Weekly Claw · episode 28
4 Sept 2026
GPT-6 Astra arrives, OpenClaw 2.0, Jensen Huang on AI and work, NYC’s classroom AI pause
What this episode covers
- GPT-6 Astra leads a busy week of model launches and benchmark comparisons.
- OpenClaw 2.0 brings a new interface and changes to everyday agent workflows.
- Andy and Henry discuss Fable, Qwen, Gemini, Cerebras and other releases from the week.
- Jensen Huang’s comments open a discussion about AI, work and how people use these tools.
- The closing discussion examines classroom AI policies and proposals to pause AI development.
Published record
1051 published segments
Speaker not identified
So first of all, welcome to the weekly claw episode twenty-eight.
the agreement landed, the models answered.
I'm Andy Lauppe, your host, Andy ML on Discord in
the OpenClaw community and others.
My co-host Henry Mascot. screen name is H I
And we are your hosts.
Him that him is actually
like I you don't know what him means?
So him is just an acronym for my name,
right? It's like Henry and then Ethan is like my middle name and then a mascot.
That's that's what him is. This is basically my name.
What is ML? It's just like machine learning?
It's the s no, it's the same as yours.
is after my middle name McLaren and L is my last name Lauppe.
it's laughter. I didn't yeah,
I didn't know your middle name,
so I th I think Okay. It's the same thing.
Okay. A ML. Anti money laundering,
Yeah. That's great.
Andy.
Fantastic. I love it.
Well, welcome back to everyone for sure.
We're glad to have you. we're thrilled that the audio is working
We appreciate it. Yeah, we're excited to be back.
Thank you. Nvidia made Hugging Faces deal official.
Anthropic ship some safeguard tiers.
OpenClaw came back from seven quiet weeks.
And New York decided the classroom was not ready.
we've got some slides for you.
We've got a video to review that is not represented in the slide because we're cool
like that.
Henry, you know, what what are you thinking?
You said it was a big week.
Yeah, yeah, yeah, hundred percent.
No, it was a big week, definitely.
but you guys, I mean, I I kinda like started talking about this last week.
I'm a bit like I'm already I mean it's a big week,
right? lost I think it was like four or five big model launches.
Astra is supposed to be the mod biggest model launched to date.
Before that, that was Fable.
So yeah, so I think it's it's a super cool week.
It's exciting. Yeah.
Yeah, agreed. Agreed. This episode is brought to you by Herald Labs,
an applied AI product lab where humans and agents build together.
Entity is mission control for agent teams,
and the lab is building the practical layer behind the hype.
Build with humans, ship with agents,
find it at labs.theherald.co.
What happened this week, mate?
Right. there you are. Okay,
cool guys. obviously one of the things we tried to do is cut down a little
bit of so I I put a lot more things in the agenda in terms of what happened this
week. but so like the trade-off is I spend a lot less time sort
of like analysing and and s like describing what what went on.
but I cover a little bit more ground and and this was feedback from again
one of my friends that, you know,
listens to the show on YouTube post.
And it's like a dope engineer and he's like he he just prefers to have
a breadth of what happened in the week,
versus just going deep on one thing,
right? He wants to grok all the technical stuff that happened.
So yeah. just in case you you didn't get the memo last week when
I flipped to this this this method.
but what we're doing is obviously this you can see the URL up there,
it's weeklyclaw.ai the the slides are probably up already on the website,
so you can always sort of
Open the slides, click through,
and then you can see the full story yourself.
All right. I think the biggest thing that happened this week is
NVIDIA bought Hugging Face. you know,
Hugging Face is like a 12, 10 to 12 year old company.
so they've had their run. so I think the CEO's sort like got a bit tired,
or the team is starting to get get a bit tired of kind of like running
the business and they were looking for a home.
and in the summer.
Clem reached out to Jensen to say,
Hey, can can you c can we can we
do something together or and then Jensen,
yeah, Jensen thought it was super cool.
I think the initial price was gonna be like seven or nine billion.
that didn't quite work and then ended up being like twelve point something billion.
there's nothing else on my stage anymore.
it's it's gone away.
Yeah. but yeah, but that that's Hogging Face.
Hugging Face. just a interesting commentary.
So Hogg Hugging Face, Clem and Jensen really right now are kinda like I think
two of the really sort of like best AI CEOs really in the world on the in the West.
I mean, kinda like in terms of what I mean is AI has a really bad rep in the West,
right? It's like in America, like everybody kinda like I mean
the Americans just Bernie just
Pastor proposed the bill, yeah.
And then we're gonna talk about it little bit,
to ban AI. So it's super unpopular in in a lot of circles.
and the CEOs aren't aren't very charismatic.
So Hugging Face is a company that is super interesting.
They bought a company early a few months ago called Pollen Robotics.
And these guys make like really cool,
cute robots, right? So they make robots that make you really love robots,
right? And and it's not threatening.
You don't think you're gonna be terminated or just like really cool.
Robots. the first one was Reachy.
Reachy Mini. I almost bought that.
so Reachy Mini is super cute,
can talk to you. and then last week they put out Docky or something.
Dock something, right? That's the one that looks like a dock that works.
Did you see that, Andy? Yeah. yeah.
I saw videos of it jumping around.
I wanna say Marquez did a video on it,
right?
Yeah, I didn't see the video, but I bought one.
I paid for one. But you pre-order and like it's six months.
It's like it comes out in December.
So but anyways, the reason I'm mentioning them is like because that company
is investing in doing cool stuff that makes AI look non-threatening,
super cool, fun. and then the on the on the other hand,
Jensen is like super optimistic,
it's like out there just talking about the great things happening about AI.
I believe you're gonna review one of his his video on the G20.
So yeah. I mean, each
Yeah, yeah.
time I watch Jensen's video when he talks about AI,
and I'm super technical, I've run an AI company for almost a decade,
I'm always like, wow.
You know, which is why they call him the professor,
right? He's like such incredible insights.
Okay, cool. so that happened.
I mean obviously maybe I'll let Andy.
I don't obviously I'm not a I'm not a fan of Anthropic at all.
It's like gotten worse over the last few few months.
So I didn't use Fable at all. And I actually don't have an anthropic sob,
but Andy is is an anthropic stand.
So maybe tell us a little bit about Fable,
right? Like how's that going?
I
I I'm I'm a reluctant fan. I've been happy with a lot of the outputs
and I built a production SAS using Claude Code.
and so I have some loyalty to it just so that I don't break the
SAS as I'm continuing my development.
But I have I'll I'll tell you I did make the first commit using Codex this week
and it was totally seamless. I will say,
yeah, I mean Fable is smart.
And it's it's insightful and it it's able to and I and I I assume we'll
get the same from Astra, but it's able to read between the lines
a little bit more like a person would do.
And so you end up with yeah,
I do a lot of analysis of call recordings.
I do a lot of like in my sales process for my business,
I'll take you know, a transcript of a call with a
like a cold calling agent and the the prospect and then the notes from
the agent and then you know,
research and I'll have it I'll have it build a a a prep deck for me
and then I'll sit on a call with the with the client and then I'll take
the the transcript from that meeting and add it to the conversation
and and the insight that it draws is often kind of beyond my expectations.
It's it's less and less
I'm I'm less and less impressed by aggregating data from s stuff that
I just gave it, right? what I want is like,
you know, I'm I'm dealing with funeral homes and I'll get
I'll get insight like you know the this funeral director
just purchased this funeral home four years ago.
They're probably at the end of,
you know, three year contract cycles.
And based on this comment, you know,
they're
you know, inclined to shop even though they say they don't have a budget.
Like it's it it goes a little bit further and I I appreciate it.
5.1 is faster and it's at least a smarter,
smarter. So and I'm not telling any of you anything you don't know about Fable,
but I'm expecting the same from from Astra.
Yeah. Please do do let us know as well in the comments,
yeah, if you've if you if you're currently using fable.
I mean I was I was gonna open the benchmarks,
but I think we can open the benchmarks when we get to Astra.
Because I mean he he basically kinda like killed some of the benchmarks over Fable
Five, but then Astra just like massacres it everywhere,
right?
Yeah, yeah, yeah, yeah. It it's interesting
so
that that they've solved some of the cost issues,
right? The cache reads are cheaper in five point one.
Yeah. Yeah. Like it's seven seventy five percent.
the workloads twenty five percent cheaper.
Yeah.
Yeah, seventy five percent, like the
cost is seventy five percent cheaper.
Yeah, I saw that. Yeah.
Ye yeah,
yeah. And and and they're even saying highly agentic work could be
But but but but is is that is that
up to forty five percent cheaper.
them is that them solving a problem or is that just anthro anthropic just like
batting the dust and just like cutting their margins,
right? Because you think it got yeah,
It's a reasonable question. I I I don't know.
I don't know what I don't know which one it is.
Yeah. Anyways.
There's a lot of room
to solve problems in their harnesses.
I mean, Claude Code, as we know,
is like two tokens for every one.
Blue dead. Correct.
the results are are good. They're not better,
but they're good. but why does it use twice as many tokens?
Well, maybe the models were too verbose,
who knows?
Yeah. I think that's definitely what people complain about.
I mean OpenClaw, I mean obviously this week yeah,
OpenClaw got the biggest revamp.
I got a lot of love on X. you know,
I got I think I I wrote one tweet 'cause I logged into the control UI.
I don't know if you started using it,
but I think that's like the most exciting.
I still have three calls on on on on OpenClaw.
and so Scotty, which is one of my engineering agents.
I loaded up load up the control UI.
I could select any theme I wanted.
I mean he had themes before, but this is like a lot more cool.
it's definitely super mirror. I reckon everybody who's right here
in the community has updated, right?
so well, I'm not telling you anything you don't know.
but yeah, but that was my that was exciting for me.
That's the control UI is in this.
I've got it pinned now. and you know,
you could I could do multiple sites,
I could break multiple sessions.
And and yeah, so I think that's super cool and you know,
I could have GLM on here, I can have Seoul on here,
I can have Kimi, I can have like different sessions and different models
in front of me. And so I thought that was like definitely very cool.
I think a tweet I wrote about it also it's like got a lot of love,
maybe like thirty, forty K views.
so that was definitely cool.
yeah. Thoughts on OpenClaw and then we can run.
You know, I have not updated yet because I've I've got too many
now production workflows running in the
Yeah, yeah.
that first LTS version. I'm looking forward to testing it.
I'm I have a lot of faith in the team.
I I like what Peter's done taking the methodology of the team
and and you know building OpenClaw with OpenClaw,
doing their multiplayer mode and I'm excited
Yeah. Yeah.
to see what it can do. I really do think he's
His innovations are in the building process and less in the product anymore.
And they're personally I think they're brilliant and I'm excited to
see what comes from them. but yeah,
I I don't want to break anything today.
Yeah, I mean obviously I can see kind of like Jasing is on there.
And I mean I mean one of the feedback I was kinda giving to Hans a couple
of days ago was I I mean I am super technical,
right? I'm probably one of the most biggest power users of agents that I know.
but I did struggle a little bit to like understanding all the things happening
on the new UI. and I do think that and and and I don't know how what
the guys in the urgency I mean this is definitely feedback.
I don't think OpenClaw on Hermes,
which is again what I use, which is what like the book book of the Power users use.
I don't think it will cross the chasm anymore.
so my point of view is that the likes of Instinct,
I sent you an invite for that last week.
I sent it a few friends, Gronkbots,
I've got I've got that as well.
I think those experiences make you understand why OpenClaw on Hermes couldn't have
crossed the chasm. Like my friends,
because I remember in January when I started using OpenClaw,
I sent it to all my friends.
technical, non technical, and like very few of them got started,
right? Because you have to download the thing,
you have to install it. And like if you're just like a random business
guy or you're not like you're not technical.
You're not gonna do it. so yeah,
so setting up the grog bots and like what the experience looks like made
me realize that hey, listen, there's no way.
There was just no way like regular people could have used this.
Like it's just too complex. so a friend reach
And it and it was really never meant for them.
No, it wasn't it wasn't.
Yeah. So so yeah, so I mean but I think that there's a lot of effort right now,
even on the Hermes side and the OpenClaw side,
I I think the the the the beauters wanted to cross the chasm,
right? I I think that they wanted to they want more people to use it.
and so my feedback was if you want more people to use it,
the less you make them have to deal with.
Like the it has to be super if you my because I got Grogbot early access,
I I logged onto it.
It was too simple. Like there's literally only five pages.
There's only five views. So I didn't take it too seriously because
it was too simple. but you know,
after it came out of the mainstream and everybody lost it,
I realized that shit, that's actually how consumer products should look.
It should be like super simple,
nothing to configure, you don't have to do anything,
it just works. and so yeah, so the feedback I was kind of like given with
a new control, the new kind of like open core URLs.
I guess there's lot of things happening.
Maybe we should have two modes.
There's a mode for normal people where they don't see anything.
It's just like super simple, and then you can toggle to like an advanced mode
and then you can see everything.
Like it's like, you know, a chat GPT type view.
you know, but that's if the plan is for it to keep crossing the chasm.
But if the plan is for it to be super users like us,
then then that's fine. this feedback is also the feedback I've also given
to the Harmings team a few times or in Icon X.
because Hermes, the desktop app app is just like
Too bloated. Like I can't use it.
Like it's too complex. Okay,
cool. moving on.
It
is it is it worth mentioning that this this version touched literally every part
of the product and had sixteen hundred PRs?
Like, I mean it's just worth mentioning,
Yeah, you should mention that.
I think, right? It's
Yeah.
a hundred and six releases in two hundred and thirty days and more than sixteen
thousand PRs. it there there's a whole video
I
think it was something like 50%.
Yeah.
It's like 40% of all the commits of the lifetime of the project,
right? Something like that. Some ridiculous number like that.
Yeah. Yeah. Yep.
It's it's basically
like a rebirth of the whole thing,
isn't it? Yeah. Anyways.
Yeah. Yep. It's exciting.
Exciting. so yeah, so moving on to I mean this isn't a big deal.
Qwen 3.8 Max got a checkpoint.
so it's not a it's not a big deal,
guys. I'm gonna move on from this one because we've had way more time.
But but checkpoint is like it to just launch a new version.
Something to note here, and I tweeted about this when this came out,
is Qwen is owned by Alibaba, right?
Libaba is one of the biggest companies in China,
right? So one of the biggest like the equivalent of Google or Microsoft in China.
I th and they have even though they sell this model,
I think that they get a lot more commercial value in their
own ecosystem by launching these models,
right? So what I mean is OpenAI would launch a new model because they're gonna sell
it. They're gonna sell it to me and you.
But Alibaba has like already like hundreds of millions,
if not billions, of users. That just launching any good enough model
in their ecosystem is massive value.
They don't have to sell API. And so I'm seeing,
and so I thought this was a really interesting thing happening that this company
is always launching a new model every week.
They launch a new model every week,
literally. not the ones we know.
They launch a new video model,
a new harness, a new world model,
a new audio model. They launch something every week,
right?
it's only been two weeks this since they launched like 3.827B,
right? And then this is so so and so I was advising,
and I was kind of like saying on X like,
hey, Google and like Microsoft and all the big companies should actually start doing
this. The only ones that need to wait for months to launch a new model
is like Anthropic and Open Air,
because that's their only business.
But if you own a big ecosystem,
then you should launch quickly because then if I use GBrain,
if sorry, if I use Gmail, then I get value immediately.
If I use
YouTube, then I get value, right?
Because your model is integrated like across your products.
and funny that I said it, Google launched 3.8 this week,
right? Like let's move to the next slide.
and so Google since they
had the Yeah.
So
I'm trying.
Google since they have the yeah,
yeah, you need to move to the next thing.
Yeah, but Google since they have the new kind of I mean two weeks ago,
a bunch of did I not have that here?
but yeah, but Google two weeks,
Google two weeks ago or three weeks ago,
all the big kind of like Jeff Dean,
who's the chief scientist, left the core team.
Demis, who's the CEO of DeepMind,
became the chairman. So they basically so they basically had like a shake
up on the Google log. and and I think that
That has probably led to where we are.
So I think in the last three weeks,
Google has launched three models.
They've gone from three point six to three point seven to three point eight.
Right. And so I I think I think there's been a change in ideology within
the within the business in terms of like how their AI team should operate.
And I think they've chosen to do iterative like launches.
They're no longer gonna wait like five months and then launch something big.
They're gonna launch if they get a new checkpoint and it's better,
they're gonna launch it. They get a new checkpoint.
Which is very similar to what Alibaba does.
And so this is how I think. so I think in feature we're gonna get like
the labs are gonna divide into two.
The ones that are already big c part of big companies,
Alibaba, Google, Microsoft, they're gonna launch newer models very quickly.
And then the ones part of like startups,
open AI sm are gonna kinda like take their time and launch.
Cool. let's see,
let's play the videos, right? so there was there is
the video model, all all us and then there's a cerebral thing.
Let's let's play those videos super quickly.
You type, you get the clip, it keeps going.
Step in anytime, the world keeps running.
It can be many kinds of real.
See, he started firing forward.
You're not watching a texture.
And then you wanna watch the second one?
Do you guys see the video
on Yeah? Cool.
Play the second one. But yeah,
but these are I guess like I should have put the slides of Ash so
so so there there are a few world models that launched this week.
okay.
So this is this is yeah. Go ahead.
This
exp this experience, by the way,
in real life is mind blowing. Fifteen hundred tokens per second of like Opus four
point six level vision enabled inference is amazing.
my developer and I were talking about it and I said,
Let's let's just try something.
And he said, Why don't you have it do a a Reno Air Races video game?
And so I it took me longer to type the prompt and then this thing just started
ripping. I mean, we were hitting
The 429, my I used the DeepSeek harness to ride it.
Honest. Yeah.
And
yeah, and the DeepSeek harness is really willing to keep going.
Like it just is like a dog with a bone.
And we would hit the 429 errors from Cerebras
more often than than issues.
I mean, it was just ripping. So we're at maybe 400,000 tokens of produced material.
I haven't played it yet because we were prepared.
prepping for the podcast, but it it went really well.
Shit, I need to so I'm looking at my phone because yeah.
I mean the I mean the models I was just gonna say,
I mean the
the model's legit, right? The benchmarks say it's legit.
I we ran it on my thirty ninety.
It was fast, it was smart. I was just like,
This is a s major unlock. And then Ornith one point five came out like
two days later at twice the tokens per second on the same hardware and
I was using that, but
Yeah.
I Cerebras puts Qwen real way ahead and it's cheap.
I think you you you you fell in love with with on it.
so I'm actually trying to see the cause I benchmark it,
right? So I have I have obviously my own internal benchmark that that
is made out made up of one sixty one sixty tasks,
right? Across like engineering and and business and like across everything I do,
right? and and so every single model that comes out I benchmark it.
And so I'm actually trying to see it and it takes hours for this benchmark to work.
Like maybe it takes like two hours or so for like a model to finish,
for every level, right? but I'm looking at this just before the call,
right? I just for the the show,
I I kinda like s started the benchmark for it.
And it's already done for non low,
medium, high. I think it's finally it's at the end of high at the moment.
So it's done all the if effort levels,
all the thinking levels.
Yeah, I'm trying to see what the result is,
but that's that's a bit slow. So it's already done.
It's so fast that it's what used to take hours.
It's like done in minutes. so yeah,
so if you're not using that, definitely check it out.
I put my card in in yeah.
There's a f the there's there's a free
tier and it's it's, you know, plenty useful.
Yeah, it is. okay. And let's look at the other slide before we do those two.
These two are kinda like pretty big big news,
right? And we can end with this we can end with two of them.
There's one more slide with with a bunch of stuff.
Where where are you thinking?
Spark.
Yeah, it's switched now. All right,
cool stuff. so yeah, let me talk about these.
obviously we talked about three point eight.
So Google launched three point eight.
I saw three point eight in the wild at two thousand eight hundred tokens per second.
Did you see that? It's ridiculous.
there, but but apparently this is like a ridiculously fast model.
so the the thing that the three point eight has in common with the the
H3 Max on there it is. I mean,
obviously the H3 model is is a mini max video model.
and then this video platform file took it and fine-tuned it and it became a max.
Mm.
And then they fine-tuned it again.
They did like three versions. So the new one is H3 Max Turbo.
So there's a max turbo. because this is using the wrong image,
but there is a max turbo because max came out last week or two weeks ago.
So what came out this week is a max turbo.
and the model
Is a video model that is like so fast that it takes you 15 seconds
to type the prompt. It takes the video model 10 seconds to generate the video,
right? It's like it generates 15 seconds what a video in like 10 seconds.
so it's it's an incredibly fast model.
I don't need it, right? Because I have like Gemini,
I have a ton of Gemini credits,
and Gemini has also has a super fast model called Omni.
but I'm considering like swiping my card.
for like file to just have access to it.
it's incredible.
Okay.
obviously Muse Spark launched.
you know Muse Spark, I guess the the cool thing about it is it's like
it feels like Zuckerberg is vindicated.
because a few months ago he had like started logging collecting data from
all all the employees' laptops to be able to train his model.
So yeah fast forward like two months or three months later
it looks like that worked out and and they have a frontier model.
this model benches very similar to Fable and like a few so like benchmarks,
but I haven't used it. I haven't tried it yet,
so I can't I can't say how good it performs.
and then I guess the final news which was in the kind of like
a previous slide, is so like two big things.
one of them we're gonna s like discuss a little bit,
which is like New York.
banned
AI in in in in Slack from Slack K two to K eight,
right? free school,
Yeah, preschool. Preschool to middle school.
yes. Yeah.
Yeah, end to middle school.
Yeah. So we're gonna talk about it,
right? I mean I mean Andy agrees.
I don't agree. So that's definitely something to kind of litigate a little bit.
And then the biggest news obviously of the week is Astra.
Andy can you click on the can you click on the thing so we can actually look
at the benchmarks a little bit,
yeah.
I I've actually
my gosh.
linked ill il link the No this is this is actually not it.
Can you scroll down to the end?
I click on on the footnote on six.
I think this is just this was the article before the actual launch.
yeah, but Astra, right? I mean,
Astra broke so Astra, you know Sam Altman and I think we talked about this last
week. So Sam Otman went on Times and and basically said that he thinks before
the end of the year that they would have Astra.
and then the model that has been kinda like hacking the whole internet.
I can see Zach no, not even Zach,
who is it? Someone mentioned the German thing today,
'cause there's a hack as well.
They found like that. Basically the model that hack talking face is from this
s like pre-trained, right? So they train the whole
yeah. So that they they the the this pre-train is like a new kind of like class
of models that were trained on on that was trained on hundred thousand GPUs.
right. I think they're it's like I don't even know how I think this
is hundred thousand GPUs.
This either may be a one gigawatt watt cluster or like a two gigawatt cluster,
but it's like a it's a big model.
It's a really, really big model.
I can't remember what the parameter count is,
but I remember I think I've heard somewhere,
saw somewhere that it was probably like a 10 trillion parameter model,
something like that. but it's a really big model.
so yeah, so Brockman says it's AGI.
Sam says the next one is gonna be kind of like AGI.
open AI is so confident that they're gonna milk this model and they're gonna kinda
keep push training it and get a a few more things out of it for a while that they've
because they they they said it's that they're scared,
right, because of what the model has done.
So they shut down pre-training.
So open AI at this moment is no longer pre-training new models.
but they are still pushtraining this one and guardrailing it.
And so from this we're gonna get an Astra five point six point five or like,
you know. it's a big model. This is like the dumbest
Checkpoints. I think internal lead you already have like really strong
checkpoints. yeah, but on the benchmarks,
like Andy, I I can you did you can you scroll to the table?
There's the place that has the table with Fable and and and and
all the coast things. but yeah,
but if you look at the metrics,
right? There like there are only two metrics,
or two or three places that Fable 5.1 is better than in F.
Two of them are from artificial analysis.
And everybody agrees that artificial analysis,
like their benchmarks are dumb at this point,
right? 'cause I think even like Muse Spark 1.3 scores higher than like Astra.
Like, who does that, right? so it's like a really kind of like
old benchmark that no longer applies.
So everywhere, basically every single place it kind of like massacres table.
And it's like so efficient, like you know,
it's like, yeah, this is the table I'm looking for.
So for every single, you know,
for professional, you know, automation bench,
it's you know, it's like forty-one percent.
feeble five point one is thirty-one.
right? If you go to coding,
time out bench four, you know,
fifty-seven. Yeah, no, this is yeah,
so this is the one that's an analysis.
So all their in all their thing,
it performs badly, but like, so yeah,
deep SC SWE seventy four.
Right, that's like, you know, miles ahead.
and then the one when they launched,
the one where it's like the first thing that that five point
one actually launched was that the first thing on their benchmark
is Tamil Bench Science. you know,
it's a huge jump from February five,
mm, if you compare February five point one to February five,
but then now compare that with like Astra Astra is like on a different level.
I think RKGI was solved. So,
you know, I think I that's the first one on the on the on the on the
the stable list like saturated AKGI3 is done.
so yeah so it's a great model.
I I've watched a few videos.
I think every single influencer that I've kind of like watched their
v video talking about it has like been incredibly impressed.
they think it is better. I mean there are one or two people who still prefer
so like Fable for setting things.
Fable still wins for like everything kind of like front end visual.
but everybody still like agrees that this is like a
much, much better model. Yeah.
All right, Andy. I think with that I'll hand over to you.
This took way longer than I wanted it to.
No no, it's great. I don't have a slide for this one today,
but we typically will do a signal from the outside.
I like to review these YouTube videos,
interviews, podcasts, and this one was really interesting
to me because the the perspective is different,
right? We're talking to the G twenty and so we're not talking to a room full
of technicians.
we're talking to politicians and yeah.
So I want to start with the moment that made me stop the video and scrub back.
Jensen Huang is on stage in Chapel Hill with Howard Lutnick,
the commerce secretary, in front of trade ministers from most of the G twenty.
About twenty minutes in, Lutnick asks him a big picture question.
Where are we going? And Huang says,
We'll hit what people call AGI in the next couple of years,
and then quote.
I would argue that we're practically there today.
And that line got headlines,
but it's not really the interesting part.
The interesting part is what he does 90 seconds later,
which is take the claim and deflate it on purpose in front of
an audience of policymakers. He says,
even when we get to AGI, it's not as if every company's problems are going
to be solved. He walks it out.
You don't connect.
your company to an AI service and then sit back and watch productivity happen.
He says that's just not going to happen at all.
So this is the guy who sells the compute.
He just told the room of ministers that the compute is not the bottleneck.
and that's what this video is about.
We'll back up maybe 12 minutes.
there's a stretch in here where Huang gives a definition I hadn't heard stated
this cleanly about anybody at his altitude.
Lutnick asks him about AI as software versus the physical stuff,
factories, robots. And Jensen answers that AI started with large language models
and then says what made AI useful was putting an exoskeleton around the LLM.
And he calls it, of course, the agent harness.
And he lists what it is: retrieval,
working memory, tool use, the ability to collaborate.
And then he does something clever.
He takes the same agent and puts it in a body.
Well, now it's a robot.
Put it in four wheels. Now it's a self-driving car.
Put it in a manipulator. Now it's a pick and place arm.
Surgical robot. Autonomous drug discovery lab.
Same idea. His phrase is LOMs with an agentic system around it.
And the agentic system doesn't really care whether the tools it's holding
are digital or physical. I like this because it's a hardware CEO saying
the model is kind of the commodity layer.
The harness is the product.
Which brings us back to the AGI section.
Because he reaches for the same word again,
and I don't think it's an accident.
Jensen runs a thought experiment.
Everyone at Nvidia hires out of school is brilliant.
He says these are people from Stanford,
Harvard, MIT. And even so,
we still have to invest quite a bit in quote,
surrounding them with context.
He then closes the loop.
All of that surrounding the context,
the purpose, the relevance, the access,
that's the harness around the AI.
We we all know this, but you know,
imagine telling this to policymakers,
opening the the door to them, right?
So twice, 10 minutes apart, he mentions harnesses,
once about engineering, once about management,
and the claim that they're the same activity.
Even at AGI, you spend your energy the way you currently spend it onboarding
a PhD hire. That matches my experience uncomfortably well.
The model is rarely the thing that's failing anymore.
It's that the agent doesn't know what our repo does or why somebody made a call,
made a decision in the the development of it in 2024,
or which of three plausible answers is the one that won't get an engineer paged
at two in the morning. He follows it with.
A split that's four seconds long and the best compressed thing in the video.
The tasks are going to be automated,
but our jobs, the purpose, remain.
He says a job is defined by purpose,
context, and meaning. And inside that job,
there's a lot of typing and talking.
Well, the typing and talking is what's going to go away,
but calls the idea that all jobs getting eliminated as nonsense.
He says it's a blunt word from a man.
being otherwise very diplomatic.
He literally calls shenanigans on the Doomers,
which is kind of fun. Now, he's the single most interested party in the room
on the question. So we're we'll hold on to that.
And before I get to it, the practical bit,
it's the most concrete 30 seconds on the tape.
Jensen says Nvidia uses AI
across the whole company and he names names,
off-the-shelf models from Anthropic and OpenAI,
plus cursor, and then a lot of custom work off their own on the top.
His advice to the ministers is to use off the shelf wherever you can.
And then he says you can't outsource all of your intelligence to somebody else.
He's saying it to countries. It reads the same to an engineering organization
Buy the model, build the harness,
because the harness is where your context lives,
and your context is the part that nobody else can sell you.
And then he makes the architecture argument and hold your wallet,
because of course it's a sales pitch.
His claim is that AI infrastructure has gotten so expensive that specialization
is now a risk. He puts one gigawatt at fifty to sixty billion dollars.
And says NVIDIA is building toward 100 gigawatts by the end of the decade.
Models and algorithms change way too fast,
so specialized silicon can get obsoleted mid-build by the fungible thing that runs
everything. The tension underneath is real.
Algorithm churn does outrun hardware specialization cycles,
but notice who benefits from the framing.
And notice that Jensen has quoted a considerably higher per gigawatt number more
recently than this one.
So treat 50 to 60 as the floor,
he's comfortable saying out loud to customers.
Which brings me back to what I flagged.
The middle of this interview is Jensen making a policy argument.
And I want to lay it straight before I say anything about it.
He tells the ministers that the worst outcome in this industrial revolution
is that you don't take advantage of it and get left behind.
That, he says, is quote the single worst outcome.
He argues for regulating actual and practical harm.
Instead of hypothetical and theoretical harm.
And he claims that advancing the technology is what made it safer,
pointing at hallucination and grounding as problems that got solved
by moving faster rather than slower.
And he tells an airline story.
There was a period when carriers spent their advertising dollars on being safer than
the other carriers. And it turned out passengers didn't want to hear about it.
They wanted safety to be handled.
What they wanted to hear about is where they could go with it.
It's a good analogy. as the AI says,
it does a lot of work. the reason I lay it out rather than co-signing
it is that this argument is being made by the vendor to the customers,
and the customers here are governments deciding how much to buy.
You think the technical claim about grounding is basically right,
but I I mean I mostly do.
You still notice that the conclusion lands on build faster and buy more.
So take the harness material.
that's the part that survives contact with your Tuesday afternoon.
And and that's what I have to say about that video.
But I'm glad I'm glad I'm glad that Jensen was invited to the G two
to provide this context to policymakers who literally have no idea what
any of this stuff does. So
We such an incredible video. It was it was an incredible I learned
so much by watching that video,
I was like it's it's incredible.
It's an incredible
Well
video.
he's he's an incredible teacher.
He's got he communicates very clearly and his obviously his knowledge
on the topic is incredibly deep.
He's not you know, he's not a CEO pretending to understand his product.
So yeah, very
Yeah.
glad to hear from him.
Yeah. Is it you just think, you have AGI,
so what happens? A GI will do all the work.
Of course not. Like we have me and you have agents,
but like we have more work to do,
dude. Like what most people don't understand
Yeah. No.
is like again when you deploy agents in the world and I think everybody
in in the audience understands this,
when you deploy agents in the world then I mean it doesn't reduce your work,
like to some extent it might increase it.
But like 'cause then you
You know, depending on where you are,
you have to like review the work,
right? It's like, right, imagine if we didn't look at the slides and
we showed up here, like then, you know,
it's like we'll be super super surprised,
right?
Yeah, imagine.
Yeah, so so all of a sudden if if I'm able to do twenty software projects
and the past I could only do maybe five in a year and I could do twenty,
like, yeah.
That 20 still requires a lot of resource.
Like I need like humans to be able to look at that.
Like I show the quality and like help people get on board it.
Like, yeah, it's yeah,
there's all this like magical things like,
hey, we're all gonna lose our jobs and the AIs are just working
the company's like how? the the exciting thing for me about the future
is always this like concept. I'm like,
I'm sci-fi puted. I wanna I I kinda like live in a world where one
day we're kinda like all in the stars,
like
You know, where when you're flying,
you're no longer just flying from one continent to the other,
but you're flying from like one planet to another planet,
right? It's like you you know,
but you need robots to build those things,
right? You need to like build like super fast things very quickly.
human beings would always we would always move up,
like we move a level higher, right?
And then you know, we're no longer talking about you know going from
San Francisco to New York. We're talking about going from from Earth
Earth land to like Mars land, right?
And like we're going talking about going from one galaxy to another.
Like there's so much more so much more to do.
that I feel like people who feel like,
hey, this is where we are now,
we should keep these jobs we have today.
I think it was Sam's video. maybe next week we can look at Sam's video,
but it was I don't know if it was in his particular video or like another one,
because there are quite a few of them.
Elon spoke as well. the anthropic CEO Tom Brown also spoke in in those.
people always just tend to
You know, over index you wanna keep that this job you have.
Like dude, like fifty years ago there was no product manager.
There was there's no like podcaster.
Yeah. Alright.
No. No, it's different.
And
yet I think we're doing the AI thing now.
Yeah, this is you.
Let's
do it. Okay. so we've got this news,
right? New York City signed the mayor signed in
to I guess regulations, the one year AI moratorium.
Henry and I sort of both both agree that it's a mistake,
but as you might remember, I have seven kids.
We homeschool them and
Part of the reason for that is the way that the schools in
our area handle technology at younger ages.
And though I don't think it's wise to literally blan it ban it from the classroom,
generative AI.
can become a crutch pretty quickly.
and so my my argument is
AI literacy is important, but so is literal literacy.
Like being able to actually put thoughts together that are original,
that are understandable, being able to communicate clearly is
not something that you necessarily just inherit,
but rather is taught in schools.
And so, you know,
it's not a permanent ban, it's a scoped pause to let standards and rules and
privacy frameworks catch up. I mean,
I'd be lying if I didn't say I tried to build like a family AI portal that could,
you know, let my kids have age appropriate access to models and stuff.
I've definitely played with it.
But yeah, I don't know.
I mean one year in policy, one year is a very short timeline.
I'm I guess to those of us that are AI and agent pilled,
that's like twelve years. But you know,
you Henry pointed out like
Most members of of the citizenry are are looking at chat GPT and seeing
it the same way they saw it a year ago.
And I I suppose that makes sense.
So
hasn't changed for them.
yeah, and and it has real exceptions.
high school AI literacy modules,
supervised pilot, the bands just through grade eight.
I think grade seven and eight is kind of when that's that's when this tool's gonna
be useful. So I don't know, they didn't get the line too far off.
attention and human teaching have
Developmental value for sure. And th this pause allows schools
to build verified equitable infrastructure before rolling out AI.
what would change my mind?
maybe some evidence that supervised classroom AI improves learning
equitably without dependence or introducing privacy harm.
maybe. I I I don't actually don't think mine in mind would be that that hard
to change, but what what do you think?
yeah. No, no, obviously I I don't think it's a good idea,
right? but also d you know, I I get where they're coming from,
right? I get where they're coming from.
They're trying to be they're not trying to be Elliot doctors,
right? There might be some risks attached to having AI this young in in school.
There might be, right? It's it's not proven.
It's not like they came out with like they said,
Hey, here are twenty studies across the world that shows that this is super
bad and we should do it. It's not data driven.
Right. This is I mean depending on who you're talking to,
like the different kind of conspiracy theories around it.
And like this might be protecting teachers,
'cause like, you know, we're going to a place where there's less humans like
teaching things and just much more technologies like guiding people through it.
so yeah, so it depends on where you I get where they're coming from.
But I I definitely do disagree with it.
and and obviously one of the one of the points is on the slide,
right?
I am AIPO. I don't think I'm I I'm fairly s like you know I am
at the edge of these things, right?
So my kids will have AI, right?
They will have access, right? They you know,
they you know, I would teach them,
right? You know, and so on and so forth.
But there are families who aren't like that,
right? They they don't know anything about AI.
They and so the the school is probably where they would learn some of those things,
right? Or where they would s like understand it.
and then obviously if you take that away,
then the rich
It's like upper class kids. Their parents are still gonna teach them yeah,
they might even have private tutors,
right? Even if they go to public schools,
they will these things will be taught at home.
They will have access to it. so you know,
so yeah, so that's number one.
and obviously I posted about this in a group and a group of men with a few friends,
and I gave an example with China,
right? China hasn't banned, I mean,
China is obviously a cute communist country.
They ban everything all the time,
but they haven't banned AI in like even like primary schools,
right? What they did do is to say you can't use it without supervision,
right? So there always has to be a teacher.
but but what I what I suspect needs to happen,
which is why I think this is a lazy move,
is that the the promise of AI and the promise of kind of like I mean,
the AI might not be there yet,
but I'm sure there are people working on technologies that make it like thus.
is like because when you're in you when you're in school.
Everybody's taught the same thing,
but not everybody understands the same way.
They're like kids with ADHD, kids with X Lit,
kids on a s different spectrum.
So not everybody's in the same place,
right? So that the the kind of like promise of AI,
even from like a super young age is you get personalized learning that
is able to sort of like, you know,
meet you where you are and then,
you know, follow through and and have real time feedback.
Or you don't understand this and then it figures out stories of how it's
to make you better understand it.
And so
What I think should be happening is in schools we should
be redesigning how education works.
and and the example I gave I c I don't think the way schools work today
was the way it worked like fifty years ago,
hundred years ago. The curriculum has changed,
No. yeah.
the technology has changed.
So of course it will change, right?
so yeah, but I get it, I get it.
New York is choosing not to be early adopters and there will be other states that
will figure these things out. and this is not this is not news.
Norway banned AI from
Again, I don't remember the screw,
but like a few like a month or so ago,
Norway did the same ban. So, I think over the next few weeks or months,
this is gonna be pretty kinda like common,
right? You're gonna have people who are you know,
who I mean, I think like America half of America is already kinda like s not half,
like eighty percent of Americans don't like AI.
So Bernie just came out with like let's ban AI yesterday.
Let's bend data
So
centers. He he wants to not build data centers,
Let's ban everything. He's like,
let's ban superintelligence.
yeah.
Let's ban AI. Let's ban it. Let's not let them be the AI anymore.
Let's just stop it. so I think this is just the start,
right? And this is also why I don't think it's a good idea,
'cause when you start getting into s like this ban regime,
it's only kinda like it's a domino effect.
And so you're gonna get like more things getting banned.
yeah. Yeah. Cool stuff.
I don't know what you guys think on the in the on the on the thing,
you guys let us know. whether you
Yeah.
are you're a pro or or a con.
It's it's interesting to me.
I I like to think of what the the mayor in New York is doing as just a PR stunt.
I mean, this is a guy who's gotten terrible press.
He's probably a terrible mayor,
that's my opinion, but there's a lot of evidence.
Mm-hmm.
But the other side of this is like the public schools have
in in many cases somewhat abandoned data driven educational methods
methods. yeah.
Yeah, a long time ago. A long time ago.
Hundred percent.
And and
so the way the way, for example,
most public schools in the United States teach reading is the data
on the method is terrible. The the data on the method says that it's
it barely works. And what they find is that kids who are not naturally
drawn to reading don't really learn to read very well using that method.
And real methods exist and the teachers don't like them because it's they're
too much work or you know they're they're
Well w for whatever reason. I I'm not trying to editorialize.
I realize I'm being really bad at that effort,
but there will be data on
on education with and without AI and and with certain methods and other methods,
and that data is gonna tell us what's gonna really work.
And the question is whether or not the the schools are gonna follow the data anyway.
Yeah. Yeah. Not not
So we'll see.
I mean the data data like you said,
even like I don't think American public schools are doing very even with STEM,
right? I think there's been a few states that have like been super like pass like
some punitive policies around STEM.
you know, special education. It's like,
yeah, it's all right. all right,
cool. Andy, let's let's keep it moving.
I think we we're over time.
Yeah,
yeah. this this episode's also brought to
you by Heritage Telecom. While the AI industry bundles chips,
models, and seats, Heritage does the thing it's good at.
Unified communications as a service and VoIP phone service
for businesses that need their calls to work.
Independent, boring, reliable,
no telemetry except for all of your call recordings and all of your CDR records.
Find us at heritageel dot com.
And that's basically it, right?
Three things to watch
So
for after the show. you know,
we're excited to see Astra roll out.
I'm I'm interested to check and see if I have it yet.
Henry already does. API
Yeah.
users, like let's let's get our open clause connected to it.
we'll see what what independent benchmarks say once they have it.
And then let's see what regulators and companies say about Nvidia
and Hugging Face Agreement before it closes in 27.
And lastly, whether Visko can demonstrate a persistent world outside
the launch reel while New York measures the consequence of
a year-long classroom pause. So that's the week ahead.
We'll be back Friday, September 11th,
4 p.m. Eastern. Follow the weekly claw at weeklyclaw.ai and join
the Discord through the link here on your screen.
we're really looking for users to provide you know,
interested users to provide feedback to us on the the segments and content.
you know, please by all means join us jo join us and join the conversation
and we'll look forward to seeing you next week.
Thank you guys. Let's just go do Astra.
Have a good one.