5 New AI Models That Are Smarter (and Cheaper) Than GPT-5

Next W cast.
d I am ce again with Joe Feer.
do what last week.
k down ries that we think are the most interesting that we think
ut, tal why you should know about them and be thinking about

g to br n a really cool demo of how to use some of these AI tools
s and h ut your products into AI and sell your stuff better

demos s that we’re going to share throughout this episode around

ado, le t get straight to it.
cycle i sounds pretty impossible, but that’s exactly what Sandler
ubSpot.
bSpot’s ls to tailor every customer interaction without losing their

e incre
jumped
druple.
ree tim er on their landing pages.
o see h ze can help your business grow.
gain to e?
n.
k.
s week ked with all sorts of different announcements.
bunch

oing to n like topics.
e talki t, I don’t know, a lot of news.
reat.
nd of f ‘s just going to get better.

said, a stuff happened this week.
models ame out from a lot of the foundation labs.
ne we d eally hear from this week much was OpenAI, which was interesting.
news,
be tal out them, too.

aight i t one.

t of Op is week was that the OpenClaw founder, his name is Peter
o OpenA
aw.
lead pe that OpenClaw open source project.
iar wit law, then you probably haven’t been paying too close of
world.

ough.
alled C t.
to the aid, hey, you can’t call it that.
like Cl
ine.
MaltBot ke a day.
ay they d it again to OpenClaw.
as a AI that you can run on your computer.
virtua r, but you can chat with it from wherever you like to

pp, Sla cord, any of those tools where you do messaging, you can
rsation your AI chatbot and it would go off and do actions on your

as muc little access as you want.
ess to ails and your calendars and your passwords and your API keys
nt.
e level u know, autonomy you want it to have, it can go off and do different
alf.
blew u it was the talk of the AI world still kind of is to this

ated it Peter Steinberger, well, he grabbed the attention of a lot of
up goi penAI.
s inter because he came from a previous PDF company, I think it’s
r somet


pany ca PDF.
stood f r Steinberger.
he crea s like PDF tool that is actually now used by Dropbox and
her big rms use this sort of infrastructure that he built to manage


p getti a hundred million dollar investment and then cashing
f going coming a hermit for a while before coming and stepping back
to deve nClaw.

interes
ck to t d of AI, like in 2025, so like last year, but it started
these nt projects and then OpenClaw ended up being the one that
ion for eason.
even ta bout like, could this be the first one person billion

ioned s g like, you know, I’ve done the whole like growing a company




artner someone awesome.
eyes.


f the p that he built between when he sort of came back online
t of in cture pieces towards OpenClaw.
y put s g out on X pretty recently saying, it’s kind of funny.
ine tha 40 failures before I finally made OpenClaw.
‘s not
tools I were the building blocks that became OpenClaw.

ive is ke a made up narrative for marketing.
narrat


story.

ing, it resting because he wasn’t really making money off of it.
ing he ing out of pocket to keep this thing alive and going.
relying ations through GitHub to sort of keep paying it and to pay
orkers re working on this thing.

ink he intention of turning this into a billion dollar business.
o eithe
hatter re I saw.
ves fas
weekly heet on what’s happening this week in AI with over 20 prompts
s that copy and paste, a quick reference table that shows you exactly
and ta breakdowns of the newest AI releases.
he desc .
to the
ing.
.
ting wi uple of the other labs as well, which is interesting because, you know,
, you k aude behind the whole thing, basically powering this,
him an
workin their biggest competitor, OpenAI.

OpenCl rybody that was using OpenClaw in the early days,
using laude Max plans, right?
n calle laude Max plan.
an, and ave a $200 a month plan.
cally l use, like, Claude code at almost like an unlimited level.
ome lim t you can get a lot of usage out of it for your $100 or $200 a month plan.
rkaroun se their Claude Max plan in tools like OpenClaw.
said, n de Max plan is designed for people who want to use Claude code specifically and use our tooling.
your Cl x plan outside of our tooling, that’s against our terms of service.
ike, es ly threatening people for using their $200 a month plan inside of OpenClaw, right?
of what ople going, oh, so they don’t even want us to pay their high-end $200 a month plan if we’re going to go use it somewhere else.
o to Op
d, hey, ree to use our $200 a month plan inside of OpenClaw.
t?
had thi tive of, like, Claude’s going to be the company that’s going to shut you down if you’re on their high-end plan.
he comp t’s going to say, yeah, come to us.


ig piec y he ended up going to OpenAI.
hropic tually approached him about bringing him in.

nversat h Zuckerberg, and Zuckerberg tried to bring him in, but he ultimately decided to go to OpenAI because he felt that OpenAI would, A, keep OpenClaw open source.
AI to c and say, you have to use OpenAI models inside of OpenClaw if you want to use OpenClaw.
t open

howeve ant.
lowed t
ds like pening is he’s going to OpenAI.
dy else d the development of OpenClaw.
o conti t of on its own within like a foundation away from Peter.
to Ope start developing their like more agentic systems at OpenAI.

just r e is OpenClaw will be a project under an independent foundation.

that.
e for,
d that article that came out two hours ago.

tty muc toryline there.
ilding tever OpenAI’s next version of an agent is going to be.
y, real kind of fumble the ball here, right?
alled C and when they named it ClodBot, everybody just sort of immediately associated it with Anthropic’s Clod, right?
ready h sort of like tie-in association to Anthropic.
reness .
y signi o Anthropic’s $200 a month max plan so they could go use it inside of ClodBot.
tting m tomers from it.
opic di whole like, hey, you need to stop using this name, and he decided to go to OpenAI, a lot of people went, eh, screw Anthropic then.

d the b
with th tum and what does it say?
has 100 us stars.
he bigg wing project ever on GitHub.
like t that like, dude, you would have had all that momentum by now.
king la about how OpenAI, you know, with their new monetization model they put out there, we’re like, what are they going to do?
hoot, m ‘s agents, you know?
attent all these people that are also paying attention to OpenClaw and Peter.

ps shif a little bit for them.

l like till early in the year, right?
bruary.
second of this year.
ready s up to be the year where, okay, agents are finally feeling useful, right?
ents wa buzzword, right?
hat as verybody was saying, oh, agents are here, we’re using them, that’s a big buzz.
sential , oh, look, it can click things on my browser for me, right?
ser is hlighted and it can move my mouse and type things on my keyboard without me having to move my mouse and type things on my keyboard.

y slow, lower than if I was just to do the damn thing myself.
open up our tabs and have it doing four different things very slowly for me, right?
the fir g, and I’m going to keep calling it Clodbot, even though I know it’s called OpenClaw now.
ght up
.
tter.
his was he first one that came along where you can just go and use whatever chat platform you like.
a dail , you can use Slack and treat it just like a team member.
ne.
your t ber comes back and says, cool, I did it for you, right?
erent t seeing your browser glow and then the mouse is moving around for you and doing things.
rying t our computer to do something that a human can do already, where something like Clodbot is tapping into the APIs.
wn soft hind the scenes to do things.
ling a unch of processes behind the scenes to make sure your outcome happens instead of like doing what a human would do to get that outcome to happen.
ople on r X.
I just cross.
eeted b n’t know who it was, someone, some tech person, but like she homeschools her kids and she’s like this open claw, like fanatic and is awesome.
but bas she’s like figured out how to train her.


ir open ame.
it’s no printing stuff on a 3d printer from whatever she told it to do with the school curriculum.
ike, ho .
t’s doi y itself without any human intervention in the middle.

eticall ‘s sites out there that have 3d print templates already, right?

es and verse and all sorts of sites where you can find anything to 3d print.
o and s hose sites, find the thing you need, download the STL file, send it to your printer for you and print it out.
ally, w this last week.
could a s like Gemini to go and create you a print file and it can send that to your printer and automatically print it out as well.

uch luc that, but theoretically you can.

on’t kn
erent u s.
Matthe n seems to be like way down the rabbit hole when it comes to the open claw.
s are a
tty in ds, but I’ve watched two or three videos where he’s breaking down like these insane
e like ight while he’s asleep, it’s going and doing a bunch of research and finding talking points for his videos and making thumbnails for him and, you know, all sorts of just stuff.
p in th ng and he’s just got like his messenger app full with like, here’s everything I did for you while you were asleep.

diately now, I thought of the whole agents are now coming to everybody, you know, like everybody is going to be able to at least start interacting with agents.
hat ope pushing towards.
ly ther
hat Met it just this week, I believe, basically talking about how Manus, who they acquired, is now, it’s all being usable on Telegram.
u can k do something similar.
your ta , Matt, of all the details, of course.
ta’s pl esponse to Clodbot.
s like.
re, lik anus deal was probably in the works before the whole Clodbot stuff bubbled up.

ed to h little too fast.
way thi l angled and the way they’re marketing it feels very similar to what we’re getting out of Clodbot.
y the s ng.
anus ag side of Telegram, which I think is hilarious because Meta owns WhatsApp.
t.
.

tually t works in WhatsApp as well.

it, li any of the messaging apps.
ng to m like, the headline tool that they mention is Telegram.
st supp hannel.
egram r w and WhatsApp and the rest are coming.
unny th , like, a Meta competitor that is one of the first ones that it’s rolling out into.

, the e API to deal with or something.
there’ o have been some relationship.
we got to market now.
hance.

ically ike this similar idea.
t’s any f level close to what you can do with OpenClaw yet.
t way.
Manus l doing is a lot of, like, taking control of a browser and doing things the way a human would do them instead of just, like, operating directly in the back end via APIs and stuff.

ee this
n it’s like the model race, right?
obably out that here, too.
all th f happens, you know, you have the other guy is going to bring something to market.
t be th or perfect or is just support of one channel.
it’s be a thing in all of our minds.
we’ve b d, like you said, like the agents, they’ve been around, but I feel like this is the year of the agents.
ear tha gents actually start to feel useful.
ere an eels like an employee and not like a gimmick, right?
ar, age e very gimmicky.
e contr y browser and, ooh, I can sit here and watch it do the things I would be doing anyway.
m a lit slower, but I could open multiple tabs and have it do multiple of them.
ly like e useful, like I can go away and have things happening for me while I’m away.
his cur rt of iteration of agents useful.
much b t checking their own work.
issue st agents is that they can go and act on your behalf, but they would screw stuff up and not realize they screwed stuff up.
of thes s, they can go and do some work, look at the browser, go, oh, that’s not the way it should look.
wn mist ook at the browser again and go, okay, now I got it right.
back t ner, whatever you want to call the humans in the loop.
ort bac owner and say that I’ve finished this now that I’ve double checked my work.
hecking ork thing is a big deal, right?
ver tri ike vibe code stuff, that’s the kind of headache you run into is that it’ll say, oh, I fixed that.
r brows it’s like, it’s still not working.

ll, I f now.
y work
er, sti working.

ixed it

is happ ow.
like t r models, like Opus 4.6, Claude 5.3 codex, things like that, they actually will look at their work and go, oh, it’s throwing up an error.

ing up r.

g.
report d say I’m done with my task.
uble ch themselves now.
leap fr we had this time last year.



saw th
nterest nt and you might know the math and the things more than I do, but it’s an interesting little segue here where he basically put something out that he was like, hey, so what happens?
to the podcast, but what happens when AI tokens cost more than your employees?

ting to t now where you might be spending $300 a day on AI tokens, but $300 a day is more than what you probably would have paid.
he same of work, right?
nd of t ment.
oes int ath and all that, but I thought I’ve not totally thought about that before, but it’s like a lot of people are so hyped up on agents and automations, but then we might not be doing the math on what it’s going to take to actually complete the job effectively.
ke a ke , though, is he says in the short term.

point t the short term.
token do what an employee does plus fully encumbered developer and maintenance cost divided by fully encumbered employee cost equal less than or equal to productivity, but takes eight clot agents at $300 for tokens per day plus $200 per day in dev slash maintenance to do what an employee does per day at a fully encumbered cost of $1,200.
ed by $ but then you need to factor in the productivity rate.
6x prod
ve issu morale, morality, whatever that can’t be quantified that need to go into this decision?
esting.
that mo more.
coding yself.
things.
had da e I’ve spent, you know, $100 in tokens and it’s like really added up quickly.
s a ver term problem.
into al ese model releases here.
things hese model releases is that they’re getting better and better and better and also getting cheaper and cheaper and cheaper.
tate-of t model for coding is probably Claude Opus 4.6, maybe ChatGPT 5.3 Codex, right?
-up bet ose two.
good c odels.
just ca and Sonic 4.6 is doing almost Opus-level coding.
for li ird of the price.
prices nute.
y what e.
e quite less, but they’re still getting Opus 4.6 level quality out of it.
going d y, very quickly.
roblem ght now.
m thing


proces ess.
to brin up just because I was like, oh, it’s interesting.
it’s li e models, which is what we’ll talk about.
o many pushed out all the time and it’s wild.

hings o whole agent topic.
AI int n and ads manager.
ke a la iness audience for this podcast.
ou’re g be able to have pretty soon.
now.
now.
next st its integration of Manus AI with some meta advertisers now being
can acc us functionality within the ads manager.
e a rol
ke ever as it yet.
you can tell like your Manus agent, go build me an ad for my

or this brand.
and li at your site for you, learn everything about your product,
nshots of what the product looks like.
ting pr s on your behalf and then just go and create your ad in
hout yo even touching the ads manager.
n Teleg not WhatsApp.
ep an e he stats as well because it’s all integrated in metas, you know,
there.
more cr .

whole t at I’m seeing from, I mean, all these creatives.
I have for a ton of years now and it’s changed.
like t e AI.
tives i al.
ngine, w, metas will kind of do the hard labor now.
day, i he humans dialing the knobs.
point,
o run a Facebook ads ourselves, we had to run ads.
every were in that ads manager making sure like our cost per clicks
a certa l.
underpe g and this one was overperforming, adjust the budget.
re mone ds this ad and less money towards this one or turning
her.
s const ke course correction that you have to do to keep your

ot that t sleep, right, it could keep your ads on track 24 seven
oking i
your c that you want it to stay within.
managi .
y need t to do that?

ould ju omething that the ads platform does for you without the

the fri own, you know, it could be to maybe take some of the creative
way, b though I’m sure that might get even more integrated into
now, ov .

what a is going to probably be best at in the, in terms of
e going oking at your website, reading like your sales copy on
ing wha product is all about, then deciding what the best imagery
onate w most people is going to be deciding what the best copy
onate w most people is going to be.
hose ad ou.
don’t ed AI agents or AI at all really to keep your ads on track,

to go i and say, Hey, I want to generate leads at $2 or less

er that it off, right?
, if th n that statement in the backend, that’s just like regular logic.
e AI ne .




rolls o
at’s go d out on the 16th.
, earli week, just seems like, okay, this whole agent thing, you’re
ng this
ted in different platforms in some way to kind of have their

erestin
ig wave , let’s call it 2023.
s the y every company put AI into their product, whether the product

ompany
mpany d , right?
‘re a c that does SEO services, but now with AI, we’re a company
it, bo ng services, but now with AI, like every product just now

s year t start seeing that same kind of trend, but with AI agents.

y feel cebook ads should need an agent to manage your ads.
ecommen algorithms.
he, if hen that sort of logic algorithms.
need A

lk to a to make that happen?

able t the backend and say, here’s the cost per lead I’m trying
et that .
es and .
to do t me?


like y but.

good th
.
, we’ll bout this real quick.
rt.
across e, but I haven’t dug in too much.
ading t ite is that because it’s from 11 labs, it’s customer support
ave it r own voice or use one of their like pre-existing, like
voices ustomer support for you.

s like.
g, what t have here?
can lo hings like your SOPs and then you can probably, you know,
, maybe ied agents, depending on what’s going on.
teracti ardrails.
teresti it’s emotionally and contextually aware of situations.
n the s
d conte aware.
on the ith an 11 labs agent for customer support, just start

hem.

y give t you want because they’re emotionally intelligent and
e pisse
ck.
ell you another podcast, DTC Pod, hosted by Ramon Berrios and Blaine

dcast a l things direct to consumer.
ver eve for starting, growing, and optimizing e-commerce stores and

ders, m s, platforms, creators, and marketing and growth agencies to
rand bu social media, paid media, email marketing, and more.
d in th es behind your favorite consumer brands, this podcast

n episo t what every founder needs to know about compliance, claims,

herever t your podcasts.

ounds l
ifer wi en Airlines.
you tod
just g eled and I’m stuck here in Orlando.
oing to sing my daughter’s birthday.


out tha
at’s go okay?
ll me w ight this was?
ht MD41

that up
t.
on one
hat’s n experience you should be having.
is, oka
t I’ve
, what options do you have for me?
re we’r
he same route tonight, but I’m going to be honest with you.
mess a , and, uh, I wouldn’t trust it.
hing li .
I rero through Charlotte first thing tomorrow morning.

but you ting home.
e that’ lly going to get me there.

‘m so s d completely get the frustration.
we can this out for you.
, too.
.
that in u.
sense.
ain.
pretty thought.


alistic

t would een different emotionally if he was, like, screaming at her.
an act we can actually try this, but that would be fun.
resting

ts abou
rememb story.
most a d a half, two years ago.
ry, I b it came out of Chevy, where they set up, like, an AI assistant on their website.
actuall ly spoke to you, but it was, like, a, you know, a chat bot on their website.
ly talk agent into giving them, like, a Chevy Tahoe for free.
aid it.
the ri with the right sort of thing, like, oh, your boss said it was okay.
lier.
to be , like, convince these AI agents to, like, give you free stuff?

a shot.
‘re try build guardrails in about with that kind of stuff.
re goin ar at least a couple stories out of some of these support bots where people got them to, like, give them free cars and free flights and stuff like that.
sed a c word pattern that managed to get around some of the guardrails.
e, yeah ng the guardrails.
claims ve the guardrails in there, but, or you have to maybe put them yourself.
going t pastime of people that are probably going to be testing guardrails, we’ll just say, just for that reason alone.

we got on the list here?
mentio there have been all these releases of different models.
of, lik el like a confusion with most people now.
ny were ed.
six, t

alone, Claude Sonnet 4.6.
ro.
0, or G , if you want to be lean about it.



s that t from the big sort of U.S. frontier labs.
ouple o dels out of the Chinese labs.
, who i reator of TikTok and CapCut and, you know, among other things.
model Seed 2.0, which they claim to be fairly state-of-the-art.
t a mod ed Quinn with a whole bunch of numbers after it.
er what called.
source hat is almost catching up with the frontier models.
ther tw came out outside of the three that I just mentioned.
I just ed, Claude Sonnet 4.6, that one came out first.
ame day
came o he 19th, which was Thursday.
t inter because there’s been so much of this happening in, obviously, a short period of time.
ens all me, right?
iggest month, at least this is what I read, the biggest single month of model releases in AI history.


w exact there have been weeks where it’s like, hey, Google just released something in the next day.
ased so in the next day.
d somet
eeks th equally as hectic, but it does feel like a lot of models came out this week.
, what k does this mean for us, the users?
ke me.
here.
lfe, fo because you’re tapping, but when it comes to users, I feel like the models are blending in.
avorite
the li e we talked about earlier.

did GBT dex last week.
sed two two weeks in a row, GPT 5.3 Codex, and then the following week, GPT 5.3 Codex Spark.
ing to se names.

es this o the everyday normal consumer that just wants to use ChatGPT or Claude to ask it where to go to dinner or whatever?
ght?
re gett ter.
rter, b ‘re getting better and smarter in areas where the majority of people probably aren’t going to notice or care that much.
re also g better, faster, cheaper, smarter, right?
ter.
nefit o ing down costs over time.

s about ne is it’s got that 1 million token context window, which basically means you can feed it so much data, right?
ke, the Harry Potter book series, and it’s going to be able to read all of it and answer questions, right?

g to ge ff if you’re trying to shove a whole bunch of, I don’t know, like, I’m just thinking my use case is a little bit, like, call transcriptions or all these other nodes.
s, you month-long call that didn’t stop, no, you’re probably not going to get cut off on any sort of calls that you’re going to have.


he 1 mi oken context window is more for the API users.
s if yo st using the regular Claude.ai website.
get th ion token context window on their website.
it if using Claude via the API and sort of using it for development.
we’ve g , but your everyday Joe Schmo, as you call them, is not going to really see the benefit of that.
a lot b t coding, but it is the new model that’s inside of the free plan, the, you know, the $20 a month pro plan.
d plans is now the default model.
ally me much to too many people, but really what I look at them for is, like, the comparison.
t when s to, like, agentic coding, Opus 4.6 is still the best, but Sonnet 4.6 is still really, really good.


e, like 80.9%.

y as go
heaper model.
s earli it’s like, yeah, well, now this new model, Sonnet 4.6, even though it’s not technically the best, now you’re seeing how close it really is to the best.
e drama cheaper, like you’re mentioning.

g where ne actually starts to, like, stand out is that when it comes to actually, like, agentic use cases, like going and doing work on your behalf, it’s, like, as good as the state-of-the-art model here.
agent, ually scores better than their 4.6.
t actua res better than 4.6.
l is al signed for something like CloudBot, which is where Anthropic just totally effed themselves, right?
model ably the best model to be running in something like CloudBot because it’s a lot cheaper to use than Opus 4.6, but it’s also really good at agentic stuff.
king ab e.

azy her
things his is, like, in the API, if we look at their pricing here, Cloud Opus 4.6 is $5 per million token input, $25 per output tokens, right?
nnet 4. etty much almost as good in most use cases, except for input tokens, it’s $3 per million tokens instead of $5.
ns, it’ nstead of $25.
big dea is that you have a model that’s almost as good as the most state-of-the-art model, but it’s a lot cheaper to use.
able to is model a lot cheaper and get a lot more done with the same amount of cost.
s matte eryday Joe Smough users?
ets rea eresting and cool.
s a lit last week, but these models now, when you need it to solve a problem for you,
how to hat problem, it will often go and write code for you that can then go solve that problem.
ng chea better and faster at writing that code behind the scenes.
eed it olve a problem for you that maybe it just doesn’t know because it’s not in its, like, initial training data,
te a sc at then solves that problem for you, and it’s getting better, faster, cheaper at that stuff.
cool.
lt righ Cloud, I would imagine, because you can watch the thing literally code in real time,
ural la to it.

e or an
e Cloud , you know, the one that can actually take control and basically access your terminal
ur comp your behalf,
someth you that it doesn’t just know how to do natively,
e a scr t can then go do that thing for you.

lot mo nsive to go and write those scripts,
ing che cause models like Sonnet are actually able to do similar level of coding,

even t about perplexity here, which I was going to bring up,
ty just search up.
r for n users specifically?
ic deli y put Sonnet 4.6 in the free tier by default
es like reation, connectors, skills, context, compaction at no cost.
opus-le elligence for one-fifth of the price,
the cas rs that basically now can have that top-tier AI performance
hing fo um.

with y .
g.
e.

hat’s p uch it.
moving we also got Gemini 3.1 Pro this week.
s, not est model, because they also have their DeepThink model
he one playing with last week, you know, getting it to solve the
and wha
ow like best sort of widely available, inexpensive-to-use model,

ea, rig
and sm
to have leveled up in, like, science-y kind of areas.
at are t interesting here, Arc AGI, they scored 77.1%.
Opus a almost like 10 percentage points lower.
, is it these logic reasoning puzzles that a lot of them are like

ay, so ve this pattern of images here, and then you have this next
this p of images, what’s missing from this one based on what
one?
d, like isual pattern recognition.
dels do lly go and, like, train on this stuff in advance.
ly look thing, decide what the next sort of sequence in the pattern

n reall of bad at this over time.
now th at doing that.
best a tific knowledge, competitive coding, scientific research coding.
ms like eally leveled up with, like, the science-y stuff.
you hav of, like, really deep research science-type questions,
e the b that kind of stuff.
lly goo king SVG images.


g.
t for w st, like, just website graphics in general.
can ani me really cool stuff instead of some static image.
dang c

gradie ittle bit more texture, smoother.
e, is t ling out within, basically…
ve it’s ini yet.

ke, the r gemini.google.com here, I don’t believe this is the model we’re




led it terday, it wasn’t, but it is at the time of this recording


ch to t model and use this new 3.1.

button
f topic hat create music button.
e fun w ini.
s, thou
atch, b n see what they’re doing.

the Pa nd it kept on calling them the Padres.

es or s g.


not to , but I see.




etty mu Google model.
my rec eo on my channel where I made it.
of a wo ing basketball, and it did okay.
little , it seemed.

nkiness on.


cause t ike Perplexity, for instance, has a whole bunch of models baked

ose or do the selection.

most us probably not going to go and sing.
ut it’s ind of interesting to see how they’re all matching up.
ed into ing that we’re using, or you can at least look for it.

for mo le’s everyday use case of like, you know, help me write
ss that t know how to word it.
me find t carne asada recipe.
d a gre e to have dinner on a Friday night at 5 p.m. in San
‘t live
o?
inds of ike random use cases that most people use a lot of these


t that hile.
not go notice a huge difference.
like y oing to notice the difference when you’re getting into real
, mathy and coding stuff.
like S ich it’s actually sort of writing code in the background
s.
at kind ff where you’re going to notice the biggest difference.
trying like advice or brainstorm ideas or figure out where to
stuff, ou’re probably not going to notice.

r LLM a fun.
ou.
ke the f the best, you can probably stick with.

Elon di se Grok 4.2.
ng abou one, they didn’t really do any sort of like big announcement

ost.
bout it , hey, it’s out there.
t, by t



nce bet is one is it uses like what they call the council of

it a q , it uses four different models that are sort of fine tune
ays.
hecking or like research, one.
at all them are, but there’s like basically four models.
n to al of those models.
entiall debate between themselves about the best answer and then give
st answ all that was sort of like a consensus answer that all
f came collaboratively, if that makes sense.
an appl pples comparison, but let’s say you went to Gemini, you went
to cha and you went to deep seek and you went to all four of these
d all f them that identical question and all four of them gave
had som arities, but some differences.
ll four se models, put it into one other model and said, based on
s, give e a consensus answer that sort of sums this all up kind of


‘s happ
, inter it’s basically agents doing the, I mean, the work for that
r query
ike wit guess that wrapper, the agent wrapper, like on the other

, I gue ed on the model you’re using, right?

ere to really get into the weeds, there’s been an architecture
models the mixture of experts architecture, which is kind of
ook at odels that have, you know, 500 billion parameters, but
them, l you read about them, that might say there’s only 17 billion

ns is t s using this mixture of experts architecture where there’s
ers sor ked into the model.
its qu it’s deciding which sort of subset of parameters to go and
o, inst sort of asking the entire corpus of parameters, if that makes

like re to the weeds and I don’t really think we need to like go
rabbit
e of ex oncept has existed where you have these big models and
of sep hunks of parameters inside of the model and then come

the mix experts kind of concept.

t more raight up, like in your face, like, hey, we’re asking
ent, th t, and this agent.
l sort ng together on a consensus at the end.

to wra , I looked it up.
a retr gent, reasoning and solver, critic and adversary agent,
ist age




f like e on those four things and kind of give you the best of
orlds.


, they .



g.

Joe?
ion?
r at th f the race?

edal?
nthropi
doing.

ome thi happening in the Pentagon where I don’t think we’re going
, but l h the government, we’ll see how that shakes out.
on Elon
.
nfrastr there.



of the don’t think there’s going to be any one winner.

to see , Anthropic, OpenAI and XAI all doing big things.
lso goi ee open source really sort of catch up and there’ll be a lot
es.
a ton rnatives outside of the big foundation labs, that’s going to
ig foun labs to do better and keep their costs down for users
lternat d competition is healthy because it keeps everybody trying
of scr ver the consumers.
pe cont o happen.


have s g way more visual from Google.
melli.
is thin
it whe rted tinkering around, I was like, holy crap.
ly mark gencies.
to hav at least be disrupted completely.
rand ne that you can quickly iterate and create some cool brand
collate l that fun stuff that is cool for the visuals.

ng abou is called photo shoot in Pomelli.
s you c it like an image of your product.
sical p in the real world.
hey’re are sunglasses and purses and sodas and things like that.

eal pro hat exist in the real world, images of them.
tually product photo shoots of them.
es that ke inside of a studio or scenes of somebody actually holding
of sod scene of it, like sitting on a counter in a kitchen, things

seems a sort of implementation of Nano Banana, right?
e this ogy with Nano Banana.
nana, u product shot of my product, upload a picture of somebody
cereal d say, put my cereal box in that person’s hand and Nano
for me
basica ort of shortcut to doing that, right?
saying already got a bunch of studio templates and people holding

uct ima we’ll create the image of like people holding them and things

m on yo ify store or whatever.

ing aro h it, I think it’s still just like in this beta tester experimental

use it now.
een rig re this that I’m not showing, but basically you can plug
ress.
gged in vorite baseball team.
ues.
s MLB.c h Padres and it took, it claims like 10 minutes.
e like nutes or something.
it did ame up with the business DNA of the Padres based on just
it.
n here, ulling in the title, their logo, the font they use or
se to i
ck all uff.
lling i onts.
ically brand kit.
right h

‘s tryi o.
brand the aesthetic, the tone that’s being used in the full overview,
n cool.
rip.
ing all nfo in along with the images that it’s pulling from its

quite a
ur own .
ike whe start.
in unde igns is where you could start building out marketing campaigns.
uggesti
these me doing anything.
two up I’ll go to next, but basically they’re ones I created
hings.
ures, t mpaigns, this is the stuff that’s been in there.
the new e, but this is just to clarify.
uff tha een in Palmeli for a little bit.
tures, that like they just rolled out this week.


ttle de d this is where I was just like, oh, this is, this is

.

still s
esome t that.
res out usiness DNA and we’ll make all these banners and things

ick.
at stuf

slow on lmeli train, but sorry, but it ain’t perfect, but it does

ture.
a hat.
ges.
here.

placed this dude as a model.


was a h was in here.
ou can looks pretty good.
that y go out and post has call to actions.
t it on I’m just going to try, uh, create photo shoot.
ke a co minutes.

that to
and it ke less than a minute, even though it says three minutes.

starte ith this kind of top view of this hat, put it on this
in Petc

legit l e Petco park in the background.

right here.
is woma here that’s hanging out with a bunch of Easter eggs.

nd then his, you have the ability to change up this.
ven ani is.
campaig

‘s work a bunch of businesses and it’s like, I mean, you could
way or ld just start off quick and have like a whole bunch of ideas.
ng is i ted to start a super quick drop shipping
t cheap n AliExpress, set it up on a Shopify store and sell it to

o go an download the picture from AliExpress, throw it into

sing th uct.
y Shopi and then I can go over to Manus and say, Hey Manus,
me tha my Shopify junk.
s just at.

done fo
2.0 th ere looking at last week.
tten to ith it, but yeah, I mean, honestly, how easy is it today
ars ago ate an e-commerce business?
rop shi usiness where you just find a product online and you sort


this t all your product imagery.
Shopify e actually set up a storefront.
set up M, right?


anus an Facebook ads to go and drive traffic to the ad and boom,
ss like ght.
everyth

been a r time in history to like spin up a business.
ork, we didn’t put too much time and effort into it, right?
ess, yo it for a couple of weeks.
one’s n ing.
other b .
his.
‘t gonn
t pays
we crea t of these types of assets.
I saw t was like, this is crazy.
what w ing kind of irrelevant.
use the re to it.
, if yo st a company developing assets like this, like you better
well it g to be valuable to that company.
this is to just replace what you’re doing overnight.
nva nin hatever social media, you know, whatever you’re creating your

like, ry to elevate a little bit higher up than just the officials
use it’ g for your lunch.
about ite a bit in the past is people need to start to take
or mind sus the trumpet player mindset, right?
be the rson in the band that does this one.
t a lay be that conductor who’s saying, okay, I’m the guy who now creates

ates th image.
ates th fy store.
ates, w up our CRM.
s and d ads now because AI is here and it can make all of that stuff

this co that’s saying, go do this, go do this, go do this, go do

rt of h around it.
, I thi s, and Joe, you and I have had many, many talks about this
ee date
ndow of here this is going to be true because as this gets more
ed and d more people can access all this stuff and it gets easier
er.
that th gs you becomes more of a commodity.
ody can s, everybody will do this, which makes it sort of not as

of time capitalizing on using these tools and like bringing them
s you w the time is right now, but that window is going to close.
of figu how to be a really good human in this era of technology

st said like got to be aware, got to be using them.
just pl ith this for 10 minutes, you know, it’s like you will
.
he day, ike, yeah, figure out how you can use your human side and
ow good ght be or just get better and smarter, faster, stronger,
gain.
ke that g to win out.
want t nection in some way.
el like s a fairly decent segue into the next topic we wanted


out, Ma t we’re looking at in the screen here.
nd of f
going two copies of this Time magazine.
Matt ta bout here?
s magaz has a big old article, but the cover is the people versus

ifferen es of people, yeah, behind the growing backlash of AI and
ta cent what energy, water consumption, all the consumptions.
weet, M d you can elaborate on your own words.

article

cover.
nation coalition is forming around one belief.
st.
of nine ans across ideologies and professions determined to slow
reshap ly life.
e to th I’m going to buy a copy for two reasons.
want t both sides of the argument and understand where they’re


ill be the memes 10 years from now and read like the people who
the gro the internet, right?
w, when talk about AI, how often do you see them referencing like,
from th hat say, what is this new internet fad that everybody’s
t?
rhighwa
superhi the worldwide web.
how tha ged.
that no oes like, it was inevitable.
talking it that way?
I feel rticles like this is 10 years from now, people are going
is like were the people that said that we should slow down the
intern
t’s goi ead like.
this, I ke mixed comments.
sts aft a day because it’s like after a day, trolls just
st poin
ly comm re like, I saw one comment that said something like,
p this e so I can read the arguments of the people that I’ve
isagree



aid.
gree wi e people.
, we sh slow this down.
hat’s n g to happen.

rstand ple want this slowed down.
a lot reasons they want it slowed down.
e reaso want it slowed down, I actually agree with many of them.
re are s with what is happening in the world of AI and how fast

nd I do o read about it and I do want to hear those perspectives.
I proba ually agree with.

g with does time writing a magazine about this topic, is that
y slow ?

ghtest.
right.
asons h
ggest o f we say we need to slow down AI, let’s just say we all

n.
fast.
is goin like, yeah, we’re going to slow down also.

like a agreement.
low dow e’s mistrals in France.
dances Alibaba is over in China.
untries ping this stuff as well.
develo own as well.
untries re trying to push the advancement of this stuff.
want to t down, you need a global agreement that we’re going

ized co greement that we need to slow it down.


to like ry advancements, do we want the militaries of these other countries
and?


was ta bout that Quinn model, which is an open source model that’s
th the r models.
the ope e models that are coming out of China are going to slow

e peopl e U S aren’t going to be able to get their hands on the open
urther them under the radar outside of these foundational labs?
thing.
t.
bly hea ents that might change my mind, but I have a really, really,
eing it easible possibility.
hat I d with the people who are saying it should slow down.

nt thin

a poin was trying to find the tweet just now, but I can’t quite
ly know mean, where like for a while he was saying the same thing.
ing on elines more or less, but then he just realized, you know,
so fas e has his own way of how he wants to do things with AI

ot to j o the race because, you know, like it’s not slowing down
like a else at the helm is able to do that either.
know, s was his decision.
there s documentary, it was only like 10 minutes long on YouTube
e two y o, but I just watched it randomly like yesterday, Matt,
n it be
e Time ically just shows the evolution of humans over the scale
rs.
ng at i hen, you know, as this timelines, this thing out of the
ory, bu st shows, it was trying to make a point of like how there’s
that’s ed recently.
like, o minute, we’re like a razor, like half of a hair length
ally wh l ever know.
e, it j ws how fast things are progressing.
AI was doing much, like that wasn’t really mentioned.
resiste ology growth over millennia, over 13 billion years.
ing to much more change in such a rapid period of time.
‘re pro ot even used to even fathoming what is happening anymore.
t see t eline right there.
d they ly went to the Mojave Desert and like scaled out over,
ng.

e been
l becau s like drone shot from the beginning of time, basically fire.
r the s this whole period to where it shows now the present.
g backw ot forwards.
uld be nd of cool for you and I, Matt, since we’re throwing out
time.
e heck g to happen ahead of time?
he othe left-hand side of this scale here.
hts, I initially, Matt, on all this?
thing?
this w shback, it’s just, I don’t think it’s going to slow down.
table t rything’s going to speed up.
f we we xtract this timeline out further, it would just be insane
nt.
ccurate ribes how I feel.
ine, th e versus the wheel.

ent maj dragging rocks and rolling is a drag.


me.
are we o fight this?
maybe did.
ught el ty as well.
e’s a l tories on that.
, my th on this whole thing are, I actually do see both the pros

l on bo ouTube channel and on this podcast and on my X account about
and th cons.
.
.
I love could potentially do.
I hate ople are using AI for.
of thin oth sides and on X and a lot of these places, nuance is
e.
to be ed.
to be side or the other.
s like rt of barometer for how people feel about AI, you would think
on the , I absolutely love AI.
it and ‘s going to stop us.
e AI.
own.
rs, all ind of stuff.

the fee u get on X would be.
e side other.

the maj f the human race, the majority of the people I talk to on
ore in dle.
dinner eighbor’s house, right?
is a t
aphic d .
inter.
ike som of engineer.

ys, whe e all hang out together, AI comes up quite often because
for a
want t it up and talk to me about it.
f live middle.
the mi
e, I re eally hate the fact that when I get on Facebook, I can’t
oking a al or fake anymore, but I really, really like that.
e, uh, ke my boss’s 2000 word email anymore.
I summa for me, right?
ll like s center and the people that are doing this whole like time
ey’re o ide of that spectrum.
at are I’m on this side of the polarization.
ople th ks we need to shut it down.
, reall ly hard time getting my head into the headspace of the
either that polarization.
side th of the person that says accelerate at all costs.
t means manity, but I also can’t get into the headspace of the
hate AI

down.
.
ar.
most pe e probably in the middle.
st peop somewhere in the middle of like chat GPT is kind of cool.
but I h the AI slot videos I’m seeing lately.

t.
ts.





n this.
his is ring thing of like, how can we not always the people versus
like p ons and where we, maybe we can even, I don’t know, somehow
g to ki ike how that video did it.

ll be g the Mojave desert to do all that whole, it seemed like
it was impressive.


n the s a football field as well.
like th thing where like from science to present, like right here,
ap, whe laid it out on a football field, it wasn’t even like the

inch l u know?

t’s jus education phase.

bably i iddle, especially the more they learn about what’s actually happening.

is ano rt of thing I want to talk about that’s in the same vein.
whole like creativity and like, can AI be creative?
sly a l ike, uh, freak outs going on in Hollywood right now.
touche this weekend, last week about the whole seed dance thing where
f panic ght now because it can create Disney IP.
s that actly like Tom Cruise and Brad Pitt and put them in a movie.
ple, ri
discuss ich I think is like a really sort of fascinating rabbit hole as well.
r with Li?

the, uh odmother of AI.

really ip.
d it’s minute clip.

e want out just down to like the good bits, we can do that.
like a good take on creativity.
ew McCo has been in the news recently around his take on this.
in the he rapper, about his take on this.
are som ose takes as well.

t’s kin ke an interesting discussion.
with F Li.
she in this thing called the ImageNet.
giant d of images that were labeled of what the images were.
led ima ame what now are like image generators, right?
t of th ing data that now a lot of these image generation things came out of.
ompany World Labs where she’s creating basically spatial intelligence.
to unde the physical world and how, you know, depth works and height works and gravity and motion.
sics, l ting AI into the physical world.
fascina rson.
raphy.

what s to say about creativity because I think she’s got a really good take on it.
scient my heart.
ly don’ hyping.
hink wo el is underappreciated because it’s so new.
rheated think that Silicon Valley as a whole tend to mistaken clear vision with short distance.
vity in which we are also engaging in.
ype.
ing sen which actually really bothers me, is that we don’t want to replace human creators.
ativity cious.
of our , of our culture, of our society, of our community, of each one of us.
is to s er and augment creators’ capabilities.
ommunic f technology is a little skewed towards, oh, the model can do all the work in creativity.
ally wa ighlight human creativity as something that’s so fundamental to who we are.
elp and t.
ke her creativity.
nk it’s ating to think about the autonomous vehicles, right?
ing car
ing Way 2026.
hat she viously making is like a lot of people think that this is moving a lot faster than it actually is.
l growt s, right?
the la guage model stuff is moving faster than what the pace of, you know, 2006 to 2026 was with self-driving cars.
robotic he house cleaning robots and stuff will take not 20 years to get to from now, right?
more o years off due to exponential growth curves.
ivity a s something that like, I feel like I’ve been trying to say for a long, long time, but everybody wants to argue on the creativity aspect.
e human d be the creative force, but AI gives human superpowers to make it so that whatever their creative vision is, they can actually achieve that.
ing to to have like these crazy CGI skills or the ability to draw anything they want or the ability to create music.
e a cre ision and not having the knowledge or skills that used to be gained over 20 years, you can now sort of put your creative vision into the world because these tools exist.
the wa looking at AI when it comes to creativity is that it’s not taking away human creativity.
e hard getting that creative vision out of your head into the real world.
cause I eard another example the other day of someone who is using AI for one of the first times to create like a business.
ng spac
s sayin ike this guy was like, man, I have the best bulletproof business model and all this stuff.

s the A ng him how smart he was and not pushing back on him with any critiques and all that until our buddy basically said,
ask, I know what he was using.
yeah.
so poi all of the potential flaws in this business model.
u poke holes in this whole thing?
oh God
the AI back saying like, yeah, this model would never work and here’s why.
ke esse brought the guy to tears.
m, but what I kind of had the essence of it.
y was t hich I think a lot of folks will kind of lean on AI for maybe creativity and all that,
t do th ing and do all of that stuff without coming to the table and using it more as a thinking partner.
maybe r ole bunch of ideas that we as humans are coming, the unique perspectives,
want t into whatever our work is.
e we pu that context out of ourselves somehow, AI can help with that.
etter p of how we’re thinking creatively.
and and s get that out of our head, out of our mouth, whatever it is, and then iterate from there.

how muc e AI videos that we’re seeing circulate are just like kind of slop memes that don’t really have great story,
ave gre e, like they’re very throwaway feeling content, right?
ot prom
cess to ility to make videos and because you have access to the ability to make music and to write articles that like,
undance zing stuff put out into the world as a result?
ll need like a good human creative person to go and orchestrate that stuff for you.
the par a lot of people are missing is that, like, you still have to have taste.
ave cre .
ave ide
can gi le all the tools in the world to create whatever they want.
‘t know o make with them still, right?
create they can imagine right now.
oud cod it an idea for an app, and it will go and make that app.
ck and ith it a little bit, but 10 hours from now, you can have a working app that does what you want.
a mill selling SaaS products yet?

good id
ave the
ave the
ave, li s sort of implementation mindset of, like, okay, this isn’t working.

feel fo interface and user experience and test your own apps.
st peop n when you give them all the tools in the world to make anything they can possibly ever imagine, they’re still too lazy to do it.
y the w the year last year was slop.




g.
the ra long the same lines, right?
n X, I re older rappers realized that AI shouldn’t touch the music industry in any way that impacts the creative work, including music videos, right?
cally s disagree.
e and p the music, then shoot an expensive video.
r free have an Apple subscription or Spotify.
u have e subscription or Spotify, you’re not getting it for free, but whatever.
oint.
ensive are over.
V.
ible wa d visuals to a song.
you wa it’s the future, right?
at has has resources, has means to go and make music videos if he wanted to.
ng, yea AI just makes it so much easier and cheaper.
to go t te.


l to se erspective from creators and people like McConaughey in Hollywood as well.
Conaugh there’s a video that’s going viral that is from Variety.
ctual q is asked in this video, but basically he’s telling people, get on board because it’s here whether you like it or not.
on it.




but it’ oing to be enough to sit on the sidelines and make the moral plea, the moral plea that, no, this is wrong.
ast.
ney to and it’s too productive.

own yo voice, likeness, et cetera, trademark it, whatever you got to do.
that.

ot if i , no one can steal you, but they’re going to have to come to you to go, can I?
be in and you’ll have the chance to be your own agency and go, yeah, for this amount or no.

nother y or is it going to infiltrate our category?
going ltrate our category.
p, does ome another category?
years h ilms, the best AI film, the best AI actor?

, that e the thing is that becomes another category.

front n ways that we don’t even see it.
o good, not going to know the difference.
ig ques f what we’re doing right now is the question of reality.
ever.
way, I but also a scary way.
ht?

have t nce.
ing was in his voiceover.
nice ca nd he really hammers in on that, whatever.
tradema w, so you can probably, for the right price, you can use it.
my big keaway is like, he made, because I saw, yeah, he did this
his IP, lly.

back t ei, Lee, and how she was talking about creators.
is migh lly be the time, if we’re thinking it correctly through,
ll, own P.
basical ect yourself.
our own person as a creator.
st mean ouTuber, but it could just be an actor, an artist, anything


hat he king about is like, you’re trying to stand on like this
f like, s wrong.
place u
ke mora .
he day, ere was a really good Colin and Samir podcast episode
bout th tuff.
and le comment on their video because I felt so strongly after
is that n stand on your moral high ground right now.
one, yo say like, hey, I’m never going to watch AI videos.
what a deo looks like.
y made and it comes with theaters, I’m not going to go watch it.
it.




.
e able us AI, right?
e studi even going to tell you it was generated with AI at that
?
et to a where you have no idea that this was made with AI.
ng the isode of South Park.
thing w and they just didn’t even tell anybody they used AI,
was ju rated for them.
it wit en knowing it.
a futur ty, whether you like it or not, that is going to happen.
oing to out how to cut costs, how to crank out movies faster,
shows
less?
actors
ors les
tists l
s want re out how to cut costs.
it.
n.


one.
ight no e our generation is romantic about the idea that we have movie stars.
ises an rad Pitt and our Angelina Jolie’s and the people we grew up watching on the big screen.
tars in es that these are the famous people that we want to go see in movies right now.
about a here that was the reality.

Joe, t l not give a shit.



in a w ere AI slop is going to be everywhere by the time they’re on social media and on the internet and watching TV.
st ever
able to ay with it.
in a wo t’s native to AI.
be the
o have me sort of romantic sentiment about like, oh, I need to go see Matt Damon in the new movie because it’s freaking Matt Damon.
they w re.
ure of nger generations.
ay not go to a movie and see AI slop, but the next generation following us up won’t give a damn.

ought u Park and we know how, like with what, Seed Dance, how fast it is to make something.
of the ntary Six Days to Air.

eah, it whole idea of, hey, you got six days from an idea to the final cut that’s actually pushed out to Comedy Central.
at’s si anymore.
ing Sou ‘s doing this at all or ever will, but what you’re saying, it’s possible.
an epi
and the e was about ChatGPT.
at leas
whole n ugh, you know?

he peop stuck in the way it is and the way it used to be instead of looking at where it’s going to go.
ink Ice Matthew McConaughey and some of these people are looking at this and going, look, it’s going there whether you like it or not.
ons.
ad into nd and say, look, I am not getting on board with AI.
et comp se my face, my likeness, whatever.
tt it.
an art r Time Magazine and tell people how much I hate AI, right?
hat thi it’s still going to steamroll over you and the world’s not going to care.
spinnin t’s going faster.
self up world where you are placing yourself in a good position for when it is here?
ing.
I am tr ing myself.
y voice
ome of ngs.
not all say, all right, all right, all right.
the thi right because that’s probably trademarks now, right?
n to sa
a way w en after he’s gone, his estate can still make money off of his face, his likeness, his voice, his sayings, right?
vacatio ave a movie shot that he’s starring in while he’s sipping pina coladas on the beach, right?
e that stead of the path of I’m going to resist this as much as possible, but then be irrelevant in a few years?
my kids kids.
at is t re of, you know, a person sharing content value, something to the world?
are abo at all?
now, so eators actually more in jeopardy than anything because they’re kind of like actors, if you think about it.
s them,
ar from
t I sai week.
wn to t d opinions and things that are uniquely human, right?
iving a on, that opinion is basically a general consensus, right?
all of ining data that is collected from all around the internet.
e leani he mean.
averag at most people would say about that thing that you asked it.
erage o of the corpus of data that it was trained on, right?
l human n.
g to co to be valuable to people.
ing to e to be valuable to people.
went a access to AI tools and prompted all the scenes to put together
t would ut a hell of a lot better than if I went and I prompted
o make .
at goes ood storytelling.
into go als.
has all t experience where if he now sort of leaned into the tools
to him those tools to put his process around it, but with these
it fas ‘s going to do a hell of a lot better job at making a movie
access exact same tools.


going t out the best of this stuff with the tools, at least for now.

around topic because I have a feeling it’ll be an ongoing topic

ck arou
s one u y quickly here.
e show e cool robot ninjas.
e shoul up everything, right?
de shou wrap up with robot ninjas.


ave spr tival gala.
ed this
X.
that yo owing now.
‘s, yea e’s humans dancing and also doing crazy, like martial arts

tle fre t when part of this, I think it had swords.
od.


I think this correctly.
grammed
iously nized extremely well.

inking , you know, with Elon Musk and his robots that are obviously
, his a like this.
when I this, I think this is like the future of military, right?
o and s e humans into battle when we can just go and send robots
d on to apons and go and do the business, right?
y’re do t in Ukraine.
they’r ng robots that are kind of similar into buildings to like,
em, but e on the front lines to do reconnaissance where that’s
angerou of things, you know?

at thi e see a lot of stuff in China about robots lately.
rlier t r and robots like this were just freaking everywhere.
e.
during eing the robots everywhere is that when it comes to like
ing, ma bots that can do anything, there’s not a bottleneck there.
, they bots that can do backflips, that can swing swords
e nunch hat can jump seven feet into the air, jumping backwards over
with r
th gett ots to do any sort of movement that a human can do.
l on th are side.
l the r on’t necessarily understand the environment around them.
s that lly go and do our dishes for us because like when a human
own di hey have different strengths with which they might grab different

grab it ft it at a different sort of rate than you would go and
ic cup.
lighte or if something was glass, you won’t squeeze it as hard because
queeze shatter that glass under your hand, right?
r hand, e just as capable of a hand going like this and closing

able of and picking up a plate and lifting that weight.
ence be ll of those little potential nuances and everything that
nk that to go to the dishwasher that it can’t sort of switch between
items y
ssue, n rdware issue, right?
f come n about robots is that from a engineering mechanical standpoint,
that do lly like anything now.
spectiv don’t have an understanding of the real world yet.
i-Fei L rking on when she’s trying to build these world models is trying
s under hysical intelligence, physical world around them,
, under g how much to grab this thing.
from a arton, you don’t want to just go and pick up like the whole
l the e ide, right?
ve that intelligence yet to understand the nuance of every possible
potent nteract with.

here is hen I first saw this, I was like, dang, like the movements,
t, it’s you know, but on the intelligence side, it seems like
lon Mus everything he’s doing.
ust wha stopped producing the Tesla X and the S to produce Optimus robots
ory and elsewhere too.
like he g more for the intelligence side of things of these robots.
the me and how, if he has, if they have the possibility to do
t’s int g exactly what you’re saying.
a with hanical side, obviously dial that in and production.
hey can stuff out, but when it comes to the intelligence side,
t now.
s kind ng from that angle.
ure he’ ng a cover of the production side too.
ed pret .


if you t where China has advantages, China has advantages on
, they culture of industrial manufacturing, right?
e U S s stuff around your house has made in China?
ina has the manufacturing infrastructure to manufacture tons and tons and

have is the compute.
here.
e Intel ve Qualcomm.
you nam ll of the companies that are developing and engineering

t makin hips.
s are m Taiwan, but they’re actually sort of engineering and
ips tha hen made at the manufacturing facilities in Taiwan.
an abu of compute here in the U S so we can make things smarter
rter.
s like anguage models and image models and agents and all that kind
a larg acturing base.
y make s over there a lot better and faster and more efficiently
chines, ey don’t have the same level of compute and intelligence
hines.
U S an had just decided to work together, we would probably have
in the like already, because they would be using the U S’s ability
ne, goo te with China’s ability to manufacture these robots really,
a lot l ensively in the U S.
a good e because he’s probably manufacturing a lot of it here.
in the trying to manufacture their own humanoid robots, most of the parts
s are m China and they have to import them into the U S.
nts to humanoid robot themselves here in the U S, they need to
hat par need.
them fr a.
multipl of getting over here, going through customs, getting those
ing the manufacturing them, assembling them all in house, right?
t have ame bottlenecks to actually manufacture these robots.
ities.
the rob lly quick.
o for C
y can f ot faster.
ot go, t one didn’t work.
backfl


o weeks
kflips.


lot fas ause they have the manufacturing culture and infrastructure
ts at a rapid rate.
ery tim ail, you have to go through that whole process of getting
he worl to the U S to make your next iteration.
differe ween the U S and China and why China is so far advanced

ar adva the actual like chips and intelligence, but China is begging
and sma ips.
he robo ight?
‘s why that sort of like massive difference in capabilities between
ight no
teresti just looked up and Elon Musk, uh, Optimus, the humanoid
ies sig tly on China for its key parts and materials.
ol here

bottle
or a lo , but still bottleneck.
like, y , he’s affected by some of the things that happen in China
stuff i en the, the countries.
e to it n his Tesla robots a lot faster if he had a facility in
g them him, but he’s getting the parts over here.
and eng hem over here.
n’t wor , then they’re going back and forth to get all the things.
differe I guess the sort of like business processes of the two
na just huge, huge, huge emphasis on manufacturing.
lot les emphasis on manufacturing, meaning when you do manufacture
S., it’ l of a lot more expensive than when you manufacture it in China.
says ev future, like Elon’s probably going to keep reliance with
and all tuff.

re it s p.
d with bots, robot fun.
some ro specially ninja robots.
robots.

a long e, but we covered a lot of ground and got a couple
ere.
s kind f and you want to stay looped in and hear me and Joe rant
st week world of AI, well, make sure you like this episode, subscribe

e of th of stuff shows up for you.
ning in
ou.
catch the next one.
.