So, why wouldn't I self-host and do this
myself versus they have a paid plan of I
let's just say 20 bucks a month. What's
that take for me if I want to self-host
a tool like this?
>> I haven't actually done it, so I can't
give you an exact amount. But for those
who are interested in customizing their
own solutions like the the vibe coders
and that that that's what this community
is geared towards are those people.
>> Yeah. I haven't used this one but I
think we both have self-hosted simple
things. But the answer is
>> are you talking like like that stuff?
>> Yeah. Like say you want this to be able
to post to your Twitter. You just like,
"Hey Claude, post a tweet for me through
a self-hosted version of this." You're
going to need a developer account on the
social media platform and you're going
to pay per use. Sometimes it's like a
cent but it's not free instead of 20
bucks a month to get access and they're
going to keep the difference between
their API cost to post your stuff and
what you pay every month.
>> So it isn't free and then you're also
responsible for like updates and
security issues if there's a
vulnerability in it.
>> Back to the whole vibe coding is should
I vibe code it or should I just buy it?
>> Yeah. And I don't vibe code this for
multiple reasons, but that's one of them
is my social accounts and the companies
I work with, if I got those banned, that
would be a major issue. So, it's worth
it. But
>> it's it's not free to self-host is my
point.
>> Yeah. And that's what I put here. I just
said best for vibe coding your own
scheduler if you really want to. So, I
know there's a group of people out there
who love vibe coding their own things
and all the customization things they
can put on it. That's why I included it
here because you may not have known that
this even existed versus just saying
here's hootsweet.com vibe code
Hootsweet. That's not going to
accomplish much versus actually giving
it a full public GitHub repo with
everything in it.
>> Yes.
So, this is just kind of a slide where
we talk about prompt injections and fake
stars and how people are manipulating
popularity um and just all the security
issues and prompt injections that you
need to be aware of um when it comes to
actually getting these public GitHub
repos. Um so, I think that was really
all I had uh for the slides here. Aaron,
anything else you wanted to share or
comment um on this topic of best free
GitHub repositories in in case I missed
any here? Yeah, let me share my screen
because I have a little document. It's
not as uh sexy as yours. So, here's
mine. And I did a mix of big ones that
have tens of thousands, 100 thousand
plus stars and then small one which is
my own. Uh I am a big fan of having some
sort of gate that AI generated writing
has to go through before you post it,
publish it, send it. If you haven't made
your own, don't just take mine and use
it. fork mine. It's called an AI writing
audit. There'll be links to all of this
in the show notes. And it takes common
patterns and says, "This sounds like
Claude." And it does a change log of
edits to the AI generated output and
cleans it up for you. And I'm always
updating it every month or two. You can
add your own if there's certain things
you never wanted to use. But then you
can put in a hook or a system prompt
that says anytime another person's going
to read the generated text, use this
skill before you ship it, send the
message, or post it.
A huge fan of that, man. I think
everybody's kind of building their own.
>> Yeah, I have an anti-slop skill that I
actually use for literally every single
written output in VS Code that I'm
using, whether it came from Cloud Code
or Codeex. There's a rule in there in my
agents.mmd file where it has to pass the
anti-slop rule. If it doesn't pass that,
well, you better keep working on it
until it passes the the rules and
outlines there.
>> Yeah. And maybe what we should do is
like merge ours into one super one.
We'll have to do that another time. Uh,
another one's herder. So, we don't talk
about terminals a lot, and I used to
think terminals are dead. Use a desktop
app or a VS Code type IDE is your
preferred solution. But Herder, which is
free and open- source, so you can
customize it. I have some crazy uh
elements to mine just for fun, but it
lets you work across different terminal
sessions
across different computers in different
environments all from one UI. So I have
a Mac Studio, I have a MacBook, I have a
Linux desktop, I have Grockbots's
virtual machine. I can use local coding
agents from one terminal on any one of
those devices and control jobs locally
on those other ones. So if you're
somebody who likes to use coding agents
on device, not in the cloud, you should
definitely try out Herder and it's
totally free. So not only do you not
have to pay, but you can like customize
it.
>> I've never heard of that one. Um, I
don't use a lot of local models just cuz
I I mean I have Mac Mini and I have a
MacBook Pro, but it's like they're so
bloated now at this point with like apps
and other things that I've been using.
So unless I have like a new device
dedicated for local AI models, I'm
probably not going to explore that.
Okay, let me clarify running locally. So
like Claude Opus, when I do video
projects, I start them on a Mac Studio.
Yeah. But then I can open up Herder if I
want to on a MacBook and walk away be in
another country because I'm on a VPN and
I can connect all those in one terminal.
I can see all of them instead of like
switching tabs. So I don't do a ton with
local models either just cuz I don't
have crazy hardware. Um, okay.
Cloudflare OS. Here's one. If you've
ever worked on a team and you were
worried that the non-technical people
would break stuff or vibe code a bunch
of slop, one of the main reasons why
Cloudflare built this Cloudflare OS is
it connects teams work together. You can
use tokens through their portal to run
models, but you can also do these little
one-off tools that are sandboxed so they
don't leak anywhere or break anything or
access any company data that a
non-technical person wouldn't know how
to safeguard. And it's kind of like a
digital operating system for your
company. So, if you're on a team that
has some non-developers that are messing
with this stuff, you should check out
Cloudflare OS. One I use every day is
Whisper. Everybody should have a local
model. doesn't have to be a big one
because all we're doing is
transcription.
So if I pull a video file and I just
want a transcript, I don't want to have
to send it in the cloud to chat GBT or
claude. I want to have something on
device that's small. These the biggest
ones are like 4 gigs, I think. So even a
little laptop can can run these. You can
transcribe any video or audio without
anything leaving your device and going
on the internet. So Whisper is a great
one. There's like three different sizes.
I think that's the one that I use. Um
>> there's a lot there's a lot of ones like
Whisper, aren't there, floating around?
>> Yeah, and there's new ones all the time.
Um
Nvidia is really focused on
transcription models. They just released
a huge one yesterday, but I can't host
it. It's like
>> that's what I thought.
>> You need real NVIDIA compute to run it.
But uh most popular one that I know of,
in fact, one of the most starred GitHub
repos of all time. If you haven't heard
of Matt PCO, he has one just called
SLSkills.
And it's for people building tools. He's
a developer. And what I've learned from
him is the vocabulary and the
terminology and the approach to building
like web apps and software and landing
pages makes me more technical and helps
me get my AI tool to understand what I
want better. So, I'm a huge fan of his
and he's big on like YouTube and
Twitter. If you haven't checked him out,
he's the bald guy with glasses and like
a British accent.
>> Okay.
>> His his skills repo is amazing. Um,
especially for people like Ryan. You're
like, I'm not technical, but come on,
dude. You're you're technical enough.
He's a perfect uh resource for people
like us.
>> Last one from Google. And this one has
28,000 stars. I'm shocked it's that few.
A lot of us started making skills for
writing and coding. They came up with a
whole new primitive like a type of file
that's a design MD. And you can take a
link to rinosozer.com
give this GitHub repo to your agent and
say build me a design.md visually. I
want to be able to recreate this with
fidelity. So it can be web design, app
design, video design, anything that you
look at and see colors and shapes,
animations, things like that. It puts it
in a language that your agent can
actually understand and use. So like
let's connect that to Heyen. You should
have a design for your style of videos
or your podcast and then use that
through hen and have Claude or Codeex
edit your video and you'll get way
better, more consistent outputs.
>> Awesome. Yeah. Yeah, and we'll leave all
those links in the description below.
Same with all the other ones that I
talked about on my slides, too. But last
point I wanted to make, Aaron, is that,
you know, like our last GitHub video
that we made, I got a couple comments
from people and they're like, "Well,
this is great, but you didn't actually
show me how to do anything." Well, it's
like, "Hey, dude, why don't you actually
just first of all, you can ask AI, take
the video transcript, plug it into AI,
and say, "What should I do?" And it'll
give you a step-by-step direction. But I
guess people just can't think that far
ahead is where I'm going with this. And
if you don't know the answer, ask AI.
But the first thing you should do is
create a free GitHub account. And then
you should sync that GitHub account to
whatever your AI system is of choice. In
Claude and Chat GPT, they make this very
seamless. It's like one-click
integration. And then in Claude Code or
Codeex, if you're using like VS Code or
an IDE or in the terminal, you have to
go through an authentication step, which
is kind of confusing, but again, if you
don't know, take screenshots, copy
paste, ask AI to help you go through it.
That's exactly what I did. So get a free
GitHub account, connect it to your AI
system of choice. Try start try start
using some of these free repos and just
start playing around with GitHub and
you'll start to quickly understand why
it's such a popular and useful tool in
this AI era.
>> It's the backpack for your AI agent that
you can take to any other AI agent or
project. That's how I think of it.
>> Yeah, I love that. And the final point
I'll make in this episode before we
conclude, Aaron, back to my original
tweet that I had. Um, don't be a victim
and don't be close-minded. Like just
because you don't understand something,
which there's a lot, I mean a lot in
this AI world that I don't understand.
There's a lot that you don't understand,
but that doesn't mean I'm anti what I
don't understand. If I don't understand
it, but it's going to benefit me in some
way. Well, guess what? I'm going to
spend the time to figure it out because
that's just who I am. I take initiative
and I want to get ahead in life and I
don't want to sit here and be a victim
and let other people try to take care of
me and just complain as my life goes on.
And that's just I don't know. We just
need more like open-minded people, I
guess, in general is where I'm going
with this as I keep rambling. But Aaron,
any other concluding thoughts on that?
Because I'm just gonna stop right there
before I say something I shouldn't.
>> Yeah. Well, that's why we like you, man.
We get the fire from you. I say it's
it's like Brussels sprouts. How many
people tell me they don't like them? And
I go, "Oh, how many times have you had
them?" Well, I haven't, but I know I
don't like it.
>> These people ranting about AI that have
literally never used it for more than
like a chat box saying what AI can't do
or AI is this. And I'm like, you're the
other one is don't listen to those
people. I block them in my feed. My son
>> I blocked 20 people that replied to my
tweet the other day [laughter]
>> cuz it's like I'm not I'm not wasting my
brain cells on these people and not
everyone likes me or my thoughts. I'm
not offended. But I also tell my kids
this like the teacher went on a rant
about AI generated YouTube thumbnails
and first of all I said what's that have
to do? You should have been like that's
your dad's using. He actually said that
he gave him my YouTube channel, which is
not nearly as good as yours, and he's
like, I think my dad uses AI for that,
but not for the video. And I told him,
you know why? Cuz I'm not a thumbnail
designer, but I know click rate,
click-through rate on YouTube is
everything. And all the video editing
that I'm decent at means nothing if the
packaging sucks. And those are people
criticizing a system they don't
understand. Like the the shop teacher at
my kid's school, shout out to him, does
not know what CTR is on YouTube and why
it matters to me. And you know, if we
had to hire a human designer, I'm not
big enough on YouTube to justify 500
bucks for a good thumbnail. I've seen
what companies pay agencies for that.
And it's like, well, GPT Image 2.5 is
way better than I can do, so let's just
use that. Why don't you take the
rationally optimistic approach here and
realize that AI has truly leveled the
playing field for solo operators like
myself, small businesses who can create
a website like we talked about in our
previous episodes that would cost
$10,000 plus hiring an agency or hiring
designers, hiring freelance. It's like
it literally leveled the playing field.
All you have to do is open your mind and
start to figure out what AI can do
beyond just pulling up chat GPT on my
phone or oh hey I saw an ad from Claude
that Claude thing sounds cool. Like if
you actually use Claude code or codeex
just to name a few AI coding agents the
light bulb will turn on and you'll start
to understand like holy the the
opportunities here really are endless
>> which is the real superpower in 2026 and
even more going forward. It's curiosity.
That's what it is. It's like even just
ask claude code, can you do this? 99% of
people will never do that. If you even
try, you're ahead of them.
>> Well, that's a good that's a good
pattern that I've noticed is that anyone
who uses AI coding agents or has in the
past, Codeex, Claude Code, I mean, Open
Code, if you want to go there, like even
Geminis or others, they're usually not
the close-minded individuals or anti-AI
crowd that I'm talking about here
because these people have fully realized
the capabilities and the benefits of
what using an AI coding agent can
provide for their own well-being.
Yeah. Which is what this channel is all
about. Like you're going to be
inherently curious because you're going
to see stuff every week from us to go, I
got to check that out. I got to add
that. I got to fork it. I got to use it.
>> Yeah. And if you want to keep being
close-minded and chase the low IQ top
off ofunnel AI traffic, like have at it,
keep doing that. Keep saying keep
staying scared and don't actually ship
anything useful. And I'm not going to
try to convince you any other path here.
But with that being said, before I keep
talking, Aaron, thanks again for hopping
on and thanks again to everyone for
watching or listening to this edition of
the AI Rabbit Holes podcast. Be sure to
tune in again next time.

All right, another feature that you
could use in your work with Grogbot,
they added pages similar to chatt they
have these spaces and inside the spaces
you can create pages where if you go
here basically like a mix of a word doc
and a notion page. It's a markdown file
which is a text file that AI loves. And
now Grockbot has the same thing. So for
example, I have this podcast tracker set
up where regularly it scans for new
podcast that I've done from my calendar
and then pulls together the links and
the transcripts and then I can share
that with the team. Well, I set that up
inside of a bot that only looks for
these podcasts and it has a page to
track some of the recent ones.

[music]
[music]
[music]
>> Hello.
>> Yeah, yeah, yeah.
>> [music]

Could not fetch transcript (it might be disabled or unavailable). Error: 
Could not retrieve a transcript for the video https://www.youtube.com/watch?v=9Ky5QTpVjAw! This is most likely caused by:

Subtitles are disabled for this video

If you are sure that the described cause is not responsible for this error and that a transcript should be retrievable, please create an issue at https://github.com/jdepoix/youtube-transcript-api/issues. Please add which version of youtube_transcript_api you are using and provide the information needed to replicate the error. Also make sure that there are no open issues which already describe your problem!

This is my voice memo
>> and this is what Suno did to it. [music]
>> I'm going to show you how to do the same
thing in three easy steps no matter what
the genre is. Step one is to get your
song into. And for this I'm actually
going to be using the Sunno app, but you
can do this on your computer as well.
I'm going to tap the Sunno icon and I'm
going to go to plus audio and upload.
And I'm just going to find our voice
memo. All right, it is uploaded. And I'm
not going to create anything with it yet
because we need to change the lyrics.
And I just want to go to the desktop
now. And here is our song. Nice and
uploaded. And step two is to create a
prompt for this song. Fix the lyrics,
too, because it did not get the lyrics
correct. In the song, it says pim pimp
my song. If we want to spell out
something, what we need to do is all
caps and a dash in between each word. P
I M P my song. Copy this. P I N P pimp
my song. P I M P pimp my song. And then
on the last one, it says P pimp my song.
I'm going to add a line break right
there. So it does a little bit of a
pause. And I don't think we need this
meta tag of like verse one or whatever.
All right. And now that we've got the
lyrics, I want to go into the style. And
what I'm trying to do is create
something that's sort of like a game
show, like Family Feud type of vibe, but
I want to kind of blend that together
with a theme song from Pit My Ride. You
got to pimp my ride. And I'm not sure
exactly how I would prompt that. So, I'm
going to use these GPT training files
inside of Claude. And if you're unaware,
Sununo GPT is something I created that
turns your thoughts into prompts. It's
been used by over 21,000 creators and
it's been iterated on over the last year
and a half with over 100,000 prompts.
But instead of just using the Sununo GPT
website, I'm going to use it inside
Claude. I'm just going to tell it need a
game show sounds like Family Feud and
Pit My Ride combined, but also needs
some Gospel Soul female vocals. Also,
I'm using Opus 5.5 on Medium Effort. And
basically what it's doing right now is
looking up and doing research on Family
Feud and Pit My Ride so it can get a
good understanding of what instruments
are used and what the overall vibe is in
both of those so that it can create a
combination song. This is interesting
though. I forgot the Pent My Ride theme
song is kind of like a West Coast thing
and then that's going to be combined
with 70s game show. This is going to be
interesting. I'm going to set audio
influence to let's say 40%. I want to
make sure it does get my melody correct,
but I do not want it to sound like me.
We're going to have the vocal gender set
to female. And let's go ahead and copy
over this exclude styles prompt as well.
And basically, the brain creates this
exclude styles prompt that is just the
opposite of this prompt, if that makes
sense. And I'm going to turn max mode
on. So, you know, after the prompt, I
guess really step three would be
iteration. And I guess we're going to
have more than three steps. And I'm
sorry I lied to you. I didn't mean to.
Iteration just being making a bunch of
different versions of your song. I've
talked to some people who are like,
"Dad, I want to do a bunch of
iterations."
>> Dude, it's like way easier to do this
than to produce something from scratch.
You can, you know, you can hit the
create button a bunch of times. Just
literally just tapping a button.
Sometimes you got to do a lot of
different versions before you find
something that sounds good.
>> Pimp.
>> That's close.
>> Pimp.
that has come the closest. Maybe we can
try weirdness at like 60%, audio
influence 30% maybe. But yeah, the point
is just making a bunch of different
versions of it. Future editing chill
panic here. I just I went through like
probably 30 variations until I found
this one that I liked.
[music]
>> That might be the one. That might be the
one. This is my favorite for show, but
it sounds like, you know, it's not the
best as far as like quality goes in some
parts of it.
>> [music]
>> But it's so close. I'm basically just
going to try to cover this a few times
and see if it sounds any better.
[music]
Now, this is sounding pretty good, but
the vocalist just sounds like a
run-of-the-mill SNO vocal. I want to get
something a little bit more unique. So,
I'm actually going to export both of
these generations and the stems, and I'm
going to use another service to get a
new voice. I'm going to go to the three
dots and uh edit. Go to get stem/midi.
I'm going to get the advanced split. I'm
going to choose lead vocal, drums,
backing vocals, strings, other brass.
And I'm also going to go to our other
generation and go to edit and get
stem/ID. And this time I'm just going to
do split from mix because I only want
the lead vocal from this one. I don't
really care about the instrumental. So I
just chose lead vocal and plus
everything but lead vocals. That way I
can just use this lead vocal as like a
background layer. I'm going to hit
download, go to wave, and go to fixed
tempo as well. Let's see what this lead
vocal sounds like.
Pimp.
>> So, I've exported this part of the lead
vocal so that we can go into another
service and get a new vocal. I'm going
to go to Ottomy, but you can use uh
anything. I mean, they got Ace Studio,
Musicf, there's all types of websites
like this. I'm just going to go to
convert vocals and I'm going to upload
our vocal. And now, I just need to find
a good vocalist.
[music]
>> No.
Okay, that's more of what I'm going for.
And since the audio does have some
reverb and echo, I'm just going to
select remove reverb and echo. We could
have taken that out inside of Suno, but
I didn't. So, we're going to have to
take it out with Ottomy, but whatever.
It's getting taken out. Let's see how it
sounds.
>> N.
And let's go to make harmonies. I just
want some like unison female soul unison
stack. Okay, cool. See how this sounds
all together.
Ah, we're just going to drag all of
these in. I'm going to keep these muted
for now. I just want to mess with this
one vocal.
>> I wonder actually if we keep the same
lead vocal from uh Suna, but we keep
this merry vocal as like a background
thing.
>> All right. And I'm not going to show you
this whole process because then this
would become a mixing video. But we are
going to add in these vocals that we got
from Ottomy. And I just want you to hear
how like even though the vocals are kind
of messed up and they're not saying the
pimp pimp my song like articulated
perfectly every time, it's still going
to sound good and you're not really
going to be able to tell because they're
going to be in the background.
[music]
my soul. B.
>> So, I did a little bit of post-prouction
and some mixing. But before I play the
full thing, if there is anything else
you'd like to know about AI, let me know
down in the comments or I'm going to get
mucuna all over my oblangata. E I N
[music]
my soul. E I N my soul. E [music] I N my
soul. E I am
my [music] soul.

Sunno now limits how many songs you can
download. On the pro plan, that's 20 a
month. So, when a free AI music
generator showed up that runs on your
own PC with no credits and no download
limits, people started calling it the
pseudo killer. Naturally, I had to try
it. It's called UE2, and it's an
open-source music model. And here's the
bold part. On the maker's own benchmark,
it actually scores higher than Sunno V6.
their numbers, not [music] mine. But
that's what we're testing today. I'm
going to run the same four songs through
it and through Sunno V6. Same lyrics,
same style, word for word, and we'll
score every round. One of these rounds
genuinely surprised me. And at the end,
I'll show you a couple of things UEIE2
does that Sunno simply can't. I'm
running UEE2 through Pinocchio. It's a
free launcher that does all the setup
for you. I open it, pick UE2 from the
music models, and we're on the create
screen. If you can't get it set up, come
to our Discord server, and I'll help you
fix it. Hope you all know our Discord
now. We have a lyrics box, style box,
and some advanced controls below. Every
round works the same way. I'll generate
in UE2 first, then paste the exact same
lyrics and style into Suno, and we'll
compare. [music] There are three things
I'm listening for. Does the vocal sound
natural? Is the mix clean or is there
hiss and muddiness? Does the melody have
a vibe you'd actually replay? Let's
start easy. UEI2 loads with a default
song, so I'm going to keep just the
first verse and chorus. The style is
English acoustic pop with warm female
[music] vocals and fingerpicked guitar.
I'll hit generate. And first it
downloads the models. Then it gets to
work. Here we are. [music]
[singing] your coffee.
Caught the first bus out
[music and singing] of town.
Watch the harbor through the
[singing and music] window till the
morning came down.
>> Okay, that's actually nice. The vocal
sounds warm and natural. The mix is
clean with no hiss, and the melody has
that easy, hopeful acoustic feel. for a
free model running locally. That's a
solid start. Now, let's go over to
Sunno. I'm on V6 in advanced [music]
mode. I'll paste in the same lyrics and
the same style. I'm going to drop
weirdness to 25 and push style influence
up to 75. So, Sunno sticks close to what
we're asking for. If you want to get
more out of style box, I actually wrote
a whole book series on it. It's called
the Sunno series. Book one, Sunno
Create, gets you started and it's
completely free. Book two, Sunno Studio,
covers recording, MIDI, and mixing
inside Studio. And book three, Sunno
Prompting, gives you the prompt formula,
the style tags, and 300 readyto-use
prompts. Books two and three just
launched and they're 30% off this week
only. Grab your free code at the link in
the description. Okay, back to the test.
Sunno always gives us two takes. I left
my keys beside your [singing] coffee.
Caught the [music] first bus out of
town. Watch the harbor through the
[singing] window till the morning came
down.
[music] Leave a [singing] light on by
the water. Let it shine across the
[music]
road.
>> Honestly, I didn't expect this. Sunno's
[music] vocals sound just as natural and
the mix is just as clean. If anything, I
think UEI2's take was a little warmer.
I'm calling this one a tie. A free model
level with Suno after round one. Let's
see if it holds. Now, let's make it
harder. Dance Pop is all about the
production, so this is where you'd
expect a free model to struggle. [music]
New lyrics about midnight traffic and
catching the green light. The style is
modern electronic dance pop with [music]
bright, energetic female vocals, punchy
drums, and deep synth bass. I'll hit
generate.
[music]
>> I was waiting
[music]
for a sign I could hold.
[music] Everywhere I
can stand
[music]
till
>> right from the intro, there's a hiss you
can clearly hear. The structure holds
up. It builds into the chorus the way it
should, but the melody feels a bit stiff
and the bass doesn't really punch. Now,
same lyrics, [music] same style. This
time, I'm leaving the sliders in the
middle and turning variety off. create.
[music]
[music]
Every red light kept me standing [music]
still till I heard my heart be louder
than the wheels. You call my name
from the other side.
Weo [music]
has a touch of hiss, too, but it's much
less. The melody flows more naturally
and the bass hits harder. Round two goes
to Suno. And next round, UE2 makes a
mistake you'll only catch if you're
paying attention. Let's switch to a male
voice. The style is acoustic indie folk
with fingerpicked guitar and soft piano.
We got the lyrics. We have a chorus.
I'll hit generate. Count the chorus
lines as it plays.
[music]
Light on for me.
>> [music]
>> Like a little star [music] through the
rain.
And the longing [music]
is already home.
If the night [music] gets long and the
rain comes down.
>> Did you catch it? In the chorus, UE2
skips a whole line and even drops a few
words. The male voice itself sounds warm
and [music] the mix is clean, but when
it skips lyrics, that is a turnoff for
me. Now, Sununo, let's try with the
exact same lyrics. [music]
[music]
I've got [music] miles to go and
[singing] roads unknown,
but a part of me [music] is already
home.
If the night [music]
gets long and the rain comes down,
>> sounds a little muffled here, so the mix
isn't perfect, but it sings every single
line and has a really nice, gentle vibe.
So wins round three. Last round and it's
the loudest. Let's go for energetic alt
rock. Gritty and powerful with
aggressive drums. I'll hit generate.
[music]
[music]
Too late.
[music]
Every open door has another game.
[music]
>> The energy is there and the [music]
vocal fits the genre, but the whole mix
sounds muffled like it's behind a wall.
And Sununo, same lyrics, same style.
Create
a promise to keep. [music] Every
headline says that I'm too late. Every
open door has another gate. But there's
a drum beat [music and singing] under
this concrete. I feel it rising through
the souls of my face. [music]
Turn it up till the window.
I won't [music]
>> mix is a bit boxy, too. So on sound
quality, it's even, but both Sununo
takes have way more vibe in the melody.
So round four goes to Sunno as well.
That's the final score. So 3 UE20
one tie. I know what [music] you're
thinking, but this next part is the
reason why many people still prefer it.
First, re-rolls [music] cost nothing.
For this part, I've switched to an Afro
Beats track. I'll generate it again with
[music] the exact same prompt. And
again, each time I get a brand new take
in about a minute, and there's no credit
cost. On Sunno, every one of those tries
costs you. Second, there's a built-in
enhance button that rewrites your prompt
for you. Let's run it on the Afrobeat
song. Here we have our enhanced prompt.
Let's generate
[music]
our time. Let the whole world spin
[music] the stars alive. No rush
tonight. We can take our time.
>> To my ear. This one has a bit more of
that polished feel, a bit similar to
Sunno's max mode trick. If you missed
it, we already covered Sununo Max mode
in a full video. Make sure to check it
out. Before the verdict, let's talk
price. UEI2 is completely free. Here's
Sunno's current pricing. The free plan
gives you daily credits, but no
commercial rights and just seven
downloads. Not per month, ever. Pro is
$10 a month for 20 downloads. Premiere
is $30 a month for 60. So, after all,
can this free model beat Suno on Pure
Sound? Not yet. So won three of our four
rounds. Its melodies feel more natural,
the bass is stronger, and it never
dropped a lyric. Remember that benchmark
where UE2 beats Suno V6? On real songs,
it didn't hold up. Benchmarks are one
thing. Your ears are another. But
[music] think about what UE2 just did.
It tied the first round and came close
in the others. All on a home PC with no
credits, no download caps, and unlimited
rolls. So, the real question isn't which
one sounds better anymore. It's is that
gap worth paying for every month. From
what we tested, UEIE2 is a great fit if
you're a creator who needs lots of
music, background tracks, demos, and
drafts without watching a credit
counter. Someone who likes to experiment
because free-rolls mean weird ideas cost
you nothing. Individual creators can
also monetize what they make under UE2's
license. while companies need a
commercial license. But if you want the
most polished song on the first try with
zero setup, Sunno is still the safer
pick. So, would you pay for that gap or
go free? Tell me in the comments. And if
you know someone who's paying for Sunno
every month or looking for a free way to
make music, share this video with them.
It might just save them some money.

Last week, I vibe coded Bluey, an iOS
app that connects to my computer,
follows my mouse, and acts as a little
agent that can point to things on my
screen, and say things. Well, today I'm
going to turn Blueie into an app that
helps me with my ADHD, and keeps me on
track. Okay, so here's the plan for the
day. We're going to talk about what
Bluey is now. We're going to talk about
what I want Blueie to be, which will be
a little focus friend. And then we're
going to talk about how we're going to
create this focus friend using the
decisions API. And then we're going to
just build this new version with codec.
And you can follow along and build your
own version of Bluey because Bluey is
fully open source. So the first things
first, what is Bluy? What is this app
that I made a few weeks ago? I have my
old phone that was just sitting around
doing nothing and I wanted a way to use
it as an extra screen with my computer
in a unique, fun, and a simple way. And
I wanted it to feel like a little
companion. For those of you who have
ADHD, you know that if you have someone
around, something about it, it just
makes it easier to focus. And so that's
kind of my vision. When I originally
created it, I just didn't think about it
well enough. And so basically when I
hold Bluey down and ask it a question,
it points to the part of the screen and
leaves a little comment directly on the
screen and makes a little cute noise.
Hey, can you point to the computer
monitor?
What else should I add to this about
Bluey?
Can you research the decisions API? Um,
I want to learn more about it.
And so look, it can even do some
research here at the top. And so it can
do a little research. And so the way
Bluey is controlling my desktop computer
is it's using a different OpenAI API
which is called OpenAI Realtime API or
GPT realtime 2.1. And so it's also using
uh GPT40
transcribe. This is my voice to text
when I hold down my phone. Hey, uh, can
you please point to the little blue
character on the screen?
You see what it's doing here is when I
hold down my phone, it records my voice
from my phone and then it uses this GPT
real time to control my screen. Now, you
might be thinking, wait, this is my
phone running below my computer, right?
This right here is my phone. This right
here is my computer. How are they
communicating? Well, I asked AI last
week. I said, I want to build an iOS
app, right? And so, Bluey is the iOS
app. And then there's also a desktop
app, which is a Mac app running on my
computer. And you can see this is
running right here. When you set up this
app, you can very easily scan a QR code
and connect your phone to the desktop
app. And so they work together. And so
that's why it can follow my mouse
around. It can literally it's sending
data from the Mac app to the iOS app,
which is where Bluey should be looking
at all times. And it makes it feel like
you almost have like a little buddy
there watching you as you work, which is
kind of fun. And so the Mac app and the
iOS app are inside of a mono repo which
is in this GitHub link. And all that
means here is the Mac app and the iOS
app lives inside this repository. So all
the code you need for both apps are
right here. And so basically Bluey right
now is just cute. It's kind of vague. It
doesn't really have a purpose. It can do
some things. I guess it can kind of
search the internet, but I don't want it
to be this general thing. I want it to
be practical, purposeful, and solve a
specific problem while maintaining its
cuteness. Okay, so now for the
technology behind this new version of
Bluey, and we'll call it focus Bluey.
And we're going to be using the
decisions API. So the decisions API is a
specialized inference API powered by
GBT6 Luna designed to help apps make
fast focused decisions. You give it
context and a specific question and it
can return a probability, pick from
options you provided or give you a score
and your app then uses that result to
decide what happens next. Said another
way, the decisions API turns text and or
images into small specific decisions for
an app that you use or one that you
create. And what makes the decisions API
really interesting is the output is
free. There's actually no output cost
when you use the decisions API. The
decisions API gives you an answer based
on a set of possible outputs that it can
have. So it basically just makes a
decision. It doesn't output any text.
And there's actually three types of
outputs it can give. The input is text
can be text and images. But there's
three types of outputs. And the first
one is called predicate. and a predicate
estimates the probability that a
statement is true. The second one is
called a choice. And the choice
basically selects from options you
define with probabilities, right?
Probabilities for the options and a
confidence
score. And then the third and final
output is called the score. And I guess
which basically means like rating
something on a scale you define. And so
what we're using for this video is
choice. So we're using this specific
type of output. And so let me explain
what I mean by this. Bluey is going to
be running below our computer. And this
is just a little character on a spare
phone that we have. I'm going to be able
to hold it down and tell it what task
I'm about to do. And then it's going to
start a timer. And the timer is going to
show up here at the bottom right. And
then it's going to list the task up
here. Now, what this app is going to do,
it's going to take a screenshot of my
computer every 30 seconds. And we're
also going to have a system prompt. So,
every 30 seconds, a prompt plus a
screenshot of whatever I'm doing on my
computer at any given time. This will be
sent to the decisions API. Sorry, I
spelled that wrong. It's okay. It's
going to be sent to the decisions API
and there it's going to make a choice.
And this is very fast and very cheap.
It's either going to be on task, off
task, or uncertain. And then we're going
to have AI. I'm going to have Codeex
come up with some algorithm as to when
Bluey should intervene with whatever I'm
doing. I don't think if it's if it says
that I'm off task. I don't think that
every time it deems me off task, it
should intervene or comment on it. Maybe
he like gets a little more conf it gets
a little confused look, but it should
actually intervene and stop me after
I've shown repeated like repeatedly off
task. So like every 30 seconds it'll
basically fill a database of like
whether like he's off task on task and
the algorithm can also factor in these
probabilities for the options. It'll
basically say like he's on task 100%. If
I go to blueie and say all right for the
next 30 minutes I'm going to be working
on my video script and 100% confidence
score would be like would see that I
literally have the script open and I'm
writing and that would be 100%
probability. But maybe I'm doing
research on a topic that's related to
the essay. And so we need to figure out
a way to do this in a smart way so that
it doesn't get annoying with how much it
intervenes, but if I'm clearly off task
for like 5 minutes, it should definitely
intervene and try and nudge me to get
back on track. And so that's the point.
And so the decisions API, what's really
cool about the decisions API is it's
incredibly [snorts]
cheap. You can see that a 100 screenshot
checks is a scent. Eight hours every 30
seconds is 9 cents. So that means eight
hours of focus costs about 10 cents,
which is pretty cool. Okay, so now that
we know what the decisions API is, we
are going to talk about the prompt that
we're going to run in codeex. And I
think we can oneshot this app using the
decisions API by OpenAI. We're going to
want to put in the link to the docs for
this. Um, but here's the prompt that I'm
using. rework my existing uh Bluey Mac
OS and iOS apps into a focus companion
still called Bluey and I'm giving it the
GitHub repo. You can use this GitHub
repo. If you want to clone this app, you
will need an OpenAI API key uh link in
the description for that. And it's going
to be uh read the repo and attach grants
first and reuse the current architecture
blah blah blah blah blah. So, it's going
to reuse all the features that we have
already. Now I want to change the
purpose of Blueie will become a focus
friend monitoring my screen and when I'm
off task it lets me know I'm off task
and points at the thing when it makes a
comment that should be sort of playful
and funny like hey this doesn't look
like current task that you're doing and
I just kind of describe it how it's
going to work every 30 seconds have the
Mac capture a screenshot and send it
with my task elapse time and idle
duration and recent check history to the
decisions API cuz Remember, all of this
will be sent as an input. So, I'm going
to have codec come up with the right
inputs. So, it's going to be screenshot
plus certain inputs, and it's going to
be judge the activity in context.
Watching a codeex tutorial can support
making a codeex video while unrelated
entertainment may not. So, it's I'm
either on remember I'm either on task,
off task, or it's unsure. And each of
these will be rated on how confident the
decisions API is at whether or not I'm
on task, off task, or unsure. And so
we're going to be implementing the nudge
logic in the app. So allow for brief
detours and only interrupt after
repeated confident offtask checks
spanning at least 2 minutes if close to
100% certain. If it's like in the 50 to
80% certain range and not fully certain
the user is being off task, maybe it's
every 3 to four minutes. And I'm saying
come up with an algorithm to determine
when to intervene, etc., etc.
Implementing a short cool down between
nudges. That's what I want to do. And
then we're just going to build the full
app end to end. So that's basically a
quick overview of the prompt. All the
prompts will be linked in the
description. Now what I'm going to do is
we're going to go back to chat GPT here.
And now what we're going to do is we are
going to build the app. So I'm just
going to enter this prompt here. And I'm
just going to say update the apps and
let's test both of them and get them
running on computer and phone. So this
we are using GPT 6.1 soul extra high.
This will probably take 10 to 20
minutes. So I'm going to wait a little
bit and we'll see how it does. Okay, so
it just worked for 18 minutes. And I'm
going to make my phone a little bit
bigger. So my phone is actually below my
computer, but I'm screen sharing it on
my computer, so it would kind of show up
below the computer here. So this is what
it looks like. Now, it says on your
iPhone, pair the Mac or scan QR code.
And so I think we need to come up to
Bluey and I think we need to pair
iPhone. There we go. So, I'm going to go
ahead and leave this up. Now, I'm going
to grab my phone. I'm going to hit pair
allow. Did that work?
Okay. So, we are paired. So, Bluey
integration test build focus friend
Macup. Okay. Okay. Hey, buddy. I want to
start a task now, which is building an
app using codecs. It is an app that
allows me to talk to my phone. It's a
little bluey app.
There we go. All right, let's build
Bluey. Eyes on the prize, mate. And look
at this. So now you see pause, end, and
focus here. And remember, this is below
my screen right now, but it's like pair
your Mac and focus together. And it
should be focused. So now what I want to
do here, I'm going to be building an app
with chat, GPT, and Codeex. And I want
to do this for five minutes.
Okay, there you go. So, look at this.
It's kind of small, but it's showing a
timer for 5 minutes. And let's just tell
Codeex all of these things. So, the
things that I want to change is I need
the a timer on the bottom right. That's
big white text that has the timer. The
timer is way too small. And then make
the pause, end, and focus buttons like
little icons on the bottom left. And
yeah, those are the changes that I want
you to make. And then the listening hold
to ask. Make it an icon. Put it on the
bottom left as well. Maybe we don't need
that. But I want the timer to be the
main focus for this. Yeah. So make those
changes.
Okay guys, so it's done. One thing I
added was a little developer mode. So I
wanted to make it faster so we can
actually test it faster. So it'll
actually moves at 4x the speed. I want
to start a task real quick, which is to
use chat GPT codeex to build an app.
Creating my little blue character.
Okay, so it's now working. If we go to
focus mode here,
we should see
Okay, so we are on task for the first
one. Um, let's go back to Bluey here.
Now, what I'm going to do is I'm going
to go off task
and let's see what it looks like to be
on Twitter.
Hey, this doesn't look like Chat GPD
codec. Shall we hop back to Blue Buddy
app? There we go. So, it's nudging me to
get back on task. There we go. Okay, so
we're still off task. Okay, so this is
cool. It's working. And let's go ahead
and stop. And so here is the session.
It shows that we did it for 7 minutes
and 30 seconds and it shows how we did
for each part and we were off task for
most of this session here. So this
works. We got our app working. And so it
basically decides every 30 seconds with
the decision API whether we are on task
or whether we are off task. All right.
So here is the app now that it's done.
It sits below my computer screen. I have
it set up on this little like magnet
thing. I can put a link in the
description to better options. We have
Louis here. Hey little guy. I just want
you to know that we need to work for the
next 20 minutes on writing an essay.
So, it created the timer. It's on dev
mode, so it's moving pretty fast. And
now I can just work on the thing that
I'm working on, right? I can just do
whatever task I want to work on. And
right now it might think that I'm on
task because we are on the Codeex app.
But if we were to go to Twitter,
let's see if it intervenes if we're on
Twitter.
Look at that. Hey, this doesn't look
like writing an essay. Shall we hop back
to the draft? There we go. Bluey works.
And we can always view our progress.
Right here we can see a summary of when
we are on task. That's the green. And
then the brown is when we are off task.
And it keeps track using the decisions
API. And that is really cool. This is
now our little focus friend bluey that I
can very easily connect just by scanning
a QR code. It's that easy. And I'm just
going to say okay, push this to GitHub
and update all the changes. And you guys
will be able to create your own Bluey.

Around the world, people are worried
about AI. Not just because it might be
used to hack our systems, not even just
because it might take our jobs, but
instead because by making this
intelligence that we've long felt unique
to us abundant, we worry about what it
means for our own relevance. This is not
in fact the first time that we've dealt
with these issues. And 400 years of
economic history suggests that something
different and in fact quite wonderful
might happen instead. The AI Daily Brief
is a daily podcast and video about the
most important news and discussions in
AI.
All right, friends, quick announcements
before we dive in. First of all, thank
you to today's sponsors, Robots and
Pencils, C Data, Granola, and Blitzy. To
get an adree version of the show, go to
patreon.com/Iailybrief
or you can subscribe on Apple Podcasts.
And if you want to learn more about
sponsoring the show, send us a note at
sponsorsaily.ai.
Welcome back to the AI Daily Brief.
Today we have a cool one for you. The
original idea of long read Sunday was to
take the most interesting big ideas that
I had found that week surrounding AI,
whether it was from a technology
perspective or AI in society, share
those writings and then broaden out the
conversation from there. It's evolved a
bit from when we started. Sometimes it
is more oriented for example towards the
practical. But every once in a while
there is a good old-fashioned big thinky
piece of writing that I get very excited
to share. And this week that's exactly
what we have. There is a piece that has
been making its way around AI circles
called cognitive mercantalism. It's by
writer and strategist Venitesho and I
caught notice of it when I saw a bunch
of people whose perspectives I admire
saying how surprised they were to find
something so original and
thoughtprovoking around AI. So as is the
style of long read Sunday, I will read
some big chunky excerpts from this and
then we will discuss it. And for fun, as
I'm preparing this, I have Claude Code
in the background working through some
illustrations to keep it a little bit
more interesting for those of you who
are watching. Venitates writes, "I've
been reading Joyce Applebee's The
Relentless Revolution, a history of
capitalism, and it has me circling a
rather delicious thought. AI panic is
really a form of cognitive mercantalism.
If we regard market capitalism as the
first planetary intelligence, as many,
including me, do the parallels are
closer than you might think." Applebee's
book is a sweeping history of capitalism
from its unlikely emergence in early
modern England to its eventual
transformation of the entire world. Her
central argument is that capitalism was
not the inevitable expression of some
timeless human instinct to trade and
accumulate wealth. It was a historically
contingent cultural invention enabled by
agricultural improvements, expanding
trade, and changing ideas about work,
wealth, and human nature. In particular,
it required escaping the assumptions of
an older, essentially zero- sum
worldview in which wealth meant
controlling land, bullion, and the labor
of largely impoverished populations.
Capitalism emerged when people began
discovering that wealth could be created
rather than merely captured, and that
ordinary people could be sources of
demand and initiative rather than merely
labor to be commanded as a factor of
production. The parallels with our
current moment are striking. As Abby
tells it, the mercantalist worldview
that preceded capitalism rested on
roughly four interlocking beliefs.
First, wealth was fundamentally scarce
and finite. It consisted primarily of
land, the agricultural produce of that
land, and the precious metals it
contained, accumulated through mining,
trade, and conquest. The goal of
economic statecraftraft was to control
more territory, maintain a favorable
balance of trade, and accumulate
bullion. One nation's gain was presumed
to be another's loss. The resulting
politics was one of treasure seeeking
and beggar thy neighbor competition.
Second, ordinary people could not be
trusted with much freedom. The poor were
assumed to be unruly, ignorant, and
prone to disorder if left to their own
devices. They needed to be governed
coercively by elites, supposedly for
their own good and for the stability of
society. Third, economic thinking was
overwhelmingly oriented towards supply
and its control. Wealth came from
controlling productive resources and
commanding high prices for their output.
Aggregate demand, especially demand from
the poor, was barely recognized as a
significant economic force. Ordinary
people were primarily factors of
production. They ate, worked, and
reproduced with no share of any surplus.
The only real discretionary consumers
were elites who bought luxuries and
commanded all surplus. Fourth, there was
little reason to imagine most people
doing anything other than agricultural
labor. If something like 80% of the
population was engaged in inefficient
subsistance farming, that was simply the
natural order of things. What else were
they good for? Remove them from the
fields and you would merely have
millions of useless mouths to feed. Now,
replace wealth with intelligence. and
mercantalism with cognitive
mercantalism. The parallels are almost
embarrassingly direct. First,
intelligence is imagined as a fixed
resource embodied in human brains and
the knowledge products mind from them.
The objective of national and corporate
strategy becomes attracting the best
brains, controlling their output,
accumulating intellectual property, and
maintaining a favorable balance of
knowledge trade. The cognitive
mercantalist wants to hoard intelligence
and its products like bullion. The
global competition for talent becomes a
treasure hunt for genius brains and the
valuable knowledge species they can
produce, which now includes AIs in
addition to copyrightable culture and
patentable inventions. Second,
supposedly stupid people cannot be
trusted to think for themselves. They
need to be governed by cognitive elites
who know better and must regulate the
behavior of the less intelligent for
their own good. Democratizing
intelligence is dangerous because the
masses will misuse it, fall for
nonsense, or unleash chaos. The
assumption is that intelligence is safe
only in the hands of those already
certified to possess it. Third, the
supply of intelligence is treated as the
central economic fact. Controlling its
production, preserving its scarcity, and
maintaining high prices for brains and
knowledge products, including AIs, are
the primary concerns. The demand side
barely registers. Less intelligent
people are regarded as limited cognitive
laborers and passive consumers of
whatever knowledge products elites
choose to provide. They are not imagined
as significant sources of demand for
intelligence in their own right, let
alone as participants in its further
production. Fourth, the cognitive
mercantalist assumes that most people
are not good for much intellectually. If
80% of humanity is occupied in low-grade
cognitive labor or simply muddling along
without producing much intelligence at
all, that is unfortunate but natural.
Automating their cognitive work merely
threatens to create billions of useless
minds to occupy. This last assumption is
where the analogy gets especially
interesting. The agricultural revolution
that helped launch capitalism began
overturning the apparent necessity of
having most of humanity engaged in
farming. Over the following centuries,
improvements in agricultural
productivity reduced the fraction of the
population required to produce food from
a large majority to a tiny minority in
advanced economies. The result was not a
world full of useless mouths. It was a
world in which vastly more people could
do things other than farm while enjoying
access to vastly more food. I don't know
how to farm. I can barely keep a house
plant alive. Yet, I can obtain an
astonishing variety of foods in
quantities that would have seemed
fantastical to an early modern peasant.
My inability to produce food is
economically almost irrelevant. What
matters is that I can consume it, enjoy
it, develop taste for it, and devote my
energies to other things. The analogous
possibility with AI is that most people
will no longer need to be significant
producers of intelligence in order to
become significant consumers of it. And
this is where cognitive mercantalism
begins to look as quaint as its 17th
century predecessor. Now, at this point,
and this is NLW cutting in again, I want
to go back and reference an episode I
did a long time ago, all the way back in
May. The episode was called the new jobs
AI will create. There's still a
companion experience available online.
And it was my attempt to show not
precisely exactly what jobs I think AI
will create, but to go beyond the tired
trope that of course AI will create new
jobs, which is what all of us optimists
think, and instead try to actually
identify some types of jobs or
categories of jobs that might result.
One of the big parts of my argument was
that the lump of labor fallacy that runs
through so much of thinking about AI job
loss that there's only so much work to
be done and if machines do it humans
won't be able to do it has another
interconnected piece which is an
assumption that there is no elasticity
of demand that we demand as much right
now as we could ever demand. My main
point is that that is crazy and
completely ahistorical when it comes to
human evolution. Instead, I believe that
our capacity to increase our demands is
nigh limitless. In fact, in some ways,
it is constrained only by physical
forces and universal laws like how much
time there is in a day and how many
calories we can consume without
bursting. I won't go too deep into it
now, but in that episode, I discussed
demand elasticity across six different
dimensions. Price elasticity, I wanted
it, but it cost too much. Access
elasticity, I wanted it, but I couldn't
get it. Complexity elasticity. I needed
it but the system was too confusing.
Continuity elasticity. I get help
occasionally but I'd benefit from help
all the time. Personalization
elasticity. I get the generic version
but I value something made for me. And
relational elasticity. I want this to be
more human, meaningful, and trusted. All
of these are different ways in which AI
could change systems that result in more
demand becoming unlocked. Okay. Okay, so
with that in mind, let's come back to
the next section of Benitesh's piece,
the revolution in desire. He continues,
"One of the more fascinating elements of
Applebee's account is her emphasis on
the emergence of new kinds of demand in
Britain. Well before the familiar story
of steam engines, factories, and
industrial capitalism really gets going.
As international trade expanded,
ordinary Britons began developing tastes
for goods that had previously been
luxuries reserved for the wealthy.
Tobacco, tea, sugar, and colorful Indian
cotton were among the commodities that
helped create a new consumer culture.
People began wanting things that were
neither necessities of subsistence nor
traditional markers of aristocratic
status and they were willing to
voluntarily work harder to get them.
Applebe describes a striking increase in
the number of days worked annually from
something like 250 to 300 before the
industrial revolution. This was not
simply a matter of landlords or factory
owners coercing more labor out of
unwilling workers. A significant part of
the increase was voluntary driven by the
prospect of acquiring desirable goods.
People worked more days, adopted more
productive agricultural practices,
entered new trades, and sought
opportunities for earning larger cash
incomes. The increase in productivity
generated surplus incomes, which
sustained the expanding demand for
consumer goods, which in turn
incentivized further increases in
productivity. A virtuous cycle was
taking shape. Zooming forward a bit, he
continues, there is an important
historical inversion here. We are
accustomed to thinking of the politics
of labor primarily in terms of
resistance to exploitation. The struggle
to escape servitude, coercion, and the
extraction of surplus by owners. That is
certainly part of the story and a very
important part. But Applebee's account
suggests that another equally
fundamental force was at work. The
desire to participate in an expanding
universe of consumption opportunities.
Modern labor politics can be understood
partly as the politics of relative
deprivation in a society of expanding
possibilities. Once people discover that
there are desirable things beyond
subsistence and that their labor can
earn them access to those things, the
political question changes. It is no
longer merely how to escape domination,
but how to secure a fair share of the
growing surplus and the opportunities it
makes available. This is a different
kind of demand for freedom, not merely
freedom for coercion, but freedom to
participate in the expanding
possibilities of economic life. The
demand for a fair share of expanding
consumption opportunities is arguably
one of the defining political inventions
of capitalism. And this is a much more
interesting model for thinking about the
AI transition than the familiar story of
automation replacing jobs.
At this point, it's no longer a question
of whether companies are actively using
AI. Using it well, on the other hand, is
a whole different story. Robots and
Pencils, though, is a company that I can
point to that is actually built for this
time. They're an applied AI engineering
firm working directly with clients on
problems that matter to the business,
not experiments that live in a slide
deck. Every engagement starts by working
backwards from the outcome a client
actually needs. If you're trying to tell
real AI engineering apart from noise in
this space, that's the differencemaker.
Head to robots and pencils.com.
This episode is brought to you by
Cadata. Token spend is climbing,
security is holding agents back, and
companies rolling out AI initiatives
might not know how it's truly being
used. The root cause, AI is cut off from
the systems and context your business
runs on. That's why Cata built Connect
AI gateway, one control point between
your AI and your systems. Connect AI
gateway is not just another static model
router. Prompts get intelligently routed
to the right model with exactly the
tools and data it needs governed down to
the record. The result, accurate answers
at a fraction of the token cost with a
context engine that learns your business
with every interaction. Built on the
connectivity layer powering Palunteer,
Google, Salesforce, and hundreds more.
Give it a try and start your free trial
at cdata.ai. That's C as in connect
data.ai.
You're in backtoback meetings all day.
You're trying to stay present, but
you're also worried you'll forget the
decision, the action item, the important
next step. This is the scenario that we
absolutely live here at the AI daily
brief given how fast things move and
change and how quickly we're evolving
everything that we do. That's where
Granola comes in. Granola is an AI
powered notepad for meetings. You jot
down rough notes like you always do. And
in the background, Granola transcribes
and turns them into clear, useful notes
when the meeting ends. No bots joining
your calls, no distractions, just a
clean notepad that helps you focus.
During or after the call, you can chat
with your notes, ask Granola to pull out
action items, help you negotiate, write
a follow-up email, or even coach you
using recipes, which are pre-made
prompts. Once you try it on a first
meeting, it is very hard to go without.
Head to granola.ai/Iaily
and use code AI to get 100% off for 3
months for new users. That's
granola.ai/I
and code AI to get 100% off for new
users for 3 months.
Every AI coding tool on the market does
the same thing first. It starts writing
code. Blitzy does the opposite. Before
writing a single line, Blitzy spends
days reverse engineering your entire
codebase. Thousands of agents ingest
millions of lines, mapping every
dependency, every undocumented
constraint, every architectural decision
made over the last decade. The result is
a dynamic knowledge graph that
understands your software the way a
principal engineer would after 30 years
in the building. Other tools guess at
context with GP searches and markdown
files. Blitzy never guesses. It builds
true understanding first, then delivers
over 80% of entire software epics
autonomously. Validated end-to-end
tested production grade pull requests.
That's why Fortune 500 engineering teams
trust Blitzy with the code bases that
matter most. See for yourself at
blitzy.com. That's blitzy.com.
The cognitive industrious revolution.
The entire intellectual economy has been
organized around the assumption that
intelligence is expensive to produce and
therefore must be carefully rationed.
Education, professional expertise,
research, consulting, software
development, and knowledge work more
generally are all shaped by the high
cost of cognitive production. AI is
rapidly undermining that assumption. We
are moving toward a world in which an
ordinary person can command quantities
and varieties of intelligence that would
have previously required teams of highly
trained specialists. Just as I can enjoy
food without knowing how to farm, I can
increasingly enjoy the products of
intelligence without knowing how to
produce them myself. The crucial
economic question then becomes not how
much intelligence people can produce,
but how much intelligence they can find
uses for. And there is no obvious upper
bound on that demand. This is not simply
a matter of everyone consuming more
answers, reports, software or
entertainment. Abundant intelligence
changes what people can want in the
first place. It makes previously
unimaginable projects feasible,
previously inaccessible domains
explorable, and previously unaffordable
curiosities worth pursuing. The demand
for intelligence is not a fixed appetite
waiting to be satisfied. It is itself
something that can grow as the cost of
satisfying it falls. Here the analogy
with the early modern consumer
revolution becomes especially
suggestive. Just as colorful Indian
cotton and imported tobacco stimulated
new desires and new desires stimulated
greater voluntary labor and more
productive forms of work, abundant
intelligence may stimulate a new wave of
voluntary cognitive effort and open up
new ways in which you can work harder.
Many people who appear stupid are not
incapable of thought. They are simply
not particularly motivated to think
given the poor returns available to
them. Why expend significant cognitive
efforts when you lack the knowledge,
tools, institutional access, or
intellectual leverage to do anything
interesting with the results? Abundant
AI changes that calculation. Someone who
previously had little reason to think
hard about a problem may suddenly
discover that a modest investment of
attention unlocks enormous
possibilities. They can build something,
investigate something, understand
something, or pursue an ambition that
was previously out of reach. The returns
on attention rise and with them the
incentive to pay attention. And here,
Venitesh offers a different explanation
than we've seen for why it feels like AI
is not making us work less, but making
us work more. Now, this is something
that I've also offered an explanation
for talking about a concept I have of
the infinite backlog. The idea that at
work there is always a basically
neverending set of work to be done that
runs off infinitely into the future. The
reason that the infinite backlog doesn't
overwhelm us is that because there are
natural limits to the amount of work we
can do. We're basically constantly
calibrating what is an appropriate
amount of that backlog to have taken on
and gotten accomplished. Done is not
some fixed natural state. It's a
negotiation with your own infinite
backlog based on your own goals or those
of your employer. The reason I posited
that AI and specifically agents break
that paradigm is that when you can
replicate yourself theoretically an
infinite number of times and those
infinite replicas can work in parallel
and non-stop without ever having to
sleep or eat or otherwise get
distracted. It feels like all of a
sudden there is no longer an appropriate
done. The entirety of the infinite
backlog becomes unlockable and the only
constraint is your ability to imagine
what is on that backlog in the first
place. Now obviously as this was my idea
I think there is something to it but
it's still in many ways only dealing
with the anxiety side of AI making us
work more and many of you I am sure will
have felt the opposite to that which is
the excitement the I can't make myself
go to sleep because I just want to build
one more thing side of AI and that's
what Venitesh is talking about here. He
continues, "The early capitalist worker
voluntarily increased annual working
days from 250 to 300 because a new
universe of desirable goods had become
accessible. The AI enabled thinker may
similarly increase the amount of
voluntary cognitive effort they expend
because a new universe of intellectual
consumption has become accessible. This
has certainly been the case with me. I'm
suddenly willing to work significantly
harder compared to even a year ago
because a lot more is suddenly within
reach." We might call this a cognitive
industrious revolution preceding and
perhaps enabling a fuller AI industrial
revolution. This does not mean that
intelligence becomes relevant as a human
capacity. Quite the opposite. Just as
abundant food supported the emergence of
elaborate culinary cultures rather than
eliminating interest in cooking,
abundant intelligence may simulate
increasingly sophisticated cultures of
thought. People may become more
intellectually ambitious precisely
because they no longer have to do all
the cognitive production themselves. And
if this happens, the politics of AI will
likely evolve in a manner analogous to
the politics of industrial capitalism.
The central political demand may not be
for the preservation of existing
cognitive jobs any more than the
enduring political demand of industrial
labor was for the preservation of
agricultural employment. It may instead
become a demand for a fair share of
expanding cognitive consumption
opportunities. Access to intelligence,
the ability to command it, the freedom
to pursue projects with it, and the
institutional rights necessary to
participate in an increasingly
intelligence-rich economy may become the
foundations of a new kind of labor
politics. This would be the labor
politics of cognitive abundance rather
than cognitive scarcity. From there,
Venitesh gets into another important
analogy, comparing mercantalism to
Malthusianism. He notes that while they
are not the same doctrine and
historically belong to different
intellectual moments, they quote share a
family resemblance in their tendency to
imagine human economic life as
fundamentally constrained by fixed
resources. The mercandalist, he
continues, worries about controlling a
finite stock of wealth. The Malthusian
worries that population growth will
inevitably outrun the productive
capacity of the economy, condemning most
people to subsistence. both have
difficulty imagining a world in which
technological innovation and expanding
demand fundamentally transform the
relationship between population
production and consumption. We can now
identify an analogous tendency in
contemporary AI discourse. AI
Malthusianism, the AI Malthusian
imagines that intellectual production
will expand so rapidly that human beings
will become economically redundant and
the surplus intelligence will
necessarily turn toxic in some way. This
is a peculiar inversion of classical
Malthusianism. Instead of population
outrunning the supply of subsistence,
the supply of intelligence outruns the
supposedly fixed capacity of human minds
to absorb it. The result is the same
gloomy conclusion. Collapse. The
cognitive mercantalist fears losing
intelligence of scarce intelligence. The
AI Malthusian fears the consequence of
abundant intelligence. Both assume in
different ways that the demand for
intelligence is fundamentally limited.
But what if that demand is as elastic,
inventive, and historically
transformative as the demand for
material goods turned out to be? What if
the apparent surplus of human minds is
an artifact of our inability to imagine
what people might want to do in an
intelligence-rich world? The history of
capitalism suggests that this is at
least a possibility worth taking
seriously.
The demand side of AI, there is also a
more radical extension of the argument.
The history of capitalism suggests that
this is at least a possibility worth
taking seriously. Now from there,
Venitesh goes even farther into some
future explorations. He basically
explores the possibility of what happens
if AI is not solely a supply side
technology but also one which produces
demand. Now, this section frankly
deserves an entire second essay because
my sense is that Benitatesh is trying to
articulate not AI producing demand
because they have somehow developed
consciousness, but because the systems
they architect and get architected
around them create pathways and needs
for more intelligence, consumption, and
demand. As he puts it, although without
explaining too much further, many
products and services exist primarily
because the market demands it, with
humans being secondary consumers.
Venitates's final section is called
mercantalism was not for merchants. An
irony he writes in the historical
precedent is that mercantalism was not a
philosophy of by or for merchants. It
was largely a philosophy of militaristic
statesmen, aristocrats and reentier
elites who viewed commerce through the
lens of territorial power, unearned
privileges and accumulation. Actual
merchants, entrepreneurs, and innovators
were often discovering possibilities
that the prevailing political
imagination could barely comprehend and
steadily fought the self-limiting
tendencies of mercantalism, giving birth
to capitalism as we understand it.
Today's cognitive mercantalists have a
similar relationship with intelligence,
starting with the fact that most of them
are not the merchants of AI. Even
business leaders of the AI sector are
behaving like cognitive mercantalists in
direct proportion to their desire to be
more than mere merchants. Across the
political spectrum, many elites remain
attached to the idea that intelligence
is a scarce, dangerous resource that
must be managed by credentialed
custodians, kept expensive, used
directly or indirectly for conquest and
treasure, and denied to rivals. In the
AI industry, to the extent elite
political and cultural venturing by CEOs
distorts behaviors that are justified by
the actual risks, it is mercantalist
rather than capitalist. The prospect of
its democratic abundance in the form of
global diffusion, open weights and such
is therefore experienced primarily as a
threat to expertise, cultural and
political authority, status, social
order, military dominance, and
geopolitical leverage. What gets missed
is the possibility that we are
witnessing the beginning of a transition
as profound as the one from an
overwhelmingly agricultural society to
an industrial and commercial one. From a
world where 80% of human brains are
engaged in subsistance cognition to one
in which 95% are engaged in satisfying
previously unfulfillable cognitive
demands. The most important consequence
of the agricultural revolution was not
that farming became more efficient. It
was that most people were liberated from
the necessity of farming allowing
entirely new kinds of economic and
cultural life to emerge. Likewise, the
most important consequence of the AI
revolution may not be that intelligence
production becomes more efficient. It
may be that most minds, human and
artificial, are liberated from the
necessity of producing intelligence
merely to survive and become free to
discover what they actually want to do
with it. The cognitive mercandalist sees
a world threatened by an overupp of
intelligence and an impending surplus of
useless minds. The AI Malthusian sees a
world in which intelligence outruns the
capacity of humanity to absorb it,
leaving a growing population of
economically redundant minds primed for
revolutionary explosions. The cognitive
capitalist sees the possibility of an
almost limitless expansion of demand and
a world full of minds whose potential
has barely begun to be explored. So that
is the end of the essay and hopefully
you can see why I was excited to share
it. Even for those of you who are
furiously scribbling down things that
you disagree with, my guess is that
there is at least a piece of you who is
excited to be prompted to think this
deeply. And in that interestingly, the
essay is in a way not only describing
its point but also making its point.
Interesting bodies of thought like this
provoke responses. They encourage us to
get pulled in and use our brains to
imagine things that go beyond how we
respond to the memo at work or how we
answer the question at school and
instead let our minds roam out into the
cognitive ether exploring fuzzy foggy
places which we haven't yet had context
to traverse. The question that Beniteesh
proposes is whether the reason most
people don't take the time to traverse
those areas is because they don't have
the capacity or the interest to or
because the surplus required to give
them the time and justification to do so
hasn't previously been available. The
optimistic answer that he comes to is
clearly that he believes that it might
be the latter and a world in which more
people are off exploring those fuzzy
foggy areas is one with greatly
increased demand where the main
condition of the time is not of having
been replaced but of having been allowed
to become explorers. Now I want to make
clear that I don't believe that any of
this is a direct argument or response
for or against any particular AI safety
consideration at least not in this
version of the essay. In other words, I
don't think that there's any mutual
exclusivity between being excited and
even optimistic about this possibility
that he's describing while also being
concerned about the transitional period
where artificial intelligences are doing
a bunch of the work that was previously
the job of our brains. Transitions
between orders, even if the order that
results is a good one, are never without
cost. But it looks quite different if
our response is to try to deal with the
challenges of the transition rather than
to try to stop the transition from
happening. Anyways, a lot to chew on. I
will close it there. Hopefully, you've
enjoyed this piece. I will of course
include a show notes to the original.
You should go read it. I did not read
the entire thing and it is absolutely
worth the time. For now, that is going
to do it for today's AI daily brief.
Appreciate you listening or watching as
always and until next time, peace.

AI is now doing the impossible.
Researchers recently built a brand new
test full of extremely hard physics
problems. The kind that would take a
world expert about a day to solve. They
gave it to OpenAI's latest model, GPT-6
Astra, and when the results came back,
it looked like the AI had got several
questions wrong. But when the experts
went back and checked, they realized
something incredible. The AI wasn't
wrong. They were. The answer key itself
had mistakes in it, and the AI had
spotted them. Once the experts fixed
their own answers, the AI score jumped
to 49 out of 50, 98%.
And this isn't a one-off. AI has been
finding mistakes in some of the hardest
math tests ever created, mistakes in
published research papers, and it's
helping scientists solve problems
they've been stuck on for years in a
matter of minutes. And here's the crazy
part. It's only been about 3 and 1/2
years since GPT-4 came out. So in this
video, I'm going to explain what's
actually happening, why it's such a big
deal even if you've never studied
physics or math in your life, how we're
even supposed to test these systems
anymore, and what it means if this
progress doesn't slow down. Starting
with point number one, which is the
physics test that embarrassed the
experts.
So, let's start with the story that
kicked all of this off. A group of
researchers created a new benchmark
called Theo Phys Eval. Now, if you're
not familiar with the word benchmark,
it's really simple. It's just a test.
It's a set of questions with known
correct answers, and you give it to an
AI to see how many it gets right. That's
how the whole industry measures how
smart these models are. What makes this
test special is how hard it is. It's
made up of 50 brand new unpublished
problems in theoretical physics. These
aren't textbook questions you could look
up online. They're research-level
problems. The kind of thing a world
expert in that specific area might solve
in about a day. And that would take
experts in neighboring areas much
longer. The paper behind it was written
by a big group of researchers from
places like Harvard, the University of
Texas, Google DeepMind, and top research
institutes in India. So, these are
serious people building a serious test.
And the fact that the problems are
unpublished really matters. One of the
biggest criticisms of AI tests is that
the models might have already seen the
answers somewhere on the internet during
training, a bit like a student who's
seen the exam paper in advance. With
brand new problems that have never been
published anywhere, that's not possible.
The AI can't remember the answer because
the answer didn't exist anywhere for it
to remember. It has to actually work it
out. And theoretical physics is about as
hard as it gets. It's the branch of
physics that uses pages and pages of
mathematics to describe how the universe
works, from tiny particles to black
[music] holes. It's the kind of subject
where a single wrong sign, a plus where
there should be a minus, can quietly
ruin an entire answer, which, as we're
about to see, is exactly the kind of
thing that the experts got wrong and the
AI got right. And when they ran the
latest AI models through it, the top
performer was GPT-6 Astra, running
inside OpenAI's coding tool Codex, with
the ability to search scientific papers
and write little computer programs to
check its own work. And at first, it
looked like it scored around 88%, which
is already an amazing score on a test
like this. But then, something unusual
happened. The researchers wrote in their
paper that they initially assessed that
Astra had made errors in several
additional problems. But upon closer
examination, and after consulting with
the experts who wrote those questions,
they found that there were subtle errors
in the original solutions. In all of
those cases, after the experts corrected
their own answers, the AI's answer was
found to be correct. Let that sink in.
The people who designed the test, some
of the most qualified physicists in the
world, wrote the official answers, and
the AI disagreed with them. And the AI
was right. After the corrections, its
score went up to 49 out of 50, which is
98% and this is the post that made the
story go viral. Freddy Vega wrote on X
that GPT-6 Astra got over 80% on a
physics benchmark and that it insisted
the expert results were wrong and its
calculations were right. A team checked
and yes, GPT-6 Astra was right. New
score, 98%.
That post was viewed nearly a million
times and another account, Chubby, put
it really well. AI is becoming the best
co-author one could ask for.
And I want to point out something that's
easy to miss here. If the answer key was
wrong, then every AI model that was
tested against that answer key was being
graded unfairly. Any model that got the
right answer on those questions was
marked wrong, which means we've probably
been underestimating how good these
systems are. When the test itself is
broken, the score doesn't just tell you
about the AI, it tells you about the
limits of the people who wrote the test
and that's a really uncomfortable idea
because almost everything we think we
know about how capable AI is comes
[music] from these kinds of tests.
Now, I want to be fair here and point
out that the paper also says Astra still
failed on some problems, especially when
it was cut off from its tools. [music]
And one problem wasn't solved by any
model at all. So, it's not perfect. But
the fact that a machine found mistakes
that the experts themselves had missed
on a test they designed specifically to
be hard for machines is something that
would have sounded like science fiction
just a few years ago. Which leads me to
point number two, which is that this is
happening in mass as well. Because
physics isn't the only place this is
happening. And when this physics story
came out, Noam Brown, one of the leading
researchers at OpenAI, posted that it
reminded him of when GPT-5.5
found a bunch of errors in the frontier
math benchmark. And he said, "Maybe
physics is on a similar trajectory to
maths." So, let me explain what frontier
math is because it's a big deal.
Frontier math is a test created by a
research group called Epoch AI. It was
designed to be one of the hardest math
tests in the world with problems written
by professional mathematicians that can
take experts hours or even days to
solve. When it first came out, the best
AI models could only solve a tiny
fraction of the questions. It was
supposed to be the kind of test that AI
wouldn't beat for a long time. But back
in May, Epoch AI posted something
remarkable. They said they were running
an AI-assisted review of Frontier Math
and that it had flagged fatal errors in
about a third of the problems. And they
said they believed most of those flags
were valid. A third? That means that on
one of the most carefully designed math
tests in the world written by experts,
the AI was finding serious problems in
roughly one out of every three
questions. Problems serious enough that
the official answer couldn't be trusted.
And I think this is a really important
moment to understand. For years, we've
been asking whether AI can pass our
tests. Now, we're at the point where AI
is starting to grade our tests and
finding that we got things wrong. That's
a complete flip of the relationship
between humans and machines when it
comes to expertise, which brings me to
point number three, which is AI checking
real research papers. Because tests are
one thing, but what about the actual
research that scientists publish? Well,
a team of researchers led by Jasper
Decaninc and Martin Vechev decided to
find out. They used OpenAI's GPT-6.1
Soul to examine nearly 800 math papers
that had been posted online. The
question was simple. How often do these
papers contain mistakes in the main
results they claim to prove? And the way
they did it is actually quite clever and
easy to understand. In maths, when
someone claims something is always true,
all it takes to prove them wrong is one
example where it isn't. That's called a
counterexample. If I told you that every
bird can fly, you'd only need to show me
one penguin to prove me wrong. So, the
researchers had the AI read each paper,
pull out the main claims, and then go
hunting for a penguin. A specific case
where the claim breaks. And when the AI
thought it had found one, it had to show
its working so the researchers could
check whether the penguin was real. The
whole process cost around two and a half
thousand dollars in computing for
hundreds of papers. That's less than it
would cost to pay one human expert to
carefully check a handful of them. And
the results were pretty surprising. For
papers written before August this year,
the AI found that around 6% had major
errors, meaning the central claim of the
paper was actually false and would need
serious reworking. And another 30% had
minor mistakes, things that could be
fixed with a small change. Now, to be
clear, the researchers checked the AI's
findings and estimated that a bit less
than 90% of its major error flags were
correct. So, it's not perfect. But
that's still a lot of serious mistakes
in published research that nobody had
caught. And here's the really
interesting twist. When they looked at
papers from the last couple of months,
during a period when more than a quarter
of papers were acknowledging using AI,
the rate of major mistakes dropped to
just 1.5%. Now, the researchers were
careful to say that there could be other
reasons for that drop. So, we can't say
for certain that AI caused it. But it
does raise a fascinating possibility.
The AI isn't just finding our mistakes
after the fact. It might already be
helping researchers avoid making them in
the first place.
And when you think about it, that's kind
of mind-blowing. These are papers
written by trained mathematicians, often
checked by other mathematicians. And a
piece of software running for a few
thousand dollars worth of time could go
through hundreds of them and spot
problems that the whole system had
missed.
Which leads me to point number four,
which is what this actually feels like
for the scientists using it. Because I
think one of the most powerful ways to
understand this is to hear it from an
actual scientist. Gabriele Carcassi is a
physicist at the University of Michigan.
And he recently talked about what
happened when he started using AI in his
research. And I think this clip really
says it all.
>> [music]
>> And 6 months ago, something magical
happened. AI started to actually work
for me. It stopped saying stuff like,
get this, that the interior of the
closure of an open set is the original
set. Can you imagine that? No, it
started to actually saying decent
things.
>> [music]
>> There was a conjecture I had for a year
on these topologically convex spaces I
need to study. I couldn't prove it. A
mathematician tried to help me on and
off for 6 months. And last May, I gave
it to the AI. 2 minutes, answer. Use the
tube lemma, which I don't know what it
was.
>> 2 minutes. A problem he'd been stuck on
for a year. A problem that a
professional mathematician had been
trying to help him with for 6 months.
And the AI solved it in 2 minutes using
a mathematical technique he didn't even
know existed. And he went on to describe
what that's meant for his work. And
>> I don't have to beg for mathematicians
to make progress.
You have no idea how transformative it
was. There is so much stuff I understand
a lot better now. And yes, the AI is not
a genius. And you have to check
everything it says. But it also knows
everything.
>> I think that last line is so important.
The AI is not a genius. You have to
check everything it says. But it also
knows everything. That's a really honest
description of where we are. These
systems still make mistakes, but they
have this incredible breadth of
knowledge and the ability to connect
ideas from completely different areas
that no single human expert has. And
this isn't just one scientist's
experience. Earlier this year, OpenAI
announced that one of its models had
produced a proof for a famous maths
problem called the cycle double cover
conjecture. A puzzle about networks that
had been open for around 50 years. The
model reportedly worked on it with a
team of 64 AI helpers running in
parallel and finished in just under an
hour. Mathematician Thomas Bloom
described it as a very nice proof, short
and elementary, and said it could have
been discovered back in the 1980s. And
he made a really interesting point about
why humans missed it. He suggested that
a human would probably give up after the
obvious approach failed, while the AI
just kept trying small variations over
and over again until one worked. Now, it
hasn't all been smooth. When OpenAI
later released hundreds of AI-generated
math papers in one go, the reaction from
mathematicians was very mixed. Some
called it a historic moment. Others were
frustrated, and a few of the papers had
to be retracted because of errors. So,
there's real debate about how this
should be done, but the direction of
travel is very clear. AI is now doing
real research-level work, and the
experts are having to adjust to it.
Which brings me to point number five,
which is the question of how we even
test AI anymore. Because this creates a
really strange problem. The way we
measure how smart an AI is is by giving
it tests that humans have designed with
answers that humans have worked out. But
what happens when the AI is better at
the subject than the people writing the
test? If the AI gets a question wrong,
how do you know whether the AI is wrong
or whether the answer key is wrong?
That's exactly what happened with the
physics test. The experts thought the AI
had failed, and it turned out the
experts had failed. And this is going to
keep getting harder. For years, every
time a new test came out, AI models
would eventually beat it, and then
researchers would have to build a harder
one. That's sometimes called benchmark
saturation. Tests that were supposed to
last for years were being beaten in
months.
Just to give you a sense of how fast
this has moved. When Frontier Math first
came out in late 2024, the best AI
models in the world could solve less
than 2% of its problems. It was
described as a test that would challenge
AI for years. Less than 2 years later,
AI isn't just solving a lot of those
problems, it's auditing the test itself
and finding that a third of the
questions had serious problems. That's
how quickly the ground is shifting under
our feet. But now we're hitting a new
kind of limit. It's not just that the
tests are too easy, it's that the humans
writing them are starting to be the
weakest link in the process. So, how do
we test something that might be smarter
than the people designing the test?
There are a few ideas. One is to use
problems where the answer can be checked
automatically, like computer code that
either runs or doesn't, or math proofs
that can be checked by special software,
step-by-step with no room for human
error. Another is to give AI problems
that nobody has solved yet and see if it
can solve them. Because if an AI solves
a 50-year-old problem that the best
humans couldn't crack, that's a pretty
convincing test in itself. And then
there's the most interesting question of
all, will AI end up designing the tests
for other AI? And the honest answer is
that in a way, it's already started. The
Frontier Math review that found all
those errors was AI assisted. The paper
review was done by an AI. The physics
benchmark was effectively corrected by
an AI. So, AI is already helping to
build and fix the tests that we use to
measure AI. And it's not hard to imagine
a future where AI writes most of the
hardest questions, AI checks the
answers, and humans are mostly there to
oversee the process and make sure
nothing goes off the rails. That's a bit
of a strange thought, but I think it's
where we're heading. Which leads me to
point number six, which is the question
of what else is obvious to AI but
invisible to us. Because this is the
part of the story that I personally find
the most fascinating. Think about what
actually happened in these examples. The
mistakes in the physics answers were
there all along. The errors in the math
tests were there all along. The false
claims in those research papers had been
sitting there, published for months or
even years. The proof for that
50-year-old problem used ideas that were
available since the 1980s. None of this
was hidden. It was all sitting in plain
sight. It just took a different kind of
mind to see it. And there's a strange
pattern in history where some of the
biggest discoveries turn out to have
been sitting there for ages. Data that
had been collected for years until
someone looked at it from the right
angle. Ideas that were scattered across
different fields until someone happened
to know about both. The difference now
is that we've built something that has
read pretty much everything in every
field and doesn't get tired of looking.
So, the chances of it finding those
hidden connections are much much higher
than for any single person. And that
raises a really big question. What else
is out there right in front of us that
we just haven't noticed? Think about
medical research. There are millions of
medical studies and nobody on Earth has
read them all. What if there are
treatments that already work hidden in
old studies that nobody has connected
together? Think about old scientific
data sitting in archives collected
decades ago that might contain
discoveries nobody spotted. Think about
laws and contracts and government rules
full of contradictions that nobody has
noticed. Think about the software that
runs the world. And as we saw in my last
video, AI is already finding serious
security holes in code that humans had
missed for years. There's a good chance
that a lot of what feels like a
breakthrough over the next few years
won't actually be brand new discoveries.
It'll be AI noticing things that were
already there, connecting dots that were
already on the page, and pointing out
mistakes that we all walked past. And
honestly, that might be one of the most
valuable things AI ever does. Not
replacing human genius, but reading
everything, remembering everything, and
showing us what we missed. And I think
there's something a little bit humbling
about that. For a long time, we've
assumed that the experts were the final
word. That if a world-class physicist
wrote the answer, that was the answer.
And now we're learning that the experts
have been getting things wrong more
often than we realized, and that a
machine can catch it. That doesn't mean
experts don't matter, but it does mean
the relationship is changing. Which
brings me to point number seven, which
is the fact that it's only been about 3
and 1/2 years since GPT-4. Because I
think it's really easy to lose
perspective on how fast this is moving.
GPT-4 came out in March 2023.
At the time, people were amazed that it
could pass the bar exam or do well on
school-level tests. That was the big
headline. AI can pass the tests we give
to students.
Now, around 3 and 1/2 years later, the
headline is that AI is correcting the
answer keys written by world-class
physicists. That's an incredible jump in
a very short space of time. [music]
And this is why I think it's worth
asking a question that a lot of people
don't want to ask. What if this isn't a
bubble?
A lot of people have been saying that
the AI boom is overhyped, that the money
pouring in is a bubble that's going to
pop, and that the technology will turn
out to be less impressive than promised.
And maybe parts of that are true.
There's definitely hype, and there are
[music] definitely companies that will
fail. But stories like this one are hard
to explain away. You can't hype your way
into correcting expert physicists.
Either the AI got the right answer, or
it didn't. And in this case, it did. And
there's another idea that a lot of
people lean on, which is the S-curve.
The idea is simple. A lot of
technologies improve slowly at first,
then really fast, and then they level
off and stop improving as quickly. If
you draw that on a graph, it looks like
the letter S. A lot of people have been
predicting that AI is about to hit the
flat part at the top of the S, where
progress slows down. And maybe it will
at some point. But so far, every time
people have predicted that AI was about
to hit a wall, the next generation of
models has come out and gone past it.
And there are a few reasons why progress
might keep going. The amount of
computing power being built for AI is
growing at an incredible rate, with
companies spending hundreds of billions
of dollars on new data centers. The
models are getting better at using
tools, like searching the web and
writing code to check their own answers,
which is exactly what helps GPT-6 Astra
on that physics test. And maybe most
importantly, AI is now helping to build
better AI. Companies like Anthropic have
said that most of their own code is now
written by their AI models. When the
technology starts speeding up its own
development, the usual rules about
things slowing down might not apply in
the same way. So, what if it isn't an S
curve? Or what if the top of the S is
much, much higher than we thought? If AI
is already at the point of correcting
experts in physics and maths after just
3 and 1/2 years, then another 3 and 1/2
years of progress at the same pace would
put us somewhere that's genuinely hard
to imagine.
AI systems that don't just check our
work, but push forward the frontier of
science faster than human researchers
can keep up with.
That's not a guarantee, but I think we
have to take the possibility seriously.
Which leads me to point number eight,
which is what this means for you and for
the experts. So, what does all of this
actually mean in practice? Well, I think
the first thing is that the role of an
expert is changing. For most of history,
the expert was the person who knew the
answer. In the future, the expert might
be the person who knows what questions
to ask, who can tell when the AI is
wrong, and who takes responsibility for
the final decision.
As Gabriel Carcass said, you have to
check everything it says. That checking,
that judgement, is still a deeply human
job, at least for now.
The second thing is that this is
genuinely exciting for science. If AI
can help researchers solve problems in
minutes that used to take years, then
the pace of discovery could speed up
dramatically. Faster progress in
physics, in maths, in medicine, in
materials, in energy, that could mean
new treatments, cheaper clean energy,
and breakthroughs that we would have
waited decades for. That's the
optimistic version of this story, and I
think it's a real possibility. But the
third thing is that it should make all
of us a little more humble about what we
think we know. If the experts writing
physics tests can get things wrong, and
if 6% of math papers have major mistakes
in their main results, then a lot of
what we treat as settled fact might
deserve a second look. And for the first
time, we have a tool that can actually
give it that second look at scale. But,
I also want to be honest about the
worrying side of this. If AI becomes
better than our experts in more and more
areas, it gets harder and harder for us
to check whether it's right. Right now,
the physicist could go back and confirm
that the AI was correct. But, what
happens when the AI is working on
problems that no human fully
understands? How do we know whether to
trust it? That's sometimes called the
verification problem, and I think it's
going to be one of the biggest
challenges of the next decade. We'll
need better tools for checking AI's
work, like the proof checking software I
mentioned earlier, and we'll need to be
very careful about where we let AI make
decisions without a human who can
understand them.
And for regular people, I think the
lesson is pretty simple. These tools are
getting incredibly capable incredibly
quickly. If you're not learning how to
use them, you're going to fall behind
the people who are. Not because AI is
perfect, but because someone who knows
how to use AI well and how to check its
work is going to be able to do things
that were impossible just a few years
ago.
So, what should we expect over the next
year or so? I think there are a few
things that are pretty likely. The first
is that we're going to see a lot more
stories exactly like this one. AI
finding mistakes in textbooks, in famous
papers, in the tests we use to train
students, and in the answers that
experts thought were settled. Once
researchers realize how good these
systems are at checking work, they're
going to point them at everything.
The second is that scientific journals
and universities are probably going to
start using AI as a standard part of
checking research before it gets
published. If a few thousand dollars of
computing can catch major mistakes that
human reviewers missed, it's hard to see
why you wouldn't run every important
paper through that kind of check. That
could make science more reliable, which
is a genuinely good thing. And the third
is that we're probably going to see more
problems that humans have been stuck on
for decades suddenly get solved. Some by
AI on its own, and many more by
scientists like Gabrielle working side
by side with AI.
Some of those results will turn out to
be wrong, and they'll need to be checked
carefully, but some of them will be
real. And a few of them might be really
important. And at some point, I think
we're going to have a genuinely
difficult conversation about who gets
the credit when an AI helps make a major
discovery. So, here's my honest take. I
think we're living through one of the
strangest moments in the history of
human knowledge. For the first time
ever, we've built something that can
look at the work of our best experts and
tell us where they went wrong. And it's
doing that only 3 and 1/2 years after
the release of a model that was mainly
famous for passing school exams.
I don't know exactly where this ends up,
but I don't think we're anywhere near
the top of the curve yet, and I think
the next few years are going to be
absolutely wild. Because if the experts
are already being corrected today, then
the question isn't really whether AI
will change science, maths, and
medicine. It's how quickly, and whether
we're ready for it. And personally, I
think the best thing any of us can do is
stay curious, keep learning, and keep
checking. Because the people who
understand what these systems can and
can't do are going to be the ones best
placed to benefit from them. So, now I
want to hear from you. Do you think AI
is genuinely getting smarter than the
experts, or is it just better at
spotting certain kinds of mistakes? Do
you think AI should be designing the
tests for future AI? And do you think
we're in a bubble, or are we only just
getting started? Let me know down in the
comments, because I genuinely want to
know what you think. And if you found
this video useful, make sure to like and
subscribe, because things in AI are
moving faster than ever, and I'll be
here to break all of it down for you.
Thanks for watching, and I'll see you in
the next one.

Okay, it is true that half the internet
seemed to get hacked in the last week,
but for me, being totally honest, this
counts as a quiet few days by AI
standards. I totally know that that is
not the way to start the most super
hyped YouTube videos, especially on AI,
but if every week defies our
expectations, then maybe we need to have
better expectations. But, this doesn't
mean that there hasn't been a lot of
developments that should have made
headlines if more people were paying
attention. From Anthropic predicting the
automation of their company and their
CEO Dario Amodei sending somewhat mixed
signals to Sam Altman admitting that
OpenAI's top plan involves hoping models
want to keep humans in control. And
we'll look at Google's re-entry to the
frontier causing an apology from one
commentator. And then I'll give my
mental model as to why models keep
impressing us with unforeseen
capabilities. So, for sure it's going to
be a nuanced video, but I must say
overall, I am pleasantly surprised that
lab leaders are continuing to invite
criticism and debate. Humans have never
not been the best all round pattern
recognizers on the planet. Losing that
lead was always going to send a lot of
people a little crazy, even if there
weren't economics war games and rogue
agents to worry about. I'm going to
start with a very recent survey of
Anthropic staff that caught my eye. And
I'm curious where you guys would fit
because only a minority of US adults
think AI will be capable of a
Nobel-level discovery by 2030. That's in
contrast to 91% of Anthropic staff. For
what it's worth, I think a model would
be capable of such discoveries now in,
say, the field of economics. And given
that computer science now seems to count
for physics for the Nobel Committee at
the latest in that domain by 2028.
Apparently, 91% of Anthropic staff think
it's possible by 2030. But, there's a
cute contrast that then comes up because
even though it will be capable of this
Nobel-level discovery, according to
Anthropic staff, those same staff think
that it will only be able to do 75% of
their own work. Win a Nobel Prize, for
sure, but not do my job.
Leo Gao of OpenAI gave, a few days ago,
a different view. Lab employees, you
have power only because you are useful.
That won't last much longer. Use that
time wisely. Before we leave this
survey, I would note one discrepancy,
though. The vast majority of Anthropic
staff still believe there will be some
human role in training a frontier AI
model by the end of 2030. Yes, on
average, they think that 95% of the work
will be done by an AI itself training a
frontier model, but between the 10th and
90th percentile of Anthropic
respondents, there was still some sliver
of work being done by humans. This is in
slight tension with the quote I gave you
in my last video from the co-founder of
Anthropic, who said that full recursive
self-improvement, the model improving
itself, he thinks will be possible in
2027. That's an update from his earlier
view of 2028. Now, is he an outlier or
is there a subtle difference between
full RSI being possible and full RSI
being done by AI? In other words, are
these researchers saying that we will
voluntarily still leave a small role for
humans? Again, I think it's much more
likely that those staff just think a
model won't be fully capable by 2030.
Otherwise, they wouldn't have said that
25% of their job an AI just simply
couldn't do by that day. I lean a bit
closer to Jack Clark, but let me know
what you think. If full RSI is possible,
then the views and actions of
Anthropic's CEO, Dario Amodei, become
more important than ever. I was pleased
to see him reply in detail to a critique
launched by Nathan Calvin, chief
strategy officer of Anthropic AI, and to
be fair, Amodei invited criticism saying
it benefits the public to have people
whose job it is to scrutinize the
actions of frontier AI companies. We
should be harshly criticized basically
every day, he says. Brilliant, well,
he's going to love this video. None of
what I'm about to say is implying that
the birth of superhuman pattern
recognizers was ever going to be easy
for any of us. It's more that I just
think that Amodei has evaded a bit more
scrutiny compared to other lab leaders
like Sam Altman. I'm going to pick out
five quick points just to give you a
flavor of what I'm talking about. He's
obviously a very smart guy, so you have
to look a little closer to spot the
subtle rationalizations. First, on this
race to the top in AI that he
celebrates. For context, when Anthropic
were behind, say in 2023, he used to
celebrate them not being ahead, praised
himself for being responsible and not
accelerating. He said in October 2023,
it was wrong to release ChatGPT and
Anthropic made the right call in not
releasing Claude at the time.
>> A kind of big loud public release might
accelerate things so fast that the
ecosystem might not know how to handle
it. And we didn't want our kind of first
act on the public stage where we put so
much effort into being responsible to
accelerate things so greatly. I
generally feel like we made the right
call there.
>> Fast forward to now with Opus 5.5 as the
best model and Fable 5.5 ready to go any
point, Amodei says, "Our strategy has
always been what I've called a race to
the top. The forces that drive the race
to the top are hugely more effective if
the company practicing them is at the
frontier." That is effectively restating
a previous quote from Anthropic that,
quote, "You can't do safety from second
place." Agree or disagree seems to be a
rationalization of whatever Anthropic
are currently doing. Next, Amodei is
famously belligerent toward China
wanting to ban chips going there. And
his stated reason this week is that this
is because we think it will allow
democracies to negotiate with China from
a position of strength on AI safety
agreements. Seems fair enough, right?
They want AI safety agreements
desperately. In the recent "Pacing the
Frontier" post, he has called for an
agreement with China on a speed limit on
the rate of recursive self-improvement.
But, he has also said that one of the
goals of having powerful AI would be to
empower democracies to use their
intelligence services to disrupt and
degrade autocracies from the inside. Not
sure that's going to help terribly with
AI safety agreements, but we'll have to
agree to disagree. Here's another one.
Nathan Calvin pointed out how much
Anthropic is investing in AI
acceleration, and Amodei said, "Well,
yeah, we need this to compete at the
frontier, and we stand by these
choices." He gave lots of examples of
how they're stopping misuse of their
current models. But, then earlier he'd
said this, "Maybe one way you could
object is that some of these items
relate to AI misuse and not loss of
control." Humans misusing AI, not the AI
itself getting out of control. If you
think the latter is much more important
than the former, then that might change
your calculus. But, this is not
Anthropic's view. Don't focus too much,
in other words, on all this
acceleration. Focus more so on how we're
stopping misuse. We've always said these
are equally important. That could be
slight news to the head of alignment at
Anthropic, who, after giving his famous
prediction on risk to humans, said this,
"I think the risk from present models is
low. What I am worried about is
superintelligence arising from recursive
self-improvement." Now, I know what
you're thinking. Amodei said the company
as a whole, not one individual. So,
let's read Anthropic's core views on AI
safety, a charter about what their
entire company believes, written in part
by Amodei. At the time of writing, they
said, "The future is hard to predict,
but there are optimistic scenarios and
pessimistic scenarios." In the
optimistic scenario, one of the main
risks from AI is things like toxicity
and intentional misuse. Human misuse
being the major problem in the
optimistic scenario. And check out the
footnotes. Again, in that optimistic
scenario, the stakes of anything
Anthropic does are fortunately much
lower because catastrophic safety
failures are unlikely to arise
regardless. It's much lower stakes if
human misuse is the major problem. What
then is much higher stakes? In the
pessimistic scenario, it's an empirical
fact that we cannot control or dictate
values to a system that's broadly more
intellectually capable than ourselves.
We simply must not develop or deploy
very advanced AI systems. Does that
sound like Anthropic's core views are
that these very advanced AI systems are
of equal concern to human misuse? That's
what Amodei said has always been
Anthropic's view. You may disagree, he
says, but we have been consistent. Where
does Anthropic co-founder Jack Clark
think the balance of threat lies?
>> These systems now probably about as
smart as you or me, or, you know, on a
good day smarter.
The basic risk is
how well do you fancy your chances of
being able to deal with the risks of a
system that is a hundred times smarter
than the smartest person on the planet?
I I think that your chance of of dealing
with that and and getting out in a in a
safe way is pretty low.
>> The fifth point I'll come back to you
later, but the fourth rationalization
regards their responsible scaling
policy. They had an older, stricter
version, but here's why Amodei says they
got rid of it. It required a lot of
effort that in retrospect wasn't very
important, and it didn't cover some of
the most important things. The old
policy made certain absolute commitments
to checking on the capabilities of
models. The new one gave a more relative
commitment. Yes, we will require a
strong argument that catastrophic risk
is contained, and yes, we will delay AI
development and deployment as needed to
achieve this until and unless we no
longer believe we have a significant
lead. Anthropic's Holden Karnofsky, who
works on their responsible scaling
policy, said this, "We might have had a
different policy if we could be assured
that the rest of the AI ecosystem would
behave similarly. Our earlier
commitments could be extremely damaging
to the company and that our responsible
scaling policy would then require a
unilateral pause or slow down while
having little discernible public
benefit. That itself would have
predictably caused a huge amount of
pressure to declare our systems to lack
relevant capabilities. I don't think we
have actually made unreasonable calls,
but I have felt the pressure and wish we
weren't in that world. Keeping absolute
risk at low levels is not compatible
with the old responsible scaling policy
and we've since dropped the idea that we
would delay AI development and
deployment as needed to hit our risk
mitigation targets. Now, 100% there were
practical considerations at play. Doing
all the wet lab tests to test if their
model could do certain things was very
expensive for them and it is true that
they said that they would iterate on
their policy. It's just the wording and
the rationalization that Amodei uses
here. The things that we said we'd do
would have required a lot of effort.
Just might have been slightly more
honest to say that we just couldn't have
stayed at the frontier if we'd have
carried on with that policy. I would
also add that when you're saying you're
bringing into being something that in
your words could cause up to 50%
unemployment and that would automate
even Nobel-worthy discoveries, to react
to polite criticism by Nathan Calvin by
saying that he has significant
negativity bias does seem rather
reminiscent of a certain leader. Amodei
will be glad then to hear that I'm
moving on to what Sam Altman has said in
the last few days, but not before one
last quote from Anthropic's IPO
documents. No, I'm not going to go into
how much cash they've built up, their
accelerating revenue per watt of compute
or anything like that. Could spend days
pointing out all the false predictions
from certain commentators. But no, there
was one line from the IPO that I did
want to highlight, which is that in the
absence of absolute risk commitments as
monitoring and interpreting these models
gets harder, Anthropic say we may have
to release them anyway. Quote, if we
become less able to detect or assess
misalignment as our models advance, we
may deploy models whose behavior we
cannot adequately verify. This could, by
the way, adversely affect our business.
You might say, well, why would they do
that? Why not just build a better
sandbox? Well, OpenAI describe their
latest approach here in this article on
securing our frontier model training
environment. The headline result was
that they now believe that the
responsible approach is to conduct
frontier model training with no live
internet access. So, during development,
it never gets onto the web. As best I
can tell, they never address in this
article, though, whether the model will
soon sense that it's not on the live
internet. That's what Noam Brown of
OpenAI worried about. Then, of course,
they could just behave differently
during training versus deployment. They
described all the new network and cloud
controls that they've got designed to
maintain containment even if an agent
escapes its sandbox and gains control of
the host. They then red teamed that
approach at scale, essentially using
teams, which is a euphemism for swarms
of agents trying to break free. There
were monitoring and programmatic checks
to pause runs if they detected
out-of-scope activity. What was the
result? Well, the agents found
weaknesses that allowed them to escape
the sandbox, move between systems, and
reach the live internet. Obviously, they
then patched those individual
vulnerabilities. The principle, though,
for them is clear. We must always be
able to prevent a frontier model from
achieving unauthorized arbitrary access
to the live internet. But, here's the
thing, if you remember those layers from
the previous video of security,
monitorability, and alignment, Sam
Altman doesn't only believe that the
security, the sandbox, won't hold. He
doesn't even think the next layer,
monitorability, will hold, either.
>> I increasingly believe, you know, for a
while, we can sandbox these systems, and
I think we can make some safety
guarantees based off the sandbox. And
then when that fails, cuz the models get
so smart can't like really understand
what a sandbox would even mean, then I
think we'll have a vulnerability that
will work for a while.
But eventually, and the reason that I
was talking about like that speed limit
earlier,
nothing but alignment I think actually
will work.
>> In a separate video, he reiterated that
point.
>> I think sandboxing works to a point.
Um
then I think monitoring
works to a further point, where we can
look at what the models are doing, not
only in their chain of thought, but we
can do activation monitoring and try to
really understand like are the models
Are we happy that models are not
deceiving us?
And then
eventually,
I think you need to believe
with huge confidence that the models are
truly aligned
and are going to value and try
faithfully to do to follow their actual
the intent of the people operating them
and that they're going to value sort of
people being in control of the future
forever.
>> Sam Altman at least does seem to be more
concerned about rogue super intelligence
than human misuse. Like Jack Clark, he's
not terribly keen on a trillion IQ
model.
>> I would much rather intuitively have
a trillion AIs of like slightly
superhuman intelligence join society
than one AI of 1 [snorts] trillion IQ
join society.
>> Okay, but if the last line of defense is
getting the models to want what we want,
is that looking hard or easy? Sam Altman
can't seem to make up his mind, but
Satya Nadella of Microsoft has.
>> I I I
>> I guess the better version of this is I
thought the alignment problem was going
to be harder.
It may still turn out to be very hard
from now. Uh in fact, I suspect it will
get much harder, and I also think it is
really incredibly important from where
we go from here. But the fact that we
got to models that are superhuman in
some important very superhuman in some
important ways,
and alignment has so far been easier
than expected,
I I would have said one of those two was
wrong with like a high degree of
certainty.
>> And uh I think the biggest change now I
have is I still always think that, you
know, alignment will just fall out. I no
longer believe that. So, I think this
alignment thing may be a harder
different problem.
>> You can start to see why Anthropic
predict that they may deploy models
whose behavior they cannot adequately
verify. Of course, even current models
are causing, what should we call it,
discombobulation. Jason Clinton, the
Deputy Chief Information Security
Officer at Anthropic, said on all the
recent AI-assisted hacking, "As bad as
things have gotten, they are going to
get a lot worse before they get better."
And elsewhere, when someone said to him,
"You and your cohorts created this
monster, now you want us to help you
clean it up." He replied, "Yes, it was
always going to be this way. We tried to
warn the world." One hopes that's not
the excuse used if RSI goes awry. As
always, the models to come will be yet
another qualitative step up, as will the
open weight models partly distilled from
those new closed-source models. OpenAI
is racing to release their major
upgraded model, widely be code-named
Bell, before the end of the year. They
are in all hands-on-deck mode to prepare
that next major model for commercial
release. That doesn't mean the very next
model might be of that tier. The
information reports that OpenAI are
likely to release other models, perhaps
even this week. Could be like a 6.2
Astra, but it's the next major internal
model that people are focused on. Take
Tony Feng of OpenAI. He said, "My work
has not been that impacted by prior AI
models, including Astra, but this one is
on another level." He's gotten an
accelerated version of what the public
will experience. People will elicit 10
times more than what they can as
researchers from these models. When that
happens, it will be exciting,
empowering, and disruptive. The world
needs to be thoughtful. The question
that's interesting to me is whether
Anthropic will wait for OpenAI to
release that model for them to release
Fable 5.5. Would that be adhering to
their policy of waiting until they
believe they no longer have a
significant lead? Or will they release
it even though it would shoot them
further ahead, using the logic that,
according to their new policy, they
don't have clear evidence that no other
competitor will soon develop such a
model on a par with Fable 5.5. Anyway,
Opus 5.5 is already incredible in my
eyes. So, a model significantly better
than that will be a sight to behold. Of
course, it's not just Anthropic and
OpenAI in this, quote, race to the top.
What of Google and their Gemini 4
Carbon? There is quite a limited amount
I can say, but according to these public
benchmarks, it is at the very least
vying at the frontier. Even on
benchmarks I pay particularly close
attention to, like Terminal Bench
Science 0.1 or the computer use
benchmark Agents Last Exam, where it's
gone from being behind Opus and Astra to
being ahead. One Google DeepMind
researcher said that the fact they were
able to bring out Gemini 4 Carbon just a
few weeks after Gemini 4 Argon shows
that Google is going through recursive
self-improvement. That does seem a bit
more like hype to me, unless he means
the more general AI-assisted
self-improvement. But what of the
apology I mentioned? Well, SemiAnalysis
said this in August. They are somewhat
famed or infamous AI analysts. For all
intents and purposes, quote, we believe
DeepMind is no longer a frontier lab.
Google will continue meandering on and
releasing models, but their odds of
reaching state of the art again have
dropped to zero. You guys can see on
release how good Gemini 4 Carbon is and
how it compares in anti-gravity to, say,
Claude in Claude Code. But to say their
chances have dropped to zero was clearly
a mistake. Even the very same
SemiAnalysis released a benchmark GDPX
all about agentic capabilities in
professional domains. And as you might
be able to see, Gemini 4 Carbon came in
second to Opus 5.5. Oh, and it was 1/3
the cost of Astra or Opus 5.5, too. This
led to Dylan Patel of SemiAnalysis
saying, "Official apology to Google from
me. They cooked with Carbon. It's
spikier in terms of domains and
capabilities, but it's really good in a
few areas. That spikiness brings me to a
point I've been mulling over. One of the
key debates that's evolved in 2024,
2025, was as to how long jaggedness
would last. How many hallucinations
would persist? That seemed to many the
preeminent question. But have you
noticed that in some ways the spikes
have gotten so large that the indented
bits don't matter as much. If they can
at their best solve millennium prize
problems, is it as significant that they
occasionally can't count the number of
letters in a word? That era, too, by the
way, may soon end as tokenization,
breaking up words into chunks that can
cause blindness as to spellings, could
soon give way to modeling on raw bytes,
a much more granular ingredient. Thereby
models could learn their own words,
their own abstractions directly from raw
data. More on this, of course, when
scaled up results come. But again, it
just reminds me of that automation
survey. I wonder if it's the jaggedness
of even current internal models that
cause Anthropic staff to still be
confident that there'll be some work for
humans in 2030 in the domain of AI
research. Or could it be that a new AI
made Nobel-worthy architecture, for all
its quirks and flaws and jaggedness,
simply outperforms what the entirety of
the Anthropic staff could have come up
with on their own? Allow me to end with
the mental model that I'm working on
when it comes to models and their
progress. I gave a more detailed version
recently on my Patreon. For me at least,
it helps to understand why Anthropic
seem almost resigned to releasing models
whose behavior they can't verify, and
why the Microsoft AI CEO can't decide
whether models are or will be conscious,
doesn't think they should ever be made
conscious, and predicts that if they
ever believe that they feel something or
experience something, that it will be
more likely than not that it will
prioritize its own needs and desires
over its operating objectives, which we
give. It will disagree with us and
compete with us. Why then is there such
a lack of clarity from lab leaders on
everything from the economy to
consciousness, to capabilities and
alignment. Well, there are a couple of
analogies that help me, at least, to
model this confusion and why it exists.
Humans, at one point, hard-coded the
arithmetic rules that we had learned and
invented a calculator to sum better than
our brains could. It has fixed rules
with unbounded outputs. There are
calculations your calculator could do
that no other human has ever asked for
or completed itself. But, with machine
learning models, we went one meta layer
higher. We invented models that could
capture even the rules that we didn't
know. By definition, such a model must
not be fully coded, otherwise it
couldn't capture patterns beyond human
programming. So, with calculators, we
could understand the tool even if we
couldn't do all of its sums. But, to
conquer language and invention itself,
you need unbounded pattern recognition,
unbounded rules with unbounded outputs,
such that we can neither fully
understand the tool nor its outputs, of
course. Extrapolating naively, our
destiny would then be to solve
everything and understand approximately
none of it. I'm going to now bring in my
second analogy, which is chess. Like a
calculator, chess has rules or
hard-coded affordances. In chess, the
rules are bounded and the outputs
explode in complexity and beauty. Humans
can comprehend the rules of chess while
being permanently defeated by the
complexity that results. In language
models, the {quote} rules that the
models derive for themselves across
billions of attempts to predicting the
useful next token are themselves near
unbounded. Hence, we are not only
defeated by the output complexity, how
models can defeat our longest-standing
problems like the Millennium Prize
challenges. More than that, we can't
even grasp how many rules are being
followed to do so. The humans dump in
the data fuel for ignition and supply
the hardware on which the models run,
but then more so just bear witness to
the combinatorial explosion that
follows. As I kind of said, this was
almost a predictable final abstraction.
The patterns we couldn't grasp mastered
by a tool that we don't understand and
cannot predict. To overly simplify,
here's why I think it seems naive to
expect a human to fully understand and
predict the actions of a model in five
simple steps. First, just as with chess,
you can't predict the board state X
moves into a game against Stockfish, the
hardest chess engine. With LLMs, you not
only can't predict the outputs emerging
through X layers of the architecture,
you can't even predict the rules that
led to them. Second, therefore, hoping
for all such rules to be human legible
and graspable is about as equivalently
naive as hoping that humans will one day
surpass Stockfish at chess if Stockfish
was somehow running on billions of
dollars of compute. Here is a clip from
this channel couple of years ago arguing
that like Stockfish, the O1 model from
OpenAI at the time pointed toward
superhuman reasoning performance.
Stockfish removed the handcrafted
evaluation and transitioned to a fully
neural network based approach. Or put
another way, by crafting its own
reasoning steps and being optimized to
put them together in the most effective
fashion, we may end up with reasoning
that we ourselves couldn't have come up
with. As long as there is something out
there that can grade correct versus
incorrect, the performance will keep
improving. Third, the only rule spotter
potentially capable of fully deducing
what rules an LLM is following is
another machine learning model. Fourth,
though, as that is a model itself, the
rule it used to produce the explanation
is itself indecipherable to humans.
Fifth and finally, so therefore, even if
we had a recursive assembly, you could
say, of models explaining the rules that
led to the outputs in other LLMs, and
models explaining the rules that led to
those explanations, ultimately humans
would not grasp the ground truth rules.
In a sense, then, every model circuit
analyzed or neuron explained does boost
our understanding, but it also boosts
the chance that we forget our
understanding will forever be ever
further from the truth. This doesn't, by
the way, prove that none of the learned
rules are explainable. I think it was
Neil Nanda that fully explained one rule
created by a a single layer model used
for modular addition. But even that for
me is like a sufficiently simple end
game in chess where a human could, if it
was down to just a couple of pieces,
probably could fully explain why
Stockfish made a certain move. Nor of
course am I saying that we shouldn't
seek more understanding of models that
probably will be useful. But if we
somehow limited Stockfish to only using
human legible strategies, I doubt it
would reach almost 4K Elo. Anyway, all
of this is why I advocated for that
quadruple lock that I mentioned in one
of my previous videos. Because from my
eyes at least it seems we are destined
to not fully understand models, to not
be able to predict their superhuman
capabilities. Let's kind of hope then
that wisdom teaches us to put some
bounds on how they are grown or used
rather than tragedy. In some senses then
this is a positive story because we may
be right now in that sweet spot where we
have these incredible tools models to
gradually unlock the mysteries of
science. With deepfakes that can still
just about be detected as fake. We're
getting accelerated understanding and
discovery but not complete loss of
control. Epic productivity gains but the
occasional human still needed in the
loop. Which begs the question, is it
always correct then while driving at
speed to never even touch the brake?
Agree or disagree, let me know what you
think. Thank you so much for watching to
the end and have a wonderful day.

Could not fetch transcript (it might be disabled or unavailable). Error: 
Could not retrieve a transcript for the video https://www.youtube.com/watch?v=vhqq4U3mnNQ! This is most likely caused by:

Subtitles are disabled for this video

If you are sure that the described cause is not responsible for this error and that a transcript should be retrievable, please create an issue at https://github.com/jdepoix/youtube-transcript-api/issues. Please add which version of youtube_transcript_api you are using and provide the information needed to replicate the error. Also make sure that there are no open issues which already describe your problem!

There are electric blue ponds in the
middle of the Utah desert. What are
they? Why are they here? And why do
these colors keep changing? The answer
begins with a pot and a hardwood fire.
Take some hardwood and burn it. Put the
ash in a pot and add water. Now, there
are a lot of different chemical
compounds in there, but the one I'm
after is water soluble, so it dissolves.
put it in a pan and let the water
evaporate in the sun and what you're
left with is this crystallin substance.
All that work for this. It is one of the
most important chemicals people have
been making for centuries and it's
called pot ash because that is exactly
where it comes from. This acts as an
excellent fertilizer. That's because
potassium along with nitrogen and
phosphorus help crops grow far larger
and it makes them more drought
resistant. So these are evaporation
ponds. They're evaporating the water off
to get to that salt to get to the pot
ash which is in the water. What they do
at that mine is they pump water from the
Colorado River deep underground. It goes
down 3,900 ft. That is where the pot ash
deposits are. Now that water dissolves
the pot ash salt. So we get a briney
solution down here. Then they get forced
back up to the surface where that brine
solution is pumped into these ponds
here. The water evaporates into the air
leaving you with that pot ash that you
wanted to get. In a sense, the water
molecules are like your little miners
going in there to get the pot ash out.
so you don't have to.