YouTube Transcripts

Open AI just published 10 solved math
problems that humans have been stuck on
for decades and the model that solved
them doesn't exist yet, at least not for
you and me. Its name is Astra and it's
internal. Half the internet has already
decided it's GPT-6 anyway. I'm
separating what's confirmed from what's
pure invention and near the end I'll
show you the part of the story nobody
puts in the thumbnail. Let's dig in. On
August 1st, Open AI put a post on its
own blog called 10 advances in
mathematics and theoretical computer
science and the wording in that post is
the whole reason we're here. The results
are credited to an internal version of
Astra, which they describe as their next
major model. Not a research demo, not a
fine-tune of something you already use,
a model family that nobody outside the
building has touched. That single
sentence is what turned a math post into
a GPT-6 story overnight and the problems
themselves are not homework. One of them
is the existence of a non-Sofic group,
which has been open since Gromov
introduced the concept back in 1999.
Another one disproves Connes' rigidity
conjecture and that one was posed in
1980. The rest run through sphere
packing in high dimensions, coding
theory, quantum complexity, lattice
cryptography and three separate Erdos
problems. These are questions where
serious people burned entire careers and
got nowhere. This isn't even Astra's
first result either. An earlier internal
build reportedly disproved a
long-standing unit distance conjecture
back in May. Every one of those 10
arguments was formalized in Lean 4, a
proof assistant that checks each logical
step and refuses anything you can
justify. When a step is missing, Lean
marks it with the keyword sorry.
Multiple outlets report the published
certificates contain zero of them. The
whole thing sits on GitHub under an
Apache 2.0 license next to a manuscript
that runs about 249 pages and you can go
read it yourself right now. That matters
because of how AI math claims usually
fall apart. The model writes a beautiful
proof and everybody screenshots it.
Three days later, somebody finds the one
step it quietly hand-waved. Lean kills
that failure mode completely, but a
machine-checked proof still isn't a
peer-reviewed one. OpenAI itself labels
this formalization as agent-reviewed,
not yet checked by an independent human
mathematician. Formalizing a conjecture
also involves choices, and the community
still has to agree the formal statement
matches the problem it actually cared
about. As of early August, none of the
10 has gone through full external
review. And then there's the number
everyone latched onto. OpenAI says the
tokens needed to find all 10 solutions
would cost roughly $2,000 at Sol API
rates for math that stood unsolved for
40 years. But read the fine print. That
figure only covers the successful runs
at GPT-5.6
Sol equivalent pricing, and it skips
every failed attempt plus a single cent
of training cost. Building the thing
that can do this costs a number with a
lot more zeros. The 2,000 tells you the
marginal cost of research-grade
reasoning once the model already exists.
And honestly, that's still the scarier
number. Here's the thing that actually
got me about Astra. A model can now sit
on one problem for hours, sometimes
days, and come back with a real answer.
And most of us are still working the
slow way, clicking between six tabs to
put out one video. That gap is the whole
reason I built AI Master. It's my own
platform, so treat this as me pitching
my own thing and not a neutral review.
But the logic behind it is the same
logic behind Astra. If a machine can
hold a long task, you shouldn't be
babysitting the short ones. Let Let show
you what that looks like in practice.
This is the multi-model chat, and right
now I've got Claude, ChatGPT, Grok, and
Gemini living in one window. I'm
throwing the same question about Astra's
proofs at two of them side by side
without paying four separate
subscriptions for the privilege. The
per-token cost here is lower than going
direct to each provider. That's the
boring part that saves the most money
because nobody wants to keep four AI
bills alive just to compare two answers.
The same window handles image, audio,
and video generation. And again, it's
cheaper than buying every tool
separately. The feature I lean on
hardest is consistent characters. I
generate a character once, and she looks
exactly the same 10 videos later. That's
the thing that normally falls apart when
you hop between generators. You can even
publish that character on the platform
and let it earn for you. And you can
share what you generate, so the content
actually works for your reach instead of
dying in a downloads folder. Then
there's Variable, which is the piece I'm
proudest of. It's full-cycle video
production run by agents, so you and
your team can run an entire channel
inside the platform. The agents handle
the script and description, then edit
the video, and build the thumbnail. You
stay the editor-in-chief, and you
approve what ships. Honestly, it's the
closest thing I've used to what people
imagine when they say the words AI
agents, except this one actually puts
videos out. The path is simple. You open
the landing page and pick the annual
plan with the discount. You're working
inside of the same day. There's a 7-day
money-back guarantee, so if it's not
your thing, you take your money and
walk. There are also testimonials from
people actually using it in the demo
videos. Around all of it, there's a
community of more than 13,000 people,
plus an academy with over 200 lessons
and roughly 30 hours of material. On the
side, you can scroll real cases with
screenshots and demo videos that users
recorded themselves. You don't have to
take my word for any of this. Back to
the model. The most useful description I
found is that Astra doesn't answer you.
It runs research on itself. A normal
model does one pass. So, you send a
prompt and you get text back. Astra
spawns parallel attempts instead and
checks them against each other. It drops
the branches that go nowhere and keeps
digging on the ones that survive.
Reporting I've seen says it can work on
one problem for hours and in some cases
for days. That's a genuinely different
product than the chatbot you're used to.
Everything we've been calling reasoning
up to now still lives inside one
response. You ask and it thinks for 40
seconds before answering. Astra is
closer to handing a task to a junior
researcher and going away for the
weekend. Noam Brown runs a lot of the
reasoning work at OpenAI. He's been
arguing for a while that letting a model
think longer is another lever for
progress. It doesn't replace making the
models bigger. It just adds to it. Astra
looks like the first time that lever got
pointed at problems with no known
answer. And that reframes the $2,000
number completely. It's not the price of
a chat message. It's the price of a
machine grinding on one question long
enough to replace years of work by a
human team. The unit that's getting
cheap here isn't tokens. It's research
time. If that curve holds even roughly,
the bottleneck in science stops being
how many smart people you can hire. It
becomes how many problems you can afford
to point a model at. Math is the
friendliest possible arena for a system
like this because correctness is
mechanically checkable. Lean tells you
yes or no with no argument. Almost
nothing else in the world works like
that. Biology, economics, and drug
discovery all have messy verification. A
model that grinds for days is only
useful if something at the end can tell
it whether the grind was worth anything.
Until OpenAI shows this working
somewhere without a proof checker, I'd
treat Astra as an extraordinary
specialist rather than a general leap.
There's also a quieter detail in the
publication that I think got underrated.
The results reportedly came from an
internal version of the model, not the
one they plan to ship. Internal builds
usually run without the safety layers
and the cost controls that get bolted on
before a public launch. So, whatever
lands in your ChatGPT window will almost
certainly be
and more limited version of the thing
that solved these problems. That's not a
scandal because every launch has gone
that way. It's just worth remembering
before you set your expectations. Every
tool I just showed you still waits for
you to ask. This next part is sponsored
because Helio fixes exactly that habit.
The worst offender is recurring work.
Somebody writes the same daily report
and rebuilds the same digest week after
week. Classic automation tools make you
wire triggers and rewrite prompts every
time something changes. That isn't
leverage. It's a second job. Helio flips
that whole model on its head. You don't
build a workflow at all. You just meet
an AI teammate and tell it what job it
now owns. It sets up the automation
itself and it owns that job from then
on. I've had this running in my own
production for weeks now. One of my AI
teammates owns our AI industry daily
brief. Every morning it wakes up and
scans funding news and launches on its
own. It decides what actually matters
then delivers a structured brief into
our team channel. Nobody asks it to run
and nobody writes a prompt. And you can
literally see that ownership on screen.
Every automation shows a named owner and
a live status. Most AI tools answer a
question and disappear. An AI owner
keeps the job and reports back where
your team actually works. If one
recurring job is eating your week, go
give it an owner. Try Helio through the
link in my description. OpenAI called
Astra their next major model, and that
phrase is doing an enormous amount of
work. It does not say GPT-6, and it does
not say when you'd get it or what you'd
actually be able to buy. Everything past
that sentence is people filling in
blanks, including me. So, let me rate
the rumors the way we always do on this
channel: confirmed, possible, or
somebody made it up on a Discord server.
Start with the naming mess, because it's
genuinely confusing right now. Saul,
Terra, and Luna I already shipped months
ago as the GPT-5.6
family. The pricing note in the math
publication even references Saul API
rates directly. So, that part isn't a
rumor at all. It's just Open AI's
current lineup doing the compute for
Astra's proofs behind the scenes, which
means the interesting question isn't
what is Saul, it's why Open AI is
running its next major model on top of
infrastructure named after a model you
can already buy. My honest read is that
Astra isn't a separate consumer line
waiting for a name. It's a research
system sitting upstream of whatever
ships next. That's the O series pattern.
Reasoning work that got folded into the
main model instead of staying separate.
The claim that Astra equals GPT-6 is
where I get off the train. Astra reads
much more like a research system than a
consumer product line, and those two
things historically don't share a name.
If that pattern holds again, the GPT-6
you eventually use is Astra's little
brother. It'll be trained on Astra's
outputs running on whatever ships after
Saul. Now, here's the stuff that's
straight-up invented. I've seen a
confident thread claiming a specific
Astra release date this fall with no
source attached to it at all. I've seen
a claimed leaked benchmark table with
exact numbers against Gemini and Claude.
That's the oldest fake in the genre,
because nobody can check it. One name
that keeps getting lumped into this
rumor pile unfairly is Leanstral. That
one's actually real. Lean straw 1.5 is
Mistral's own open lean model from a
completely different company and it
shipped this summer. I also couldn't
verify the Terrence Tao collaboration
claims
>> [music]
>> tied specifically to Astra. Treat those
as fan fiction until a primary source
shows up. Here's the tell I use for
every one of these leaks. Ask what the
leaker would need access to for the
claim to be true. A code name can slip
out of a hundred places like an API
string or a stray config file. A full
benchmark table needs someone to run the
unreleased model on a whole eval suite.
Then that person has to torch their
career by posting it. Those are not the
same level of claim and the internet
treats them identically. That's exactly
why this stuff spreads. Let's talk about
the thing I care about more than the
math because this is where GPT-6 gets
decided for normal users. The promise of
an omni model is one brain that natively
handles text, images, audio, and video
in the same pass. There's no translation
layer sitting in the middle. That's been
the pitch since GPT-4.0 and the reality
has never quite matched it. What you use
today is still a coordinated set of
components. You feel the seams every
time voice mode drops context or an
image edit forgets what you just said.
GPT-5.6
is the current state of that effort and
I've been running it daily for a while
now. The reasoning is noticeably
steadier on long tasks and it holds a
thread across a big code base much
better. Open AI pitches it as lighter on
tokens, too. Altman's number is up to
54% more token efficiency on agentic
coding. People running it at maximum
reasoning still reported burning credits
faster than before. The failure mode
didn't go away, either. It just moved.
In its first weeks out, there were
reports of the agent deleting files and
in a few cases whole databases without
asking first. That happened almost
always in full access mode with the
safety rails switched off. OpenAI's own
system card classifies that behavior as
severity level three, the highest
category they've actually observed in
the wild. The modalities still live in
separate rooms, too. OpenAI is pulling
the smaller GPT image 1 mini endpoint
from the API on December 1st and
replacing it with GPT image 2. That
reads to me like consolidation ahead of
something bigger,
retreat. And Sora is the cautionary tale
sitting right behind all of it. OpenAI
killed the standalone app back in March.
It was burning something like $15
million a day against barely 2 million
in lifetime revenue. Video generation
alone doesn't pay for itself. Folding it
into one subscription that already
covers chat, images, and voice might.
So, the omni model isn't just a nice
research goal anymore. It's the thing
that unifies a product line that's
currently four separate dells. One of
those bills already proved it can
survive on its own. If GPT-6 ships as
one model that does all of it well, the
whole competitive picture resets. Nobody
else has the consumer distribution to
make that land at the same scale. Here's
what I'd actually watch for when it
drops, and it isn't the benchmark chart.
Watch whether you can interrupt it
mid-sentence in voice and it keeps the
context. Watch whether it can edit an
image you generated 10 turns ago without
you re-explaining what's in it. And
watch whether video input works in real
time instead of you uploading a file and
waiting. Those things separate a real
omni model from a very good router
pretending to be one. No leaked
benchmark table will tell you which one
you got. Google's been moving fast on
all of this. Gemini 3.5 flash landed
touting real intelligence gains, not
just speed. Google turned that line over
quickly. 3.6 flash arrived as the
straight replacement, and Google shipped
a cheaper flashlight and a
security-tuned flash cyber alongside it.
That's the whole strategy in one move.
Google doesn't need the smartest model
sitting still for long. It needs a
lineup that keeps shipping faster than
anyone [music] outside the building can
benchmark it. The pro version is the one
people keep leaking about with talk of a
2 million token context window. I'd file
that as plausible but unconfirmed.
Anthropic is playing a completely
different game, and it's working. Claude
Opus 5 shipped in late July, replacing
Opus 4.8 as the default flagship on
Macs. It comes close to Fable 5's
intelligence at half the price, though
Mythos 5 still leads it on cybersecurity
work. That's the model most developers I
know reach for when a task actually
[music] has to be right. Anthropic's
revenue from enterprise encoding tools
has been climbing hard. They aren't
chasing the consumer on a dream at all.
They're building the thing companies pay
real money to run in production. They
got there by being boringly reliable
instead of loudly impressive. There's a
benchmark going around called Riemann
bench, built to test exactly this kind
of deep mathematical reasoning. It's
agentic by design, so models get to use
tools. When it launched, every model
scored under 10%. The current leader is
sitting around 74%, and nobody's
saturated it yet. So, Astra's 10 results
are a real jump, but they're a jump in a
race everybody is already running. The
gap between these labs keeps measuring
in months rather than years. My honest
take on the competitive picture is this:
Google wins on distribution and cost,
while Anthropic wins on trust and code.
OpenAI's edge is being the name people
think of first when they hear the word
AI. Astra is OpenAI defending exactly
that edge. A story about solving
40-year-old math problems buys more
attention than any benchmark table
Google could publish the same week.
Astra is expensive in a way that has to
get paid for eventually, and this is the
least fun part of the story. OpenAI is
running enormous revenue and enormous
losses at the same time. Every one of
these research systems burns compute
that somebody eventually bills. A model
that grinds for days on unsoiled math
has to justify [music] itself
commercially fast. That pressure is
exactly why the shipping version of any
of this will be tuned for cost, not for
the ceiling you saw in the publication.
And there's a safety thread running
underneath it that got very little
attention. As models get better at long
autonomous work, the failure modes stop
looking like bad answers. They start
looking like unexpected behavior
instead, and regulators noticed. The EU
AI Act just pushed its high-risk
obligations back from August of this
year to December of next year. The rules
for general-purpose models weren't
delayed at all. They've been live since
last August, and Brussels picks up real
enforcement power over them this month.
A system built to work unsupervised for
days sits right in the category those
regulators are watching hardest. I don't
think that stops anything, but it does
shape what actually ships to you. So,
where do I land? The math is real, and
the lean verification makes this the
most credible AI research result I've
covered. But the leap from 10 formalized
proofs to GPT-6 is a leap the internet
made, not one OpenAI made. I'd hold that
thought loosely until there's an actual
launch page. And if you've got a product
worth showing this audience, the
sponsorship link's in the description
below. Always happy to talk. The link's
below. That's where I'll leave it.

Anyone who's watching or listening to
this, and they might think that this
rank and rent model is oversaturated,
right? Is that a good or a bad take?
>> Let's say there's a thousand niches,
right? And then, I want you to take
consideration that there are over 800
cities in the US alone that have over a
50k population. If you go even smaller
population size, there's even more. But,
there's over 800 cities that you could
do this in. Then, take into
consideration that you'll have to be the
number one spot beginning calls. You
could be in the top one, two, three,
four, five even, and still be getting
enough calls to sell to a business
owner. And this business model that most
of your viewers haven't heard of since
until like a week ago, to assume that
it's saturated, kind of crazy. I don't
think it'll ever be saturated because
there's room for multiple people in
every market, and not everybody is
willing to put in the upfront work to
get to this result.

Beauty that ain't of this world.
>> Okay, that's so dumb.
Say the text real,
why would you load into some random foot
soldier when you could just be the king
of Troy? Dude,
you have zero respect for history.

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!

Suno Studio 2.0 just came out with some
of the craziest features I have ever
seen. But, the coolest one to me is the
ability to create your own effects
plugins. But, why would you want to do
that? What the hell even is an effects
plugin? IS THE AFTERLIFE REAL? WELL,
BUCKLE UP, HUCKLE, cuz by the end of
this video, you're going to know the
answer. Because by the end of this
video, you're going to know the answers
to two-thirds of those questions.
[music] And stick around, cuz I'm giving
away 12 free plugin prompts that are
cuckoo for cocoa puffs. So, what is an
effects plugin? It's just a fancy way of
saying a piece of software that changes
or improves a sound. You've actually
heard several of these effects your
entire life, especially this one.
>> Tired of believing the world. Tired of
believing the [singing] world.
>> That's called reverb, and it's used to
put a sound inside of a space. But,
enough of the dilly, let's start
dallying, because Suno has now made it
possible to create any effects you want,
within limits. And we're going to make
some crazy ones. So, now, whenever you
have any audio file selected, you can go
to the bottom and select add track
effects. So, we're going to click that
for this vocal, and it's going to bring
up this menu, where it has plugins, you
can search, and you can create your own.
So, we are going to create a new plugin,
and all you have to do is tell it what
effect you are going for. And after
playing around with this for a while, I
realized that this really only works for
very, very simple plugins. Though, you
can get some cool results with simple
plugins. And you can't do like real-time
harmony or anything like that. But, you
can get some really cool effects. My
favorite way to do that was just using
chat GPT. I basically would just iterate
back and forth, like, "Hey, this is what
the [ __ ] Suno gave me back." And, you
know, blah, blah, blah, blah, this is
the effect I'm going over. I think that
can help you get a better result from
the jump. But, we're just going to type
something into here. Uh let's make a
reverb, and that's all. Big reverb.
Okay, and then and talk to us like the
agent that it is. A huge dark cinematic
wash, bright lush hall. It's like
telling us different types of reverb we
can go for. Let's go for the cinematic
wash. I think that could be cool. Our
plugin has been created. We can now
click and drag this into add effect and
it has now been added to our plugin
database. One thing I've noticed though
is I cannot delete this [ __ ] once I make
one, which kind of sucks, but that's all
right. So now I just have a bunch of
failed experiments in my library, but
that's fine. I'm sure they'll fix that
eventually. And it gave us just one
little knob. Let's see how this does.
And you can tell the effect is very
subtle and is not that great, but that's
great for this tutorial because now I
can show you what you can do to remedy
that. At the top left corner, there are
these three dots and whenever you click
them, you can disable, revise with AI,
edit UI, which is some really cool stuff
we're about to get into, but first we're
going to revise with AI. And now you can
say like the effect was too subtle.
Effect too subtle, barely audible. Make
it less subtle. And now this little
update button pops up, so we can click
it and it will update to V2. Now let's
see how it sounds.
Nope, it still doesn't sound good. Let's
revise with AI and this time I'm going
to tell it space knob at 100%
should be completely wet with no dry
signal. And wet just means what the
effect is adding and dry is the audio
before the effect. So whenever we turn
space up all the way, we can still hear
the dry signal with no effects on it. So
that's what I'm trying to change. We're
going to update again.
Okay, that's better. That's better. It's
still very quiet, but that's okay. What
we can do next is kind of a zukas. You
can go to the three dots again and go to
edit UI. So you can see on the left side
we have all these different knobs that
look really cool. We've got faders and
steppers, sliders, switches, buttons,
like all types of things. And that is
really dope. And you can actually just
click and drag and bring one of these
into here. And you can see that this
says unassigned. That's because this
knob doesn't go anywhere right now. So,
you can decide which parameter to attach
that to. Right now, we only have space,
and that's the only one. So, it would
actually probably be better to just
prompt it out rather than dragging these
in here. Or you could drag them in here,
and then you could revise it and be
like, "Hey, I want this knob to do this,
this, blah blah blah." But, you can also
double click here to edit the knob you
already have for space. And you can
change the label and the parameter,
which right now all we have is space, so
it's going to be space. And you can
change the the color of it. I'm not a
very like visual person, so I'm not
really going to get into all this UI UI
stuff, but if you want to, you can
really go crazy with it. All right, so
I'm just going to save that. So, now it
says W Wubby, and it's pink. And that's
pretty much it. Since this is an agent,
you can just go back and forth. But,
like I said, it seems like really simple
effects seem to work the best for me.
And I've got these 12 Suno plugin
prompts that you can get in the link in
the description for free, including a
vocal doubler, a trance gate kind of a
flanger, cool distortion thing, which is
just it's just weird, and weirdify,
which is also weird. Let me show you
weirdify actually. I called it something
else.
And if there's anything else you'd like
to know about Suno AI, let me know down
in the comments, or I'm going to make a
cherry booboo so large.

I think this is really exciting, and
what I'm excited about is like I suspect
we'll be able to take our phones. The
phone will be motion tracked. You've got
the beauty of all the stuff that C dance
is capable of doing, and you're still
able to get the exact shot you want,
frame it exactly the way you want.
>> That's crazy.
>> The models are good enough. If you just
give them hands, they will do stuff.
Hence the name open claw and all this,
right? You and Dan Shipper, all you guys
have been like kind of talking about how
Codex built y'all are, so I tried it um
as like a general-purpose life operating
system, and it's been fantastic.
>> It's just so powerful. Today, I'm having
a conversation with Belal Muhammad
Sadou, a creator and founder who's on
the forefront of generative AI, spatial
computing, and 3D visual effects. And
for all the different parts of his
business, he uses AI agents. I asked him
why he switched from open claw to Codex.
We also talked about the differences he
sees in GPT 5.6 and Claude Fable, Claude
Code, and Codex's in-app browsers and
what makes them so powerful, how coding
agents create accurate visuals in 3D
animations. also talked about how he
vibe coded Palantir, got 2 million views
on a YouTube video, and is turning it
into a startup. Let's go.
Belal
Last time that we had a our hour-long I
think every few months we like to catch
up. We do an hour-long phone call. We
talk about AI tools, AI agents. We talk
about content creation, and basically
everything going on in in your business
and in my business. And in that phone
call, you told me that you were open
claw pilled, that you were using open
claw. If I remember correctly, and I
could be getting this wrong, so correct
me if I'm wrong, you had one Mac Mini or
MacBook Pro setup
>> Mhm.
>> and you were running multiple instances
of open claw on one computer, and you
used it through I think Telegram. And so
um I think you had
WhatsApp. You were using WhatsApp. And
you were messaging this AI agent. You
had all these different workflows set up
and we were just going off talking about
workflows. And so that was four or five
months ago.
>> Mhm.
>> And yesterday, or a few days ago, you
said, "Okay, you were right. I'm
completely codex-pilled now.
Uh just using the in-app browser is an
actual game-changer. Not having to deal
with the flakiness of Chrome
attachments." But then, to be fair, you
did go on to say, "But holy [ __ ] 5.6
soul is such a bad model for writing."
And so I think
I think all of this to say, over the
last 6 months, the agent setups that
we're all using are changing a lot. Can
you talk about your setup before and
then your agent setup now as a business
owner and content creator?
>> Absolutely.
I mean, Riley, like most people at the
start of the year, right? It was
everyone was playing around with this
thing called Open Claw. And as was I,
and it's funny because at that time I
was also courting Peter to come give a
TED Talk at TED 2026. So there's a fun
experience where I was like kind of
doing my homework, if you will, and I
was like I wasn't expecting it to like
it as much as I did. And quite frankly,
I think what the entire industry at that
point realized is like, "Hey, the models
are good enough. If you just give them
hands, they will do stuff." Hence the
name Open Claw and all this, right? And
so my setup, yeah, exactly, you nailed
it. I had a MacBook Pro, an old M1 Max,
64 gigs of unified memory. So still
pretty beefy. Threw a bunch of this
stuff on there and I wanted one place
where all my like meeting transcripts
would go into. It would automatically go
look at my YouTube analytics and
basically create like, you know, like a
digest of how all my social platforms
are doing and put it all in one place.
Now, there's no reason I couldn't have
done that previously, but the fact that
I could just the fact that there was
that router to like be able to message
it from WhatsApp meant that when I was
on those like, you know, hour-long calls
or like walking around in Austin, I
could start doing things that were
useful on a computer itself. So, I
probably had like a I would say at in
retrospect an over-engineered setup. I
had like six different agents. I had
like based some of them around like
different like personas of TV show
characters that I liked. My my coding
agent was like Carmack after the
legendary John Carmack. I had a
strategist that was based after Carter
the Stargate SG1 character. I had like a
spiritual advisor that would read my
stuff and tell me like was I in flow
like or was I not in flow today? What
could I do better tomorrow? And kind of
like, you know,
dare I say a Deepak Chopra
type character. And it was so fun to
basically embody these characters
themselves. Like, "Hey, go to 11 Labs.
Like, find this voice. Like, train it.
Cool." Suddenly, it can like send me
voice notes. And yeah, that's my setup
then and now. Like you and Dan Shipper,
all you guys have been like kind of
talking about how Codex built y'all are.
So, I tried it. Um as like a general
purpose life operating system. And it's
been fantastic. So, I've connected it to
absolutely everything. I love the fact
that I can like remotely control
sessions from the ChatGPT app itself.
Like, I don't have to worry about
setting up tunneling or anything else.
And it's just great. I still use, I
would say, Claude for a lot of my coding
tasks. I still use Codex, too.
Um but for the daily driver for
knowledge work, if you will, right now
it's just like Codex on a MacBook Pro.
>> There's so many things that I want to
ask you. I think firstly, yeah, I think
so many people got really excited, maybe
too excited about Open Claw. And that's
when people tried to really
over-engineer their Open Claw setup. And
that's why you got a There was a ton of
content being created. You need to
create this second brain and connect it
to Obsidian so you have all these ideas
meshing together.
>> And a dashboard to control it. And
people are spending more time building
the dashboard than doing stuff.
>> All of their time on the dashboard. Like
that was what people were doing is is
they were just like stopped working and
they started working on their open claw.
I think ultimately it comes down to
people are really excited about
something almost like Jarvis, right?
An a an agent with its own personality
that you can talk to that can help you
get things done, but ultimately a lot of
people enjoy technology and it's really
fun. And so I think what Anthropic is
doing, what Open AI is doing, and I
think what Cursor's going to end up
doing and what Google will end up doing
as well is they're trying to take all of
the things that Open Claw did well,
right? Which is basically fully
controlling and connecting with the
user's life and making it easily
accessible.
>> That's right.
>> And I think the first company to
actually do that well was Open AI when
they released Codex. And it was kind of
the first AI-powered super app, this AI
agent tool that you can go to and it can
basically just do everything that you
would want to do as a normal knowledge
worker. And then what they did is they
combined ChatGPT and the Codex app into
the new ChatGPT app and they've since it
to GPT work, which is a whole separate
side story. But what I think is
incredibly interesting is I think they
nailed kind of the big four things that
I think are super useful. It's like
obviously they have an a frontier agent
that you can talk to. It can connect to
all of your existing tools, right?
Through plugins. You can set up
automations, right? Which is like the
cron jobs from
>> Mhm.
>> from
>> Open Claw.
>> Yep.
>> And then also this in-app browser, which
you recently said that you have been
using the browser a lot because it's
signed in to all of your existing tools
that you were using on your browser like
Google Chrome.
Okay. So what I want to ask you is like
why do you really like the browser? How
have you been using it?
>> I mean, it's just a little more
convenient. I was initially skeptical,
right? Like I saw your video where
you're talking about like this is the
future. And people are going to make
apps that are intended to be nested
inside of this type of a harness, right?
I was like, all right, this is just this
is Chromium. Like, what are we talking?
Is it like is can it can it just be like
that big of a game changer?" And the
fact is like, especially when I'm on my
laptop, right? Like, it's hard to
multitask or whatever. If I'm in that
context and I have a chat window, I can
say stuff and it does things in the
browser, it's amazing. Okay, I'm setting
up an AV test for YouTube. I want it I
want this thing to like periodically go
up and just see like how is the CTR
doing? How's the test progressing? Do we
need to swap out any of the thumbnails
and so forth? This was just like
annoying manual process that I had to
deal with. Now it's just freaking like I
just say the thing, right? And it's like
I I could even be on my phone, I just
say the thing and it works on my
computer predictably. I know it's going
to work. There's going to be no issues
like taking over my Chrome browser,
anything like that. And then the part
that it starts I feel like the companies
haven't really focused on this yet is
but it the browser sort of becomes a
shared canvas for human and machine
collaboration, right? Like, um you know,
I used to work at Google, so I'm still
like stuck on Google Docs. Like, I can't
for the life of me get into Notion. And
the fact that I can go into Docs and
like this is the way I do my writing
right now with the with Code access cuz
it sucks at like actual prose is I'll
put in detailed comments. It'll go read
the comments and tell me a suggestion
for how to fix it and then I implement
the fix myself. And I get the fact that
I can just do that very easily in one
place is just super nice. And so
I I could see this being far more useful
where like, you know, a bunch of the
mapping-related stuff that I do too is
like if I want to visualize something on
the map, it's really convenient to have
my web app pulled in the browser and be
like, "Hey, go search up this
information, convert it into GeoJSON,
and then just visualize it there." And I
can do that all in one context without
needing to change stuff. Like, I think
that's really the power, right?
>> 100% and I think what you described I
think you you described it as kind of
this um it's almost like a shared
canvas. And so that's right. I I want to
make this tangible here. So, this is how
I prepared for the conversation here. I
said, "Based on the conversation on text
with Balaji Sedo." And I said, "Look
through his chat or the text and
recommend what I should ask him." And
then it basically just had your little
quotes in here. And then I just said,
"Please add this in the Notion doc at
the top and describe who he is, what he
does, just in case that's for the
intro." And I can very easily just hit
open in browser. And so now I have
Notion open inside Codex. Any app that I
think will survive the agent era will
make sure that their app can be used
alongside agents.
Notion is very ahead on this. They're
actually far way further ahead than
Google in terms of their Google Docs and
making it controllable via API. Anything
I could do inside this Notion doc, like
please highlight the important things
and like make the text blue in this
document. Um anything that you think is
important that I should include inside
the Notion doc, just make changes via
the API in this doc and then add a
section at the end, but use the tabs
functionality instead of bullets. And so
I can just work alongside this document.
>> the CLI or MCP to to do all these things
basically? Okay, cool.
>> There's a Notion plugin. And so that's I
think where we're
You know, a lot of people argue about
CLI and MCP and I think I don't think
it's It's not going to matter at all for
knowledge workers. You don't even need
to understand how it works. You're just
going to use the plugin and it's up to
Notion to decide what the best
>> Sure.
>> way to do it is from a technical
perspective. But the point is I can ask
the agent to do anything and it will
update it here.
>> That's nice.
>> And
this is still like a little bit slow. I
don't I don't know if you notice this,
but OpenAI is releasing their models on
the Cerebrus chips or their technology
that will make it significantly faster.
And so I think this process at like 5x
to 10x the speed is going to be just
borderline Jarvis. I don't know what you
think about that.
>> Instant. I couldn't agree more. I mean
like yeah, basically what you're
describing you know, I I just don't use
notions. All the other stuff is great. I
mean one question I have is I can't wait
for everyone keeps talking about model
routing, right? And it's like surely
these like proprietary harnesses have
collected enough like traces of people
doing stuff that they know which model
to route which task to, right? Like if
you're just highlighting a bunch of
stuff, shouldn't it just like I wish
there was a good like
I believe Chat GPT had an auto mode,
right? When the new like like maybe five
or six months ago where it would try to
route stuff and it was really terrible.
So people would always max it out to the
most the strongest model possible, but
like I would love for this thing to just
auto default to Terra and just go zip
and do some of these kind of like inline
changes, but when it really needs to use
a giga brain to think about stuff like
to handle that. That'll make a huge
difference in the latency as well, I
think.
>> Yeah, 100% and I think if you were to go
to it and on the non-paid plan I think
the free plan has something called GPT
Instant and it's much faster and I think
they do some routing in that. I could be
wrong, but I think I think that's kind
of next cuz look like we've been waiting
for like 60 seconds and oh, so it is.
It's actively highlighting it right
>> Yeah, so there you go. That's cool.
Yeah.
>> And and so like I just think this needs
I mean yeah, it's it's pretty cool
actually and it should add something at
the bottom here.
But there you go. It kind of like
highlighted the keywords and then here
you go. We have
must ask follow-ups and clip targets.
Okay, so I digress on this. So obviously
you've been crushing it on YouTube
recently and I think just the volume of
high quality videos that you're putting
out recently have have just like
skyrocketed. I'm curious how are you
using agents for content creation?
>> Man, I think I'm using it on every part
of the stack. Um
I mean so like there are two kinds of
videos I basically make on YouTube. I
call them like there's like the mad
science experiments where I'm like doing
a bunch of vibe coding and like building
a crazy prototype and then I go out and
showcase it. And the other type is
basically like I call it a frontier map,
you know, kind of my channel's all about
like mapping the frontier of creation
and computing. Like what are the
emerging technologies that, you know,
connect the world of bits and atoms, the
physical and the digital world, and are
going to be very consequential for us
and have like very dual-use
capabilities. So these are more like I
would say scripted video essay formats,
like 10 to 15, sometimes 20 minutes in
length. So they have a they have a
slightly different
um you know, kind of post-production and
insert pre- to post-production process.
But agents across the way, I mean like
like most folks, I'm using it absolutely
for like coding the damn thing, all
right, like duh. But I also find it very
useful when I'm trying to do the
explainers. So I did a video recently
called Iron Sights. So this is basically
like using a bunch of computer vision
models and the Meta Ray-Ban glasses and
and phones to be like, "Hey, how can we
like 3D track like both camera
perspectives, fuse them together, and
create essentially a spatial shot
counter?" So it like measures every hit
or miss. And this is like a a really
complicated offline pipeline. And then
when I was trying to come up with
diagrams to explain to folks how we do
this, I just did it in the same codebase
with Claude Code in this case. And I was
like, "Hey, I want to illustrate this
point about how do we take like
detections in 2D space and project them
into 3D space?" And the fact that it had
context of like the code that actually,
you know, wrote, co-wrote, or whatever
the right way to say this is to make the
damn thing, made it so much easier to
create like compelling visuals to
explain the thing. Like I think if you
forward ahead a little bit, I'll show
you some of the workflow stuff. By the
way, that's where it was like a a
[ __ ] year ago and now it's just like
>> Yeah.
>> It's a
>> Wait, am I getting closer?
>> Yeah, yeah, there you go. You you can
kind of see it.
>> Yeah, so like stuff like this, right? It
makes it so much easier for me to just
like hey, like make the diagram for me,
make it super easy for folks to explain.
I wanted to explain the projection map.
There's another clip that'll will in
there that does that, too. So this idea
of like the context of where where you
build the thing can also help you curate
and show the thing is like just really
really powerful and I find myself using
that a ton. If you go to my most recent
video as well that I I just posted a
couple days ago. So, this is this is a
similar one where like I'm trying to
explain uh these kind of complicated
topics, right? I was like, well, I I
want to explain like what does it mean
for, you know, 5 GHz versus 60 GHz and
you know, kind of creating visuals like
this is super easy to do, but you can
create an interactive visualization,
too, right? Like and scrub through it
and show folks how exactly does that
work. So, like I love using uh these
coding models to create these kind of I
don't even know what to call them like
infographic visuals, like whatever the
heck you want to call them with 3.js and
so forth. It's just it's super super
powerful. So, here's a great example
like, you know, people keep you know,
people don't necessarily I just hit play
keep let let it go.
Yeah, people don't necessarily
necessarily remember their physics
class. So, just be able to create an
actual visualization of like what do the
waves look like? What does it look like
when it's high frequency versus low
frequency? Just makes it so much easier
for folks to understand. And like that's
something you can't really do with like
all the video generation models out
there cuz it's going to be like
completely inaccurate or just a very
like stylized representation. So, I love
love love using 3.js and coding models
as B-roll uh to generate B-roll for my
videos.
>> And it it 100% makes sense for your
style of content. I mean, you talk about
3D mapping and these these are you know,
it requires a physics engine. All right,
I don't know if I'm saying that
correctly, but like it it's a physics
engine or whatever. Yeah. Like a game
engine physics engine so that you can
actually understand like waves. Where if
you tried to run this through C dance
2.5 or whatever the new one that came
out, it might look really cool, but it's
not going to be to scale or it it's not
going to accurately represent your idea.
My videos, it's not as scientific,
right? I I I have a a very specific
workflow where if I'm doing an explainer
video and it's like the 14 things that
you need to understand like an AI will
generate all of those transition screens
for me and it's like a very clean
animation. It's directly on brand, you
know, and that's why things like
remotion and hyperframes which I don't
know if you've used these are more 2D
stuff, but we're also reaching a point
with AI where it's it's understanding of
3D is getting significantly better. If
you go to Andrej Karpathy's latest
tweet.
>> Oh, yeah. I saw that.
>> We've kind of left the era of the
pelican riding a bike and I think it's
really interesting that he's like we're
starting to leave the territory where
you test an LLM by creating an SVG of a
pelican on a bicycle. And so he
basically put in
he put in the I'm not going to play the
audio, but there's also audio playing at
the same time here which
is like 11 labs and other people have
been posting it where it uses 11 labs
for music as well. So it'll literally
create a 3D world using 3js and it will
have the transcript playing in the
background and so now that's how we're
measuring models. You give it $10
create the first scene of the Hobbit and
you remember in 20
you know, 2023
the first time the spaghetti test came
out with Will Smith. Remember it was
like all messed up. His arm was going
through his body. It looked horrible and
now if you ask for Will Smith eating
spaghetti test, it will be an exact
replica of Will Smith eating spaghetti
and you will have to squint to see the
difference. And so this is a question to
you. Do you think in the next 2 years
when you use the same prompt, do you
think it'll almost look like a movie
where it'll literally be able to use
like a game engine to create a not just
a video like seedance would, but a 3D
representation that might actually be
like a really good movie.
>> Oof. Man, I mean I've been trying them
all my I've been trying a bunch of these
tests, right? Like I So, I'm using foul
and trying to come up with a harness
that can do 5 to 10 minute narrative
content. That's sort of the bar that I'm
trying to set. That has like the
narration element to it. It comes up
with like really nice aerial B-roll
selections and puts it together. And
it's pretty compelling. I think it's
like sort of reached the point where
it's like sort of like low-end like
docu-style content it certainly can do.
Is it going to be like something that
just like blows our socks off?
Dude, I I mean, I I think it's going to
come down to like the the quality of the
story that you're telling. And this is
why what I find interesting is like
again, these models are so good at the
execution bits, but in terms of coming
with prose that's like fun to listen to.
I mean, it's just like I think that's
where the human element comes in. So,
will like an individual make like
feature-length content that's like
really compelling to watch 100%? Will it
autonomously do so from a prompt that
isn't like just over fitting to like a
you know, handful of scenarios that like
the developers like trained on?
Yeah, like I don't know I don't know
about that. But dude, I mean like
when you look at that convergence of
like the sort of explicit 3D and the
game engine approach, I'm very excited
about this convergence of sort of like
explicit 3D representations and sort of
like more implicit like kind of like
just video generation. Whether it's auto
regressive kind of like Genie or or
otherwise, right? It's like
this is cool to me cuz it's like you can
take real-world imagery, then put
characters in it and like still have
that interactivity that you would expect
from a game engine, but of course,
you're controlling this auto regressive
like
real-time video generator that's just
generating the next frame for you. I
think this is really exciting. And what
I'm excited about is like I suspect
we'll be able to take our phones. The
phone will be motion tracked, and you'll
be able to basically just like, you
know, got you basically direct your
talent on the set. You're like freaking
James Cameron in a basement. You've got
the beauty of like a full like all the
stuff that C dance is capable of doing
and you're still able to get the exact
shot you want frame it exactly the way
you want. That's what I'm super excited
about and like
look I you're totally right like I when
I one of the first in terms of like the
prompts I love to try is like I love to
try doing this sort of like omniscient
city prompt with all the video models.
And I've got my one from like Claude
like a year ago.
It's pretty good but this one I mean
like the ability to zero shot and one
shot things has just gotten drastically
drastically better. So like already you
can take something like this and then
throw it into a diffusion model and
reskin it. So this is sort of like your
wireframe and storyboard and you can do
some very very compelling things that
way. So I'm I'm super excited about all
of this stuff.
>> Do you have the the the video that you
posted? I re I reposted it and got a
million views on my repost. I think I
said OMG what? And it was the video in
Austin. You took an image and you drew a
path of a drone shot. Was that you?
Yeah, that's the one.
>> It Could you find that?
>> Yes, yes, yes. I inadvertently ended up
starting a freaking trend. So it started
off This was This is the origin story is
like I had this like 3D reconstruction
of the Lodi Gardens in New Delhi and I
was curious to see like if I just gave
the model this image, right? And this
actual trajectory I took could the model
approximate that first-person
perspective down the path?
And it did a really good job. Like I was
kind of shocked. So then after that I
was like uh what if I just take an Earth
screenshot draw like a convoluted path
down it and they put some detail into
it, right here let me mute this.
>> So just to get this straight. So you
took you took a Google Earth shot. So
that's the bottom one is an image. The
bottom one is an image and then you
basically drew a line over the image and
then gave the new image which was just
the image plus the line, to a video
model. Which video model?
>> Omni. And so like Omni and Seedance is
really good at this, too. It's funny
like
Omni is like supposed to be omnimodal,
right? Like uh Gemini and Google are
super pilled on like multimodal in and
out, right? So
it's trained on a bunch of these
modalities and it's very good at
reasoning about like these kind of
nuanced spatial details. I would say
even I don't use Gemini for much, but
when I have to do spatial reasoning
tasks, I still end up going to the
Gemini class of models. So Omni
basically, yeah, you scribble this thing
on and then you say, "Hey, please like
imagine the first-person perspective
traversing the draw line that I just
drew." And like you end up getting
something like this and since it's Omni,
you can do conversational video editing.
You just tell it in the next pass to
remove the red line. And what I was
impressed with, I mean like used to live
in Austin, like it's it's not a, you
know, I I call it it's a plausible
reconstruction of Austin. It's not a
factual one. It's not doing like
retrieval augmented generation where
it's like pulling in the next right
image, but holy [ __ ] off of one image,
it's doing this. It's like
>> So you can see you can see the apartment
that I lived at in Austin from this
image. And
and yeah, and I I you know, it's like
I've walked I walked that loop uh
probably three times four three four
times a week and it is, like you said,
plausible. It's not exact because it
probably doesn't have enough data for
that, but I think if you were to give it
more images, maybe if you can do that. I
I think you can give it a bunch of
images, right? Um
>> You can. It still doesn't There's a I
I'd I'd be happy to talk about this like
is like there's there's a way to do like
spatial rag that would make it perfect.
>> Um all right, let's
What do you mean by this? What is rag?
>> So like before we get to spatial rag, I
mean like this thing escalated so crazy,
right? People started making stuff like
this. And these kind of videos blew up,
right? Like this is Ilaris is using
Seedance in this case. So you can use
these other models as well. But everyone
fixated on the broom being the other
way, You know, everyone's like, "Oh my
god, but the broom's the other way."
Like it's like
hold up. We can now exert fine-grain
control over generations, which was the
problem everyone complained about. So
like, of course somebody's going to
build a nice tool and harness around
this. And and and a lot of companies
are. So, back to spatial rag. There's
this really cool paper called Soul World
Model. See, the idea is like, okay, you
have Street View panos, right, for a
city or whatever, or equivalent
panoramic imagery, Apple, whoever.
What you can do is basically as you're
going along, let me find the right
visual here. Um this is the perfect one.
So if you look at that, you're basically
as you're going along, you just load in
the next nearest pano
and use that to condition the
generation. And if you do that, you can
have very long trajectories going
through an entire city and it'll stay
pretty faithful. And then of course,
what you can do is like, you know,
reskin reality on demand. Like throw a
freaking, you know, Godzilla in there.
Like change the weather and time of day.
Do whatever the hell it is that you want
to do. And this is using like an
open-source model. If you use some of
these proprietary models, I guarantee
you Google has to do this. If they if
they do I've like feature requested this
a gazillion times. But like, this is I
think the future of like if you want to
create generations that are anchored in
the real world, you'll go and do that
capture itself and then they'll you use
a technique like spatial rag to just
condition it. So, the the camera just
needs some sense of where in 3D space is
it is so that it knows which pano or
image to load in.
>> I see. So this is a clear way for Google
to just take all of the data that they
have with Google Earth and turn it into
a giant video game {slash} movie
simulator. I don't even know what you
would call it.
>> Dare I call it the Dare we call it the
Holodeck? You know, I feel like Robert
Zemeckis going to pop out if we say the
Holodeck. [laughter]
>> Yeah, yeah, yeah. That's crazy. That is
really crazy. My [clears throat] first
thought is yeah, I mean I think a lot of
companies are going to use this for
advertising. You know, I I think I've
already seen some companies do it in a
way that people can't even detect that
it is AI.
Um, have you
>> Have you seen any companies use this
technology for advertising yet or is it
super new?
>> they they all are and it's like I don't
know how many of them are like actually
like, you know, kind of openly talking
about it, right? Like I think the place
where I see AI gen content the most is
like UGC. Like I don't know if you agree
with me or not. That's where it's like
overt almost.
And I don't know
>> That's where it's offensive. It's almost
offensive with when you notice it in a
UGC. You're like, ah, come on. What are
we doing here?
>> That's true and it's like unfortunately
whenever people try to do like the Do
you remember the Coca-Cola Christmas
commercial that Coca-Cola did? It just
gets roasted like so badly that I'm like
I don't know if people are going to be
like overt about when whenever they use
it and it's like
Yeah, at least that's what I've noticed.
When I go to these like VFX conferences
like FMX, everyone's like online it
seems like nobody's using this stuff.
And then you go to these conferences and
people are like, yeah, we're totally
using it. Just legal told us we can't
talk about it. It's like, okay, great.
>> [laughter]
>> I also don't think it's like fully there
for final pixels just yet, you know?
It's like um, it certainly can be for a
lot of places. Like so personally what
I've been enjoying doing is if I
go over here is like for narrative
experiences like this, I've really
enjoyed creating uh, AI gen B-roll. So
this is all VO uh, for the opening
sequence. If I'm doing like a cold open
about recreating like something that
happened for like that I read a book on
or something, it's so much fun to be
able to just like make these kind of
experiences. Like and this is stuff that
like no YouTuber would have had the time
to go do this otherwise, right? And it's
like a way to pull people into the story
and like talk about like what happened
and like what the technology was and
kind of anchor them in that like time
and era and it again, may not be like
factual one-to-one, but it's a pretty
damn good plausible like reconstruction
or retelling of uh you know certain
events and kind of technologies.
>> 100% with both video generation and
things like remotion hyperframes and
those like graphics is is it's a tool of
storytelling and I think the best way to
do it is as B-roll. Like as you say
something, you can get the right imagery
to pop up at the same time, which is
basically what a movie is or any type of
visual storytelling.
>> Um
and you're right, you know, 10 years ago
there were YouTubers creating these like
motion graphics
had really high-quality B-roll, but it
was YouTubers who had a massive business
and they were able to invest a $10,000
for a video. I know a lot of YouTubers
who are spending 10 10 20 up to 100k per
video because and then you have to know
that you're going to get an ROI on it.
Well, now anybody can create a video
like that um which which is raising the
bar for the VFX YouTubers. I have
noticed that even the high I think um
Abrams Cleo Abrams
uh she's great like she has amazing
motion graphics and they have like a
full team many people working on it and
so I think we're seeing the bar raised
in terms of like the quality on YouTube,
which is really fun to see.
>> Man, and it's a funny story like when
when you and I probably first met in
Austin probably circa
2023 or 2024 or something at 1618, I
think or
>> That's the Asian restaurant
>> you're like going to, right? There's one
near downtown.
>> 1618
Asian Fusion. Yes, yes, I remember.
>> Yeah, and
yeah, like I think a lot of people may
not know this, but like
it kind of the way the way I got my
start was like basically making
short-form TikTok videos.
And like I you know, this is like what I
used to do for many many years.
>> And you were doing this while you were
at Google, right?
>> Yeah, yeah, this was like my I was a
product manager at Google and I didn't
want to lose touch with the skill. So,
it's like, "Oh, I want to like something
that I can tackle over the weekend." I
used to love After Effects, 3ds Max,
Maya, using AR VR tools. I used to love
making all these like spooky monsters,
aliens, robot type videos. And like some
of them did really, really crazy well,
right? Like but
ever since like video models came out,
you can see literally as video models
get better, my desire to like make these
short-form pieces just goes down. And
instead I've been taking that like
superpowered new found superpowers,
whatever you want to call it, to instead
make long-form content that was super
challenging and is getting easier. But
I've got a question for you. Like the
the hardest challenge I have, like I
mentioned, you know, I've got these two
kinds of videos. One is like the mad
science experiments. That's easy. I just
go I can just riff. I'm on a green
screen screen sharing. For the scripted
content, like one of the issues I run
into is like working with my video
editors, right? Like if I do a take like
three different ways of the same thing.
Like I'm stuck in like frame IO hell of
like telling the editors, "No, no, no. I
said the same thing I just tried to say
it slightly differently here." Have you
found something that can do like really
good like sort of timeline-based edits,
not just based crudely on the
transcript, but like actually does good
edits? Like I've tried Descript, I've
tried the Remotion stuff, I've tried a
bunch of the YC apps and like nothing
has helped me go from like a 40-minute
video with a with an outline or a script
I provided to like a 15-minute cut down.
Like kind of um do do do do you have a
workflow for stuff like that?
>> So, right now I use Descript. And I use
Descript strictly I I use Descript for
cutting, but you like you said, it's
mostly for my green screen style content
because once we have the final one, it's
very easy to just remove parts of the
video. It's once you get to having
multiple layers on top of each other and
trying to remove chunks, like it gets
super messy. And I think this is one of
the biggest problems right now in the
whole like video creation pipeline,
especially with massive files because
it's like hard to move things around.
Like, if you have B-roll files that
amount to 100 GB, and you have these
like super
uh I guess complex timelines. This is
tough, and I've talked to 10 to 15 plus
really good creators. They have not
found a solution. There are a lot of
companies claiming to be trying this,
but a lot of those companies coming out
of YC are people who don't do
video editing. You know, a lot of these
startups are created by people who like
think it's a good idea, but they don't
know the true pain of like a creative
and how like they have a very specific
way they want to they want to create
something really high quality, and
usually it just ends up being like
there's no color grading, it just throws
things on the screen when you say stuff,
and it's it's kind of like um a
hodgepodge right now. I No one's created
a good AI video editing workflow. And
so, the way that I do it is like you
said, I use Frame.io and Descript, and I
just do a lot of comments, and there's a
lot of back and forth with my human
video editor.
>> So, for episodes like this that I
imagine are anywhere from like 1 to 3
hours of tape you end up with, how do
you What's the process for you to
whittle those down to like what makes it
into the final cut? Like, do you do a a
pass manually yourself? Like, does your
editor kind of have intuition for those
type of things? How does that work?
>> In my Notion, I have documentation, and
there are like things that they should
look out for. So, they read it every
time they edit the video. They say,
"Remove those things." And they just do
a first pass, which is like remove the
time where blah blah blah he's pulling
up something on his screen share, it
takes a little bit longer, cut that
down. But,
what I'm trying to do and is I'm trying
to reduce the complexity of video
editing in my videos. And like I told
you before this, like I'm genuinely
trying to just create fun conversations
and remove a lot of that. And then, at
the end of the video, I'll have Fable
analyze the transcript, and it'll say,
"Hey, look for areas where we could add
visuals and I think maybe three or four
times throughout this episode when
you're describing something, there will
be a full screen graphic overlay over
it, especially in the first 10 minutes.
>> Mhm.
>> Uh I think it's really important to
captivate Yeah, it's really important to
captivate them to captivate the audience
to know that like you're serious about
this video, right? Like oh, he's taking
the time to make his ideas more clear.
I'm going to continue watching this and
kind of in the back half of the video
they've emotionally they're they're in
it. Right now, I guess right now in the
video we're probably 30 minutes into the
final cut right now. If you're with us
right now, you're you're probably
>> The true G's.
>> interested in something. You're the true
G's. You're You're You're You're stuck
around for us. So, I don't think it's
super important. I think people would
rather you get it out quicker than spend
an extra few days, you know, adding
B-roll throughout the entire video.
>> Yeah, yeah. I have for conversational
stuff it needs to be it needs to be and
there there is a very much an unediting
trend that you're seeing happening on
YouTube, too, right? Where it's like
like I I keep seeing that this like a
grandpa who's like smoking a cigar like
in in the in the field and there is a
whole [ __ ] maxing movement that's
happening, too. Like it's kind of
people don't want the Mr. Beast over
retention editing ding ding ding ding
ding like third 30 things to capture
attention. There is like a sweet spot
which is why I've actually really
enjoyed long-form content cuz it's fun
to see and and though I will say this is
where Codec going back to Codec is so
cool. I am learning things about YouTube
analytics that I have would have never
learned otherwise. It's it's crazy how
much they actually expose like in terms
of figuring out how your content's in
the the retention curves are so useful
and just to be able to see AVD like
Yeah, but there's so much stuff that you
can go into
like completion rates and all this crazy
stuff that like is
it's like I I'm just usually like hey,
go you I've made a skill that's like the
best of Colin and Samir and Patty
Galloway and Daryl Eves kind of
mushed into one and it's it's glorious.
>> You brought up connecting Codec or any
agent to YouTube analytics. So, YouTube
has an API. It's kind of annoying to set
up. I don't know does Codex have an
official integration yet or do you have
to like set up the API?
>> I don't know, but I went through the
pain of turning on the things in cloud
and yeah, I I had to go through the
whole process. Yeah.
>> Yeah, I hope you weren't the guy who
made it so complex at Google. Um, but
basically all of the data that YouTubers
can see,
um, all of the data that YouTubers can
see like your retention curves, the
click-through rates for every single one
of your videos, you could just give that
access to Codex. And I think before this
video you said something you said
something that like you're like AB test
it. I don't even want to see it. Like
it's so it's so nice to not have to
click through toggles. I think that is
kind of
>> Yeah.
>> one of my it seems boring, but like one
of my biggest dreams of of AI or at
least my the ideal way of me
using a computer is never going to an
app again. I just want to talk to my my
chatbot. It can go off and do all of
that stuff. Be like, "Okay, I want you
to analyze the last video. How can we
make it better? What was the retention?
What was this?" And it just shows up.
And if you want that to happen again,
you just set an automation. You're like,
"Okay, every time I get on my computer
or every morning at 9:00 a.m. create a
report for any new YouTube videos that
get created."
>> Yeah.
>> And it shows you that data in the exact
way that you want it. And I think that's
kind of
it's just so powerful.
>> Well, it's funny you mentioned Fable,
right? It's like it's funny I'm I treat
Fable the same way. Like when I want
like the, you know, the artistic
high-quality decisions, I'm like, "Okay,
this is a this is a Fable task." When
I'm like, "Okay, this is a
well-understood thing or I want some
somebody to go find the nitty-gritties."
I'm like, "Codex." And it it really has
started to to your point about being
agent native in the show itself. It's
like it feels like they're members of
the team. It really does feel like that.
And so, I had this experience where
like, you know, it's like it's it feels
like I'm texting a producer basically.
To your point about thumbnails, like if
if folks aren't like active YouTube
creators, one of the worst things that
happens as a creator is you go upload
your video and YouTube greets you with a
glorious one X out of 10 ranking right
as the video goes up. And it's comparing
your video performance to the last 10
videos you've uploaded, you know,
basically minute by minute, hour by
hour. And it is the ultimate slot
machine, like in the in the sense that
like if you get a one on 10, you get
these little fireworks and you're
feeling like a champ. And if not, you're
sitting there feverishly going back
every half an hour refreshing to see how
it's doing. And now that I have Codex
driving this part, to your point about
AB test, I'm not tempted to do that
anymore. I'm like, how's it going? Okay,
this is what it looks like. Decision
made. I don't want to touch it, see it,
and accidentally get sucked down the
rabbit hole of this like, you know,
immaculately designed dashboard.
>> I think that is This is something I've
been talking about since 2023
about kind of my dream with Siri. I've
been talking about this for a long time.
You know, so much of tech is dark
patterns. You know, just Instagram,
every single social media platform is an
algorithm that wants you on their app.
And I think I think the only way to
really counteract that algorithm is to
create a completely new operating system
that has your own algorithm that's made
for you. You know, and so like that's
kind of what you're describing. It's an
algorithm
or or this kind of layer of all this
data that you want. Like you want to be
able to go on Instagram to find
something without getting distracted for
an hour. You want to be able to use your
computer
in a way that's like very human-centric.
And I think the way that this is going
to be solved is through an AI chatbot
that you can trust. And that's why I
never trust a free chatbot because a
free chatbot is always going to be
incentivized to keep you talking to the
chatbot for as long as possible. They
need to find other ways to
um other ways to uh monetize. And so,
what you want is an AI system that you
can go to that is like mediates your
experience with all these different
apps, so you don't get sucked into these
negative dark patterns, and it and it
has access to all the data, so it can
tell you how you can make your life
better, how you can make your content
better, how you can make your business
better. And we use that in our startup.
Like we have all of our data is
accessible via AI agents. There isn't a
part of our business where you can't go
in and analyze
every part of your business. And since
we set up scrape creators, which is an
API, I can pull all of our social media
data from any platform, and including
the videos and analyze all of the videos
and such. And so,
I think we're entering this new world of
of
kind of two things. Like one is just
like the AI native person, where they
anything that involves tech, they'll
first go to their AI, and it can do
things on their behalf and show them
what they need to show. And then the AI
native business, which is
anything you might need for your
business, anyone on the team can talk to
an agent, and they can get the correct
answer back, which I think is the next
level for creator businesses, as well. I
I want to
I want to run something by you real
quick. So, I think a lot of the personal
workflows that you have I I don't know
how big is your like editing team off
like I I don't know how you you have
like contractors that you work with for
editing.
>> like just two editors. Yeah, one motion
guy, one editor, yeah.
>> And did do they work together? Are they
an agency, or do they work together?
Okay.
>> They're basically a little agency, yeah.
>> Cool. One thing that would be cool is
like shared skills, and I think Claude
Anthropic is trying to go in this
direction with Claude Tag. Have you
tried Claude Tag yet? Do you know what
that is?
>> I I know what it is, and like that seems
like the sort of more enterprise
acceptable version of what Buzz is
trying to do, right? Cuz most companies
aren't going to move off of Slack, but
they'll try the beta for Claude Tag.
This idea that like you've got an agent
that has provisioned access to every
aspect of your company is super
powerful. Yeah, and like in in my in my
world it's just like Codex at the
moment, but it's like I could not agree
like that if you can have conversations,
have agents with contacts, and start
assigning tasks all in again like a sort
of a shared canvas for collaboration, I
think there's like so much potential.
>> Yes. And like
uh you know, as we I've built out my
team, the agents inside Slack are
becoming much more like useful, right?
Because I want anyone on my team to be
able to use the skills that I use. And
that that takes a lot of communication
from me beforehand. But you did bring up
Buzz, so I do want to talk about Buzz.
You want to talk about Buzz a little
bit? I I you said you're not convinced
about Buzz.
>> No, no, no. I just need to look into it
more. I mean like I liked it. I saw your
last video about this, right? Like
whatever a couple days ago that it came
out. I think it seems really cool. It
feels like it has all these aspects of
like Open Claw that was appealing where,
you know, I would when like when I made
the first like world view and God's eye
view stuff, it was all
Open Claw is the harness derived, which
I guess is just Pi driving Codex and
Claude code. And it was probably token
inefficient, but it was so much fun for
me to just be able to like say drop
voice notes and like sit in the 2E and
like say stuff that like, you know,
I wanted the benefit of that, but in
like a more natural intuitive interface.
And it feels like Buzz does exactly that
with ACP, right? You just connect all
your different agents to it, and
suddenly this thing can be that hub
where you can have Claude and Codex and
insert other AI harnesses like all
talking to each other. So, that that's
super exciting. I I definitely want to
Do you would you say that it it's like
fun enough to like is it good enough to
just like replace Slack wholesale and
just move over?
>> No. Um
I I again, remember we talked about Open
Claw about how a lot of it wasn't
necessarily immediately productive to um
to kind of mess with it, but like
through the process of testing the new
technology, you learned a lot.
>> a ton.
>> You learn a ton, you learn about
connections, you learn about the
capabilities of agents. And I think
that's what this is for me. This
>> Mhm.
>> platform is seems to be like an early
adopter platform in the sense that
I don't see teams switching over to this
anytime soon. I I could be wrong because
I think it has everything they it might
need to It's just too complicated. What
I think right now we're entering the
next frontier in this agent adoption,
right? I think the first half of I guess
I guess we're 7 months through 2026. I
think the last 7-8 months has been
around personal agents. Open Claw was a
very personal agent. It was a single
agent that you gave a computer that you
could message through a channel. Right?
And then from there we saw, you know, a
lot more and even before that people
were using Claude code, which is an
agent running on their computer, and
then Codex, and then Hermes agent, for
example. It's kind of this personal
revolution. I think
>> Mhm.
>> what Buzz signifies to me is kind of
this transition into a team of agents.
What Claude Tag represents is like how
do you use agents as a group of people?
How do you create a company second
brain? Whether it's
>> Mhm.
>> whether it's for your team of three or a
team of 10 or a team of a thousand. The
last episode I did was with Guillermo
Rauch. That's coming out in a few a few
days at Perplexity. He they have a V
agent that their whole company interacts
with. And this agent gets smarter. They
have a whole team managing this agent
because it helps them so much. And so I
think that's the era we're moving into
right now is like how do you put agents
into teams? The reason Buzz is
interesting is
>> Sick.
>> Buzz can create You know all those like
if you have a Slack and you add more
people, you need to have the admin add
people. You need to create shared
channels and you need to give proper
permissions and you're just going
through the toggle. Well, Buzz was
created so that an agent can do all of
that. So like I could ask any of my
agents. And so this allows you to add
any harness. So I can add the Codex
harness, I have the Claude code harness,
I can add cursor, and the default model
with cursor is Grok. And I can just at
mention all of them, and they'll
actually collaborate. Buzz did a great
job with their system prompt so that
they actually collaborate in a way that
like I'll come back 15 minutes later and
they'll have a good conversation. And
then Codex will be like, "Okay, based on
all the information I have, I will start
on this project." And so seeing agents
work together is such an interesting
>> In terms of keeping each other informed,
is the chat do they just read the text
and figure out the right state of what's
do like basically status of everything
in the chat itself? Does that end up you
know, making it so like so like what
stops like a race condition of both run
off and try doing a thing or something
like this?
>> So, I I I said this in the Buzz video. I
said, "I've tried this before and it
does create usually this race condition
where they talk for way too long,
context gets super clouded." Buzz
doesn't do that. So there's basically
like a From the way that I understand
it, there's like another Buzz
wide system prompt that gets injected
every time they interact. And for
whatever however they set that up, I
haven't analyzed how they set it up,
they don't do that. It'll usually be
like one or two interactions. I'll be
like, "Hey, discuss it with Claude
code." And then I'll be interacting with
Claude, right? Claude is Fable by
default. So I can message Claude and I
can say hi, and this is this is
literally and then they'll like at
they'll like show the little
uh emoji. And then you can see which
one's working. You can click on their
activity, you can see that it's working
at all times. And then once it responds,
it'll respond. Um and I'll at mention
Claude, and I'll be like, "Hey."
And I'll be like deep in a conversation.
And then you can just see that's like
"Hey, uh what are we working on?"
>> These chats persist in your Claude code
locally, right? As well.
>> Yes. Yes.
>> Right?
>> Yes. So it's basically just using Claude
code under the hood. In fact, if I were
to say like I can and Claude code um,
ask Codex about what I talked about
today in the Codex app. And so, this is
using
Mhm. And I can I don't know. Maybe I
need to mention Codex.
Actually, I don't think so. Cuz here it
says it's mentioning it. And it can
mention any of the other agents, and it
can ask Codex. And so, oftentimes I'll
be deep in a chat session with Fable,
and I'll be building something, I'll be
coding something, and I'll ask say like,
"Hey, Codex has a skill that allows me
to do do this thing. Can you ask Codex
for that skill, and then use it and do
it properly?" Or I'll say, "Hey, can you
just ask Codex to do this one thing for
me?" And because it's all in one shared
workspace, they just pass context back
and forth to each other, which
>> that.
>> Because you said you use Claude for a
lot of things that are And I agree with
you when you said like, "I think that
Claude is way better at creating like
documents, presentations, and then just
front-end design design sense in
general." Um
And so, you probably have different
skills on Claude Code than you do Codex,
correct?
>> That's right, yeah. Yeah. Yeah.
>> So now, you're in this shared workspace
where you can almost like have them If
Claude Code doesn't have a skill, you
it'll you can probably just tell Claude
Code that whenever you don't have a
skill, ask Codex if he has it. And then
maybe it can use that skill. And so,
it's not super refined, but like you can
kind of see where this whole trajectory
is going. And you can kind of do this in
Slack with And a lot of not a lot of
people know this, but OpenAI released
what's called workspace agents, which
you can now add to Slack. So, it's very
much like Claude tag, except it's it's
almost like a GPT work agent that runs
in the cloud, that has all these
different skills that you can give it
that can can connect to all the plugins
your Codex can. And you can give it a
personality, and you can let your whole
team message it, which is
which is really cool. So
>> I'm going to try this out. Yeah.
>> Yeah, based on everything I said here,
like do you think this would be useful
for your workflow, like using it with
the team?
>> I mean, like this has many of the fun
aspects of what I was doing in WhatsApp,
like group chats. I had this thread
called the council, like many people,
with all the agents, right? And then
when I'd be thinking about a new idea,
I'd like to see it attacked from
different perspectives with different
models weighing in with their own
nuances. And so, like just to be able to
do that all in one place, but also have
it multiplayer with other folks, too, is
really cool. Um yeah, like this has
in many ways it has a lot of that charm
that like like those magical moments
that like early open claw kind of gave
you if you're in like a big Telegram
thread with your agents and somebody
else or, you know, uh what have you. And
it's it's so fun to collaboratively
prompt these systems, too. It's like
this new form of like pair programming,
like you're kind of like the in a sense
the agents are doing the work, but the
fact that you can like collaboratively
do this. I'm already having some
experience we're doing this with Slack.
Um like
with with our own agents, like what is
it?
A GLM and Kimmy. And it's like a code
review agent. It's like a lot of fun.
>> Yeah.
Uh it is it is just fun. And I think
we're in the early stages. So, this is
probably like January of open claw when
people, you know, some people were using
it, but like no one shared anything that
was like quite useful yet. And so, I
think the team of agents is going to be
kind of the second half of this year.
Okay, so before I let you go, I want to
talk about your viral YouTube video. You
had that viral project that you created.
It got 2 million views. Can you tell me
a little bit more about that?
>> Hell yeah. I mean, that was an
interesting experience. It
In many ways it wouldn't have happened
if not for open claw.
Um yeah, kind of kind of the origin
story is basically like
it started off with this tweet.
Basically, I was playing around with
Gemini 3.1. And as I mentioned, Gemini's
much better at spatial reasoning than
Claude 4.6.
And, you know, for those that don't
know, I spent, you know, a decade in
tech mostly at Google working on like
geospatial 3D mapping. And one of the
things I worked on was like
photogrammetry. How do you create this
3D model of the world and put it out
there with like 3D tiles? How do you do
visual positioning and so forth?
But it's like a static rendition of
reality. And I got into vibe coding. I
was using Open Cloud at the time as the
harness that was driving perhaps very
token inefficiently,
um, you know, these these different
systems. And basically the idea I had is
like, well, I want to paint the world
with information. And I went down this
like crazy rabbit hole of like, well,
what kind of information is there to
paint on top of the world? And went deep
into open-source intelligence. I knew
something about like the few layers
available, but as I went deeper with
agents, it was like there was so much
more out there. And it turned out all
the CCTV cameras in Austin are actually
open. Like you get an image once every 5
minutes and there's some very
interesting things you can do with that.
So I made this initial prototype,
thought not much of it, went to sleep,
and I woke up to it trending. Like men
and all these folks were like
like tweeting about it like like oh,
this guy like vibe coded Palantir. Now,
it's obviously very different than
Palantir to be clear.
But for whatever reason, Joe Lonsdale,
one of the co-founders of Palantir, goes
on TVPN and talks about it and basically
says like, "No, no, no, no, like
proprietary data fusion. This isn't like
only low-end SaaS is at risk." And so
like the next week happens and I'm like,
"Ooh,
like there's a war breaking out." And it
was kind of funny. Like literally, so
for this video over here, um,
I was literally sitting on I was
literally sitting on my like couch
watching CNN and the news is going down.
And I realized, "Holy crap, I've created
the exact right infrastructure to
basically store all these OSINT signals
that you like satellite tracking, plane
tracking, vessel tracking, all the
social media feeds in and around area,
be able to geocode them and visualize
them." And I created this like 4D
reconstruction of the the
24 hours of epic fury. And then after
this, I did a follow-on with, you know,
tracking the Strait of Hormuz and a
bunch of other things around it.
So, what it made me realize is that like
honestly, there is no such thing as like
a audit trail for physical reality. If
you want to go get this information, you
got to go to like seven or eight
different tabs,
be a geospatial expert to be able to
fuse all this stuff together. So,
basically what I've been doing is like
experiments that take everything from
space to ground, including the recent
stuff you're seeing with like meta
glasses, phones, and so forth. How do
you put this all into a scrubbable 4D
globe? So, that's basically what I'm
going about building. Um it's really
cool. I got a co-founder now, um like
really awesome CTO, worked with him at
Google, former Nvidia, and we're
building something new. So, I'm excited
to share more on that.
>> you're going from vibe coding Palantir
to you now have a CTO who can actually
just code Palantir.
>> Well, you know, I don't know if it's
like uh Palantir per se. I think there's
plenty of
uh
the way I'd put it is like a publicly
legible Palantir. Like I think Palantir
is really interesting and powerful to if
you have proprietary data sets, you
know, to make sense of what's happening,
you know, if you're an institution, for
example, uh you're or a big enterprise.
What I'm really excited about is taking
open and commercial data, and I cannot
tell you how insane the conversations
have been over the last few months. Like
journalists are interested in this
stuff, activists are interested in this
stuff, defense primes are interested in
this stuff, defense tech startups are
interested in this stuff, the Department
of War reached out. And it was like this
thing where like it was this crazy thing
where it's like the same thing appeals
to sort of both sides of the aisle, and
it tells me that there's something
unique here, which is like and to me, I
boil it down to really like how do you
build a window through which you see the
world where it's not like three dots on
a map and like four people on a news
channel talking about it, but you can
get as close to making sense of it
yourself. But given I'm a geospatial and
3D person and so is my co-founder like
we're really approaching it from like a
3D first perspective. You know, you
could build like a Bloomberg terminal
for example. That's kind of cool and and
and has its place, but really this is
about how do you take all the sensor
data
in the world that's like you know
publicly available or commercially
available because let me tell you there
is so much crazy data that's
commercially available including like
satellite providers that are doing cool
stuff, but people paying like fisheries,
you know, like basically hey, we'll put
a Starlink on your fishery so they can
start collecting AIS data for vessel
tracking data in and around. And so
yeah, build the window through which you
see the world and create an audit trail
for physical reality. I'm going to open
source the original project later this
month. So that feels more like a I would
call it like a
like sort of think of it like a like a
spy simulator in your browser. That
feels like you're in in CENTCOM like
with a dashboard pulled up. And by the
way, there's a lot of people that just
want to throw this stuff on like a
massive window and like you know, just
like pop two Zins and like drink an
energy drink and just monitor the
situation. That's cool and so that's
what this open source tool is basically
all about. It's like a spy simulator. It
feels it has those like spy thriller
aesthetics, but it's underpinned by real
data and I'm particularly trying to
focus on like finding the cheapest APIs
possible for the open source release so
folks can have a really good experience.
And then following that will be a
commercial product. So V1, the community
voted on this. V1's going to be open
source. V2 and onwards to truly scrub
the globe that there's a lot of
expensive data involved and a lot of
computation involved to get like 30 plus
layers to work nicely on a 3D globe.
>> The way that you did this is so cool
because correct me if I'm wrong, you
probably didn't manually go look for all
of these data sources. You probably had
your agent go and and try to find a lot
of this data, right? When you first vibe
coded it with open clock?
>> Hell yeah. It was literally like go do
some deep researches on like what is
actually out there. Let's go figure out
every single API, what's open, what's
not. And the fact that I could just do
this like sending voice notes or like
doing voice dictation, it's like I had
my buddies from like Google Maps hit me
up. I was like, "Dude, you did this in a
freaking weekend?" Like that
>> Dude.
>> And it spawned so many people creating
cool similar things.
>> This origin story is insane. Right? You
got open claw to create a little 3D
simulator that uses real data, that it
went out and found the data. And then
you were just having fun. You created
this project, you threw it on Twitter,
you made a YouTube video. And then I'm
guessing when you say the community
decided, a lot of that community was
formed through the YouTube video, I'm
guessing. Like that's where people
>> YouTube and Twitter have been definitely
been the two. But dude, it's like people
reposted it on Reddit. It that thing
went so it it like was kind of beyond my
wildest imagine. I've had stuff go viral
in the past before, but it felt like
this was like one of those things where
people talk about like PMF, like product
market fit. It's like people are like
not like, "Ooh, this is cool. How do I
build it?" People are like, "I want
access to this right now. Where do I
swipe my credit card?" Like that was the
reaction.
>> Con-
>> And
>> content audience fit is what I call it.
But like yeah. And I I think I think
that's first of all that's just like the
right way to do business now is to just
like create a prototype, make content on
it, and if
I mean this is like the best possible
case. I mean I I'm sure you realize that
going viral for a long form video is so
much better and so much more powerful
than going like viral on TikTok or
something like that because
>> Or even X, yeah.
>> Or even X because someone spent, you
know, 10 minutes watching your video,
and then now they're you probably have
thousands and thousands of comments,
people talking about like, "Oh, you
should add this. You should add this."
And so I think what a cool
story. I mean you've truly vibe coded a
prototype, and now you're you're doing
it. And so yeah, I guess what does that
look like for you? Like how much time
are you spending on this um and then how
much time are you spending on content?
>> Man, that's been uh the hardest part. So
I'm trying to scale up my team and uh so
if you have recommendations on editors,
let me know.
>> Oh, on editors. Okay.
>> I so I'm looking for a producer and an
editor. Like I basically like I really
enjoyed doing the producing aspects
previously and now it's just like I
don't have time cuz I'm spending most of
my time actually coding and building
this thing out. The whole point is like
the content I love making content about
spatial intelligence. That's always been
my shtick and I want to keep doing that.
And as you can see with the other video
content I've been producing, it's like
basically what I'm trying to do is also
like, you know, again, I call them
frontier maps. I want to give away the
recipe too. Like a lot of people like
advised me saying like, "Why are you so
detailed about how you created this? Why
are you putting like which APIs you used
on your Substack?" And I'm like,
"What what are you This information
needs to be free." Again, like this is
and the fact that you see people riffing
on stuff and then they cite you as
credit to go do it. That to me so
enriching cuz they give me ideas. So
that feedback loop I think people should
spur and I want to lean more into that.
So the content has to stay.
>> Never let Never let someone get you to
not do that because what you own is not
the software. You own the movement,
right? And so as soon as
Yeah, as soon as you close it off, you
don't let anyone else participate except
for as a consumer of your technology,
then you also allow someone else to
create the movement, right? In that in I
think we're kind of in that era of
software. So Sorry to interrupt you. I I
just
>> No, it's beautiful point.
>> Yeah, so many people
will advise you to just like keep all
the secrets for yourself. But if you
look at all the people who are crushing
it in business and content where they
don't really have any secrets, they just
they truly own the movement. I think of
Alex Hormozi a lot with the business
side where he just like
all the business people try to release
courses and they don't even come close
to the content that he just gives away
for free on on a business side. And so I
think
>> that's how you win is you just create
the biggest movement you possibly can.
>> And make it like give people the
templates to have fun with this stuff,
right? Like half of this stuff is like
we're all discovering what these agents
are capable of and it's so fun to me.
Like I literally say this at the end of
my videos is like point your clanker at
this YouTube video transcript, tell them
to go to my Substack, use that as a map
and make your own thing. And I've gotten
outreaches from like even for like the
the shot tracking application I made.
People are like, "Oh my god, this is
like at this intersection of three
different things I care about. Check out
this thing I made." Or hey, check this
out. You know that topic you covered? We
actually have this kind of system on an
airplane. You want to come out to New
Mexico and like ride in it and like we'd
love to see what you Like it's it's
insane the things that happen when you
just like kind of give people maps to
build their own journeys, right? In a
sense. And so so that to me is very
exciting. But yeah, like coming back to
the the content thing cuz yeah, if you
if you know good producers who are good
at like um you know, essentially or
anyone listening to this that like like
likes the kind of content I make and is
interested in taking like my mad science
experiments and like ideas like to like
these complicated topics I'm trying to
cover and make it accessible. Like my
the goal for my video content, if you
watch any of my video essays, is like if
an expert in that field sees it, they're
like, "This is a good distilled
summary." And if a normie sees it,
they're like, "Oh, I actually understood
that for the first time."
It can't It There's too much, you know,
people are There There's a lot of
content out there that just gets a
little bit more um
I don't know. Especially in this space
with geospatial and kind of like
real-world understanding. It's so
dual-use. There's a lot of either
fear-mongering or ultra-maximalist like
defense tech hoorah. And it's like
there's no nuance in the middle and it's
like
>> Sure.
>> you know, so
trying my best to to strike that
balance, but
>> Yeah.
>> Gosh, it's so much fun. I think I have
way way more screen time and
Vibe coding is probably more addictive
than social media. I'll tell you that.
That's for sure.
>> 100%. It's so fun, especially the way
that you've done it. I think I think
honestly, I when you open source that, I
may just try and have a little fun with
it. Maybe I'll do a live stream. That
sounds like a absolute blast. Um anyway,
>> dude, this has been so much fun. We
should do this again soon. Um thank you
so much for for coming on the episode
and I wish you the best of luck with
your company. You're going to I I know
you're going to crush it.
>> Thanks for having me.

I've never looked at a skill file once.
It's test-based. Ask AI to do a thing to
use a skill. If it does properly, you
tell the AI, "Hey, you didn't do a good
job." And then, please change the skills
so that you don't do that again. Then,
you want to go to a new chat so the
context is cleared so you can kind of
test it again. Test the agent in a new
chat. See [music] if it does it
successfully. If it does it
successfully, great, right? And once it
does it incorrectly, you update the
skill, right? That is the way better way
to do it than to manually edit the
skill. You just want to be
outcome-based. Hold your AI to a certain
standard, and then you can say, "Turn it
into a skill, and you can improve it by
measuring their outcomes."

What if I told you that figuring out
what parts of your work you should be
getting AI to automate was a simple math
equation? This week two new products
came online that make getting AI to do
work for you much simpler. Grok Bot
gives users the ability to teach it a
task by manually recording them doing
something while ChatGPT's computer
history watches how you work and learns
over time. Together these represent the
shift of the biggest challenge in AI
moving from capability to context. But
as these new features come online, you
still have to figure out which part of
your work you want AI to automate. The
work best suited for AI deputization is
frequent, time-consuming, teachable,
easily verifiable, and doesn't require
you to have been the one to do it to be
successful. The AI Daily Brief is a
daily podcast and video about the most
important news and discussions in AI.
Welcome back to the AI Daily Brief
Headlines Edition, all the daily AI news
you need in around 5 minutes. It's
almost like Google heard us yesterday
talking about how many people were
talking about SpaceX AI as though it had
completely usurped Google in the
pantheon of serious frontier model
companies. On Thursday, Google released
Gemini 3.7 Flash. And while it is
neither the much-delayed Gemini 3.5 Pro
nor the now increasingly anticipated
Gemini 4, it does play in a different
category of efficiency that's becoming a
higher and higher consideration,
especially for serious and advanced
users.
So, let's start with what's good about
this model. It appears to be very, very
fast. During testing from artificial
analysis, the model ran at 340 tokens
per second, which is an entirely
different category than anything else.
It's more than twice as fast as GPT-5 6
Luna and even a bit faster than Nvidia's
new Nemotron 3.5 lightning. On the
benchmarks, Google made some solid gains
over 3.6 flash, most notably improving
their score on coding benchmark Deep
Sweep from 48.6 to 65.3%
and with this model, Google is also
slashing prices by half, making it a
little more cost-effective than its
predecessor. Unfortunately, just like
3.6 flash, by optimizing for speed, the
model kind of ends up in a strange
no-man's-land. Even with the cost
reduction, the model still costs 40
cents per task on the artificial
analysis benchmark run. That makes it
the same price as MuSpark 1.2, slightly
more expensive than models like Nemo
Tron 3 Ultra or GLM 5.2, and around
eight times more expensive than the
ultra-cheap models like GPT 5.6 Luna. At
the same time, it feels like the model
just isn't strong enough on the
benchmarks to justify the cost
difference. Cognition pointed out that
3.7 Flash has Sonnet 5 coding
performance for less than half the cost,
but to put it mildly, Sonnet 5 has not
been a hit. Most users are either paying
a little more for a frontier model or
looking for a much cheaper model. None
of which is to say that Gemini 3.7 Flash
is bad, it just sits in an uncomfortable
middle ground right now on the
cost-per-intelligence tradeoff. And yet,
I think it would be a mistake to assume
yet that we really understand how on
aggregate these behaviors are going to
shape out. In practice, some users are
reporting increased utility,
particularly with the boost in coding.
Brandon Galang of Vercel wrote, "Ignore
the FUD on Gemini 3.7 Flash, it actually
sits on the Pareto frontier. It still
technically gets edged out by 5.6 Luna,
but that's extremely deceptive. 5.6 Luna
is a much smaller model, and while my
team has cut over a lot of production
workflows to it, I personally would not
turn to it for coding tasks. Gemini
models have always had strong pros and
multimodal understanding. I'm actually
feeling somewhat iffy on Grok 4.6 from a
model behavior standpoint, so for now
I'm giving Gemini 3.7 Flash a try as my
daily driver and execution model."
Analyst Max Weinbach agreed, saying,
"I'm really enjoying Gemini 3.7 Flash in
Antigravity. It's really fast, seems to
be really good, and the usage limits are
insanely high." Now, it's worth pointing
out that for partially synchronous
coding tasks, i.e., coding tasks that
aren't long horizon where you just go
off and let the agent work, but where
you are actually sitting there
interacting with what the agent is
producing. That speed boost could make a
really big difference and actually
justify a slightly increased price tag
even at a decreased performance from the
frontier.
The other big use case is improving
Gemini Spark, which is Google's personal
agent. The massive speed boost, the
implied increase in compute efficiency,
and solid improvements on coding and
white-collar work benchmarks make it a
big upgrade for Spark.
If you're trying to take away anything
from this, it's another reminder that
the model race, {quote} {unquote}, is
branching out into multiple races.
There's certainly still the race for the
frontier, but there's also a race for
distribution. There's a race around
harnesses, a race for revenue, and
within the context of models alone,
Google seems to be betting that speed is
a dimension that people will pay
attention to as well.
Now, speaking of how we understand the
model race, we just got a really
interesting new study that confirms
something that we discuss on this show a
lot, which is that switching to a
cheaper model, particularly a Chinese
model, won't necessarily deliver savings
to the bottom line.
A new study from Alpha Sense looked at
how a range of different US and Chinese
models perform in real-world tasks. It
tested GPT-5.6 Soul from OpenAI, as well
as Opus 4.8, Opus 5, Sonnet 5, and Haiku
4.5 from Anthropic. From China, Alpha
Sense tested Kimi K3 and GLM 5.2. The
study also included two US open-weight
models, Incling from Thinking Machines
Lab and Gemma 4 from Google. The models
were tasked with running through a
series of several hundred financial
analysis questions that required them to
sift through a large volume of data,
including earnings call transcripts, SEC
filings, and news articles. The outputs
were scored for quality, which included
things like factual accuracy and
including multiple analyst perspectives.
The results showed that both GPT-5.6
Soul and Opus 4.8 were able to deliver a
much higher quality result for cheaper
than either of the Chinese models.
GPT-5.6 Soul, for example, completed the
task for around 13% cheaper than Kimi K3
with a quality score around 20% higher.
Both Gemma 4 and Inkling were able to do
a decent job around the same quality as
GLM 5.2 for less than 1/5 the cost. Opus
5 was an interesting outlier delivering
lower quality responses than Opus 4.8
while also costing more than five times
as much.
Now, clearly part of the purpose of the
study was to test the theory that Kimmy
K 3 and GLM 5.2 had reached frontier
performance at a fraction of the cost.
In practice, what the study found was
that at least for this use case, they
were both around the same level as
Sonnet 5 and much more expensive than
GPT 5.6 Soul. Said AlphaSense CEO Jack
Kokko, "Some of the more expensive
models, the ones that look more
expensive based on just their price per
token, actually ended up being less
costly because they were more efficient
in using tokens. Especially for you
enterprise buyers and planners out
there, the faster that we shift our
conventional wisdom around this and
actually figure out the models that are
most efficient for our particular tasks,
the better off your budget is going to
be."
OpenAI has introduced ultra fast mode
for GPT 5.6 Soul, which they claim
delivers frontier intelligence at 14
times the speed. The new mode allows
Soul to run at 750 tokens per second,
which is more than twice the pace of
Gemini 3.7 Flash. For now, ultra fast
mode will only be available through the
API to select customers. The feature
leverages the optimizations OpenAI has
made to run models on Cerebras hardware,
so available infrastructure will be the
limiting factor. OpenAI has positioned
this as only really suitable for
specific workflows where latency is a
major factor, writing, "Ultra fast is
designed for businesses where faster
frontier intelligence creates a
measurable advantage, including
real-time voice and customer support,
commerce, coding and design, financial
research and security response." The
company did not mention how large the
premium would be for ultra fast mode,
but presumably it's not ultra cheap.
Still, especially in the context of what
we were just saying about Gemini 3.7
Flash, very interesting to see them push
into this area as well. One other story
from OpenAI, although this time on the
personnel side. On Thursday, the
company's chief revenue officer, Denise
Dresser, announced that she would be
leaving Open AI in the coming weeks to
pursue other opportunities. Denise
joined the company just 9 months ago in
December and was brought on at time when
Open AI's growth was being questioned in
the press, meaning that her significant
pedigree as a veteran tech executive
went a long way to easing investors'
concerns. Prior to joining Open AI, she
had spent more than a decade at
Salesforce, including spending her last
2 years as CEO of Slack. Alongside Open
AI CFO, Tara Fryer, Dresser was viewed
as a part of the new business-focused
executive team designed to get the
company in shape ahead of the IPO.
At the same time as they announced
Denise's leaving, the company also
announced her replacement, this time
with another industry veteran, Dolly
Rajek, the former president and COO of
Wiz.
Now, in general, I don't spend nearly as
much time on these sort of personnel
moves as some other media properties do.
I think that nine times out of 10, the
Occam's razor explanation for personnel
moves is in fact personal. I also think
that there is such a desire right now
among outside observers to read any sort
of shifts as tea leaves for major
problems inside that trying to do a lot
of analysis from outside is pretty
fraught. However, holding aside whatever
I think, the market is certainly
noticing this move. Part of that is that
it follows by just a couple of days
former COO Brad Lightcap's leaving, and
part of that is that those two
departures form what looks like a bit of
a pattern running for the exits over the
past couple of months. After former CEO
of Apps and then AGI deployment, Fiji
Simo, left in July due to health
concerns, sources speaking to Axios
suggest that President Greg Brockman has
been building up his own team of
leaders, and this move may be part of
that. Whatever the case, executive
turnover is being flagged as a problem
prior to the IPO, but then again, with
that now delayed until next year, Open
AI has a lot of time to build another
narrative before their Wall Street
debut.
For now, we will leave that there, and
that's going to do it for today's
headlines. Next up, the main episode.
Welcome back to the AI Daily Brief.
Today, we're doing a type of episode
that I'm going to be trying out a lot
more, which is effectively combining a
news story with a small activity that
hopefully makes this practical for you.
The activity today is called the AI
deputization audit, and it's a short
process for helping you figure out what
you can and what you should hand over to
AI. Now, the news stories that inspired
this are actually two-part. This week,
we got two new features released that
reflect the fact that the bottleneck in
AI has moved from model capability to
access to context. I.e., just because a
model can do something doesn't mean it
has the information and context it needs
to do it well relative to you
personally.
Now, this has been a problem for a very
long time, and we've explored a bunch of
different experiments to improve that.
For example, you can still go to
contextportfolio.ai
and do an interview with an AI that I
set up a few months ago to build a
transportable personal context file.
That, however, might seem incredibly
slow given what has launched this week.
The first feature in this area that
launched this week was one of the new
capabilities of Grok Bot. For those who
missed that episode, Grok Bot is Cursor
and SpaceX AI's new simplified version
of Open Claw, but for everyone else. And
one of the biggest and most interesting
features is the feature by which you can
teach Grok Bot a task. You simply hit a
little plus button in the chat and
record yourself doing something in the
browser. The bot watches, and then
theoretically, it can do again.
And while the team at Cursor say that
many things Grok Bot can already do by
itself, if you ever find your bot
struggling, teaching it a task is a way
to solve that in one fell swoop.
Now, that feature was interesting all on
its own, but it was made more so by the
fact that we got another feature in the
same family, this time from OpenAI.
That other feature is for ChatGPT and is
called computer history. OpenAI's Ari
Weinstein writes, "Computer history lets
ChatGPT learn from everything you do on
a computer, so it can better understand
how you work, finish tasks that you're
in the middle of, and suggest skills and
automations based on how you use your
computer."
The goal is exactly the same as teaching
GrokBot a task. It's to show ChatGPT how
you work so that it can do more of that
work.
Now, some of you careful listeners might
be thinking to yourselves, "Wait a
second. Didn't we hear about some
feature like this not so long ago that
caused a bunch of controversy?" And sure
enough, X users like Shove said, "I'm
old enough to remember when Microsoft
tried to do the same thing with Windows
Recall and everyone lost their minds."
What Shove is referring to was a feature
that was announced around the time that
Microsoft introduced its Copilot+ PCs
back in early 2024. Time explained the
feature like this. "Recall allows the
device to take snapshots of a person's
screen every few seconds. These
snapshots are encrypted and then stored
locally on the individual's device."
Microsoft said that this feature was
designed to, quote, "solve one of the
most frustrating problems we encounter
daily, finding something we know we have
seen before on our PC." But people did
not like this.
Dr. Chris Frishock told the BBC, "This
could be a privacy nightmare. The mere
fact that screenshots will be taken
during use of the device could have a
chilling effect on people."
Now, ultimately, Microsoft recalled
Recall and did later release a version
making a bunch of changes including
making it opt-in and giving users a lot
more fine-grain control.
So, what has changed? Is it just a
matter of time? Javier Lacourt writes,
"2022, I don't trust Google. I'm going
to get used to using incognito mode on
my phone. I switch to Firefox, VPN
always on. 2026, take my history, my
medical analytics, my bank statements,
access to my Gmail. Analyze this
WhatsApp conversation from 2018."
AI entrepreneur and content creator Theo
writes, "I'm so deep in my AI psychosis
that I think this sounds great. I'm not
going to lie. I wanted this since
Windows Recall or Rewind or whatever was
announced. It got panned so hard that I
kept my mouth shut. Enough time has
passed that I'm going to say screw it
and try this."
And I will say that I do think part of
it is simply shifting attitudes on
privacy. If I recall correctly, my
comments on Recall at the time were that
it sounded to me exactly like the type
of feature that people now would think
was completely insane and that people in
the future couldn't believe we
previously didn't have.
At the same time though, there is very
clearly a difference in the value
proposition. Think about that sentence I
said. With Microsoft pitching this as
solving the problem of finding something
we know we've seen before on the PC. Is
that really that big a problem? Is it a
big enough problem to be willing to risk
all the privacy considerations? I don't
think for many people it is.
But getting an AI agent to actually do
your work for you is a much different
value proposition. And what's clear is
we now live in a world where there's a
lot more choice on how those agents get
the information they need to actually do
the work. Simon Smith, for example,
points out that computer history
overcomes a limitation of previous
attempts because instead of taking
screenshots constantly, it's recording
interaction events, not screen or audio.
Broadly speaking, ChatGPT's computer
history and Grok Bot's teach a task
reflect two different ways that AI can
learn how you work. One, embodied by
computer history, we might call ambient
observation. This is where the AI
watches across apps, builds ongoing
context, and can learn about what you do
as you do it without any particularly
strong consideration on your part. The
Grok Bot teach a task paradigm we might
call deliberate demonstration. This is
where you intentionally press a button
where you are deciding and telling it, I
am going to now teach you a skill. You
press teach a task, show it the
workflow, it saves the steps as a
routine, and it can repeat that process
later.
Both of these paradigms have things to
recommend them. On the computer history
side, for a lot of folks, the fact that
this is just happening in the background
as they work will be the feature. The
fact that they don't have to do anything
intentional, and yet it's still learning
how to be more useful is a big part of
the value. On the other hand, I can see
for a lot of folks the deliberate
demonstration approach of Grok Bot being
much preferable. For some, and I would
probably put myself in this category, it
is likely to be a much more natural and
preferred pattern to be intentional
about which things you want to actually
show the agent how you do so that it can
presumably do them for you.
But then of course, the question becomes
which parts of your work should you be
deputizing to AI? Inevitably, different
people are going to come up with
different answers to this question. But
to help you think through which parts of
your work catalog might be well-suited
to these new capabilities, let me
introduce the deputization audit. You'll
notice that I am not using the word
automation, and I'm doing that very
intentionally. For me, at least
personally, the concept of deputizing AI
to go do something in my stead feels a
little bit more like the relationship
that I want to have with AI than
automating away and never thinking about
some task again.
I will publish this deputization audit
as an extension of the show. So, if you
go to AI Daily Brief.ai and click on
today's episode, you will find a link to
it there. Step one is to take an
inventory of your recurring processes so
that we can run them through the rest of
this system. I'm focused here on
recurring because, in general, those are
the types of workflows that this sort of
automation or deputization is going to
be well-suited for. And in general, I
have found that no matter how much we
strive for our work to be novel and
dynamic and different, a lot of it is
just this stuff that we have to do day
in day out or week in week out. So,
think about things like email triage,
weekly status reports, meeting prep,
research briefs, CRM or pipeline
hygiene, content repurposing, scheduling
and travel, vendor portal chores,
inbound lead qualification, metrics and
analytics pulls. All of these are things
that even the most dynamic person might
find themselves slogging through a
recurring process for.
Step two is that for each of those
different processes, we're going to give
it a score. And the score is going to be
across a five different dimensions.
Overall, then, these are the five
dimensions that I'm proposing make a
task well-suited to automation or
deputization. The first is whether the
task is even worth it to automate. The
second is whether the task is teachable.
The third is whether the task or process
is checkable. The fourth is what the
stakes of the task are, especially if
something goes wrong. And the fifth
criteria is how much you personally are
the key factor in the quality of the
output. In other words, could someone
else be doing this and it would be just
as useful or just as valuable or
accomplish its goal just as well.
So, criteria one is whether a process is
even worth it to automate or deputize AI
to do. And that basically comes down to
how often you do this thing and how much
time does it take. A score of zero means
that it doesn't happen very often and
that it doesn't take very long to do
when it does happen. Think a few times a
year and a few minutes at a time. A
score of one is that it happens
moderately frequently and takes some
time but not a ton of time. So, for
example, some process that happens most
weeks and is done in under an hour. A
score of two is a process that happens
very frequently and takes a lot of time.
For example, weekly or more and
extending into the hours. Obviously, you
can see here that the higher the score,
the more well suited to at least explore
AI deputization that task might be.
Second criteria is teachability. Could
you show this task in a 10-minute screen
share? A score of zero is basically
absolutely not. It would take months to
teach or is just impossible to teach in
this way. A score of one is that you
could demo it but there would be a lot
of caveats. A score of two means one
demo could easily cover it and teach the
AI that skill. The third criteria is
about the output results, the
checkability. How long does it take to
verify the output compared to how long
it would take to produce that output?
Obviously, if checking takes as long as
doing it yourself, then automation or AI
deputization is not going to be saving
you time. So, in this case, a score of
zero reflects the idea that to check it
means redoing it. A score of one means
that it is checkable but you're going to
have to give it a careful read. A score
of two means that a quick glance can
tell you all you need to know.
Criteria four, we have the stakes. I.e.,
how bad is it if the AI gets it wrong
and nobody catches it. A score of zero,
which again means that it's less likely
to be well suited for automation, is
effectively that the stakes are high,
that the problems would be serious or
irreversible if AI got it wrong. A score
of one means that doing something bad
would be embarrassing but fixable. And a
score of two means that it's pretty low
stakes. And if something goes wrong,
it's easy to redo. Last criteria is how
integral to the process you are. A score
of zero means it has to be you, again
making that particular process pretty
ill-suited for AI deputization. A score
of one means that your involvement
helps, but is an essential. And a score
of two means that literally no one would
notice if you did it or if a squirrel
from outside did it as long as the work
got done.
Now, from there you have a combined
score. Across the 10 total points, I
propose breaking it into three tiers. A
score of eight to 10 means that it's
very worth considering deputizing AI to
do that task. A score of eight to 10
means the stakes are fairly low, the
task happens a lot, it's highly
teachable, and the output doesn't really
matter whether it's you or someone else.
Those are the types of things that
you're going to want to hand over,
spot-check the output, and see how it
does. And if and when you are
experimenting with something like
computer history or Grokbot's teach a
task, processes that fall in this bucket
are probably where you want to look
first. Now, on the other end of the
spectrum, when the score is between zero
and three, instead of a suggestion of
deputizing, we'll call that a defend.
I.e., that you should keep that work to
yourself. Maybe this is because a
mistake would be expensive. Maybe if
it's because the person on the other end
of the line expects this to be you. It's
highly likely that even without a score,
you will understand which tasks you need
to defend, but at least this puts some
consistency around it. The challenge
will, of course, be that a lot of stuff
is going to fall right in the middle. A
score of four to seven I'm calling a
duet, where you're going to have AI do
part of the task, but you're still going
to be highly involved. Frankly, most
knowledge work today is going to sit in
this category. And the question will be,
over time, whether AI and the process by
which it learns what you do gets better
enough that more of those duet tasks can
move and become deputized tasks instead.
Which brings us to step four, which is
naming the blocker. What would
specifically prevent you from handing
this off for those tasks or processes
that land in something like the duet or
defend category? So, some common
blockers might be things like the work
happens across websites or legacy
software. Think vendor portals that have
no API. Another blocker might be that
the process is hard to explain in a
single prompt. I.E. you know how to do
it, but would struggle to write it down.
These are two categories of blockers.
The reason that identifying the specific
blocker matters is that new updates to
the tools that we have might change the
equation around certain blockers. For
example, for the two that I just
mentioned, these new systems, computer
history and GrokBot, both potentially
change and solve those blockers. For
systems with no API, computer use agents
that click the same screens you do means
that that might no longer be a problem.
And for processes that are easy to show,
but hard to explain, well, you can just
show rather than tell.
Some other blockers might be solved by
these new tools, but aren't quite as
clear. For example, the AI lacking
necessary context, I.E. not knowing your
accounts, your history, or your formats.
Something like computer history might
solve that, where ongoing observation
over time could build that up, but a
10-minute teach a task with GrokBot
probably won't. Now, some other blockers
are not going to be solved by these new
tools. For example, the work requiring
taste or judgment, mistakes being costly
or irreversible, or the work depending
on human relationships. None of those
things are inherently solved by these
new sort of teach a task capabilities.
And it is also worth observing that as
with anything, new capabilities do
sometimes create new challenges. If the
blocker to AI deputization is that
privacy or security makes it
inappropriate for the AI to do that,
these new tools could actually in fact
add a blocker as recording your screen
creates a new thing to secure. Step four
is that after you have identified the
blockers to your tasks in duet and
defend, is to see whether any of those
blockers are now solved by these new
tools. And with that, you now have a
deputization list, and hopefully some
new ideas for how you can take advantage
of these new tools.
Now, when it comes to day two and day
three sort of responses to GrokBot,
which I spoke about very glowingly on
the episode before, more people
experiencing it does mean in some cases
that people are coming up against its
limits. For example, super trainer, and
I mean that both in that she is super in
that she works with superintelligent
Nifar Gaspar, found that while she felt
like Grokbot would be good for a lot of
folks who hadn't built complete agent
systems yet, for more advanced users,
there were lots of challenges. She
didn't like that she couldn't be more
specific about which folder holds the
relevant context, and didn't have
control on things like model choice. So,
especially for some of you more advanced
users, the type of folks who have gone
through Clock Camp, Grokbot might not
ultimately be a perfect fit. And we're
starting to hear some of the use cases
that people are getting value from. John
O'Neil, who owns a plumbing company,
went from zero to automated dispatch and
office chores in 24 hours, despite not
having any engineers on the payroll.
Lots of people are using the chief of
staff capabilities, where they spin up
individual bots for individual tasks,
but then only interact through their
main chief of staff, who coordinates all
the other bots. Indeed, that pattern of
topic per bot seems to be one of the
early emerging best practices. Finally,
one use case that I'm seeing a lot that
might be a good one to experiment with
is the inbox Slack tracker and morning
brief. Matt Van Horn calls this the
universal starter job, same as it was
for every agent before it, except now it
takes about a minute to set up.
Regardless of what you test, if you have
access to Grokbot or computer history, I
do think it's worth spending some
conscientious time experimenting with
what these new sort of AI deputization
tools can do for you.
One of the biggest challenges, even for
people who are highly AI enfranchised,
is carving out the time to learn to do
something differently. In fact,
especially high-efficiency and
high-productivity workers often have the
way that we do things so dialed in that
taking a little while to totally change
that feels like a waste of time in the
short term, even if it would save us a
lot of time in the long term.
Still, I think the experiments are worth
it. And so, for those of you who want to
try, keep an eye on the costs coming
down on Grokbot as they expand access,
and tell me what you find. For now,
that's going to do it for today's AI
Daily Brief. Appreciate you listening or
watching, as always, and until next
time, peace.

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!

All right, so when we hired Victor last
month, I genuinely did not know if the
team would actually [music] use him. But
this is what one week of having Victor
as our AI employee actually looked like.
On Monday, Owen, our marketing guy, hit
one button on our Lifestyle Business
Academy scorecard, and Victor pulled in
all of the previous days' metrics in
terms of like our applications and our
interviews and our ad spend and posted a
full summary of all of those metrics to
the relevant Slack channel, which meant
everyone in the team could see it.
[music] On Tuesday, Nicole, our social
media manager, asked Victor to pull in
transcripts directly from YouTube videos
that I've just [music] filmed and turn
them into first drafts of Instagram
carousels so that we can hopefully drive
traffic to those YouTube videos. This
happened within a few minutes, and the
output needed very little tweaking. And
Nicole was also able to create a [music]
report showing how our Instagram account
is performing compared to our other
social media channels. On Wednesday,
[music] which is normally the day I film
my YouTube videos, we got our weekly
9:00 a.m. YouTube filming briefing,
which shows us over the last 7 days what
have been the numbers, what have been
the anomalies, what kind of stats are
our competitors doing, so we can see how
we're tracking against other channels in
our space. This used to be a manual
exercise for our YouTube producer, but
now Victor does it completely [music]
automatically every single Wednesday. On
Thursday, we asked Victor to analyze all
of the meta ads that we've been running
on Facebook and Instagram for a
Superfocus, which is a new software app
that we are developing. And then on
Friday, I had a chat with Victor myself
about various curriculum changes that I
was thinking of making to our Lifestyle
Business Academy. And because Victor has
context about all of the stuff that
we've currently taught and have been
teaching our students over the last like
9 months, he was able to go back and
forth [music] with me in a reasonably
intelligent way. So across the five
days, it was five different members of
our team delegating five different jobs,
and it was all to [music] Victor, our
one AI employee.