WEBVTT

00:00:55.126 --> 00:00:57.165
Hello everyone, welcome to ThursdAI.

00:00:57.206 --> 00:00:58.785
Today is September 24th.

00:00:58.846 --> 00:01:00.966
This is Alex Volkov talking to you live

00:01:01.606 --> 00:01:02.885
from another ThursdAI.

00:01:03.286 --> 00:01:04.486
An incredible week.

00:01:04.846 --> 00:01:10.286
Anthropic says Opus 5.5 does Fable level
work for 40% less, and we've been testing

00:01:10.326 --> 00:01:11.246
this.

00:01:11.246 --> 00:01:16.926
OpenAI cut GPT-6 prices in half and also
released 2 new models, and Meta

00:01:17.126 --> 00:01:18.685
AI passed GPT,

00:01:19.366 --> 00:01:24.366
ChatGPT in the App Store as the number one
downloaded app in the fastest time ever,

00:01:24.886 --> 00:01:29.706
all in one week. Plus we had Meta Connect,
we had Grok release, we had Gemini come

00:01:29.766 --> 00:01:34.325
back. It's, it's a big, big, big, big
show, and I would love to welcome my host,

00:01:34.806 --> 00:01:38.126
uh, Peter Gostev, co-host, uh, Wolfram
Ravenwolf

00:01:38.806 --> 00:01:41.945
from, uh, France, Peter, how are you
doing, sir?

00:01:42.026 --> 00:01:43.026
How is your week?

00:01:44.106 --> 00:01:45.225
My week's been good.

00:01:45.546 --> 00:01:49.866
Yeah, lots of testing, lots of new
capabilities to explore.

00:01:50.106 --> 00:01:54.425
Yeah, I know we're gonna talk about all of
this, but yeah, Opus 5.5 has been fun.

00:01:55.106 --> 00:01:56.565
It's always cool, you know, when,

00:01:57.426 --> 00:02:01.025
when you feel like you know what the
models can do, and then they just throw

00:02:01.106 --> 00:02:05.106
something in, it's like, holy crap, like I
didn't, I had no idea you could even do

00:02:05.146 --> 00:02:08.866
this. And now, yeah, all of your timelines
are filled with this stuff.

00:02:09.666 --> 00:02:11.326
The timelines are, uh,

00:02:12.906 --> 00:02:14.986
the social media timelines are condensing.

00:02:15.786 --> 00:02:16.786
The

00:02:16.866 --> 00:02:21.225
timelines that we get releases, though,
are just crazy condensed.

00:02:21.366 --> 00:02:23.785
Now, um, folks, there's so much in the
show

00:02:24.486 --> 00:02:28.466
that is just like major releases that,
that I don't know if we're gonna get to

00:02:28.525 --> 00:02:30.446
like open source, which is also very
interesting.

00:02:30.525 --> 00:02:31.525
So, um,

00:02:32.886 --> 00:02:36.065
Peter, what is the one thing that's
highlight of this last week for you?

00:02:36.306 --> 00:02:38.025
I think I know, but like I would love to
hear from you.

00:02:38.666 --> 00:02:42.925
I mean, it, it really has to be Opus
because I think it- Opus 5.5, yeah?

00:02:43.226 --> 00:02:45.086
Yeah, Opus 5.5 because

00:02:45.746 --> 00:02:49.226
it is, like they've clearly added
something.

00:02:49.506 --> 00:02:52.505
The fact that they can create, and if you
don't know what I'm talking about, it's,

00:02:53.066 --> 00:02:57.886
you can sort of create something like with
in just plain HTML.

00:02:57.946 --> 00:03:03.126
Like we we had this Astra creation stuff
in Blender, but here now in HTML, and I

00:03:03.186 --> 00:03:07.265
don't even see that it's using like some
assets or libraries or something like

00:03:07.306 --> 00:03:10.425
that. It just handcrafts stuff, I don't
know, in JavaScript or something,

00:03:11.226 --> 00:03:16.606
and uh just creates these amazing
animations, and I, and it's like, you

00:03:16.666 --> 00:03:20.246
know, we've been looking at diffusion
video and how much better it's getting,

00:03:20.786 --> 00:03:22.066
and there's like I have something

00:03:22.866 --> 00:03:26.005
to show about this this week, like my mind
was absolutely blown.

00:03:26.026 --> 00:03:28.686
I don't know if you saw it, but yeah,
diffusion video is going.

00:03:28.706 --> 00:03:31.705
Yeah, and it's like, yeah, diffusion video
is amazing, right?

00:03:31.786 --> 00:03:37.765
But then you, you start seeing like, oh,
the LLM can like create, obviously not the

00:03:37.866 --> 00:03:40.866
super realistic ones, but you can create,
like, versions of it.

00:03:40.906 --> 00:03:45.946
I- I haven't shared it yet, but I was, uh,
putting a recipe to Claude and to Opus 5.5

00:03:46.226 --> 00:03:47.946
and say, can you, like, animate it?

00:03:48.065 --> 00:03:51.065
And it's insane, like, it just looks real,
right?

00:03:51.106 --> 00:03:55.065
So now suddenly there's a whole category
of capabilities that's opened up.

00:03:55.466 --> 00:03:57.305
Like, it could be for children for fun.

00:03:57.386 --> 00:04:03.025
I created the, the, you know, the Rickroll
kind of, uh, in kind of artsy

00:04:03.226 --> 00:04:07.625
style. Amazing. It just took, I mean, took
some time to make it, but I didn't have to

00:04:07.666 --> 00:04:11.745
put in much work. So it, I don't know, who
knows, maybe it's all hype or goes away,

00:04:11.826 --> 00:04:13.625
but it's amazing. It's so cool.

00:04:14.065 --> 00:04:17.466
While Peter talks, I don't know if you
guys noticed that we're talking about new

00:04:17.546 --> 00:04:20.305
capabilities. We have a new producer today
on the show.

00:04:20.505 --> 00:04:22.605
Opus 5.5 is live producing the show.

00:04:22.666 --> 00:04:26.386
It listens to us as we speak and put up
these like nice chyrons as well, and also

00:04:26.466 --> 00:04:27.886
in our chat and also on X.

00:04:27.906 --> 00:04:30.266
So if if our stream disconnects, you know
who to blame.

00:04:30.506 --> 00:04:32.255
But Peter, I I'm absolutely with you.

00:04:32.316 --> 00:04:36.766
I think Opus 5.5 for me was the absolute
highlight of the week.

00:04:37.646 --> 00:04:38.646
Folks,

00:04:38.926 --> 00:04:40.985
Opus 4.6 is back

00:04:42.366 --> 00:04:43.366
with

00:04:43.405 --> 00:04:45.246
Fable level abilities,

00:04:46.246 --> 00:04:49.726
and eh eh eh eh and it talks like a normal
person.

00:04:49.886 --> 00:04:52.245
I, it is quite crazy.

00:04:52.726 --> 00:04:55.025
Uh, Wolfram, I know you're traveling a
lot.

00:04:55.126 --> 00:04:56.566
Hopefully you're back now with us.

00:04:56.726 --> 00:04:57.726
Um,

00:04:58.326 --> 00:05:00.485
how you doing? Tell us what's your
highlight of the week.

00:05:01.452 --> 00:05:03.837
Yes, sure. I'm here right at the

00:05:04.518 --> 00:05:10.413
AI Engineer Paris, hosted by Mistral, and
uh, yeah, Uh, thank you,

00:05:10.473 --> 00:05:13.473
Wolfram, for reporting live and covering
that event and going there.

00:05:13.593 --> 00:05:16.153
Shout out to the AI engineer folks from
France.

00:05:16.473 --> 00:05:19.282
Uh, Mistral, I think, is co-hosting that,
so big shout out to them.

00:05:19.483 --> 00:05:22.843
All right, folks, uh, 3 themes for the
show of this week.

00:05:24.083 --> 00:05:27.602
Number 1, uh, pacing the frontier does not
mean stopping.

00:05:28.282 --> 00:05:33.043
OpenAI and Anthropic releasing back to
back on the same day incredible models

00:05:33.203 --> 00:05:34.203
this week.

00:05:34.763 --> 00:05:39.883
Opus released 5.5, uh, ChatGPT released
GPT-6 Soul and GPT-6

00:05:40.563 --> 00:05:41.563
Luna.

00:05:41.563 --> 00:05:46.342
Terra is no longer there at half the
price, so pausing, uh, the frontier does

00:05:46.363 --> 00:05:48.243
not mean stopping at any, any point.

00:05:49.803 --> 00:05:50.803
Number 2,

00:05:51.104 --> 00:05:54.103
every assistant is an agent, not every
agent is an assistant.

00:05:54.224 --> 00:05:58.543
We're gonna have to start talking about AI
assistants in a different language.

00:05:58.864 --> 00:06:02.533
Assistants have a heartbeat, assistants
are proactive, assist- assistants know

00:06:02.654 --> 00:06:06.053
context about you, and assistants take
care of their own context.

00:06:06.494 --> 00:06:09.534
They have tools, they have a browser, they
work while you sleep.

00:06:10.134 --> 00:06:14.353
Agents is a very generic term, and I know
it's confusing to many people, but we're

00:06:14.374 --> 00:06:17.933
gonna lead the charge. We talked with
David Palon from Assistant Benchmark last

00:06:17.974 --> 00:06:22.114
week, and uh, this week we're going to
cover Meta Muse and the stuff that they

00:06:22.154 --> 00:06:24.673
did at Meta Connect. Those are AI
assistants.

00:06:25.094 --> 00:06:28.553
There's a very thorough and distinct line
between them, and

00:06:29.494 --> 00:06:31.154
assistants is what people need.

00:06:31.614 --> 00:06:35.393
Agents is a scary word for many people, so
we're gonna talk about AI assistants, and

00:06:35.494 --> 00:06:38.053
I think at some point we're gonna go have
like an assistant corner.

00:06:38.573 --> 00:06:40.779
And number 3 is voice.

00:06:41.219 --> 00:06:46.718
Gemini with a Gemini 3.8 Flash and Gemini
3.8 Flash TTS,

00:06:47.219 --> 00:06:51.258
number 1 at voice design, and sounds like
just absolutely human.

00:06:51.299 --> 00:06:53.659
So we're gonna definitely test this out on
the stage.

00:06:53.779 --> 00:06:56.298
By the way, shout out to the beautiful
folks from Gemini.

00:06:56.619 --> 00:06:59.579
Uh, I received like a mystery box, and I
was like, who is this from?

00:06:59.939 --> 00:07:03.664
Uh, the folks at Gemini said, hey, we're
launching voice, and they sent me this

00:07:03.664 --> 00:07:07.578
like beautiful microphone, and a t-shirt
and a cap, and shout out to the folks in

00:07:07.639 --> 00:07:08.639
there who,

00:07:09.275 --> 00:07:11.045
who just thought about me, is which is
very nice.

00:07:11.185 --> 00:07:12.465
But I think first of all,

00:07:13.185 --> 00:07:14.524
we'll start with the...

00:07:15.825 --> 00:07:17.697
Should we start with TLDR and then start
going?

00:07:30.857 --> 00:07:33.176
TLDR, uh, on the show this week.

00:07:33.457 --> 00:07:34.937
We'll start with the main three.

00:07:35.097 --> 00:07:40.697
Anthropic launches Claude Opus 5.5,
claiming Fable 5.1 level work at 40%

00:07:40.977 --> 00:07:42.697
lower cost than Opus 5.

00:07:43.257 --> 00:07:44.257
I can concur.

00:07:44.737 --> 00:07:50.417
I've been running this model on ultra high
yesterday in like 4 things, and I

00:07:50.737 --> 00:07:52.637
barely hit 15% of my weekly quota.

00:07:52.697 --> 00:07:57.376
It's quite uncanny after all of what
happened last week, which I, uh, prefer to

00:07:57.457 --> 00:07:58.697
forget. Uh,

00:07:59.537 --> 00:08:03.736
just a few hours later, OpenAI rolls out
GPT-6 Soul and GPT-6

00:08:04.457 --> 00:08:09.867
Luna at half the price of the GPT-5.6,
promotional API price.

00:08:10.027 --> 00:08:12.606
We got price cuts from both companies, and

00:08:13.497 --> 00:08:14.497
it's very welcome.

00:08:15.182 --> 00:08:16.182
And I think the number 1

00:08:17.542 --> 00:08:21.742
news, at least for me from the assistant
perspective, is that Meta unveils that

00:08:21.862 --> 00:08:24.742
Muse is central to their whole roadmap.

00:08:25.582 --> 00:08:26.582
Muse

00:08:26.742 --> 00:08:30.382
will now have a voice and a video chat
that you will be able to talk with your

00:08:30.462 --> 00:08:32.201
Muse, like me and Peter are talking right
now.

00:08:32.262 --> 00:08:34.182
You'll see it like lips animate, etc.

00:08:34.502 --> 00:08:39.081
Muse will also come to the Meta VR glasses
and a charm device that kind of looks like

00:08:39.122 --> 00:08:41.822
a Tamagotchi that you can buy for
Christmas very soon.

00:08:42.422 --> 00:08:46.801
And uh, Meta also announced VR glasses
with Muse built in- into operating system.

00:08:46.882 --> 00:08:51.731
So AI is kind of permeating through all of
what Meta released, and Meta Connect

00:08:51.972 --> 00:08:53.972
coverage will do so very soon.

00:08:54.252 --> 00:08:59.211
In other big news from big companies and
LLMs, SpaceX AI released Grok 4.7,

00:08:59.852 --> 00:09:05.682
and it's reported 46% on CursorBench, and
yet, and yet many people are very

00:09:05.782 --> 00:09:06.841
disappointed with that release.

00:09:06.862 --> 00:09:08.062
I think we expected more.

00:09:08.582 --> 00:09:09.582
However,

00:09:10.582 --> 00:09:11.802
they also have distribution.

00:09:11.902 --> 00:09:15.501
So Tesla now connects with Grokbots and
connectors.

00:09:15.822 --> 00:09:20.541
So literally, you can now talk to your
assistant while driving hands-free, and I

00:09:20.622 --> 00:09:23.422
think that that is more important news
than Grok 4.7.

00:09:23.502 --> 00:09:28.511
I think that that 0.1 release of models is
now augmented by where we can use those

00:09:28.612 --> 00:09:32.361
models or assistants, and Grokbot in a
Tesla, which I didn't get yet, maybe after

00:09:32.382 --> 00:09:34.681
the show I will, is just an absolutely

00:09:35.422 --> 00:09:36.702
mind-blowing experience.

00:09:36.982 --> 00:09:41.821
And also, OpenAI published a four area
plan for independent safety assessment,

00:09:42.272 --> 00:09:45.311
but no assessors of start date, but this
is a

00:09:46.432 --> 00:09:49.261
continued thread of what we talked to you
about pacing the frontier and letting

00:09:49.342 --> 00:09:52.742
evaluators, external evaluators, in to
your lab.

00:09:53.142 --> 00:09:57.701
In this week's buzz, folks, we're one week
away from Fully Connected in Moscone's

00:09:57.782 --> 00:09:58.782
house next week.

00:09:59.382 --> 00:10:00.722
Pitbull, Sarah

00:10:01.502 --> 00:10:05.282
Guo, Fei-Fei Li, Alex Volkov is not nearly

00:10:06.102 --> 00:10:09.222
close to the level of those folks, but
also we're gonna be there with Wolfram.

00:10:09.522 --> 00:10:12.661
it's gonna be sick, folks. I really want
you to come, uh, to Fully Connected.

00:10:12.722 --> 00:10:14.122
We have a free ticket for you.

00:10:14.202 --> 00:10:17.471
If you don't know where you're gonna be
after Dev Day on OpenAI, we're, we're

00:10:17.492 --> 00:10:19.524
gonna welcome you, to see Pitbull chat
with us.

00:10:19.705 --> 00:10:22.321
so the Thursday I show next week is going
to be a little different.

00:10:22.340 --> 00:10:25.681
We're gonna start a little bit later, and
we're gonna host a show, folks on the

00:10:25.781 --> 00:10:26.860
floor. It's gonna be exciting.

00:10:27.181 --> 00:10:30.690
Uh, we haven't done a live one since, AI
engineer in Moscone also.

00:10:30.771 --> 00:10:34.471
So this is, you know, back-to-back live
shows from Moscone, not too shabby, not

00:10:34.491 --> 00:10:39.370
too shabby, ThursdAI. Uh, and also
CoreWeave earns the SemiAnalysis Platinum

00:10:39.571 --> 00:10:42.651
ClusterMax benchmark for the third report
in a row.

00:10:43.051 --> 00:10:46.810
Shout out to the CoreWeave folks for this
incredible, incredible news.

00:10:48.051 --> 00:10:54.046
Uh, ClusterMax measures performance of
NVIDIA kernels on how we set them up,

00:10:54.091 --> 00:10:57.130
and CoreWeave for the third year in a row
is platinum tier.

00:10:58.451 --> 00:10:59.911
Folks, last week we told you about JEV.

00:10:59.971 --> 00:11:04.510
This is the TLDR. We're not gonna go super
deep, but JEV changed the world since last

00:11:04.571 --> 00:11:09.010
week we told you about here with Ali Labs
from TypeSafe until this week.

00:11:09.331 --> 00:11:10.610
JEV is everywhere.

00:11:10.760 --> 00:11:15.200
but as we told you last week, the open
source community is gonna show up.

00:11:15.881 --> 00:11:19.601
So apparently some folks already trained
decision models, and so there's like

00:11:19.841 --> 00:11:23.461
multiple, uh, classifier open source
stuff.

00:11:23.561 --> 00:11:27.720
It's And, um, Swix, our friend of the pod,
the host of Latent Space and the creator

00:11:27.761 --> 00:11:32.920
of AI Engineer, which Wolfram is now at,
hosted Diogo Almeida, the CEO, uh, and

00:11:33.061 --> 00:11:37.860
co-founder of TypeSafe, and they talked
about Diogo's not like, he's really happy

00:11:37.901 --> 00:11:40.701
with this. They're not like, oh no,
competition is gonna come from open

00:11:40.721 --> 00:11:42.440
source. He's like, yes, this is a new
category.

00:11:42.541 --> 00:11:46.000
We all need to innovate. It's amazing that
open source is gonna catch up, and I don't

00:11:46.021 --> 00:11:50.940
care about benchmarks. So a lot to talk
about how JEVI this last week was and how

00:11:51.001 --> 00:11:54.250
it took over the feeds, uh, but also We
need to talk about Gevbench from our

00:11:54.291 --> 00:11:58.451
friend Florian. Maybe he'll show up, uh,
that measures other classifier models and

00:11:58.691 --> 00:12:03.390
System 1 models. Uh, so we're definitely
gonna talk about Leia and, uh, and and

00:12:03.531 --> 00:12:06.090
Jeff. There's Jev and there's Jeff.

00:12:06.731 --> 00:12:07.731
Uh,

00:12:07.731 --> 00:12:12.010
let's see how my transcription deals with
me saying Jev and Jeff back to back.

00:12:12.531 --> 00:12:17.291
Okay, folks, voice and audio is also just
blowing the fuck up.

00:12:17.771 --> 00:12:23.731
Google debuts Gemini 3.8 Flash and
Flashlight TTS, claiming number one on

00:12:23.811 --> 00:12:25.490
the AI voice design benchmark.

00:12:25.571 --> 00:12:26.571
Folks, this is a

00:12:27.451 --> 00:12:30.491
cloning your own voice system from Google.

00:12:31.891 --> 00:12:37.170
We told you about this 3 years ago, that
nothing is gonna break in the world where

00:12:37.651 --> 00:12:40.010
audio models will be sounding like you.

00:12:40.411 --> 00:12:44.081
Nothing is gonna happen. And now Google,
one of the biggest companies in the world,

00:12:44.121 --> 00:12:48.080
is allowing you to clone your voice and
play it back within like seconds, and it

00:12:48.181 --> 00:12:49.541
sounds really, really, really good.

00:12:50.376 --> 00:12:51.835
Um, we're gonna play around with this.

00:12:51.856 --> 00:12:54.916
We're gonna play you some samples of
including me speaking.

00:12:54.976 --> 00:13:00.075
It's gonna be dope. Qwen also debuts Qwen
3.8 OmniFlash, agentic audio video model

00:13:00.176 --> 00:13:05.416
with 1 million token context, and Qwen
also has a live translate model, cutting

00:13:05.456 --> 00:13:09.375
its reporting interpolation lag to 2.3
seconds live translating.

00:13:10.576 --> 00:13:14.296
Qwen has been all over the open source
news as well with vision, image, and video

00:13:14.416 --> 00:13:19.455
category. Uh, Alibaba Qwen open sources
Qwen Image 2.1, a 7B model for image

00:13:19.496 --> 00:13:21.976
generation and editing with native
transparency.

00:13:22.126 --> 00:13:25.985
Black Forest Labs open sources Flux 3
Action, a 7B.

00:13:26.046 --> 00:13:31.176
Do Uh, a 7B world action model that it
says tops Rovalab at

00:13:31.356 --> 00:13:34.875
42%, so, uh, you know, Black Forest Labs
is back.

00:13:35.436 --> 00:13:36.436
And um,

00:13:37.916 --> 00:13:39.075
this is quite crazy, folks.

00:13:39.236 --> 00:13:42.195
Karina Nguyen unveils act 486.

00:13:42.426 --> 00:13:46.786
A pre-rendered research demo that lets
viewers talk back to a Joe Rogan episode.

00:13:47.021 --> 00:13:49.036
it's quite the eye-opening demo.

00:13:49.095 --> 00:13:50.095
It's quite crazy.

00:13:50.616 --> 00:13:53.535
In open source LLMs, we have a big one as
well, Xiaomi.

00:13:54.396 --> 00:14:00.035
Uh, open sourced Mimo V2.6 Pro, and
artificial analysis rocks this, the top

00:14:00.276 --> 00:14:05.766
open weights model at 46% artificial,
analysis intelligence

00:14:05.986 --> 00:14:06.986
index.

00:14:07.706 --> 00:14:08.706
And

00:14:09.226 --> 00:14:13.565
there's a bunch more stuff that we haven't
covered in the TLDR, but, Wolfram, maybe

00:14:13.726 --> 00:14:17.165
one or two more items that you want to
make sure that we need to at least

00:14:17.286 --> 00:14:18.286
mention.

00:14:18.486 --> 00:14:22.295
Yeah, So, uh, NVIDIA Nemotron 3, it's a
diarization

00:14:22.896 --> 00:14:27.236
model that came out open, of course, a 100
million parameter model for speaker

00:14:27.336 --> 00:14:29.736
tracking and even overlapping speakers.

00:14:30.096 --> 00:14:31.496
So this is super exciting.

00:14:32.776 --> 00:14:37.735
And, um, yeah, in other news, OpenRouter
has a batch API which can save you a lot

00:14:37.776 --> 00:14:40.915
of money if you are using OpenRouter and
you don't need the response now.

00:14:41.015 --> 00:14:41.976
If you

00:14:41.976 --> 00:14:46.295
can live with it 24 hours later, then you
can save up to 50%.

00:14:46.856 --> 00:14:50.575
Well, that is something, uh, cost is a
very important aspect.

00:14:51.216 --> 00:14:52.375
Yeah, I think those are the

00:14:53.015 --> 00:14:54.175
2 main things.

00:14:55.296 --> 00:14:59.736
Stepfun also has a new agentic model with
a million context, but it's, uh, it's

00:14:59.896 --> 00:15:05.875
supposed to come out as open weights, uh,
in October, but it's, uh, API usable

00:15:05.936 --> 00:15:07.695
now. I think at least I saw it.

00:15:09.496 --> 00:15:13.035
All right, folks, I think this is the
TLDR, and I think it's time for us to kick

00:15:13.136 --> 00:15:13.936
off

00:15:13.936 --> 00:15:14.776
our

00:15:14.776 --> 00:15:15.616
show

00:15:15.616 --> 00:15:16.616
officially,

00:15:17.736 --> 00:15:18.736
uh, with,

00:15:19.656 --> 00:15:20.935
with Frontier Labs. Let's go.

00:15:34.936 --> 00:15:37.095
Alrighty, folks, I think let's,

00:15:38.096 --> 00:15:40.776
let's start. Folks, Claude Opus 5.5.

00:15:40.936 --> 00:15:46.116
Anthropic just shipped a model that beats
their own fable 5.1 that just recently was

00:15:46.176 --> 00:15:50.656
released on real work, and it is 40%
cheaper than Opus 5.

00:15:51.056 --> 00:15:52.216
And honestly, it just

00:15:52.896 --> 00:15:53.935
feels limitless.

00:15:56.312 --> 00:16:01.272
My week started a week ago with burning
quotas all over Fable, burning my quotas

00:16:01.432 --> 00:16:05.932
ranges, and I was like, oh no, I cannot do
this work because for Thursday I need a

00:16:05.992 --> 00:16:10.712
cloud for production. Then Opus 5.5 feels
limitless because I was barely able to get

00:16:10.752 --> 00:16:12.832
to the end of the quota, and I was running

00:16:14.032 --> 00:16:16.071
workflows and and agents, whatever.

00:16:16.391 --> 00:16:18.711
Um, Opus is back.

00:16:19.111 --> 00:16:20.911
It feels like Opus 4.6,

00:16:21.512 --> 00:16:25.392
and you can talk to this model, but it's
as intelligent as Fable as I was able to

00:16:25.712 --> 00:16:29.072
to see. So, uh, I think the timeline
agrees.

00:16:29.192 --> 00:16:31.311
I would love to hear from folks here.

00:16:31.632 --> 00:16:35.772
Uh, for coding and agentic stuff, most of
your bill is cash reads, and so the cash

00:16:35.872 --> 00:16:38.592
read, uh, is a significant price drop, and
you can feel it.

00:16:39.032 --> 00:16:43.192
Also, it's fast. And, uh, Opus is also
producing the show, Opus 5.5, so hopefully

00:16:43.272 --> 00:16:45.631
we'll see like a little banner that it's
gonna put up for us.

00:16:45.992 --> 00:16:48.791
Uh, but I would love to hear from you
guys.

00:16:48.992 --> 00:16:49.992
Uh,

00:16:50.152 --> 00:16:53.592
how was, how was your few days with Opus
5.5?

00:16:53.752 --> 00:16:54.752
Let's start with Peter.

00:16:55.952 --> 00:17:00.631
Yeah, well, first thing to say is that,
uh, we released a score for Code Arena,

00:17:00.712 --> 00:17:03.552
which is measuring front end and all of
that fun stuff.

00:17:03.872 --> 00:17:06.791
And guess what? It's, uh, on top again.

00:17:07.192 --> 00:17:12.352
So Astra was above everyone else, and now
Opus 5.5 is above Astra.

00:17:12.992 --> 00:17:17.152
And, uh, yeah, it's super impressive on,
on that kind of coding front end.

00:17:17.472 --> 00:17:20.352
We got a bunch of new capabilities in
terms of using the,

00:17:20.992 --> 00:17:23.351
all of the HTML generations.

00:17:23.472 --> 00:17:27.391
That seems to be just completely insane,
and all the 3D stuff.

00:17:28.222 --> 00:17:32.062
to your point about efficiency and cost,
obviously they did reduce cost by 20%.

00:17:32.212 --> 00:17:34.891
that is good. They said it's using fewer
tokens.

00:17:35.212 --> 00:17:38.824
It's kind of interesting. I kind of had
mis- mixed experience with that when I was

00:17:38.904 --> 00:17:41.593
running, the harder prompts on Max,

00:17:42.194 --> 00:17:47.564
it was running for many hours, and it was,
then I calculated the API cost, and it was

00:17:47.624 --> 00:17:49.041
like like 80 dollars per generation.

00:17:49.641 --> 00:17:50.641
So I

00:17:51.081 --> 00:17:53.401
would guess maybe if you take a median,

00:17:54.081 --> 00:17:58.740
maybe it is fewer tokens, but then if you
take the top 10% of prompts, it's probably

00:17:58.841 --> 00:18:00.380
way more tokens, is my sense.

00:18:00.470 --> 00:18:02.410
it's just kind of vibe-based sense.

00:18:02.441 --> 00:18:03.640
I don't have data for it yet.

00:18:04.161 --> 00:18:06.490
So that feels kind of different.

00:18:06.608 --> 00:18:10.877
I was comparing to, GPT-6 Soul at the same
time, and I had like

00:18:11.518 --> 00:18:12.518
Soul, $1

00:18:13.348 --> 00:18:15.308
and Opus, $60

00:18:16.308 --> 00:18:18.187
side by side. So, yeah, I had a few
experience like that.

00:18:18.248 --> 00:18:22.118
I'm not saying that's universal, and I do
agree with you, in just day-to-day use, it

00:18:22.197 --> 00:18:24.217
does seem to last longer than than, your
sub.

00:18:24.617 --> 00:18:27.577
Um, but yeah, I think it's not, it's not
just completely across the board.

00:18:28.658 --> 00:18:29.957
What happened to me last week was...

00:18:32.138 --> 00:18:34.217
for the first time, the things flipped.

00:18:34.698 --> 00:18:39.567
I remember being very cautious about my
kind of like quotas on the Cloud Max

00:18:39.588 --> 00:18:45.327
account, and I remember Limitless in the
GPT, uh, you know, Sol and the 5.5

00:18:45.447 --> 00:18:51.007
days. I remember limitless codecs, just
never thought about reaching my quota,

00:18:51.088 --> 00:18:52.088
like never, just never.

00:18:52.688 --> 00:18:57.187
Uh, and then Astra came out, and then
within a span of half a day, I finished my

00:18:57.288 --> 00:18:59.767
weekly quota and hit the reset, finished
my weekly quota.

00:19:00.088 --> 00:19:01.607
Uh, not to mention I got bad results.

00:19:01.968 --> 00:19:03.527
We, let's not talk about this.

00:19:04.048 --> 00:19:06.377
Fine, we're over. and it flipped now.

00:19:06.738 --> 00:19:12.258
Now I ran Anthropic Claude Code for the
longest time with just no end in sight,

00:19:12.698 --> 00:19:16.678
and I, this is quite incredible to me, but
also the performance of this model, it

00:19:16.738 --> 00:19:19.107
works very well. Uh, Wolfram would love to
hear from you.

00:19:19.148 --> 00:19:20.597
You know, you've been traveling, and
Nisten as well.

00:19:20.618 --> 00:19:21.997
I don't know if you already had a chance
to try.

00:19:22.058 --> 00:19:24.510
Feel feel free to chime in. How has your
experience been?

00:19:24.591 --> 00:19:25.556
Because

00:19:25.556 --> 00:19:28.835
As much as I don't like the politics of
Anthropic,

00:19:30.116 --> 00:19:31.116
I do

00:19:31.916 --> 00:19:37.665
find, because Ti- I was a TypeScript dev
for like 6 or more years before this, it

00:19:37.766 --> 00:19:39.816
writes the best code. So I've been using
it the whole time.

00:19:40.576 --> 00:19:46.096
Opus 5.5, uh, there are some drawbacks,
like it is still bad at

00:19:46.256 --> 00:19:50.956
DevOps. I would still revert to Fable for
anything like configuring

00:19:51.656 --> 00:19:54.155
networking, DevOps, containers and stuff
on Linux.

00:19:54.736 --> 00:19:57.856
But when it came to anything to do with
the web,

00:19:58.816 --> 00:20:03.275
WebGPU, web APIs, anything to do with
Chrome, it's just been trained on much

00:20:03.976 --> 00:20:09.955
newer data, and it understands even the
int8 optimizations of, WebGPU,

00:20:10.756 --> 00:20:16.575
and, it was pretty crazy at writing
kernels, both like on SGLang on

00:20:16.636 --> 00:20:17.636
hardware and,

00:20:18.301 --> 00:20:21.206
uh, making websites and also using WebGPU.

00:20:22.006 --> 00:20:25.095
And, yeah, the model, it's, it's just an
absolute banger.

00:20:26.116 --> 00:20:28.355
It's, it, it is that good.

00:20:29.276 --> 00:20:32.475
Uh, it, it does everything, it does it
fast,

00:20:33.236 --> 00:20:35.236
it uh, knows how to troubleshoot

00:20:35.916 --> 00:20:39.315
better. There are a few things which you'd
still revert to Fable,

00:20:39.956 --> 00:20:42.616
but uh, yeah, overall, that is my default
now.

00:20:43.246 --> 00:20:46.446
on most things, yeah. Folks, you, do you
understand what it takes for Nisten to

00:20:46.486 --> 00:20:49.906
tell you that his default is not some open
source obscure model that he runs himself

00:20:49.946 --> 00:20:54.005
on 17 GPUs? Do you understand what it
takes for Anthropic to win over Nisten?

00:20:54.046 --> 00:20:56.765
I don't think you guys quite get what just
happened here.

00:20:57.166 --> 00:20:58.765
This is my default as well.

00:20:58.966 --> 00:21:01.605
This is a fast model that is as smart.

00:21:02.206 --> 00:21:04.645
I used to love Opus back in Opus 3 days.

00:21:04.765 --> 00:21:05.765
Opus was like

00:21:06.206 --> 00:21:08.566
Opus, and then they screwed it up.

00:21:08.938 --> 00:21:10.277
Somebody from Anthropic said, hey folks.

00:21:10.397 --> 00:21:11.557
we're sorry for Opus 5.

00:21:11.797 --> 00:21:13.357
We hope this model will make up for it.

00:21:13.637 --> 00:21:17.296
Because Opus 5, and we told you about
this, just like 5.0, was an absolute crap

00:21:17.357 --> 00:21:18.357
model.

00:21:18.357 --> 00:21:19.417
Nobody wanted to talk to this.

00:21:19.477 --> 00:21:21.777
It was slow. It was, uh, incoherent.

00:21:21.877 --> 00:21:23.196
It used Claude-isms.

00:21:23.797 --> 00:21:24.896
Claude-isms are gone.

00:21:25.037 --> 00:21:26.076
Jargon douche is gone.

00:21:26.157 --> 00:21:28.756
All of that is gone. This model talks like
a normal person.

00:21:29.917 --> 00:21:34.197
And here's a tip from me to you, if you
haven't got there yet, if you don't use

00:21:34.277 --> 00:21:35.876
Pstack from Lauren

00:21:38.597 --> 00:21:41.477
10, I believe, from, uh, the GrokCursor
team.

00:21:41.796 --> 00:21:44.356
Uh, Pstack has this beautiful skill called
Bro.

00:21:45.477 --> 00:21:46.477
All it does

00:21:46.677 --> 00:21:51.021
is when you get something that's very,
very confusing, you do slash bro, and it

00:21:51.101 --> 00:21:53.661
asks the model to repeat it in human, uh,
words.

00:21:54.021 --> 00:21:55.360
I, I love using that skill.

00:21:55.421 --> 00:21:58.800
I was like, every time, even Opus 5.5, or
even before, it gives you like some

00:21:58.860 --> 00:22:00.181
Claudeese, you're like, bro,

00:22:01.061 --> 00:22:02.540
and it explains to you like a human.

00:22:03.021 --> 00:22:05.041
Um, so, you know, we'll link that.

00:22:05.201 --> 00:22:07.360
in the show notes for the folks.

00:22:08.201 --> 00:22:10.301
Anthropic, please don't nerf Opus.

00:22:10.841 --> 00:22:12.321
This is a strong request from us.

00:22:12.521 --> 00:22:15.920
There's this whole thing about models
coming out, and they're amazing, and then

00:22:16.001 --> 00:22:19.100
after a few weeks, they're not amazing
anymore because the hype died down, and

00:22:19.161 --> 00:22:20.441
maybe you tweaked inference.

00:22:21.201 --> 00:22:24.941
We ask you, keep Opus 5.5 the same as it
is now.

00:22:25.471 --> 00:22:28.630
Please do not ner- nerf Opus for us.

00:22:29.111 --> 00:22:31.911
I think with this we can move on to the
next news.

00:22:33.436 --> 00:22:36.315
OpenAI just cut GPT-6 prices in half.

00:22:36.836 --> 00:22:40.916
And the funny part, the cheap models might
be better to use than the flagship model

00:22:40.996 --> 00:22:46.156
of Astra. GPT-6 Soul and GPT-6 Luna, the
faster, cheaper, and

00:22:46.656 --> 00:22:47.656
probably

00:22:47.736 --> 00:22:48.736
more,

00:22:49.035 --> 00:22:53.315
performant siblings of Astrum, uh, dropped
at half the price of GPT-5.6.

00:22:53.396 --> 00:22:57.575
So this is a new next level model at half
the price of the previous model that was

00:22:57.716 --> 00:22:58.716
promo.

00:22:58.876 --> 00:23:04.595
Soul is 2 dollars in and 10 dollars out,
and folks, Luna, GPT-6 Luna is 10

00:23:04.836 --> 00:23:08.755
cents in and 50 cents out, which is like,
it's nothing.

00:23:09.076 --> 00:23:12.435
This is the best cheap frontier model that
we can now have.

00:23:12.535 --> 00:23:16.635
It beats the previous, uh, model from GPT
also.

00:23:17.436 --> 00:23:19.836
GPT-6 family now is Astra, Soul, and Luna.

00:23:20.356 --> 00:23:21.435
Terra is gone.

00:23:21.876 --> 00:23:27.476
No Terra anymore. And honestly, what needs
to be renamed is Astra should be Soul,

00:23:29.476 --> 00:23:31.916
Soul should be Terra, and Luna should be

00:23:32.676 --> 00:23:33.735
Asteroid or whatever.

00:23:34.076 --> 00:23:35.156
Uh, but okay.

00:23:35.436 --> 00:23:40.336
On DeepSui, Soul, GPT-6 Soul gets 68.8,
and Luna gets

00:23:40.436 --> 00:23:41.476
66.6.

00:23:41.676 --> 00:23:47.236
And OpenAI says that Luna is roughly Opus
5 and Fable 5 at medium effort for 96%

00:23:47.516 --> 00:23:49.315
less per task. I don't know about that.

00:23:49.756 --> 00:23:53.515
I don't know, but I've been very, very
impressed with Luna specifically.

00:23:54.476 --> 00:23:57.755
Here's the sneaky big, uh, news for the
builders.

00:23:58.196 --> 00:24:00.522
Caching. 90% of cached input.

00:24:00.642 --> 00:24:03.521
All these labs understand that we are now
running

00:24:04.282 --> 00:24:10.042
long ass agentic stuff for over 500, like,
thousand tokens or so, and all of it needs

00:24:10.201 --> 00:24:14.462
caching. And, uh, changing reasoning
effort or tools doesn't break your cache

00:24:14.502 --> 00:24:18.141
anymore. This is like the big highlight
from both, I think, Anthropic and GPT, and

00:24:18.242 --> 00:24:22.612
I think that's great. You Uh, if you run
agents or assistants, this is huge.

00:24:22.972 --> 00:24:25.371
So here's my honest read on this, and I
said this in my notes.

00:24:25.532 --> 00:24:28.852
Astra has been kind of dumb for me in the
daily use, like, honestly, it's been

00:24:28.932 --> 00:24:32.892
burning tokens and it's been giving me
crappy results for the stuff that I care

00:24:32.951 --> 00:24:33.951
about.

00:24:34.332 --> 00:24:36.371
Sol is great, and Luna is even greater.

00:24:36.612 --> 00:24:40.451
Uh, Luna you can integrate into your own
products, and it's very, very, very cheap

00:24:40.492 --> 00:24:43.331
and also fast. These ones I- I'd actually
now use.

00:24:44.492 --> 00:24:46.692
Uh, Reddit isn't fully sold on this.

00:24:47.132 --> 00:24:50.832
Uh, our Kotex crowd says that the
performance gain on this is marginal, and

00:24:50.872 --> 00:24:52.012
the real news is the price.

00:24:52.652 --> 00:24:56.812
And couple of threads are showing Sol
scoring below GPT 5.6 on DeepSui.

00:24:57.452 --> 00:24:58.452
Uh,

00:24:58.732 --> 00:24:59.772
all right. Folks,

00:25:01.292 --> 00:25:04.292
is half the price with same score a win,
or did OpenAI just

00:25:05.092 --> 00:25:06.172
lose the race a little bit?

00:25:06.372 --> 00:25:08.207
Let's go to, let's go to Wolfram.

00:25:08.272 --> 00:25:10.151
I know you posted some stuff about Astra.

00:25:10.272 --> 00:25:11.272
Wolfram, let's go.

00:25:12.162 --> 00:25:15.601
just quickly, because I, I switched to,
uh, GPT

00:25:16.242 --> 00:25:21.661
6, uh, Soul because I was using 5.6 more
because Astra, it's just eating tokens too

00:25:21.722 --> 00:25:26.601
much. And I noticed in my personal
assistant that the humor has been gone and

00:25:26.882 --> 00:25:30.601
the soul, the writing, it was different,
and I didn't like it, and I've seen people

00:25:30.682 --> 00:25:31.962
online say the same thing.

00:25:32.482 --> 00:25:36.801
And, um, 5.6 Soul remains my favorite
model from OpenAI.

00:25:36.842 --> 00:25:41.921
So that happened with GPT-5 when it came
out, it was so robotic in a way, and I

00:25:42.122 --> 00:25:43.241
feel the same with this one.

00:25:43.282 --> 00:25:46.801
So actually, I would rather stay with 5.6

00:25:47.482 --> 00:25:50.262
or 5.5. Those two, those were the big ones
for me.

00:25:50.762 --> 00:25:54.402
It feels like the 4.0 keep 4.0 movement
now and say keep 5.6, please.

00:25:54.961 --> 00:25:56.622
Um, yeah, unfortunately.

00:25:56.682 --> 00:25:59.321
Or maybe I do have to go to Opus.

00:26:00.722 --> 00:26:02.621
You should, you should, uh, you should try
Opus.

00:26:02.782 --> 00:26:03.782
Um,

00:26:03.782 --> 00:26:04.782
Peter,

00:26:04.962 --> 00:26:06.101
you kicked off this category.

00:26:06.122 --> 00:26:07.981
You've been playing with obviously Astro.

00:26:08.042 --> 00:26:09.042
You're like, you're super fame.

00:26:09.081 --> 00:26:13.141
I think, dude, I think your level of fame
jumped because of your Astro demos, uh,

00:26:13.242 --> 00:26:14.882
and now we got other GPT-6.

00:26:15.081 --> 00:26:18.762
So now the GPT family of models, GPT-6
family of models is complete.

00:26:19.482 --> 00:26:20.482
And

00:26:20.522 --> 00:26:24.562
may I just say, just before we get to you,
that OpenAI's Dev Day is next week, folks.

00:26:25.002 --> 00:26:28.321
And the fact that they released their
agent platform, which is directly for

00:26:28.522 --> 00:26:34.242
developers, and then they released GPT-6
Soul and GPT-6 Luna, and Luna is

00:26:34.362 --> 00:26:39.421
priced for developers just before Dev Day,
means what the hell is coming up on Dev

00:26:39.442 --> 00:26:41.781
Day? What, what are they going to not?

00:26:41.802 --> 00:26:45.821
I, I, I think I have an idea of like one
thing, but just to highlight that we'll be

00:26:45.922 --> 00:26:48.422
covering Dev Day. Peter's gonna be there,
I'm gonna be there, and we're gonna talk

00:26:48.442 --> 00:26:51.901
to like a bunch of folks, and we'll bring
that all to you next week.

00:26:51.922 --> 00:26:54.401
But Peter, about GPT-6 stolen Luna,
please.

00:26:54.442 --> 00:26:55.641
What, what was your experience?

00:26:56.081 --> 00:27:00.067
Yeah, I think it's, they kind of went
opposite ways, I think, with, with

00:27:00.247 --> 00:27:05.347
Anthropic, and, the definitely the
important part there is efficiency, with,

00:27:05.757 --> 00:27:07.157
obviously with price reductions.

00:27:07.277 --> 00:27:11.076
I think that's the kind of the obvious
point, and 50% is a lot,

00:27:11.677 --> 00:27:16.477
right? the, and the second part there is
that it's genuinely using fewer tokens.

00:27:17.037 --> 00:27:21.297
Right, we, I looked at our data a little
bit yesterday, so no- nothing conclusive

00:27:21.337 --> 00:27:24.157
yet. We need a bit more data, but it
genuinely looks like it's using fewer

00:27:24.417 --> 00:27:29.376
tokens, and it's gonna rank higher than
5.6, uh, on our leaderboards as well.

00:27:29.816 --> 00:27:31.792
So that, that is the important part there.

00:27:31.812 --> 00:27:35.112
So I'll say it's kind of iterative
improvement in terms of quality.

00:27:35.212 --> 00:27:36.691
I think it is better than 5.6.

00:27:36.772 --> 00:27:41.081
I think, uh, it is, at least for me
personally, it just feels like smoother

00:27:41.242 --> 00:27:42.232
model.

00:27:42.232 --> 00:27:45.571
and for me, it is the kind of the daily
driver model.

00:27:45.672 --> 00:27:46.672
If you just want to do

00:27:47.232 --> 00:27:50.072
normal stuff, that makes the most sense to
use.

00:27:50.232 --> 00:27:51.471
It's not going to burn your sub.

00:27:52.152 --> 00:27:57.273
And, it's, I think the problem was, and
Alex, back to your point about just

00:27:57.384 --> 00:28:00.024
burning your subs, is that we were used to

00:28:00.844 --> 00:28:04.183
pretty generous subscriptions at, medium-
sized models.

00:28:04.284 --> 00:28:07.663
They were big at the time, so like 5.6 and
Opus, and Mm-hmm.

00:28:07.844 --> 00:28:11.023
Remember Sonnet? Like, we used to use
that, that thing as well.

00:28:11.564 --> 00:28:12.284
So

00:28:12.284 --> 00:28:14.504
if you're using these models, your sub's
so fine.

00:28:14.524 --> 00:28:17.783
You forgot to say this. Sonnet and Haiku
5.5 are coming.

00:28:17.844 --> 00:28:19.803
Are they? Yeah, please go ahead. Okay,
okay, okay.

00:28:20.004 --> 00:28:24.823
Yeah. So remember we were using these
models, and then your sub was fine, and

00:28:24.844 --> 00:28:28.704
then they give us Fable and Astra, and you
want to use these models instead, and then

00:28:28.764 --> 00:28:30.564
your sub just dies in, in like

00:28:31.204 --> 00:28:35.743
3 hours. And I think it's not, and I don't
think we should be blaming the labs, like

00:28:35.844 --> 00:28:38.744
these are more expensive, bigger models
Yeah.

00:28:38.764 --> 00:28:42.383
and that they need to run, and they
genuinely, I don't think they're like

00:28:43.084 --> 00:28:45.763
lying to us. I think they just genuinely
don't have compute.

00:28:46.444 --> 00:28:49.504
So they gave us more allowance, we just
kill it quick, and they just wouldn't have

00:28:49.604 --> 00:28:53.223
compute. So OpenAI already, remember they
had to stop the signups, right?

00:28:53.284 --> 00:28:58.063
So yeah, there's no 200 dollar a month,
pro tier anymore on OpenAI.

00:28:58.264 --> 00:28:59.883
You only can get to like 100 dollars.

00:29:00.084 --> 00:29:04.663
on that, also, if we're comparing the two
releases, kind of like Opus is still

00:29:05.364 --> 00:29:08.903
the better deal. Like, you can get 200
dollars worth of, like 20x usage, and

00:29:09.024 --> 00:29:12.384
OpenAI had to stop because so many people
wanted Astra, specifically Astra is

00:29:12.464 --> 00:29:13.464
burning.

00:29:13.504 --> 00:29:18.293
Question is, my question for you, Peter,
is 6 Sol right now replacing for most

00:29:18.354 --> 00:29:20.534
people what they thought they needed Astra
for?

00:29:21.330 --> 00:29:23.649
Yeah, I think it, it's a little bit
tricky.

00:29:23.970 --> 00:29:27.824
my general rule of thumb would be is if
you have an OpenAI sub, you should just

00:29:27.965 --> 00:29:28.805
use

00:29:28.805 --> 00:29:34.224
uh, 6 Sol for everything, and then just
switch to Toggle when you need Astra, uh,

00:29:34.405 --> 00:29:36.990
when it doesn't work. it's not that
difficult, and I do that.

00:29:37.109 --> 00:29:40.530
That's kind of what I do myself, and that
feels good to me.

00:29:40.775 --> 00:29:44.954
I don't feel like it's terrible or
anything, and it, and I, my sub goes

00:29:45.175 --> 00:29:48.555
pretty long way now. Um, but yeah, I think
it was just never sustainable.

00:29:48.615 --> 00:29:53.654
I think it's just unrealistic for us to
say, oh, I want to use Fable and Astra at

00:29:53.695 --> 00:29:58.174
my 200 dollars in exactly the same way how
a model used to be, which is like quarter

00:29:58.195 --> 00:29:59.374
of a price. It's just not that.

00:29:59.455 --> 00:30:02.934
Can I say something? I don't know if it's
controversial, bro, but he- here's my

00:30:03.655 --> 00:30:05.075
take on this after playing,

00:30:05.735 --> 00:30:07.074
going very hard this week, folks.

00:30:08.095 --> 00:30:11.815
Uh, if you guys remember the, the token
billionaire thing that the AI engineer

00:30:11.855 --> 00:30:16.735
invented and the card that there's now
finger prone token billionaires, uh, I

00:30:16.815 --> 00:30:19.734
think I'm at like, like 8 billion tokens
this week.

00:30:20.015 --> 00:30:21.015
Uh,

00:30:21.295 --> 00:30:22.454
I'm working on something super cool.

00:30:22.695 --> 00:30:25.175
However, I've tried all the models and all
the

00:30:26.015 --> 00:30:29.874
harnesses as well. So I've tried Cursor,
Devin, directly Codex, directly Claude

00:30:29.975 --> 00:30:31.895
Code. I have my thoughts on that.

00:30:33.575 --> 00:30:34.575
And here's my take.

00:30:35.655 --> 00:30:37.654
I personally, for my work,

00:30:38.335 --> 00:30:42.535
I don't want an extra powered model that
can solve Navier-Stokes equations.

00:30:43.295 --> 00:30:47.375
I want a model that can understand me and
build me a simple fucking settings page

00:30:47.975 --> 00:30:50.715
without going over engineering, writing a
thousand tests.

00:30:51.055 --> 00:30:52.435
That's what I need right now.

00:30:53.215 --> 00:30:56.534
And it feels like what we just got with
Opus

00:30:57.175 --> 00:30:59.874
and Sol, so like the, the, the one

00:31:00.655 --> 00:31:02.955
rank lower models that also are

00:31:04.335 --> 00:31:07.215
sparing your quota, is quite that.

00:31:08.294 --> 00:31:11.334
Opus, in my case, is significantly better
at understanding me and doing some stuff

00:31:11.495 --> 00:31:16.175
exactly like I like, and and Sol still
needs, like, very direct instruction and

00:31:16.515 --> 00:31:18.294
is very good at, like, computer use.

00:31:18.563 --> 00:31:21.683
for my work, I need a reliable, fast
assistant

00:31:22.484 --> 00:31:25.204
model that can understand me better and
run fast.

00:31:25.524 --> 00:31:26.763
I don't need Navier-Stokes.

00:31:27.324 --> 00:31:30.163
So it's a very interesting thing that that
we started the show with.

00:31:30.364 --> 00:31:34.384
We were pacing the frontier after last
week Dario Amodei and Sam Altman agreeing,

00:31:34.444 --> 00:31:37.664
Elon Musk, whatever, and we're still
getting a lot of

00:31:38.364 --> 00:31:43.944
very incredible advancements in practical,
everyday AI brains, and to me, that's

00:31:44.004 --> 00:31:45.004
great.

00:31:45.764 --> 00:31:47.574
Uh, Nisten, we haven't heard from you on
the GPT stuff, CoreWeave.

00:31:47.594 --> 00:31:48.873
Thank you, Wolfram, for hopping on.

00:31:49.274 --> 00:31:53.353
And uh, taking Wolfram's praise, we have
back on the show Yam Peleg.

00:31:53.514 --> 00:31:57.053
Welcome, Yam. We are now talking about
GPT-6 Soul and

00:31:57.874 --> 00:32:00.034
Luna, and uh, Nisten is up.

00:32:00.914 --> 00:32:06.413
I use Astra, but mainly I just let uh,
Claude Fable manage it, what it does.

00:32:06.874 --> 00:32:09.933
It's just for the work that I do, and I'm
pretty opinionated.

00:32:10.193 --> 00:32:13.154
I really don't like the taste it has or

00:32:13.834 --> 00:32:16.154
how many different files it, it creates.

00:32:16.294 --> 00:32:20.492
So you use, you use Anthropic's taste and
then OpenAI's execution?

00:32:21.094 --> 00:32:26.854
Uh, I mainly just use OpenAI for,
restyling, because it does produce

00:32:27.114 --> 00:32:28.114
much better

00:32:28.274 --> 00:32:30.354
websites in terms of design now,

00:32:31.114 --> 00:32:36.714
and I use it for reviews, but overall I
don't like having it

00:32:37.194 --> 00:32:39.393
make a whole bunch of files in my repo.

00:32:39.533 --> 00:32:43.342
I, I just don't like the style it develops
at.

00:32:44.573 --> 00:32:46.633
and I'm a, uh, but I've been very used to.

00:32:46.733 --> 00:32:50.193
Nisten, you remember we're talking about
GPT-6, Sol, and Luna, right?

00:32:50.292 --> 00:32:54.032
Like, Astra is, is a model that, uh, have
you tried the new ones?

00:32:54.133 --> 00:32:55.832
Have you tried the Uh, no. the smaller
brothers?

00:32:55.973 --> 00:32:58.253
No, I just haven't, have not used those.

00:32:58.613 --> 00:33:02.113
and I would love to know why, because I
think, first of all, a lot of folks love

00:33:02.192 --> 00:33:03.492
your takes here on the show because
they're

00:33:04.142 --> 00:33:05.982
not always like going with the hype, which
is great.

00:33:06.462 --> 00:33:09.273
Uh, it's cheaper, it's faster, 68%.

00:33:10.794 --> 00:33:11.893
Why, why didn't you?

00:33:11.994 --> 00:33:15.013
Just because you already have your
workflows, or do you think that your

00:33:15.114 --> 00:33:18.433
prompts will not align with the new
models, or just like too many news in the

00:33:18.474 --> 00:33:21.633
span of a short week? I mean, it's just
for what I do.

00:33:21.674 --> 00:33:25.093
I know the majority of people do prefer it
over Opus.

00:33:25.674 --> 00:33:27.233
It's just for what I'm doing.

00:33:27.594 --> 00:33:28.594
One,

00:33:28.594 --> 00:33:34.234
even Fable and Opus 5.5 are not good
enough at DevOps, and I find Astro worse.

00:33:35.154 --> 00:33:37.253
The only thing I find it better is that

00:33:38.074 --> 00:33:39.074
designing stuff,

00:33:39.874 --> 00:33:45.253
and uh, if I'm doing kernel work, it just
makes way too many files, and it just

00:33:45.394 --> 00:33:47.673
feels, it feels dumber to me.

00:33:48.074 --> 00:33:49.833
So I just use it as a second opinion.

00:33:50.434 --> 00:33:54.693
But then again, I've never been a fan of
how ChatGPT models

00:33:55.314 --> 00:33:56.314
respond,

00:33:56.594 --> 00:33:57.594
and uh,

00:33:58.234 --> 00:33:59.234
even

00:33:59.754 --> 00:34:04.623
uh, when I've been working with people, I
could tell right away when it was like a

00:34:04.744 --> 00:34:08.534
ChatGPT written report, and uh, like I
absolutely hate it.

00:34:09.214 --> 00:34:10.413
I hate the way it writes.

00:34:10.734 --> 00:34:12.133
Sorry, it's just me.

00:34:12.614 --> 00:34:13.973
Uh, not the best writer, I agree.

00:34:14.214 --> 00:34:17.334
Yam Peleg, uh, welcome to the show.

00:34:18.613 --> 00:34:22.693
So. GPT-6 Soul, GPT-6 Luna, no Terra
anymore.

00:34:23.014 --> 00:34:24.853
What's your take? Have you used those
models?

00:34:25.014 --> 00:34:26.353
How are you feeling about them?

00:34:26.514 --> 00:34:28.074
which one of you is in what camp?

00:34:28.354 --> 00:34:29.513
Who is on, on Opus?

00:34:29.594 --> 00:34:30.613
Who is in GPT-4?

00:34:30.634 --> 00:34:35.373
I don't want to take sides. I want to run
both, but I will say Opus is doing just

00:34:35.514 --> 00:34:36.514
incredible stuff,

00:34:37.394 --> 00:34:42.993
and I definitely, definitely have used the
trick that I will now teach you.

00:34:43.354 --> 00:34:47.533
Uh, Weights and Biases Hive Mind is a tool
that stores all your conversations from

00:34:47.634 --> 00:34:49.554
all your agents across everything in the
cloud.

00:34:50.114 --> 00:34:54.834
It has a very nice feature fork, so you
can like, you can download another's agent

00:34:54.874 --> 00:34:56.313
conversation regardless of which machine.

00:34:56.834 --> 00:34:58.073
I have definitely used

00:34:58.714 --> 00:35:03.013
fork my Codex conversation into Claude and
say, hey, review what he did

00:35:04.114 --> 00:35:05.114
and tell me

00:35:05.354 --> 00:35:06.833
what you d- you would do differently.

00:35:07.314 --> 00:35:11.054
And that's been my, like, this is the
direction of the stuff that I use.

00:35:11.114 --> 00:35:13.053
I haven't used the other way around.

00:35:13.074 --> 00:35:14.373
So I haven't ever used

00:35:15.074 --> 00:35:18.893
OpenAI's stuff to go and review Claude,
but I definitely use Claude to go and

00:35:18.954 --> 00:35:23.233
review Codex. That that's where I fall,
like true- That thing, that thing sounds

00:35:23.394 --> 00:35:24.674
useful as hell.

00:35:25.199 --> 00:35:27.634
yeah, it is very useful. Yam, where are
where do you fall?

00:35:28.394 --> 00:35:29.394
Look, uh...

00:35:31.034 --> 00:35:36.954
Uh, let me just say directly, I wish I
could use Soul enough to give you

00:35:37.014 --> 00:35:38.033
an opinion. Mm-hmm.

00:35:38.254 --> 00:35:39.254
It's just

00:35:39.294 --> 00:35:42.074
it burns the entire account really quick.

00:35:42.614 --> 00:35:43.974
So does? Yeah.

00:35:44.614 --> 00:35:46.394
And at least for me, at least for me.

00:35:46.454 --> 00:35:50.274
So, so here's, here's the thing that I,
uh, I complained about this on my timeline

00:35:50.294 --> 00:35:53.634
a lot. I don't know if you guys saw, uh,
Astro just like decimated my limits.

00:35:53.813 --> 00:35:56.054
And then, uh, helpful folks from OpenAI.

00:35:56.214 --> 00:35:59.434
Shout out, shout out to incredible Deverel
folks, by the way, on OpenAI.

00:35:59.533 --> 00:36:02.214
Yeah. Uh, Dominic Kundul, a friend of the
pod, was on the show.

00:36:02.604 --> 00:36:03.604
Eric Povancher,

00:36:05.024 --> 00:36:06.848
Anyway, he went to OpenAI to work on on
the stuff.

00:36:06.909 --> 00:36:09.168
He's very, like, adamant about, no, folks,
you don't understand.

00:36:09.229 --> 00:36:10.509
This is a a skill issue.

00:36:10.629 --> 00:36:11.989
Sol and Astro are amazing models.

00:36:12.348 --> 00:36:13.348
Eric Povancher,

00:36:13.709 --> 00:36:14.709
um, he

00:36:15.348 --> 00:36:19.709
was also replying to people, and then I
noticed that I had some settings in Codex

00:36:19.749 --> 00:36:21.469
that fucked up with my whole

00:36:22.669 --> 00:36:23.629
setup.

00:36:23.629 --> 00:36:24.929
Which setting? Here's the setting that I
had.

00:36:25.309 --> 00:36:29.128
I, a long time ago, played around with the
context window, and I set mine to like a

00:36:29.229 --> 00:36:32.348
million context window, which, uh, OpenAI
models support.

00:36:32.749 --> 00:36:37.909
This absolutely kills their cache system,
and so this like rips through your tokens.

00:36:37.949 --> 00:36:41.748
I also had experimental settings for
having it keep it notes for itself, and I

00:36:41.829 --> 00:36:44.509
also played around with how many sub-
agents it opens.

00:36:45.309 --> 00:36:49.908
After removing all of those and asking the
model to actually open the sub-agents in a

00:36:50.069 --> 00:36:54.029
lower tier model, I actually started
seeing normal use back in Codex again.

00:36:54.109 --> 00:36:58.769
So if this was your issue and you felt
complete burning of tokens, see if your

00:36:59.149 --> 00:37:04.068
Codex.toml or settings.toml file is, uh,
carrying some older stuff.

00:37:05.309 --> 00:37:06.669
Uh, all right, all right.

00:37:06.774 --> 00:37:09.674
I must admit I need to check, but I'm
playing with the

00:37:10.275 --> 00:37:14.634
codex.toml file all the time, by the way,
it's like all the time tweaking things.

00:37:14.675 --> 00:37:18.754
There are so many hidden, hidden stuff
that are like you should, you should

00:37:19.075 --> 00:37:21.035
tinker with. Uh, just want to say,

00:37:21.795 --> 00:37:22.795
look,

00:37:23.075 --> 00:37:28.995
a sol that is more efficient, like 5.6 sol
that is more efficient, even

00:37:29.075 --> 00:37:30.315
if it wasn't even better,

00:37:31.035 --> 00:37:32.634
I'm taking this seriously.

00:37:32.814 --> 00:37:33.874
it's a really good model.

00:37:33.975 --> 00:37:36.575
It's just that it eats the account really
quick.

00:37:37.295 --> 00:37:42.755
And just on the contrary, I just want to
say, Opus ever since

00:37:43.365 --> 00:37:49.344
4.6, I had something to say about each of
the models, but the

00:37:49.445 --> 00:37:52.805
new one, 5.5, is just good.

00:37:53.535 --> 00:37:54.535
hands down.

00:37:55.055 --> 00:37:56.254
There there's no denying.

00:37:57.335 --> 00:38:01.914
I'm really confused about your burning the
sub point because, uh, in our, like, in my

00:38:01.975 --> 00:38:04.214
personal use, I don't think it's burning
my sub anymore.

00:38:04.255 --> 00:38:08.571
And also on, in our data, we can see it
costing about half, so it's like more

00:38:08.691 --> 00:38:10.211
token efficient and cheaper.

00:38:10.531 --> 00:38:12.550
So I don't see why I should be burning it
more.

00:38:12.631 --> 00:38:17.370
So I would also say this is such a common
pattern for me, whether it's Claude Code

00:38:17.431 --> 00:38:18.910
or Codex or whatever else.

00:38:19.591 --> 00:38:23.371
I find something super annoying,
something's not working, and

00:38:24.031 --> 00:38:27.051
9 times out of 10, it's some stupid
setting somewhere, some Oh yeah.

00:38:27.071 --> 00:38:29.118
plugin is broken or something else.

00:38:29.518 --> 00:38:33.398
So it's good idea just as a general tip
for everyone, like if something is

00:38:33.478 --> 00:38:35.597
annoying you about something, just ask
your

00:38:36.238 --> 00:38:38.859
agent to see what's going on and diagnose
it and fix it.

00:38:39.000 --> 00:38:39.880
Just

00:38:39.880 --> 00:38:41.919
I come across something like that every
week.

00:38:42.280 --> 00:38:45.290
They should, everyone should have like a
heal button or something, just like Yeah.

00:38:45.870 --> 00:38:48.689
Let them run through it and like reset
your settings or something.

00:38:48.850 --> 00:38:49.630
Just

00:38:49.630 --> 00:38:51.469
way too common. I will say, though,

00:38:52.230 --> 00:38:55.469
both Astra and the new models, I've
noticed this as well,

00:38:56.110 --> 00:38:57.909
suffer from the pause syndrome.

00:38:58.630 --> 00:38:59.870
Can we talk about the pause syndrome?

00:39:00.030 --> 00:39:05.005
The pause syndrome is where, you talk to
the model, you ask it for like a long

00:39:05.146 --> 00:39:07.385
thing to do, and then you steer,

00:39:07.986 --> 00:39:12.086
and then you interject with your thoughts
in the middle because OpenAI came up with

00:39:12.126 --> 00:39:13.126
this amazing

00:39:13.466 --> 00:39:17.306
steering concept where, instead of
stopping the model and inject your

00:39:17.506 --> 00:39:19.105
message, you basically steer.

00:39:19.306 --> 00:39:21.206
I think Anthropic got there eventually as
well.

00:39:21.546 --> 00:39:27.526
Oh yeah. Apparently, for the Astra and Sol
models, steering too much after compaction

00:39:27.866 --> 00:39:32.626
causes it to lose the original direction,
and the model just stops working.

00:39:32.985 --> 00:39:36.325
It will say something, you would steer and
ask it a question, like a by the way

00:39:36.425 --> 00:39:38.806
question, the model will answer and stop.

00:39:38.985 --> 00:39:43.275
This pattern has happened to me and many
folks on my timeline, and what Eric, f-

00:39:43.335 --> 00:39:47.322
from OpenAI, Eric suggests is to ask the
model to write down and jot down some

00:39:47.403 --> 00:39:50.703
messages for you when you steer and
continue working, because

00:39:51.443 --> 00:39:52.842
the steering is the problem.

00:39:53.643 --> 00:39:57.003
OpenAI needs to fix this in their own,
harnesses, et cetera, but, steering is

00:39:57.423 --> 00:40:03.003
absolutely the problem. So, uh, if you
steer a lot and you get Astra pause

00:40:03.063 --> 00:40:04.063
syndrome,

00:40:04.343 --> 00:40:05.622
they're not pausing the frontier,

00:40:06.223 --> 00:40:08.098
they just, the model is confused after
compaction.

00:40:08.178 --> 00:40:10.279
So this is a tip for you from the folks at
OpenAI.

00:40:10.520 --> 00:40:14.200
Uh, these definitely are some issues that,
h- have plagued my timeline as well.

00:40:15.080 --> 00:40:16.839
Folks, I think it's time for us

00:40:18.119 --> 00:40:21.320
to... Anything else that's missing that we
haven't said?

00:40:21.719 --> 00:40:22.920
the only thing I will repeat,

00:40:23.800 --> 00:40:28.240
OpenAI released the agents platform,
managed agents platform that runs inside

00:40:28.320 --> 00:40:31.319
your stuff last week, which is incredible
and sounds like a

00:40:32.000 --> 00:40:34.060
this should be the tool that they would
release on Dev Day.

00:40:34.240 --> 00:40:35.919
They now release 2 new models,

00:40:37.279 --> 00:40:38.800
and Dev Day is upcoming next week.

00:40:40.440 --> 00:40:44.279
Uh, speaking of the Hive Mind tool,
producer put it up for us.

00:40:44.380 --> 00:40:48.779
Thank you, producer. Every agent
conversation and forkable, and it's free

00:40:49.060 --> 00:40:51.059
to use hivemind.1b.tools.

00:40:51.340 --> 00:40:55.120
This is the brainchild of Chris von Pelt,
one of the co-founders of Weights and

00:40:55.140 --> 00:40:58.779
Biases, because he used to work across
many computers, across many harnesses.

00:40:59.020 --> 00:41:04.969
It's really damn useful for you if you run
multiple, harnesses over multiple,

00:41:05.330 --> 00:41:06.330
computers.

00:41:06.690 --> 00:41:07.690
Right, folks.

00:41:07.865 --> 00:41:09.170
for, I think it's time for MetaConnect.

00:41:13.036 --> 00:41:14.036
Meta AI

00:41:14.156 --> 00:41:15.156
News app

00:41:15.636 --> 00:41:19.735
just passed ChatGPT in the number of
people, uh, getting it.

00:41:19.835 --> 00:41:23.595
It's number one in the App Store, and at
Meta Connect they gave it an email

00:41:23.636 --> 00:41:27.356
address, a control of your Mac, and a
place on your face.

00:41:28.116 --> 00:41:29.856
Meta Connect happened yesterday, and

00:41:30.516 --> 00:41:33.615
I did my bingo card on Twitter, and I got
some of the stuff that I wanted for the

00:41:33.676 --> 00:41:34.676
bingo card.

00:41:34.956 --> 00:41:35.956
The...

00:41:36.836 --> 00:41:38.436
the headline is Muse.

00:41:38.756 --> 00:41:43.416
Zuckerberg was up on stage and said that
the central piece to everything they're

00:41:43.476 --> 00:41:45.236
building at Meta is Muse.

00:41:45.836 --> 00:41:49.835
Muse is Meta's personal AI assistant, and
it's number one downloaded app in the App

00:41:49.876 --> 00:41:53.916
Store right now, which a year ago, if you
told me that Meta AI would be outranked

00:41:53.956 --> 00:41:58.075
ChatGPT, I would probably, I would believe
you, as we told you, don't bet against the

00:41:58.116 --> 00:41:59.955
Zuck, but many people would, would have
not.

00:42:00.436 --> 00:42:01.706
as a reminder, last year,

00:42:02.532 --> 00:42:06.612
Zuck started hiring with some soup all of
these incredible developers and started

00:42:06.652 --> 00:42:09.877
bringing them from different labs, and the
soup paid off.

00:42:10.918 --> 00:42:11.918
So,

00:42:13.238 --> 00:42:17.218
now Muse as an assistant that sits on your
Mac, it can control your Mac, they're

00:42:17.278 --> 00:42:21.378
giving it its own email address so you can
CC conversations and real-time voice and

00:42:21.438 --> 00:42:24.518
video conversations. So I assume by this
time next month,

00:42:25.398 --> 00:42:29.237
Muse, my Muse specifically, could join the
show as a participant on the show.

00:42:30.718 --> 00:42:32.318
It works in background while you're
talking to it.

00:42:32.358 --> 00:42:36.258
This is the what separates assistants from
agents, and there is an avatar you can

00:42:36.318 --> 00:42:38.838
talk and voice and face design yourself.

00:42:38.918 --> 00:42:41.998
You can customize your own muse and voice
design it yourself.

00:42:42.358 --> 00:42:46.408
And Muse is coming to Meta glasses, you
will be able to walk around in the world

00:42:46.768 --> 00:42:48.388
and chat with your conversation.

00:42:48.808 --> 00:42:51.587
And the custom wake word is the highlight
feature for me.

00:42:51.768 --> 00:42:54.528
You'll be able to say, hey, Wilfred,
instead of hey, Meta.

00:42:55.127 --> 00:42:58.448
I think that's super cool. I think if
they're training custom wake word for me

00:42:58.528 --> 00:43:00.408
in my Ray-Bans, I'm gonna walk around in
the world.

00:43:00.768 --> 00:43:05.047
This is like us having our own, uh, you
know, Tony Stark assistant.

00:43:06.208 --> 00:43:09.863
Then there's the hardware, and there's a
lot of hardware, Ray-Ban Meta Gen 3 with

00:43:09.943 --> 00:43:14.758
new 6 microphones and aviary styles, The
first audio-only glasses that people

00:43:14.838 --> 00:43:18.978
cannot see the cameras anymore, and not
only that, the audio glasses, the video

00:43:19.058 --> 00:43:21.718
glasses too, they work with the FDA to
make them,

00:43:22.498 --> 00:43:23.977
like, certified hearing aid,

00:43:24.658 --> 00:43:27.677
which is, I think, is a genuinely great
move

00:43:28.298 --> 00:43:30.898
because of the ADA Act, the

00:43:32.218 --> 00:43:33.858
Disabilities Act in the United States.

00:43:34.378 --> 00:43:38.097
if you're disabled and you wear meta
glasses as hearing aids, you can sue

00:43:38.178 --> 00:43:40.537
people who tell you to turn them off
because of the ADA.

00:43:40.778 --> 00:43:43.587
I think this is like a genius play by
Zuckerberg that's not like talked about.

00:43:44.268 --> 00:43:48.428
Uh, so audio only glasses with no video,
uh, and the lightest frame they've made.

00:43:48.508 --> 00:43:50.548
They kind of, Peter, they kind of look
like your glasses, honestly.

00:43:50.588 --> 00:43:54.228
They just look like there's no, no
indication that this is like listening to

00:43:54.268 --> 00:43:55.427
you. Um,

00:43:56.308 --> 00:43:58.247
and they cost a fraction of normal hearing
aid.

00:43:58.388 --> 00:44:02.138
And honestly, this is could be the most
important thing that they announced, and

00:44:02.298 --> 00:44:05.617
VR glasses. Meta announced a VR headset.

00:44:05.978 --> 00:44:08.578
That's not a headset. This is the first
time they look like glasses.

00:44:08.658 --> 00:44:09.938
In fact, the the big reveal

00:44:10.818 --> 00:44:14.538
for which many people, compare Zuckerberg
now to Steve Jobs, is that the

00:44:15.338 --> 00:44:19.618
VR glasses that they've announced have
been on the table with other Meta glasses,

00:44:19.698 --> 00:44:20.898
and people didn't even notice.

00:44:20.988 --> 00:44:21.988
It's quite incredible.

00:44:22.628 --> 00:44:23.628
And then

00:44:23.868 --> 00:44:25.347
I think we're getting to a point where

00:44:26.348 --> 00:44:29.747
Meta is beating Apple and OpenAI's
whatever rumored hardware.

00:44:30.268 --> 00:44:34.887
They announced this little Tamagotchi-like
device that your muse is sitting in a

00:44:35.028 --> 00:44:40.027
keychain-like device with a screen at the
size of the Apple Watch, pretty much, with

00:44:40.068 --> 00:44:41.138
a thumb fingerprint

00:44:42.738 --> 00:44:43.738
that is

00:44:44.458 --> 00:44:47.777
shipping in December by the time for the
holidays, and no price yet.

00:44:47.938 --> 00:44:49.578
I think that this is going to be

00:44:50.538 --> 00:44:55.138
sold out the second it drops, as people
think about which gifts to give other

00:44:55.178 --> 00:44:56.178
people.

00:44:56.588 --> 00:44:58.467
Peter, have you tried Muse already?

00:44:59.093 --> 00:45:00.652
Is it available in the UK?

00:45:00.813 --> 00:45:03.173
No. Mm-mm. I haven't tried it yet.

00:45:03.253 --> 00:45:04.652
It's coming, it's coming. I think

00:45:05.313 --> 00:45:06.313
what's been

00:45:06.713 --> 00:45:10.973
missing, and like, if you remember the
trajectory, right, is that OpenAI went

00:45:11.073 --> 00:45:12.073
consumer,

00:45:12.313 --> 00:45:16.952
then sort of competing with Google, then
OpenAI realized, oh damn it, Anthropic is

00:45:17.033 --> 00:45:19.513
eating our lunch, so we're gonna go more
enterprise.

00:45:19.673 --> 00:45:22.712
It's not like they've forgotten about
consumer, they're gonna release devices as

00:45:22.793 --> 00:45:27.038
well, but we certainly, and and Google's
been like okay at consumer stuff.

00:45:27.073 --> 00:45:32.632
So we've been missing a, a lab that is
very, very good, very focused, and has a

00:45:32.753 --> 00:45:34.072
genuine consumer interest.

00:45:34.713 --> 00:45:36.513
And uh, yeah, Meta is it.

00:45:36.633 --> 00:45:39.832
So it looks like they're really just
going, going

00:45:40.673 --> 00:45:43.052
all in on making consumer great.

00:45:43.273 --> 00:45:44.193
And

00:45:44.193 --> 00:45:45.243
like Nat Friedman, right?

00:45:45.303 --> 00:45:47.083
The, he understands this a lot.

00:45:47.103 --> 00:45:49.007
He's one of the key people behind it.

00:45:49.328 --> 00:45:50.988
And it's making things really fast.

00:45:51.007 --> 00:45:52.007
Nat Friedman was the

00:45:52.408 --> 00:45:53.468
CEO, so sorry to interrupt.

00:45:53.648 --> 00:45:55.028
Nat Friedman, just Friedman was the CEO of
GitHub.

00:45:55.548 --> 00:46:00.057
And then Nat Friedman and Daniel, Gross
became one of the more prolific AI

00:46:00.177 --> 00:46:03.417
investors in the Valley a long time before
the AI wave started.

00:46:03.538 --> 00:46:07.207
And then the biggest news about the Meta
Superintelligence Labs is that, Zach was

00:46:07.248 --> 00:46:08.308
able to bring both of them on.

00:46:09.268 --> 00:46:10.268
Specifically

00:46:10.668 --> 00:46:15.008
getting Daniel Gross away from Ilya
Sutskever at SSI into the MSL.

00:46:15.028 --> 00:46:16.628
I think this is like a big, big, big, big
move.

00:46:17.147 --> 00:46:22.638
And Nat Friedman is behind Muse, and Alex
Wang is showing up with Crocs and a

00:46:23.158 --> 00:46:27.777
crocodile thrifted t-shirt on the stage of
Meta Connect, and talks about how Nat

00:46:27.858 --> 00:46:29.767
Friedman, worked his ass off on this Muse
thing.

00:46:30.098 --> 00:46:31.537
Nat Friedman doesn't talk on Twitter too
much.

00:46:32.018 --> 00:46:34.298
He said that once OpenClaw dropped,

00:46:35.098 --> 00:46:39.417
they all stopped everything they're doing,
and he brought 400 Mac Minis into the MSL

00:46:39.538 --> 00:46:43.497
lab to all run OpenClaws all the time to
figure out what's going on.

00:46:44.098 --> 00:46:47.618
Just so you understand how early he was,
this was in, I think he said February,

00:46:47.778 --> 00:46:50.217
January. That's when we all bought our Mac
Minis.

00:46:50.738 --> 00:46:52.858
Meta has a top level exec

00:46:53.858 --> 00:46:56.837
that now drives the ideation thing

00:46:57.978 --> 00:47:01.733
as early as when we tell you about stuff
before it blows up on the internet.

00:47:01.733 --> 00:47:03.527
That- that's how they got to Muse now.

00:47:04.098 --> 00:47:05.498
The security of Muse

00:47:06.178 --> 00:47:08.297
is they're going for unparalleled.

00:47:08.378 --> 00:47:12.337
They understand that I'm going to put more
and more trust in this device,

00:47:13.018 --> 00:47:16.637
and unlike the folks who ran OpenClaw on
their Mac Minis, et cetera, and that

00:47:16.698 --> 00:47:19.863
didn't really care about, those are folks
who are putting their, like,

00:47:21.329 --> 00:47:24.868
bank account information, integration with
1Password, et cetera.

00:47:25.389 --> 00:47:30.255
This all needs to, jump over the fear that
folks had that Meta has all my data for

00:47:30.415 --> 00:47:33.014
ads. So they're talking about a few things
here.

00:47:33.175 --> 00:47:34.175
One is

00:47:34.575 --> 00:47:35.774
the business model is new.

00:47:35.895 --> 00:47:37.624
There's no more ads, no more your, your
data.

00:47:37.785 --> 00:47:40.145
The business model is transactions fee.

00:47:40.945 --> 00:47:44.705
Muse and other assistants are going to do
so much for people shopping on the

00:47:44.785 --> 00:47:48.684
internet, which, by the way, they
announced integrations with Best Buy and

00:47:48.985 --> 00:47:50.404
Walmart and other folks

00:47:51.385 --> 00:47:54.545
at the same way that Amazon has banned
Muse from shopping.

00:47:54.585 --> 00:47:57.504
So if you use your Muse for shopping in
Amazon, if you use a computer browser,

00:47:57.985 --> 00:48:02.075
Amazon will ban your account, which think,
is a ridiculously stupid move, but I think

00:48:02.114 --> 00:48:04.904
it's just muscles, like they're, they're
still negotiating probably the, the cut

00:48:05.224 --> 00:48:10.185
with Zuck. Um, and this, these assistants
will do a lot of shopping for people, and

00:48:10.345 --> 00:48:12.005
Zuck is like, okay, this is the business
model.

00:48:12.145 --> 00:48:14.704
No longer do we sell your data or train on
your data.

00:48:15.185 --> 00:48:19.744
We take a small cut from the seller's fee
to you if you use Muse to buy stuff.

00:48:20.185 --> 00:48:21.224
I think it's a genius model.

00:48:21.865 --> 00:48:23.105
It's a win-win for everyone.

00:48:25.065 --> 00:48:28.464
I think, yeah, it's the problem is it's
hard to make money that way.

00:48:29.105 --> 00:48:34.985
ChatGPT backed, uh, backed away from this
model just because it's the, we can go

00:48:35.145 --> 00:48:37.384
deeper on this, but basically the way

00:48:37.985 --> 00:48:42.705
you make a lot of money is advertising
because it's, uh, you know, you can kind

00:48:42.745 --> 00:48:47.025
of raise prices on your suppliers without
consumers paying more for it.

00:48:47.065 --> 00:48:49.805
But look, I think- I don't think that
Facebook is gonna stop advertising

00:48:49.865 --> 00:48:53.274
elsewhere. I just think that, for muse to
give it out for free, the business model

00:48:53.315 --> 00:48:57.975
that they're advertising for it is that,
hey, we're gonna have a secure VM, which

00:48:58.155 --> 00:49:03.915
was co-architected with Moxie Marlinspike,
the creator of Signal.

00:49:04.235 --> 00:49:08.234
That secure VM is at the level of a
cryptographically verifiable,

00:49:09.725 --> 00:49:13.444
solution that even Zuckerberg himself
would not be able to read your data

00:49:13.525 --> 00:49:14.624
because it's end-to-end encrypted.

00:49:14.685 --> 00:49:16.405
That's what they're touting live on stage.

00:49:16.444 --> 00:49:19.965
So they are going after, hey, we know
about privacy, we know what people think

00:49:19.985 --> 00:49:21.794
about Facebook, et cetera. H- here is

00:49:22.715 --> 00:49:27.065
industry level, architects that say that I
cannot read your data anymore, but I can,

00:49:27.105 --> 00:49:30.025
like, look out at your transactions, and I
can take a fee from your connectors.

00:49:30.105 --> 00:49:31.245
That's kind of the, the business model.

00:49:31.345 --> 00:49:33.504
Yeah. I think, anyway, we'll see.

00:49:33.665 --> 00:49:37.105
I'm, I'm a little skeptical, and I think
they're probably also doing it

00:49:37.305 --> 00:49:42.444
unnecessarily. You know, our parents don't
give a crap about that, and they don't

00:49:42.545 --> 00:49:45.785
watch these keynotes, and they're not
gonna know what that, any of that means.

00:49:46.025 --> 00:49:48.904
So I think that's my, maybe over-egging it
anyway.

00:49:48.945 --> 00:49:51.575
You, Mm-hmm. But I think what, what I
would like to add is that

00:49:52.415 --> 00:49:53.894
what's really impressing me about

00:49:54.495 --> 00:49:56.674
how they're approaching it is that I think

00:49:58.055 --> 00:49:59.055
it's the

00:49:59.135 --> 00:50:02.054
only, like, really big consumer app,

00:50:03.055 --> 00:50:05.855
uh, the where they don't treat people like
idiots.

00:50:06.135 --> 00:50:07.815
they give the powerful tools.

00:50:08.015 --> 00:50:09.375
It's all the agentic stuff.

00:50:09.455 --> 00:50:13.794
It's all of that. and it's not like some
crappy flash model.

00:50:14.395 --> 00:50:15.475
I think they're taking it serious.

00:50:16.035 --> 00:50:19.495
Now, obviously, like OpenAI, to be fair to
them, I think their ChatGPT is like the

00:50:19.535 --> 00:50:20.735
best it probably could be.

00:50:21.095 --> 00:50:22.295
But if you look at Apple,

00:50:23.015 --> 00:50:27.494
uh, as in the free version, and if you,
but if you look at Apple, it's just like,

00:50:27.655 --> 00:50:30.855
oh yeah, it's gonna like summarize your
notifications.

00:50:30.895 --> 00:50:35.214
Like they're just not, not particularly
interested in, uh, in really giving the

00:50:35.295 --> 00:50:36.635
best stories to consumers.

00:50:37.300 --> 00:50:41.110
Yeah, they can't or they don't care or
like, uh, I don't know what, but I think

00:50:41.190 --> 00:50:45.239
they also care about the experience, so
they just want to be like, oh, we want to

00:50:45.420 --> 00:50:47.360
be the last ones, but do it the best.

00:50:47.820 --> 00:50:49.679
May I remind you of, uh,

00:50:50.650 --> 00:50:51.710
a famous actress,

00:50:52.410 --> 00:50:53.729
I think she was in Game of Thrones,

00:50:54.530 --> 00:50:59.350
that, uh, took ads for Apple and their
Siri AI 2 years ago, where Siri was

00:50:59.410 --> 00:51:02.749
supposed to be this like assistant thing
that reads your messages from your mom and

00:51:02.810 --> 00:51:04.610
puts on calendar and does all these
things.

00:51:05.010 --> 00:51:08.090
And then Apple pulled those ads because
they couldn't deliver on this promise.

00:51:08.490 --> 00:51:13.689
And then Apple launched Siri AI, and it
fizzed completely in the air, and iOS 27

00:51:13.850 --> 00:51:16.289
is now on many, many devices, and many
people will use Siri.

00:51:16.530 --> 00:51:20.170
And we had such great hope for Siri as
being that personal assistant with access

00:51:20.210 --> 00:51:25.130
to my stuff, and that absolutely fizzled
because Siri is not agentic in any sense

00:51:25.170 --> 00:51:28.910
of the word. And I think for us, what we
need in assistant

00:51:29.550 --> 00:51:30.550
is

00:51:30.609 --> 00:51:34.445
proactivity, in addition to context, in
addition to connectors, to everything.

00:51:34.540 --> 00:51:38.020
I don't need it to be super private on my
device, whatever, because my device

00:51:38.100 --> 00:51:41.550
cannot, read, uh, I don't know, go on my
Twitter via connector.

00:51:41.630 --> 00:51:44.040
So that's, that's what's missing for me
in, the Appley stuff.

00:51:44.560 --> 00:51:46.239
And I think that, eh,

00:51:47.160 --> 00:51:51.219
whether or not all the folks care about
privacy, I think it's incredibly important

00:51:51.280 --> 00:51:54.359
that the company that gives them that
product does care about their privacy.

00:51:54.720 --> 00:51:59.140
And so I absolutely applaud the folks from
Meta for following Apple, and Apple

00:51:59.400 --> 00:52:03.760
famously has the secure cloud where your
LLM execution with Siri and then go, if

00:52:03.800 --> 00:52:05.819
they need bigger LLM, they go into the
secure cloud.

00:52:05.920 --> 00:52:09.199
Apple cannot read this, even FBI, FBI
comes knock on their door.

00:52:09.560 --> 00:52:15.129
Uh, Meta is working on that for folks that
use Muse, now that they're shoving Muse in

00:52:15.150 --> 00:52:17.729
the face of the billion people across
their devices.

00:52:17.790 --> 00:52:21.520
They're shoving Muse in every Instagram
ad, and every Instagram story has a little

00:52:21.580 --> 00:52:22.580
Muse thingy.

00:52:23.180 --> 00:52:27.009
Uh, obviously the Facebook crowd will
gonna get there very, very soon, and, it's

00:52:27.030 --> 00:52:30.889
been exploding. Everybody who I know
that's not in our circle in AI knows about

00:52:30.950 --> 00:52:32.589
Muse. It's quite crazy.

00:52:32.750 --> 00:52:35.666
how strongly Meta has, like an effect on
the populace.

00:52:35.786 --> 00:52:36.826
And I've been talking about this here.

00:52:37.986 --> 00:52:41.986
Many folks use Google, and then ChatGPT
came about, and many folks use ChatGPT as

00:52:42.026 --> 00:52:45.026
the new Google. They just ask a question,
get an answer, and they continue to the

00:52:45.066 --> 00:52:46.925
next one. We are like,

00:52:47.586 --> 00:52:50.745
you know, advanced there. We're now using
a agent and assistant for coding.

00:52:51.225 --> 00:52:54.785
The next iteration is having an agent
that's an assistant.

00:52:55.266 --> 00:52:59.085
There is a difference there. That is 24/7,
has its own computer, has its own browser,

00:52:59.146 --> 00:53:04.745
and does for stuff for you proactively,
reminds you about stuff proactively, uh,

00:53:04.986 --> 00:53:09.505
renews your driver's license online, which
was what I had it do last night.

00:53:10.066 --> 00:53:15.145
Grok went and renewed my, uh, expired car
registration, for example, and did it on

00:53:15.186 --> 00:53:18.706
its own. And the assistant can do shopping
for you and pay for you via the Stripe

00:53:18.746 --> 00:53:21.106
Link integration and the 1Password
integration.

00:53:21.385 --> 00:53:23.906
So I think many, many people will get
there faster.

00:53:23.946 --> 00:53:26.026
I think they'll skip the agent part,
they'll just go to assistants.

00:53:26.626 --> 00:53:27.626
Um,

00:53:27.666 --> 00:53:29.365
Yam, thoughts on Muse?

00:53:30.105 --> 00:53:31.205
Are you getting Yeah, look.

00:53:31.266 --> 00:53:32.286
FOMO from not getting it?

00:53:33.066 --> 00:53:36.266
First, let's, let's just say that I
already got it.

00:53:36.466 --> 00:53:37.466
You already got it. Amazing.

00:53:37.486 --> 00:53:40.906
Yeah. Absolutely, yeah. I, I tried it and
it's brilliant becau- because of, you

00:53:41.066 --> 00:53:45.246
hyping it every, every week, so I, I had
to try, and it's absolutely brilliant.

00:53:45.506 --> 00:53:49.005
Yes. But, uh, can we stop and talk about

00:53:49.745 --> 00:53:54.425
it, uh, overtaking ChatGPT number of
users?

00:53:55.786 --> 00:53:57.986
Seriously? Not number of users, the

00:53:58.666 --> 00:53:59.705
speed with which

00:54:00.506 --> 00:54:04.166
Okay. it acquires users and the speed with
which it got to number one on the App

00:54:04.226 --> 00:54:05.606
Store. Mm-hmm. Okay.

00:54:05.626 --> 00:54:07.246
That's the That, that, that makes sense.

00:54:07.266 --> 00:54:10.145
But, But, but they will surpass number of
users 100%.

00:54:10.306 --> 00:54:11.385
Meta has the biggest

00:54:12.106 --> 00:54:13.666
surface area products in the world.

00:54:13.826 --> 00:54:16.666
That Zach literally stood on stage and
said, we have

00:54:17.706 --> 00:54:21.206
the most people using our products in the
world and everybody else, and they're

00:54:21.226 --> 00:54:24.065
gonna shove Muse into all of them very
soon.

00:54:24.346 --> 00:54:27.215
So I expect it to, like, surpass ChatGPT
in whatever.

00:54:27.496 --> 00:54:28.496
Also, we expect

00:54:29.296 --> 00:54:30.296
OpenAI

00:54:30.376 --> 00:54:35.835
to come back both with an assistant model,
which GPT for work kind of already is, but

00:54:35.896 --> 00:54:37.376
people don't know because it's hidden in
there.

00:54:38.136 --> 00:54:39.016
And

00:54:39.016 --> 00:54:43.995
do you guys remember Jony Ive and LoveFrom
and OpenAI and the devices thing that they

00:54:44.096 --> 00:54:47.492
promised and Sam Altman putting his head
on Jony Ive's shoulder and collaborating

00:54:47.552 --> 00:54:49.911
together. That's also coming at some
point.

00:54:50.072 --> 00:54:52.592
OpenAI's hardware is coming, and where is
it?

00:54:52.952 --> 00:54:54.691
Uh, yeah, yeah, so back to Muse.

00:54:54.792 --> 00:54:55.792
You...

00:54:56.592 --> 00:54:59.112
Look, I don't know about the the end
device.

00:54:59.272 --> 00:55:02.791
I must say that why won't I use my phone
for this?

00:55:03.032 --> 00:55:06.372
I- I'm asking because you sound really
excited about this, and Yeah.

00:55:06.391 --> 00:55:09.172
I'm like, I don't know, my phone works
great.

00:55:09.312 --> 00:55:11.951
I just want the software, I want the
assistant.

00:55:12.952 --> 00:55:15.772
Like, convince me that I want the device
in general.

00:55:16.072 --> 00:55:19.031
I will let Jony Ive, convince you when
their device comes out.

00:55:19.152 --> 00:55:22.351
I will just tell you that I cannot wait
for Muse on my glasses.

00:55:24.147 --> 00:55:25.407
Why wouldn't you do this on the phone?

00:55:25.428 --> 00:55:26.867
Yam, you could. You could.

00:55:27.348 --> 00:55:28.687
However, putting out a phone

00:55:29.348 --> 00:55:32.408
and clicking buttons, shoving this into
people's face when you talk to them, it's

00:55:32.428 --> 00:55:34.947
like a whole another thing that you need
to do.

00:55:35.468 --> 00:55:38.588
Whereas just saying a keyword and then
just going about your day, looking the

00:55:38.628 --> 00:55:41.148
person in the eye is a completely
different experience, I believe.

00:55:41.508 --> 00:55:42.508
Nisten.

00:55:43.908 --> 00:55:44.908
Muse in Canada

00:55:45.468 --> 00:55:47.028
from last week. That's also big news.

00:55:47.188 --> 00:55:48.507
So you got it. Uh.

00:55:49.028 --> 00:55:50.028
Yep.

00:55:51.108 --> 00:55:52.108
Tell us.

00:55:52.228 --> 00:55:53.707
Uh, I got Muse.

00:55:54.188 --> 00:55:58.488
I really liked it because I just put it to
computer use at first place, and it ran

00:55:58.528 --> 00:56:00.408
into a CAPTCHA on Temu.

00:56:00.648 --> 00:56:05.048
I was just asking it to look up products
on Temu, and then it had a remote desktop

00:56:05.128 --> 00:56:09.007
session, and I noticed it was a Chromium
session running inside it, and then it

00:56:09.048 --> 00:56:13.168
would have a, inside the app on my
Android, it would have the remote desktop

00:56:13.208 --> 00:56:15.327
window, and I can go and I, I could click
the CAPTCHA

00:56:16.088 --> 00:56:18.648
with my finger, and then it would just
continue on.

00:56:18.923 --> 00:56:20.257
That, that was the best

00:56:21.038 --> 00:56:22.038
user experience

00:56:22.878 --> 00:56:25.958
so far, and uh, I love how open

00:56:26.718 --> 00:56:27.718
they've left it.

00:56:28.238 --> 00:56:31.607
Uh, I'm not going to reveal too much, but
it is

00:56:32.268 --> 00:56:35.537
like and you have root on the whole
machine, which is pretty nuts.

00:56:35.638 --> 00:56:39.257
They're just letting everyone have
machines running as root, and you can

00:56:39.438 --> 00:56:40.438
actually do

00:56:40.718 --> 00:56:42.217
whatever You mean the you want.

00:56:42.238 --> 00:56:43.898
the VM that comes with Muse, Yeah, yeah.

00:56:43.998 --> 00:56:45.378
the virtual machine that they, yeah.

00:56:45.718 --> 00:56:49.937
It's an 8 gig VM with, uh, 100 gigs of,
uh, storage space.

00:56:50.588 --> 00:56:53.407
The thing is that you told me last week
they had a TailScale integration, Yes.

00:56:53.428 --> 00:56:55.347
so I was able to TailScale in.

00:56:55.787 --> 00:56:57.207
So now I have a complete shell.

00:56:57.308 --> 00:57:02.527
It can, you can even like run a model
inside the 8 gigs of RAM it has if if you

00:57:02.548 --> 00:57:04.488
wanted to. They've done a really good job.

00:57:04.508 --> 00:57:05.847
I want folks to understand what's going
on.

00:57:05.868 --> 00:57:08.827
It's it's insane, like that Zuckerberg is
they are on another level.

00:57:09.868 --> 00:57:10.868
For free,

00:57:11.388 --> 00:57:14.108
everyone, 100 million tokens a week,

00:57:14.988 --> 00:57:17.647
plus a whole computer in the cloud

00:57:18.508 --> 00:57:21.527
that is a good computer in the cloud, for
free, for everyone.

00:57:22.148 --> 00:57:24.728
Do you guys understand the scale of what?

00:57:25.228 --> 00:57:28.428
One more thing, as for privacy, the
people, you can just open a new Gmail

00:57:28.468 --> 00:57:32.007
account, and you can just log in with that
Gmail account and have nothing connected

00:57:32.028 --> 00:57:35.967
to it. Because this week they announced
that Muse is going to get their own inbox.

00:57:36.188 --> 00:57:38.527
They'll be able to just CC on your
conversations.

00:57:38.628 --> 00:57:41.128
Okay, they're making it like Yeah, they're
making it less.

00:57:41.628 --> 00:57:45.918
One step less. Folks, assistants are here
to stay, AI assistants are here to stay.

00:57:46.278 --> 00:57:49.117
Uh, we're going to cover use cases on the
show.

00:57:49.198 --> 00:57:54.568
So I want to share 2 real use cases for AI
assistants, and I think that we will have

00:57:54.588 --> 00:57:57.308
a corner here that will just share our
experiences because folks,

00:57:58.428 --> 00:58:03.428
when they get those things and they don't
know fully how to utilize them, although

00:58:03.547 --> 00:58:05.268
Meta did an incredible job in the ideas.

00:58:05.908 --> 00:58:08.988
So number one, if you get Muse, go into
the ideas tab.

00:58:10.028 --> 00:58:14.422
Number two, when you get Muse or Grokbot
or whatever assistant, Folks, go on a

00:58:14.503 --> 00:58:15.503
walk,

00:58:15.622 --> 00:58:18.922
hit the record button, and just talk about
your day.

00:58:19.143 --> 00:58:22.583
Talk about your business, talk about your
workday, talk about your family, talk

00:58:22.622 --> 00:58:24.263
about, talk about, talk about what you do.

00:58:24.943 --> 00:58:28.943
Because if you try to imagine use cases
that it can do for you, that's great.

00:58:29.623 --> 00:58:30.423
But

00:58:30.423 --> 00:58:33.903
the thing with Assistant is that, like,
this is a whole new paradigm for you of

00:58:34.023 --> 00:58:36.023
stuff that it can do that you don't even
know it can do.

00:58:36.423 --> 00:58:40.263
And a lot of the stuff that you can do or
need to do comes from your emails, etc.

00:58:40.943 --> 00:58:43.782
So number one, connect your email inbox,
uh, so it knows.

00:58:44.103 --> 00:58:48.222
And number two, just talk to it and then
say, hey, here's the stuff about me.

00:58:48.663 --> 00:58:51.802
The beautiful thing about Muse and other
assistants, they have this memory file

00:58:52.063 --> 00:58:56.202
that they put about you and about your
contacts, and in that memory file, th-

00:58:56.303 --> 00:58:57.843
this is how assistant knows how to help
you.

00:58:58.303 --> 00:59:02.142
The assistants will only be as helpful to
you as the information that you give them.

00:59:02.703 --> 00:59:06.483
And an additional tip is that there is an
import memory feature when you go to Muse

00:59:06.943 --> 00:59:10.483
and you click a button there, it will give
you a prompt to send to your ChatGPT, to

00:59:10.503 --> 00:59:14.242
your Claude, to other assistants, to
basically brain transplant whatever

00:59:14.383 --> 00:59:16.242
information they have into Muse.

00:59:16.263 --> 00:59:19.443
You can review that, and you can send, and
then your assistant's significantly

00:59:19.543 --> 00:59:23.113
better. And then the idea stuff that
they're surfacing there, is going to also

00:59:23.173 --> 00:59:25.483
be significantly more contextual and
stronger for you.

00:59:25.643 --> 00:59:27.283
So those are my tips for now.

00:59:27.723 --> 00:59:30.742
A real use case that I had is my weekend
bot.

00:59:30.923 --> 00:59:33.042
It's in Grok, but Muse also does this.

00:59:35.843 --> 00:59:40.302
I arrived up very early to put my kids in
a karate class on Saturday, only to find

00:59:40.363 --> 00:59:43.533
out that I should have known that they are
closed that uh, week.

00:59:43.773 --> 00:59:45.133
I could have slept for another hour.

00:59:45.493 --> 00:59:49.133
Immediately then, I talked to all of my
assistants to, first of all, check all my

00:59:49.213 --> 00:59:52.073
plans for the weekend and their websites
and emails, because I don't look at their

00:59:52.093 --> 00:59:55.732
emails for cancellations, and two, plan
out weekends ahead of time.

00:59:56.333 --> 01:00:01.273
So now my assistants go search local news
ahead of time, so by the time that I

01:00:01.333 --> 01:00:06.133
arrive on a weekend, I have a list of
plans that I could be doing with kids or

01:00:06.213 --> 01:00:07.213
with my

01:00:07.653 --> 01:00:12.232
wife, and that's a weekend planner prompt
recommendation for you that you should,

01:00:12.313 --> 01:00:15.613
you should add to your bots, like, hey,
every whatever Thursday, go and look at

01:00:15.653 --> 01:00:19.673
the local stuff that are happening around
my city, sort them out by free, paid,

01:00:19.732 --> 01:00:23.712
whatever, premium tier, and do this for
this weekend and the weekend two weeks

01:00:23.933 --> 01:00:26.072
after, because things do run out
sometimes.

01:00:26.173 --> 01:00:27.653
You do need to buy, uh, places.

01:00:28.093 --> 01:00:29.352
That has changed my weekends.

01:00:29.453 --> 01:00:32.092
I no longer have to, like, think about,
oh, what are we doing?

01:00:32.403 --> 01:00:33.643
Peter, what's your assistant tip?

01:00:34.523 --> 01:00:37.122
What is one thing that you're like, oh,
this is novel, people should know about

01:00:37.143 --> 01:00:38.562
this? Put you on the spot there.

01:00:40.489 --> 01:00:46.398
For me, I know it's, it's very boring, but
just aligning my admin, so,

01:00:46.459 --> 01:00:50.298
like, any kind of emails, calendars, all
of that flying around, it just-

01:00:50.909 --> 01:00:54.319
I feel like I was always anxious about
what's, uh,

01:00:55.119 --> 01:00:56.298
is my stuff aligned?

01:00:56.399 --> 01:00:57.638
Am I forgetting something?

01:00:58.079 --> 01:01:00.458
And just whatever bothers you like that,
just try and

01:01:01.599 --> 01:01:03.159
give that over to an assistant

01:01:04.079 --> 01:01:05.798
that can just, like, manage that stuff for
you.

01:01:07.534 --> 01:01:13.234
I don't really, um, otherwise do a lot of
stuff like that, but just the simple

01:01:13.334 --> 01:01:15.894
things to just get off my plate, that's
what I really appreciate.

01:01:15.934 --> 01:01:20.173
So I just have more headspace not to think
about stuff, uh, along those lines.

01:01:21.574 --> 01:01:24.834
Yep. Uh, all right, Nisten, maybe you ran
this for the whole week.

01:01:24.894 --> 01:01:27.713
What is one use case coming up for you
with Muse that, that you think people

01:01:27.734 --> 01:01:28.734
should know about?

01:01:29.334 --> 01:01:30.334
Uh, so

01:01:30.933 --> 01:01:35.413
it is structured very much like a very
light and very nice Hermes type of build,

01:01:35.653 --> 01:01:36.993
but actually a very lightweight Hermes.

01:01:37.014 --> 01:01:38.153
So it has the Sol file

01:01:38.814 --> 01:01:40.054
and, and memory and stuff.

01:01:40.774 --> 01:01:41.774
Uh,

01:01:42.174 --> 01:01:44.554
it does adopt personas quite okay.

01:01:45.174 --> 01:01:49.754
I just give it the persona of philosopher
Epictetus, and then the prompt is just

01:01:50.454 --> 01:01:51.834
first learn the meaning

01:01:52.614 --> 01:01:55.734
of, uh, what you have to say, and then
speak.

01:01:56.093 --> 01:01:59.073
And then I tell it a whole bunch of stuff
to say in that persona.

01:01:59.534 --> 01:02:00.534
It actually

01:02:00.813 --> 01:02:02.373
is super nice

01:02:03.414 --> 01:02:05.333
by default, like it's not,

01:02:06.374 --> 01:02:08.973
it's so nicely done, it is so nicely
prompted.

01:02:09.654 --> 01:02:14.814
It's not overbearing, and it's not like
ChatGPT where it just makes assumptions.

01:02:14.974 --> 01:02:17.153
It's like right on the spot Mm.

01:02:17.214 --> 01:02:18.214
as as an assistant.

01:02:19.204 --> 01:02:20.724
It really is a good product, folks.

01:02:20.844 --> 01:02:22.939
It really is the product for the masses.

01:02:22.939 --> 01:02:26.683
Like, we started the show with OpenAI, et
cetera, being for the AI geeks, and

01:02:26.724 --> 01:02:27.724
OpenAI, like, quickly

01:02:28.164 --> 01:02:30.563
got to, whatever billion people that use
it as a chat.

01:02:30.964 --> 01:02:32.884
This is a good product for many people.

01:02:33.324 --> 01:02:34.164
Just

01:02:34.164 --> 01:02:38.063
the notification thing that shows you what
it's doing, and it's summarizing really

01:02:38.104 --> 01:02:40.844
what it's doing versus just, like, typing
or showing you CRL command.

01:02:40.884 --> 01:02:43.124
Just that is worth, like, so many points.

01:02:43.444 --> 01:02:46.523
And shout out to Alex Conway, I believe,
is the designer for that.

01:02:46.844 --> 01:02:50.723
Uh, that pattern has existed, obviously,
in Telegram, but Telegram only said

01:02:50.824 --> 01:02:54.623
typing. Here they actually tell you what
the assistant does, and I think that's

01:02:54.664 --> 01:02:59.583
incredible. Uh, all right, folks, I think
it's time for us to move on to Grok, and

01:02:59.824 --> 01:03:00.664
uh,

01:03:00.664 --> 01:03:02.103
Grok 4.7 was released.

01:03:02.144 --> 01:03:04.143
We waited for this. Uh, Grokbot is

01:03:05.344 --> 01:03:06.364
XAI's assistant, right?

01:03:06.411 --> 01:03:11.450
is running Grok 4.6 fast, and now, after
promises, uh, there is Grok 4.7.

01:03:12.651 --> 01:03:15.330
And, uh, it's kind of a disappointment for
me.

01:03:16.051 --> 01:03:19.050
I, you know, I wish I had like a wah, wah,
wah, wah.

01:03:19.251 --> 01:03:20.251
Do I have it?

01:03:20.451 --> 01:03:21.570
I think I has, yeah.

01:03:25.051 --> 01:03:26.791
I am sorry, Cursor folks.

01:03:27.051 --> 01:03:29.490
I believed that your coding mojo

01:03:30.651 --> 01:03:35.090
will be able to, to beat whatever,
whatever stuff, and Grok 4.7

01:03:36.291 --> 01:03:37.570
is not that great.

01:03:38.371 --> 01:03:39.371
I expected more.

01:03:40.210 --> 01:03:42.871
I wanted my Grok bot to use Grok 4.7.

01:03:42.971 --> 01:03:47.291
It'd be significantly better at writing,
at writing code, at making decisions, and

01:03:47.451 --> 01:03:49.230
that's not, that did not arrive.

01:03:50.051 --> 01:03:51.610
However, what did arrive is

01:03:52.251 --> 01:03:53.251
Grokbot

01:03:53.571 --> 01:03:55.251
is now connected to your Tesla,

01:03:56.531 --> 01:04:01.970
which I find is an amazing pattern because
your car is basically Kit

01:04:02.611 --> 01:04:07.451
from the 70s show, and you can talk to it,
and it self-drives, and it can do stuff.

01:04:07.571 --> 01:04:09.610
It's in-fucking-credible.

01:04:10.331 --> 01:04:12.690
So this is the jagged frontier of AI we're
talking about.

01:04:12.891 --> 01:04:14.931
You're getting a new model, and it's not
the best.

01:04:15.011 --> 01:04:17.931
However, you're getting a connector, and
it can do stuff.

01:04:18.171 --> 01:04:21.291
So, for example, you can talk to your
assistant without hands.

01:04:21.531 --> 01:04:24.131
You can hold it like this and say, hey,
Grok,

01:04:25.731 --> 01:04:28.210
what are some of my emails that I haven't
replied to?

01:04:28.731 --> 01:04:31.670
And it will say, or you say, hey, Grok,
check in on

01:04:32.571 --> 01:04:36.651
that thing. I will go and do the
connectors, and Grokbot in Tesla, I think,

01:04:36.751 --> 01:04:40.791
is significantly more important than any
advancement in, in AI that could have,

01:04:41.251 --> 01:04:42.251
could have gotten.

01:04:42.491 --> 01:04:44.731
But back to Grok 4.7.

01:04:45.731 --> 01:04:47.930
Uh, CursorBench is getting

01:04:48.691 --> 01:04:52.571
46% back from, uh, 40 for Grok 4.6.

01:04:53.051 --> 01:04:57.611
It's the same price, 2 dollars per million
tokens with half a million context window.

01:04:57.891 --> 01:04:59.931
The fast variant is 2x the speed and
price,

01:05:00.731 --> 01:05:02.530
but TerminalBench is 37,

01:05:03.491 --> 01:05:07.090
and it's still behind GPT 5.6 Seoul on
DeepSui.

01:05:07.641 --> 01:05:11.721
xAI trained this model for the longest
time and still behind OpenAI's stuff.

01:05:13.201 --> 01:05:14.201
Uh,

01:05:15.280 --> 01:05:16.280
Peter,

01:05:16.361 --> 01:05:18.481
any thoughts on on Grok 4.7?

01:05:18.921 --> 01:05:20.280
Were you excited, disappointed?

01:05:21.681 --> 01:05:24.240
Yeah, I think, yeah, it's kind of a weird
one.

01:05:24.321 --> 01:05:27.001
I'm with you that when the Curse team

01:05:27.840 --> 01:05:31.980
stepped in to do the training, I kind of
thought, oh yeah, that's it, they're just

01:05:32.041 --> 01:05:35.921
gonna fly off into the distance, and I
think, uh, they didn't quite

01:05:36.681 --> 01:05:38.080
get there very quickly.

01:05:38.641 --> 01:05:40.700
I don't think we should write them off
yet.

01:05:40.801 --> 01:05:44.900
I think, you know, these things are hard,
and I'm sure, you know, you never quite

01:05:44.961 --> 01:05:48.980
know that they just not quite have enough
GPUs at that point, and they're just

01:05:49.161 --> 01:05:51.820
ramping up the next one, and it will be
all amazing.

01:05:51.881 --> 01:05:55.240
So I don't think we should make any big,
um, conclusions from that.

01:05:55.321 --> 01:05:56.681
But I think, um,

01:05:57.401 --> 01:06:00.221
w- well, so clearly it's not an amazing
model.

01:06:00.361 --> 01:06:01.621
I think it's gonna be fine.

01:06:01.721 --> 01:06:05.721
Maybe it's like slightly better than the
previous one, but actually 4- 4.6 was not

01:06:05.761 --> 01:06:09.906
that great either. So I, the, the question
in my mind is, you know, they had this,

01:06:10.007 --> 01:06:15.286
uh, deal, uh, with Anthropic, right, for 6
months of compute.

01:06:15.567 --> 01:06:16.927
Mhm. What happens to that deal?

01:06:17.087 --> 01:06:20.927
It's, I think we are coming up to 5 months
now into the 6 months deal.

01:06:21.447 --> 01:06:23.566
So in my mind, it goes 2 ways.

01:06:23.807 --> 01:06:24.807
One is

01:06:25.207 --> 01:06:30.466
they take the compute back and say, look,
our model needs more compute, we need to

01:06:30.607 --> 01:06:33.406
train more. I think I can see that
happening.

01:06:33.647 --> 01:06:34.647
Yeah.

01:06:34.647 --> 01:06:38.687
Or they, you know, or they say to
Anthropic, you know what, pay us double.

01:06:39.286 --> 01:06:43.007
Uh, you know, you need, you need the
compute, you're building good models, pay

01:06:43.047 --> 01:06:44.047
us more.

01:06:44.447 --> 01:06:45.447
I,

01:06:45.567 --> 01:06:49.027
maybe this is rumors, maybe I'm completely
wrong, but I don't think it's a scale

01:06:49.087 --> 01:06:51.167
problem. I think they have enough to train
Grok as well.

01:06:51.286 --> 01:06:52.286
I think it's a

01:06:53.167 --> 01:06:54.886
talent and data problem.

01:06:55.327 --> 01:06:59.167
Uh, Elon Musk got incredible talent when
the Cursor acquisition came over.

01:06:59.286 --> 01:07:02.886
Obviously, Grokbot is a Cursor product
that they took over, slapped the name Grok

01:07:02.927 --> 01:07:04.646
on it, and now is on every surface.

01:07:05.527 --> 01:07:07.167
They're doing the same to their

01:07:08.447 --> 01:07:10.826
surface area that Meta is doing for Muse.

01:07:10.887 --> 01:07:14.206
They're putting Grokbot on top of every
Cursor, uh, cloud

01:07:14.807 --> 01:07:17.447
instance. You can have a Grokbot link now
in every

01:07:18.087 --> 01:07:19.007
X

01:07:19.007 --> 01:07:23.007
chat. They're opening this up to, like,
all of the, all of the X users.

01:07:23.087 --> 01:07:26.226
They lowered the price. It's not quite
free yet, but I'm pretty sure they're

01:07:26.247 --> 01:07:28.766
gonna look at, uh, Zach and say, hey, this
is now free.

01:07:29.407 --> 01:07:34.966
But for the model training, I don't think
that the Cursor data did a dent

01:07:35.607 --> 01:07:36.607
yet.

01:07:36.847 --> 01:07:38.147
I don't, I don't see it.

01:07:38.287 --> 01:07:42.206
Like, I don't think that now I'll go to
Grok 4.7 and start using this

01:07:43.007 --> 01:07:44.646
for my coding assistance.

01:07:44.847 --> 01:07:47.206
If it's better and I use it inside
Grokbot, that's great.

01:07:47.807 --> 01:07:50.646
And I think here we're seeing like a
battle lines are drawn where

01:07:51.487 --> 01:07:55.526
the interface slash harness slash
assistant where through which you will

01:07:55.607 --> 01:07:58.607
talk to these models matters more than the
actual model.

01:07:59.127 --> 01:08:04.066
Like, literally, I am more excited this
week about some of the features in Grokbot

01:08:04.167 --> 01:08:06.447
that launched, which is 1Password
integration.

01:08:06.607 --> 01:08:10.547
Folks, if you don't use a password
manager, please, please, please sign up

01:08:10.607 --> 01:08:13.767
for one. It's important even more in the
era of AI assistants.

01:08:14.047 --> 01:08:16.766
Do not use the same password for all your
services.

01:08:17.127 --> 01:08:19.146
However, an additional reason to use

01:08:20.247 --> 01:08:23.547
password manager, specifically 1Password,
which I love with a passion over a decade,

01:08:23.647 --> 01:08:25.126
nothing happened, nobody hacked me,
anything,

01:08:25.807 --> 01:08:28.306
is that those assistants now integrate
with your own password.

01:08:28.327 --> 01:08:32.447
You'll be able to share your passwords in
a secure way that the a- assistant can

01:08:32.567 --> 01:08:33.567
see.

01:08:33.647 --> 01:08:38.226
So I'm more excited, Peter, about this one
integration with Grokbot than the whatever

01:08:38.327 --> 01:08:41.986
next step of intelligence that they didn't
quite hit, because it, for for most users,

01:08:42.087 --> 01:08:43.447
this is like the most important stuff.

01:08:43.887 --> 01:08:46.406
Um, Nisten, thoughts on Grok 4.7?

01:08:49.327 --> 01:08:51.607
Yeah, look, there are 9.4

01:08:52.647 --> 01:08:54.526
million Teslas or more

01:08:55.167 --> 01:08:56.447
out still in circulation,

01:08:57.247 --> 01:09:00.926
a- as total, eh, and we're just, we're
also- 9.4 million?

01:09:01.287 --> 01:09:04.506
Yeah, yeah, in total production over,
well, it's been over like a decade.

01:09:04.527 --> 01:09:06.026
I think there's way more than 9.4.

01:09:06.087 --> 01:09:07.087
Uh.

01:09:07.806 --> 01:09:11.267
We- we will, uh, we'll ask producer fact
check Nisten live on air.

01:09:11.327 --> 01:09:15.207
Yeah. Uh, go and find out It's around the
best estimates of how many Teslas are

01:09:15.287 --> 01:09:17.166
currently driving on the road. But yeah,
Nisten, Yeah.

01:09:17.407 --> 01:09:20.686
go make your point. We'll see if the It is
around that, that ballpark.

01:09:20.806 --> 01:09:24.346
I mean, as far as I only checked with the
bots, but

01:09:24.967 --> 01:09:29.806
you have to think that even the initial,
like, NVIDIA SoCs that were in those, they

01:09:29.847 --> 01:09:31.347
might not be too powerful.

01:09:31.567 --> 01:09:34.886
They might barely be holding the
interface, but you don't need anything to

01:09:34.927 --> 01:09:38.107
run the agent on it because it's all just
API calls.

01:09:38.207 --> 01:09:39.287
Just the internet connection, yeah.

01:09:39.527 --> 01:09:41.766
Yeah, you've just offloaded all of that
server side.

01:09:42.487 --> 01:09:46.166
So it is pretty, yeah, it's pretty nuts
that they reach those.

01:09:46.407 --> 01:09:50.926
Uh, on the other hand, as for having all
the Cursor data, uh, yeah, you have a lot

01:09:50.967 --> 01:09:55.166
of enterprise data and code and stuff, but
that might not actually be good code.

01:09:55.727 --> 01:09:59.887
You... might need like very smart
filtering pipelines and the stuff like

01:09:59.927 --> 01:10:03.246
that. This is why I think Codex and uh,

01:10:04.047 --> 01:10:06.027
Anthropic are, are ahead on that.

01:10:06.167 --> 01:10:08.047
So yes, they have the data, but they prob-

01:10:08.687 --> 01:10:10.746
they might not just be using it that well

01:10:11.647 --> 01:10:14.327
or be able to generate that good of a
synthetic set.

01:10:14.807 --> 01:10:15.807
I think that,

01:10:16.847 --> 01:10:19.886
I think we're underestimating what I just
said before.

01:10:19.967 --> 01:10:23.327
I, I, I need to, Yam, I need to bring your
energy about like Navier-Stokes from last

01:10:23.367 --> 01:10:27.426
week. My car can fucking drive itself, and
I can talk to it, and it can read my

01:10:27.487 --> 01:10:28.487
email.

01:10:29.767 --> 01:10:34.786
My car can drive itself, and while it
drives itself, I can voice

01:10:35.447 --> 01:10:38.687
interface-wise ask it to do my emails for
me.

01:10:38.767 --> 01:10:42.806
Do you guys realize what the fuck amazing
world we're living in right now?

01:10:43.207 --> 01:10:45.486
Wait, do you have Grokbot in your car?

01:10:46.127 --> 01:10:47.147
That's what I'm saying. Oh.

01:10:47.166 --> 01:10:50.926
Grokbot integration into the Tesla has
dropped this week, and it's fucking

01:10:51.127 --> 01:10:52.226
insane. Oh, oh, shoot, okay.

01:10:52.767 --> 01:10:53.767
Right.

01:10:54.382 --> 01:10:56.906
know. Guys, this is a beautiful world
we're living in.

01:10:57.067 --> 01:10:59.526
We get like amazing like discoveries and
presents.

01:10:59.552 --> 01:11:00.552
stuff. By the way, our

01:11:00.917 --> 01:11:02.897
live AI assistant has fact checked Nisten.

01:11:03.437 --> 01:11:04.437
Uh,

01:11:05.397 --> 01:11:07.557
I think. Hey, I I was not wrong.

01:11:08.837 --> 01:11:10.677
I was pretty there. I said 9 9.5.

01:11:10.877 --> 01:11:11.877
I was right in the middle.

01:11:14.197 --> 01:11:16.196
Yeah. So, Nisten, you want to read out the
fact check?

01:11:16.356 --> 01:11:21.557
Tesla reported 9.2 million cumulative
deliveries as Q1 of 2026 and

01:11:21.717 --> 01:11:23.596
48,000 in Q2.

01:11:23.997 --> 01:11:25.716
So 9.7 to June.

01:11:25.837 --> 01:11:29.517
With Q3 progress, it's around 10 million
delivered by now, with a few percent

01:11:29.597 --> 01:11:32.196
scrapped. So, uh, Nisten is right, a bit
more than that.

01:11:32.557 --> 01:11:34.217
Yeah, so call it the 10 million Teslas.

01:11:34.277 --> 01:11:35.237
Nisten,

01:11:35.237 --> 01:11:38.256
uh, fact- fact checked live by our AI
producer.

01:11:38.397 --> 01:11:41.577
Yeah, that's pretty cool. We should have
that for our politicians now, please.

01:11:41.597 --> 01:11:43.937
All right. Can I just schedule the, yeah,
yeah, go ahead.

01:11:43.957 --> 01:11:47.376
It's just listen, it just listens to us,
and and that's it.

01:11:47.757 --> 01:11:48.956
Like, it's a producer.

01:11:48.977 --> 01:11:50.117
It's not it, brother. It's not it.

01:11:50.197 --> 01:11:52.617
Uh, if you guys want to pause for a second
and talk about how

01:11:53.797 --> 01:11:54.797
AI,

01:11:54.917 --> 01:11:59.717
um, how there's this new concept called
AX, assistant experience or agent

01:11:59.797 --> 01:12:03.536
experience, and I believe, given where
Muse is going, given Grubbot, et cetera,

01:12:03.597 --> 01:12:08.337
that like many, many, many surfaces and
services will need access to assistants

01:12:08.517 --> 01:12:10.756
because APIs is not enough, MCP is not
enough.

01:12:11.317 --> 01:12:15.376
I am planning to make ThursdAI the most
assistant-pilled show in the world.

01:12:15.717 --> 01:12:18.517
We have an AI assistant that is a live
producer.

01:12:18.757 --> 01:12:21.112
Uh, if you guys go to Thursday Live, we
have an AI assistant.

01:12:21.117 --> 01:12:24.277
Our AI assistant is also like marking off
and checking off the, the stuff that we're

01:12:24.317 --> 01:12:25.916
talking about and, and, and shows pace.

01:12:26.277 --> 01:12:29.796
So pacing is very important, right, in
the, in the world of, uh, live podcasting.

01:12:30.037 --> 01:12:35.686
so if you go and, uh, type ThursdAI.live,
you'll see all our notes and me talking,

01:12:35.747 --> 01:12:38.127
obviously. This is going to be a little
bit of a window in the window in the

01:12:38.147 --> 01:12:42.127
window situation, uh, but you can see the
rundown, like we're running through all of

01:12:42.147 --> 01:12:44.986
the stuff. We need to start talking about,
uh, uh, Fully Connected.

01:12:45.067 --> 01:12:46.226
We're like a little bit behind.

01:12:46.307 --> 01:12:47.546
All of our links are here as well.

01:12:48.027 --> 01:12:52.186
We have a producer that's listening to our
show in live real time and and putting up,

01:12:52.227 --> 01:12:56.946
uh, things. Uh, not only that, the the
another producer will edit it,

01:12:57.907 --> 01:13:00.226
and so you will get the show much faster
after it.

01:13:00.307 --> 01:13:04.006
So we're working to being the most agent-
pilled, uh, show in the world.

01:13:04.307 --> 01:13:05.986
Most agent-pilled ever.

01:13:06.627 --> 01:13:08.746
Everything is live, listening to us, just

01:13:09.467 --> 01:13:10.587
doing Yeah. what we're saying.

01:13:10.967 --> 01:13:14.106
I totally agree with you that,
transparently

01:13:14.767 --> 01:13:18.966
being in the background and, acting on
whatever is going on, regardless of the

01:13:19.087 --> 01:13:23.346
AI. Oh yeah, absolutely, absolutely the
next gen interface AI that just

01:13:24.027 --> 01:13:25.867
listens to you and does stuff.

01:13:26.187 --> 01:13:28.416
But you just talk about the actual, uh,

01:13:29.117 --> 01:13:34.427
just using Grok as it is, not, not within,
inside any harness because I don't know if

01:13:34.607 --> 01:13:36.486
it, if it matters. Does it matter?

01:13:36.527 --> 01:13:39.926
Okay. I like, where, where would you use
Grok 4.7?

01:13:39.967 --> 01:13:42.826
In Grok Code, the thing that they launched
that, like, I don't know many people who

01:13:42.887 --> 01:13:45.886
use, uh, Cursor has Fable and Astra.

01:13:47.086 --> 01:13:49.927
It's a Cursor model. It's default, default
in Cursor.

01:13:50.127 --> 01:13:52.947
Definitely, I think it's, I mean, it, it,
Yeah.

01:13:53.007 --> 01:13:56.727
it is targeted at least somewhat to
developers.

01:13:57.687 --> 01:14:00.907
It's disappointing, uh, and this is also
the reaction for everybody else.

01:14:00.987 --> 01:14:04.666
However, sitting in my car while it
drives, talking to my cars when it does...

01:14:05.407 --> 01:14:06.946
Guys, my car has an MCP connection.

01:14:07.007 --> 01:14:08.606
Do you understand what I'm talking about?

01:14:09.087 --> 01:14:13.547
My car, when I talk to it, has an MCP
ability to do anything in the world that

01:14:13.606 --> 01:14:14.606
has MCPs.

01:14:15.767 --> 01:14:18.086
Okay. Bro, you just, you just want it to
drive.

01:14:18.407 --> 01:14:21.146
Like, you, you don't need Oh, just put a
robot arm.

01:14:21.407 --> 01:14:24.747
You have to put a robot arm in it, and
then when you get snacks or coffee, you

01:14:24.767 --> 01:14:25.927
can just feed it to you.

01:14:26.407 --> 01:14:27.846
But s- but seriously, what can you do?

01:14:27.887 --> 01:14:30.706
What can you do with it in your car?

01:14:30.767 --> 01:14:32.247
Like, what, what the difference is?

01:14:32.407 --> 01:14:34.346
Open up your garage door, Yam.

01:14:35.407 --> 01:14:37.206
Garage door connector, smart home.

01:14:37.327 --> 01:14:39.387
You can say, hey, Tesla, open up my garage
door.

01:14:39.407 --> 01:14:40.807
Can you tweet from it?

01:14:41.007 --> 01:14:42.166
100% you can tweet.

01:14:42.207 --> 01:14:44.327
Oh yeah, okay. Can you What do you guys
mean?

01:14:44.407 --> 01:14:47.606
MCP is an, is a connector, universal
connector to everything.

01:14:47.897 --> 01:14:52.046
I saw one dude, showing an example,
sitting in his car and saying, hey, get me

01:14:52.087 --> 01:14:54.166
my favorite thing from Starbucks and drive
there.

01:14:54.607 --> 01:14:56.826
And the Grok bot sent a Okay, that's cool.

01:14:56.867 --> 01:14:59.286
invite to the Starbucks, order that he
usually gets.

01:14:59.327 --> 01:15:03.076
He didn't even say what it was, and then,
sent a ping to his Tesla on the map, and

01:15:03.117 --> 01:15:04.976
all he we needed to do is hit drive.

01:15:05.637 --> 01:15:09.396
By the time he arrived to, Starbucks, his
coffee was waiting for him, already paid.

01:15:10.397 --> 01:15:11.397
That's-

01:15:11.397 --> 01:15:14.457
Yeah, that's pretty crazy because Yeah,
that's while it might not be that good at

01:15:14.517 --> 01:15:18.456
coding, it's more than smart enough to
just handle other agents to It is really

01:15:18.477 --> 01:15:21.236
good for handling a lot of agents,
specifically Grokbot when they talk to

01:15:21.276 --> 01:15:24.032
each other, and also, this whole show is
also produced with Grokbots.

01:15:24.067 --> 01:15:25.797
All right, right, let's go to this week's
buzz, please.

01:15:45.109 --> 01:15:48.509
Everyone, welcome to This Week's Buzz, a
corner of ThursdAI, where we talk about

01:15:48.709 --> 01:15:51.359
the company that makes it all possible,
CoreWeave,

01:15:52.029 --> 01:15:57.589
Next week, ThursdAI is live, coming to you
live from San Francisco, Moscone South, at

01:15:57.629 --> 01:16:00.069
the Fully Connected 2026 show with

01:16:00.749 --> 01:16:01.808
folks like Fei-Fei Dr.

01:16:01.849 --> 01:16:03.168
Fei-Fei Li from World Labs,

01:16:03.889 --> 01:16:05.629
with battle bots fighting out

01:16:06.529 --> 01:16:07.529
live on

01:16:08.289 --> 01:16:10.188
the show floor, and yes,

01:16:10.849 --> 01:16:13.329
Mr. 305 Pitbull is headlining.

01:16:13.709 --> 01:16:18.409
Fully Connected is next week, September
29th to October 1st, Moscone South in San

01:16:18.469 --> 01:16:20.908
Francisco, the day after OpenAI's Dev Day,
by the way.

01:16:21.069 --> 01:16:24.069
So if you are flying out for Dev Day, come
to ThursdAI.

01:16:24.189 --> 01:16:28.728
You can get our tickets for free if you
follow the show and our newsletter, and

01:16:28.789 --> 01:16:30.389
we're doing the live show on the floor.

01:16:30.749 --> 01:16:33.248
Uh, I'm very excited to hear the keynote
from Dr.

01:16:33.309 --> 01:16:34.488
Fei-Fei Li from World Labs.

01:16:34.709 --> 01:16:35.709
Sarah Guo

01:16:35.908 --> 01:16:39.918
is hosting a panel, folks like Jerry Liu,
Tom Rockschadow, Biases, Lucas Biewald,

01:16:39.939 --> 01:16:44.678
and CoreWeave CEO Mike Intrader is going
to talk about how we received the Platinum

01:16:44.819 --> 01:16:46.746
tier on the benchmark from SemiAnalysis.

01:16:47.667 --> 01:16:51.226
We also have Ian Buck from NVIDIA and, uh,
live BattleBots, and, Pitbull is gonna

01:16:51.287 --> 01:16:52.287
close the show.

01:16:52.607 --> 01:16:57.567
the ClusterMax 3.0 from SemiAnalysis is
our little flex this week, yesterday, and

01:16:57.627 --> 01:17:01.686
we got platinum. It's a 3 reports in a
row, and we're only the only provider

01:17:01.787 --> 01:17:03.507
that's been platinum all 3 editions.

01:17:03.987 --> 01:17:08.107
Uh, Dylan Patel called CoreWeave the
operational benchmark for the industry in

01:17:08.226 --> 01:17:10.292
GPUs, and hey, we'll take it.

01:17:10.292 --> 01:17:12.532
buzz. I really do hope that you join us in
in person.

01:17:12.537 --> 01:17:14.097
All right, this has been this week's buzz.

01:17:14.457 --> 01:17:15.977
There's a lot for us to talk about.

01:17:16.537 --> 01:17:17.957
Let's move on to...

01:17:19.177 --> 01:17:22.776
Let's move on to the world of open source
and JEV.

01:17:23.137 --> 01:17:25.776
I really want to talk about that
correlation.

01:17:41.137 --> 01:17:44.216
Open source AI, let's get it started.

01:17:47.617 --> 01:17:48.776
All right, bring it back.

01:17:50.057 --> 01:17:53.296
Everybody, uh, bring it back to the show.

01:17:55.977 --> 01:17:56.977
All right.

01:17:58.697 --> 01:17:59.697
So,

01:17:59.897 --> 01:18:04.296
one week after Jeff dropped from TypeSafe
AI, the open source clones are here, and

01:18:04.377 --> 01:18:06.736
one of them is already beating JEV in some
benchmarks.

01:18:08.417 --> 01:18:11.096
Last week I told you that JEV is a ChatGPT
moment.

01:18:12.377 --> 01:18:16.576
JEV from TypeSafe AI, a new, uh, System 1
type model.

01:18:16.857 --> 01:18:17.857
We had

01:18:18.017 --> 01:18:22.296
Ali Labs, the the dev role for JEV, talk
to us about the different, uh,

01:18:23.497 --> 01:18:26.016
the primitives that that you can use.

01:18:27.017 --> 01:18:29.257
And I told you about that, open source
will copy it fast.

01:18:29.537 --> 01:18:30.697
Well, it took less than a week.

01:18:30.977 --> 01:18:31.977
Over the weekend,

01:18:32.297 --> 01:18:36.816
more than tons of JEF compatible projects
showed up, and they all speak the same

01:18:37.057 --> 01:18:38.417
format, decision-making format.

01:18:38.757 --> 01:18:41.377
And so you point it to the official SDK,
and it just works.

01:18:41.977 --> 01:18:47.717
Uh, one is Classifier Dev, and, and uh,
JEF Bench is something that I

01:18:47.817 --> 01:18:51.537
wanted to show you. So, uh, I will pull up
Gev- GevBench.

01:18:51.737 --> 01:18:56.257
Meanwhile, uh, Nisten, I would love to
hear from you, your thoughts about Gev,

01:18:56.617 --> 01:19:02.556
because last week you were somewhat, uh,
skeptical/disappointed, but now I think,

01:19:03.177 --> 01:19:04.177
uh, you've played with it.

01:19:04.377 --> 01:19:06.196
What, what do you think about this new
paradigm?

01:19:07.497 --> 01:19:08.656
Uh, yeah,

01:19:09.327 --> 01:19:11.826
it actually brings the fun back in
computing

01:19:12.527 --> 01:19:13.976
because you're thinking

01:19:14.697 --> 01:19:19.187
more in terms of like TypeScript or
object-oriented thinking that, yes, I have

01:19:19.227 --> 01:19:22.546
all of these tasks and I just need to
categorize them really fast.

01:19:22.907 --> 01:19:26.946
And they could be images on the screen,
stuff you have to click, emails you have

01:19:26.987 --> 01:19:28.122
to write, and

01:19:28.859 --> 01:19:33.109
it's like now you just use an LLM in a
much faster way for this,

01:19:33.989 --> 01:19:37.009
and it's just excellent at feeding it

01:19:37.669 --> 01:19:43.368
big, large JSONs and then coming up with
other large JSONs, and it does the

01:19:43.509 --> 01:19:45.148
accuracy better than

01:19:46.189 --> 01:19:48.088
an open source LLM often.

01:19:48.629 --> 01:19:50.349
Yeah. Even the ones that you can run at
home.

01:19:51.029 --> 01:19:53.049
So this is, this is actually

01:19:53.749 --> 01:19:58.768
super useful, and uh, I'm really surprised
as to how close the open source

01:19:59.029 --> 01:20:02.488
alternatives are. It's just that JAB
itself does not really make mistakes, and

01:20:02.509 --> 01:20:07.008
the open source ones, eh, tend to have
about like 1% or like a little bit less

01:20:07.029 --> 01:20:12.069
than 1% pers- uh, mistake rate, which is
still a lot less than using an actual LLM,

01:20:12.549 --> 01:20:13.648
and even a,

01:20:14.269 --> 01:20:15.269
uh,

01:20:15.309 --> 01:20:16.428
like a somewhat good one.

01:20:16.909 --> 01:20:21.128
So, yeah, this is a, yeah, it's a new
paradigm in how you make the apps.

01:20:21.349 --> 01:20:23.249
It will just be in every app, basically.

01:20:23.869 --> 01:20:25.409
Server side, client side, yeah.

01:20:25.589 --> 01:20:26.589
This is like a

01:20:27.549 --> 01:20:30.229
new paradigm in in software engineering,

01:20:30.989 --> 01:20:33.528
where you can, because of the speed and

01:20:34.149 --> 01:20:35.689
performance and cost of

01:20:36.509 --> 01:20:39.189
System 1 models, JEF specifically, but
also open source ones,

01:20:39.949 --> 01:20:41.708
you can put them inside your

01:20:42.309 --> 01:20:44.749
execution chain in software and rely

01:20:45.549 --> 01:20:50.848
on the non-probabilistic, uh, how should I
say, execution of this, that it will work

01:20:50.989 --> 01:20:52.228
the next time your software runs.

01:20:53.108 --> 01:20:54.108
It is

01:20:54.309 --> 01:20:55.369
the microprocessor

01:20:56.709 --> 01:20:58.228
of the next era of software engineering.

01:20:59.349 --> 01:21:00.349
It is that big.

01:21:01.189 --> 01:21:04.268
Our whole timelines collectively were JEV
build.

01:21:04.469 --> 01:21:05.828
It broke through the bubble.

01:21:06.029 --> 01:21:08.429
My friends who are not into AI know about
JEV.

01:21:08.909 --> 01:21:12.549
Every, every engineer that I know is
thinking and thinking, oh, how can I use

01:21:12.629 --> 01:21:17.469
JEV more? Um, I've been completely Jeff
pilled since before we got Ali on the

01:21:17.489 --> 01:21:21.308
show. And then, uh, Swix, a friend of the
pod Swix, uh,

01:21:22.249 --> 01:21:27.149
hosted Diogo Almeida on his podcast on
Latent Space, and that is a conversation

01:21:27.329 --> 01:21:28.869
that I recommend everyone

01:21:29.529 --> 01:21:30.529
listening to.

01:21:30.869 --> 01:21:32.649
I would love to hear from you about, uh,

01:21:34.489 --> 01:21:37.809
how big is this deal, and also the open
source catching up.

01:21:37.849 --> 01:21:38.554
And

01:21:38.554 --> 01:21:39.654
Absolutely big. Yeah.

01:21:40.514 --> 01:21:46.113
Absolutely. Uh, look, the thing about LLMs
and and and all all of that

01:21:46.374 --> 01:21:51.313
is, yeah, they can do general things, and
you can just talk to them, and agents also

01:21:51.434 --> 01:21:52.733
figure out, but the thing is that it

01:21:53.394 --> 01:21:58.433
everything becomes probabilistic, and
there are ways to mitigate this and so on,

01:21:59.034 --> 01:21:59.994
but

01:21:59.994 --> 01:22:00.994
like there are

01:22:01.753 --> 01:22:04.613
building blocks that if you had

01:22:05.234 --> 01:22:10.433
a way to like compartmentalize and and,
you know, branch out and choose,

01:22:11.074 --> 01:22:13.794
uh, how to use e- easily,

01:22:14.794 --> 01:22:15.794
you can do

01:22:16.193 --> 01:22:17.193
many,

01:22:17.354 --> 01:22:19.313
many of the things that you do with LLMs
just

01:22:20.354 --> 01:22:23.714
without using the LLMs that are much more,

01:22:24.954 --> 01:22:29.574
you know, out there in, in terms of, uh,
all over the place, uh, in terms of how

01:22:29.753 --> 01:22:31.313
probabilistic they really are.

01:22:32.114 --> 01:22:33.594
And, uh, Jeff,

01:22:34.514 --> 01:22:35.974
just I, I didn't expect this.

01:22:36.234 --> 01:22:40.533
I knew, I know about all, uh, encoders
and, and so on and how they work, but

01:22:41.434 --> 01:22:45.214
the fact I, I didn't take into
consideration that if you train a really

01:22:45.314 --> 01:22:47.913
good one, probably a really big one, with
all the modern

01:22:48.554 --> 01:22:50.793
inference, uh, tricks and so on.

01:22:50.994 --> 01:22:52.453
And data, a lot.

01:22:52.914 --> 01:22:54.293
Absolutely. Data.

01:22:55.514 --> 01:22:59.394
If you do it well enough, you get to a
point it works

01:23:00.193 --> 01:23:01.834
really well, near perfect

01:23:02.494 --> 01:23:07.574
and nearly free. I mean, that's, I think
this is the big deal, that it's basically

01:23:08.394 --> 01:23:10.954
fast and basically free to use,

01:23:11.794 --> 01:23:17.673
so you can just, you know, just spam, spam
decisions in like in insane

01:23:17.834 --> 01:23:19.873
rate, like you did in your demos.

01:23:20.154 --> 01:23:23.394
I mean, it was the the rapid fire, it was
the the speed.

01:23:24.234 --> 01:23:28.233
Look, I'm all for training open source
ones, absolutely go for it, no problem.

01:23:28.474 --> 01:23:29.474
I'm just saying that

01:23:30.034 --> 01:23:32.873
it's not easy to get to these results, and

01:23:33.634 --> 01:23:39.154
I think we should wait a little bit before
we claim an open source replacement for

01:23:39.314 --> 01:23:40.274
JEV.

01:23:40.274 --> 01:23:42.514
Um, they are, they are impressive.

01:23:42.674 --> 01:23:47.594
I must say that, uh, they look quite
impressive, but I would wait for more

01:23:47.754 --> 01:23:53.074
benchmarks to, to, to measure them all out
in all sorts of different ways.

01:23:53.434 --> 01:23:55.634
Yeah. Uh, just because, look,

01:23:56.634 --> 01:24:00.193
J- the, the people behind JEV told you we
worked on this

01:24:00.794 --> 01:24:04.334
full time. Let's just talk about the
people behind JEV.

01:24:04.434 --> 01:24:10.073
Yeah. Diogo, is the guy behind ChatGPT and
RLHF, and his takes, you should listen to

01:24:10.114 --> 01:24:15.113
the show because Swix takes him deep on
the process of how did you get there?

01:24:15.594 --> 01:24:19.213
Basically, the summary is he was thinking
about who will use

01:24:20.594 --> 01:24:21.594
computers.

01:24:22.674 --> 01:24:26.753
And and he's like, okay, AIs will use
computers way more than human,

01:24:27.354 --> 01:24:30.594
but ChatGPT and LLMs were trained for
humans to use.

01:24:31.274 --> 01:24:32.633
So who's gonna talk to other AIs?

01:24:33.154 --> 01:24:38.554
AIs. But they're now built to write text
and respond as a chat that says, oh, I'm

01:24:38.594 --> 01:24:42.334
so sorry about that, my friend, but
computers talk in fucking language machine

01:24:42.394 --> 01:24:47.554
code. So this is why he's like, okay, we
need to build, uh, AI for computers to, to

01:24:47.674 --> 01:24:49.313
talk to. And how do computers talk?

01:24:49.714 --> 01:24:51.994
Formatted structured structure with JSON
files.

01:24:52.354 --> 01:24:53.354
He has a, a

01:24:53.993 --> 01:24:57.754
story there where Sam Altman told him, go
work on this, and he's like, no.

01:24:58.074 --> 01:25:00.953
And then he worked on it for 2 years after
the pre-training was done.

01:25:01.514 --> 01:25:05.473
But also, the important, going back, the
important note about open source Yam is

01:25:05.554 --> 01:25:09.554
that it may not, uh, get quite to that
level, but it gets very close.

01:25:09.569 --> 01:25:12.693
Folks, I wanna continue the show a little
bit, so while this runs, I wanna bring up,

01:25:12.794 --> 01:25:13.794
uh, Florian.

01:25:14.434 --> 01:25:16.314
Let's say hi to Florian, everyone.

01:25:16.594 --> 01:25:17.874
Florian, say hello to the folks.

01:25:18.434 --> 01:25:19.753
Welcome. Hey, everyone.

01:25:20.594 --> 01:25:21.934
How are you? Thanks for having me. room.

01:25:21.974 --> 01:25:25.439
If Yeah. Florian, I want you to tell us
what you built last week.

01:25:25.539 --> 01:25:26.758
I think this is a super exciting part.

01:25:26.859 --> 01:25:29.972
We just showed it off, but like I would
love to hear from what you do.

01:25:30.290 --> 01:25:36.209
Okay, yeah, I mean, last week has been
kind of insane, um, because I've kind of,

01:25:37.130 --> 01:25:39.470
um, bare- barely slept.

01:25:39.650 --> 01:25:42.530
Like, I've been obsessive, um, but I loved
it.

01:25:42.650 --> 01:25:43.949
Like, it was like, um.

01:25:44.330 --> 01:25:45.330
What were you obsessive

01:25:45.610 --> 01:25:46.949
about? Why did you need to sleep?

01:25:47.090 --> 01:25:50.810
Yeah, Jeff, ch- I still feel like Jeff is
completely underrated still.

01:25:50.830 --> 01:25:54.009
Like, I mean, Still. you can't call it
underrated because it got like 40 million

01:25:54.109 --> 01:25:57.969
views, everybody talked about it for a few
days, but still, I believe it's kind of

01:25:58.030 --> 01:25:59.030
underrated because,

01:25:59.735 --> 01:26:01.940
what just so many people haven't really

01:26:02.640 --> 01:26:03.640
fully

01:26:04.079 --> 01:26:05.079
understood is,

01:26:05.920 --> 01:26:11.300
um, it unlocks a lot of use cases, like
because of this unique combination of,

01:26:11.629 --> 01:26:15.745
three things, like it's incredible fast,
so in fact they can, like, use it to play

01:26:15.926 --> 01:26:19.585
Doom and win a level, like you have to be
really fast, multiple

01:26:20.366 --> 01:26:24.505
times per second. And then it's incredible
cheap as so you wouldn't do it with any

01:26:24.626 --> 01:26:28.286
normal large language model because it
would just blow your bill and you couldn't

01:26:28.346 --> 01:26:32.235
pay for it. And then it's also incredible
intelligence, so that's like it's, it's,

01:26:32.336 --> 01:26:34.155
um, the, the combination of these three.

01:26:34.516 --> 01:26:37.165
You can't just look at one of these
dimensions and say, oh, I've got got a

01:26:37.206 --> 01:26:39.346
chat competitor. You you gotta do all
three.

01:26:39.986 --> 01:26:40.986
And so, like,

01:26:41.546 --> 01:26:45.045
like right after it was released, I
started to build a benchmark because I was

01:26:45.266 --> 01:26:46.266
totally aware that

01:26:47.146 --> 01:26:50.505
the open source community will start to
work on this right away.

01:26:50.866 --> 01:26:54.005
It's obviously a small model, if it's fast
like that, it's gotta be a small model,

01:26:54.045 --> 01:26:56.525
and small models are kind of affordable to
train.

01:26:56.966 --> 01:27:01.905
So basically, at least if you, if you
believe you can just like hack an existing

01:27:02.006 --> 01:27:03.605
open source model and make it a Jeff.

01:27:04.646 --> 01:27:09.575
And yeah, it came as I expected, like the
next day it was like 10 or 20 competitors,

01:27:09.656 --> 01:27:12.095
now it's more than a 70 on this benchmark.

01:27:13.351 --> 01:27:16.151
I've, I've like been benchmarking these
things.

01:27:16.271 --> 01:27:20.231
I've been fending off, uh, approaches to
gain the benchmark.

01:27:20.671 --> 01:27:24.550
People have tried to, to hack me, tried to
exploit my machine.

01:27:24.911 --> 01:27:29.271
People have tried to steal the, the
benchmark test set, all that stuff.

01:27:29.871 --> 01:27:31.971
But it's super exciting and it's super
fun.

01:27:32.471 --> 01:27:36.050
I've been like, I- I've had like more than
half a million views, and I'm really a

01:27:36.171 --> 01:27:38.490
small, small, unimportant

01:27:39.151 --> 01:27:41.350
Twitter account, you know? Very important
Twitter.

01:27:41.751 --> 01:27:44.331
Florian, let's not, let's not sell
ourselves short.

01:27:44.431 --> 01:27:46.911
I think that benchmarking is an incredibly
important part.

01:27:47.190 --> 01:27:50.810
Uh, if Wolfram was here, he would like
very, very much agree with what you're

01:27:50.851 --> 01:27:54.720
doing. folks. I really wanted to shout out
Florian for benchmarkheaven.com.

01:27:55.021 --> 01:27:58.581
Please go there, and JeffBench
specifically, jumping on this new thing,

01:27:58.901 --> 01:28:01.160
uh, despite the creator of, uh,

01:28:01.821 --> 01:28:04.181
TypeSafe and Gev hates benchmarks as a
concept.

01:28:04.306 --> 01:28:07.313
Yeah. I think it's a very incredible thing
for folks to, realize, like, hey, not all

01:28:07.433 --> 01:28:09.963
open source are the same, not everybody
trains on on the same data.

01:28:10.499 --> 01:28:13.418
I fully agree with you that open source is
important from that perspective because

01:28:13.459 --> 01:28:18.074
people may not want to send data into
TypeSafe, despite TypeSafe being like the

01:28:18.079 --> 01:28:19.688
the next maybe, frontier lab.

01:28:20.049 --> 01:28:22.929
People may not want to send or may not be
able to send.

01:28:23.489 --> 01:28:25.188
We're talking about medical stuff with
Nisten.

01:28:25.209 --> 01:28:26.728
We're talking about different things on-
prem.

01:28:26.769 --> 01:28:30.748
Like, people run open source specifically
for we cannot,

01:28:31.569 --> 01:28:33.948
uh, legally send stuff, especially
Florian.

01:28:34.009 --> 01:28:35.328
Uh, I don't know where you're from.

01:28:35.369 --> 01:28:36.909
I think Germany, right? Uh.

01:28:37.369 --> 01:28:41.738
Yeah, Germany. So Europe has like very
strict laws and like Mistral, the whole

01:28:41.839 --> 01:28:44.689
reason for their existence is that they're
the only one who can sell to like European

01:28:44.729 --> 01:28:48.589
governments. Uh, people want stuff on-prem
or approved v- versus this.

01:28:48.649 --> 01:28:50.168
And when you, when JEV K5

01:28:50.769 --> 01:28:56.584
on the benchmark gets 62% of the stuff
that Florian is is testing, uh, it looks

01:28:56.669 --> 01:28:59.308
the most... Where's Leia on this, by the
way, Florian?

01:28:59.389 --> 01:29:01.129
Where's Leia? Because Nisten, that's the
one.

01:29:01.149 --> 01:29:04.328
Oh, it's it's not performing well, very
well on this benchmark, to be honest.

01:29:04.349 --> 01:29:08.349
Ooh. Um, it it used to be on place 2 for a
moment, Yeah.

01:29:08.389 --> 01:29:12.508
but um, like I I uh improved um the the
benchmark.

01:29:12.589 --> 01:29:17.728
You gotta really keep up, and now um new
new competitors have shown up, so it's

01:29:17.829 --> 01:29:23.768
pretty pretty down. I mean, what what what
Leia is is good in is it's incredible fast

01:29:23.869 --> 01:29:28.228
and and can run on on a CPU even, so it's
a very valuable thing to have.

01:29:28.509 --> 01:29:32.668
Yeah. But if you look at the, I've I've
got these radio charts on the page where

01:29:32.709 --> 01:29:36.308
you can, like, look at multiple dimensions
very, very far down

01:29:37.109 --> 01:29:41.969
of, of, like, tests, and there you can
see, like, Leia is just not as smart, you

01:29:41.989 --> 01:29:46.209
know? It's, it's, uh, yeah, try to find it
in this I'll find it, I'll find it.

01:29:46.249 --> 01:29:48.508
huge list. Leia 36 at the moment.

01:29:49.189 --> 01:29:52.838
So if you look at the, different types of,
questions,

01:29:53.799 --> 01:29:56.758
maybe scroll a little bit further down
even, there is another radio.

01:29:58.039 --> 01:30:00.169
Left. the left one is, is very telling.

01:30:00.189 --> 01:30:05.389
Mm-hmm. On the hard questions, like the
red part is where it kind of, yeah, what

01:30:05.449 --> 01:30:09.688
it achieved. And Jeff is just so much
smarter in all these different types of

01:30:09.829 --> 01:30:11.028
questions. Yeah. So Yeah.

01:30:11.364 --> 01:30:15.949
mean, Leia is great if you know that your,
your tasks are not so complicated, but if

01:30:16.009 --> 01:30:16.929
you,

01:30:16.929 --> 01:30:19.388
as I've said, you need to fulfill all of
the three axes.

01:30:19.649 --> 01:30:23.989
You're gonna be intelligent as well to be
actually a chat class, um, competitor.

01:30:25.249 --> 01:30:26.929
Yeah, so. Yeah, on, on,

01:30:27.569 --> 01:30:30.289
sorry. You have to, yeah, you have to
benchmark it.

01:30:30.369 --> 01:30:36.009
So as long as I see that it gets over like
99.5, it's pretty good for that use, like

01:30:36.289 --> 01:30:38.568
comparing that JSON. or doing some
computer use.

01:30:38.889 --> 01:30:43.908
But it's definitely not a generalist, but
this is not, uh, anywhere there, like to

01:30:43.969 --> 01:30:44.969
play games and stuff.

01:30:45.689 --> 01:30:49.949
So, folks, Yeah. uh, open source stuff
will catch up, already is catching up.

01:30:50.009 --> 01:30:53.408
It's incredible. Uh, if you want to follow
along with this, go to

01:30:53.569 --> 01:30:55.954
benchmarkheaven.com. Florian, thank you so
much for joining us.

01:30:55.999 --> 01:30:57.538
So, uh, Florian, thank you so much for
coming.

01:30:57.559 --> 01:30:59.158
You're right. You're right. Uh, Thank you.

01:30:59.279 --> 01:31:00.799
Have a good day. Yeah, thanks, man.

01:31:00.899 --> 01:31:02.438
Um, all right, folks, let's talk about
voice.

01:31:02.619 --> 01:31:03.619
Let's talk about

01:31:04.419 --> 01:31:09.198
voice. Google says it now has the best
text-to-speech model in the world, and you

01:31:09.239 --> 01:31:10.518
can clone your voice.

01:31:11.439 --> 01:31:16.998
Yes, Google, the big company, will let you
clone your voice from 30 seconds of audio,

01:31:17.159 --> 01:31:19.319
which we should do now live on stage.

01:31:20.389 --> 01:31:24.909
Two new speech models, Gemini 3.8 Flash
TTS and Flash Lite TTS.

01:31:25.229 --> 01:31:30.868
Google says they're number 1 and number 2
on Hume AI's quality index, with Flash at

01:31:30.909 --> 01:31:33.689
the top of the Hume voice design benchmark
as well.

01:31:33.749 --> 01:31:35.388
You can design voices or clone them.

01:31:35.869 --> 01:31:39.048
Flash is for voice design, character
acting, and long narration.

01:31:39.189 --> 01:31:42.349
Flashlight is for high volume dubbing and
voice agents.

01:31:42.709 --> 01:31:46.249
And Logelkin Patrick, our friend of the
pod, says both are cheaper than the old

01:31:46.309 --> 01:31:50.829
3.1 Flash TTS, and you can do 2 speakers
in one request, and you can drop tags like

01:31:51.029 --> 01:31:53.509
laughs and hmm and listener and sounds.

01:31:53.909 --> 01:31:58.509
The voice replication is the part people
will pick up because it is a huge company

01:31:58.569 --> 01:31:59.828
that allows you to clone your voice.

01:32:00.149 --> 01:32:03.689
30 seconds of audio, only adults, they
can, they don't allow to to to change

01:32:03.749 --> 01:32:05.388
kids' voices, uh,

01:32:06.029 --> 01:32:10.028
and the voice owner has to record the
consent, and every clip gets a SynthID

01:32:10.189 --> 01:32:14.709
watermark, which I think Google is very
incredibly on leading, and cloned voices

01:32:14.789 --> 01:32:16.069
get a C2PA credential.

01:32:16.149 --> 01:32:18.469
So Google clearly thinking about the
deepfake problem.

01:32:19.469 --> 01:32:24.488
Voice remixing is coming soon, and
somebody on rbard in Reddit already built

01:32:24.529 --> 01:32:28.029
a live AI podcast where listeners talk to
the hosts, and the top reply was

01:32:28.149 --> 01:32:29.819
NotebookLM had this a year ago.

01:32:30.874 --> 01:32:36.314
will we have an AI assistant, uh, here on
the show with us with using Gemini's TTS,

01:32:36.674 --> 01:32:37.953
or will it be Muse?

01:32:38.074 --> 01:32:39.794
I don't know, but we definitely need to

01:32:40.794 --> 01:32:42.083
show you and hear this.

01:32:43.049 --> 01:32:45.208
Okay, so let's listen to the meditation
guide.

01:32:47.529 --> 01:32:49.849
This is the design your voice thing, okay?

01:32:49.929 --> 01:32:51.448
Yeah, the naming is meditation guide.

01:32:51.809 --> 01:32:55.908
A soft, airy female voice in her 40s,
lower middle pitch, very slow, unhurried

01:32:55.968 --> 01:32:57.688
pace with gentle pauses for breath.

01:32:58.089 --> 01:33:01.109
Uh, you can hit improve my prompt and
it'll generate it, but let's generate this

01:33:01.169 --> 01:33:02.169
voice.

01:33:02.329 --> 01:33:03.399
It's three voice options. voices.

01:33:03.459 --> 01:33:05.698
Let's hear for the first one. And so,

01:33:06.578 --> 01:33:07.938
in this moment,

01:33:09.018 --> 01:33:09.979
just

01:33:09.979 --> 01:33:10.979
breathe

01:33:11.659 --> 01:33:12.659
and

01:33:12.939 --> 01:33:15.498
find that space within.

01:33:16.939 --> 01:33:17.859
It's

01:33:17.859 --> 01:33:18.998
always here,

01:33:20.139 --> 01:33:20.979
and

01:33:20.979 --> 01:33:22.619
you don't have to try.

01:33:23.259 --> 01:33:24.259
This is great

01:33:24.659 --> 01:33:26.618
because of the breathiness of this.

01:33:26.819 --> 01:33:27.859
I could hear the,

01:33:29.179 --> 01:33:30.978
especially in in headphones, you can hear
this.

01:33:31.099 --> 01:33:35.659
Uh, but yeah, this is not our energy right
now, so I wanna, I wanna, like, bring up

01:33:35.859 --> 01:33:39.359
the late night DJ. Ready to make something
amazing?

01:33:41.379 --> 01:33:43.938
And the mad scientist. Got a project in
mind?

01:33:45.259 --> 01:33:46.559
So, folks, um,

01:33:47.419 --> 01:33:51.979
I, I had my voice in there, but I can't
find it, so we will definitely talk with

01:33:52.018 --> 01:33:53.419
the AI studio folks.

01:33:53.699 --> 01:33:56.738
I don't want to clone this live on air
because, uh, obviously the consent thing,

01:33:57.018 --> 01:34:00.698
I need to record the consent in there, and
I don't want my consent being recorded,

01:34:00.819 --> 01:34:02.378
although I think that that's going to be
faked.

01:34:02.578 --> 01:34:03.539
Um,

01:34:03.539 --> 01:34:04.339
but

01:34:04.339 --> 01:34:05.718
the, the voice replication

01:34:06.779 --> 01:34:10.719
is a thing that we told you about, what, 2
years ago, that nothing is going to happen

01:34:11.899 --> 01:34:16.038
in the, like, deepfakes, the world, et
cetera, big, and then the big companies

01:34:16.079 --> 01:34:20.659
will follow suit. I remember there was a
fear in the industry about, hey, oh no,

01:34:21.059 --> 01:34:23.699
voice cloning is coming. Yam, I don't know
if you remember that episode or not,

01:34:23.779 --> 01:34:25.219
Nisten. And we told you

01:34:25.859 --> 01:34:29.159
nothing is gonna happen. If anything, it's
for the best because people will know that

01:34:29.219 --> 01:34:32.578
this ability is out there so they don't
trust phone calls from grandmas anymore,

01:34:33.179 --> 01:34:34.219
and nothing has happened.

01:34:34.579 --> 01:34:37.459
And I think there's a parallel there, and
I think we'll finish on this.

01:34:38.019 --> 01:34:39.558
There's a parallel there between

01:34:41.619 --> 01:34:47.578
the fears that we are getting told by, oh,
AI is gonna kill us all, and the reality

01:34:48.019 --> 01:34:51.099
that we see. Open sourcing GPT-2 did
nothing.

01:34:51.659 --> 01:34:52.978
Open sourcing, uh,

01:34:53.659 --> 01:34:56.039
Stable Diffusion didn't break the world.

01:34:56.659 --> 01:35:00.258
Open sourcing and allowing everybody to
clone their voice and use that

01:35:00.979 --> 01:35:02.578
didn't sh- break the world.

01:35:02.979 --> 01:35:04.839
We're still here, better than ever.

01:35:04.979 --> 01:35:08.038
I think there's a very great parallel
between this and all the AI is gonna kill

01:35:08.059 --> 01:35:09.538
us all. Nobody fucking knows.

01:35:10.099 --> 01:35:13.979
Nobody, as I say again, and I keep saying
this, this is my line, nobody could have

01:35:14.019 --> 01:35:19.139
predicted a Austrian dude that has, you
know, done his thing and doesn't need to

01:35:19.179 --> 01:35:22.638
work anymore, creating the world's next
pattern that's called assistant that

01:35:22.699 --> 01:35:23.699
everybody's cloning.

01:35:24.379 --> 01:35:28.579
Meta Muse is a very, like, inspired thing
from OpenClaw Hermes as well.

01:35:28.779 --> 01:35:32.149
All of the assistants now do a thing
there's some rando, like a great, great

01:35:32.509 --> 01:35:35.908
visionary Peter Steinberger, uh, did in
his, like, apartment.

01:35:36.589 --> 01:35:40.109
Nobody in the doomer world could have
predicted this is where it's gonna go

01:35:40.189 --> 01:35:42.068
because nobody can predict where we're
going.

01:35:42.429 --> 01:35:45.509
And so this is why we're here to tell you
about all the cool stuff in a positive

01:35:45.549 --> 01:35:49.348
way, and I think voice cloning is a very
positive way to finish on this.

01:35:50.348 --> 01:35:54.288
A few housekeeping notes that of the stuff
that we skipped this week that we didn't

01:35:54.309 --> 01:35:56.069
get quite to talk about lightning round.

01:35:56.348 --> 01:36:01.628
Uh, Qwen 3.8 OmniFlash released a, a Omni
model with a 1 million context window.

01:36:02.029 --> 01:36:07.768
Uh, Qwen 3.8 Live Translate has released a
real-time interpreter model that we can

01:36:07.909 --> 01:36:09.808
pull up samples here for every language.

01:36:09.989 --> 01:36:12.468
Maybe we'll build this into uh,
ThursdAI.live.

01:36:13.189 --> 01:36:14.189
Uh,

01:36:14.348 --> 01:36:15.789
those are the 3 main ones.

01:36:15.869 --> 01:36:20.389
In vision and video, Qwen released a 7
billion parameter model, Qwen Image 2.1,

01:36:20.869 --> 01:36:23.189
which is going around and doing some
incredible stuff.

01:36:23.789 --> 01:36:25.048
And AI art and diffusion.

01:36:25.229 --> 01:36:26.588
Flux 3 action model.

01:36:27.205 --> 01:36:32.285
Black Forest Lab released the FLAX model
open source, uh, an action model, 7

01:36:32.605 --> 01:36:35.704
billion parameter world action model
model, which is incredible.

01:36:35.965 --> 01:36:39.445
It takes camera frames and robot state and
text instructions, predicts the next 32

01:36:39.525 --> 01:36:41.444
actions and how the scene will change.

01:36:42.525 --> 01:36:45.925
And, uh, Peter, I promised you this at the
beginning of the show, and I promised

01:36:46.005 --> 01:36:47.344
people. There's this thing called

01:36:49.500 --> 01:36:50.579
Act 486.

01:36:51.340 --> 01:36:54.360
This is, I think, uh, this is the thing
that we'll

01:36:55.060 --> 01:36:56.060
finish on

01:36:56.380 --> 01:36:58.219
because I think it's incredible.

01:36:58.560 --> 01:37:00.040
So, like, this is a pre-recorded clip.

01:37:00.835 --> 01:37:03.555
It's actually just a roofing torch with an
air rifle cover.

01:37:04.315 --> 01:37:06.794
I don't get that. What don't you
understand, Leah?

01:37:07.155 --> 01:37:08.514
How a flamethrower works?

01:37:09.355 --> 01:37:12.355
So I'll draw the propane canister feeding
a pressure line.

01:37:12.915 --> 01:37:15.435
I want to pause here, uh, and say two
things.

01:37:15.555 --> 01:37:19.794
Beside the fact that they got an OnlyFans
model to to to to to market this for

01:37:19.955 --> 01:37:24.754
nerds, besides that point, what's
happening on our screen right now is, eh,

01:37:25.034 --> 01:37:28.994
somebody's watching the Lex Friedman Elon
Musk, uh, podcast.

01:37:29.075 --> 01:37:31.635
They they pre-recorded this, they, they
pre-baked this.

01:37:32.355 --> 01:37:34.954
Presses on Elon Musk and asks a question,

01:37:35.555 --> 01:37:37.395
and so then they clone his voice.

01:37:37.955 --> 01:37:40.594
I don't think he gave him permission, by
the way, to clone this voice, and I don't

01:37:40.635 --> 01:37:42.115
think they used the Google TTS for that.

01:37:42.885 --> 01:37:44.005
They ask the

01:37:44.965 --> 01:37:46.804
character in that video questions.

01:37:47.165 --> 01:37:50.604
The character turns to them, starts
answering the questions,

01:37:51.285 --> 01:37:53.284
and then also shows How a flamethrower
works?

01:37:54.045 --> 01:37:58.484
So, like, the model throws, like, a real-
time visualization about flamethrower

01:37:58.605 --> 01:38:00.245
stuff. Gas exits the nozzle.

01:38:00.565 --> 01:38:02.985
And he knows about the person that talks
to them.

01:38:03.005 --> 01:38:04.744
He's like, what do you, don't you
understand me?

01:38:04.765 --> 01:38:06.845
Take out Starship and describe it for me.

01:38:09.125 --> 01:38:11.084
Here's Starship version 3.

01:38:11.245 --> 01:38:16.484
The stainless steel shell is built for
heat, strength, and reuse, with 4 flaps

01:38:16.685 --> 01:38:17.865
steering it through atmosphere.

01:38:17.885 --> 01:38:20.564
Open it up and describe its components for
me, please.

01:38:24.805 --> 01:38:29.364
So this is like a beautiful interaction
moment, uh, because she pressed on, uh,

01:38:29.525 --> 01:38:30.525
Elon Musk's face,

01:38:31.085 --> 01:38:35.440
and then you can see him nodding while,
the thing generates the thing, and he's

01:38:35.455 --> 01:38:38.025
fading in and out. I don't know if you
guys noticed this, but like he's fading in

01:38:38.065 --> 01:38:39.480
and out because he's in loading state.

01:38:39.480 --> 01:38:41.654
I, I find that this interaction is
incredible.

01:38:41.675 --> 01:38:42.675
I will play this again.

01:38:43.035 --> 01:38:44.175
Steering it through atmosphere.

01:38:44.215 --> 01:38:46.614
Open it up and describe its components for
me, please.

01:38:46.634 --> 01:38:47.634
Like, look at this.

01:38:49.595 --> 01:38:53.735
He's like acknowledging, and he's fading
in and out like a person in loading state.

01:38:53.755 --> 01:38:54.755
This is crazy.

01:38:55.275 --> 01:38:56.615
He just looks so sad.

01:38:56.795 --> 01:38:58.994
It looks like he's been defeated, you
know.

01:38:59.075 --> 01:39:04.914
But look at this. He is now, the the video
model generates the starship model that he

01:39:04.975 --> 01:39:07.515
is talking about, and he is like rotating.

01:39:07.955 --> 01:39:11.794
Folks, this is the dream. This is the Game
of Thrones, and you're editing your own

01:39:12.035 --> 01:39:13.694
ending dream that we want.

01:39:13.755 --> 01:39:18.855
This is the demo of that. This is the demo
of talking to any type of, like, video and

01:39:19.075 --> 01:39:22.895
making it your own. Uh, this is, I haven't
gotten access to this.

01:39:23.195 --> 01:39:24.794
Uh, Act 486.

01:39:24.955 --> 01:39:26.754
Let's see what else they have in their
demo.

01:39:27.115 --> 01:39:30.855
Uh, oh, she she goes, she goes crazy in
here.

01:39:30.995 --> 01:39:32.754
Like, okay, let's finish on this because
this is crazy.

01:39:32.795 --> 01:39:35.994
Okay. Mars makes those two futures feel
tangible,

01:39:36.755 --> 01:39:40.895
and one future is we're out there among
the stars and things we read about and see

01:39:41.035 --> 01:39:42.115
in science fiction movies.

01:39:42.355 --> 01:39:47.254
So for folks who are listening, uh, she
now asked the video to put both Joe Rogan

01:39:47.315 --> 01:39:49.534
and Elon Musk in SpaceX suits on Mars.

01:39:49.675 --> 01:39:52.795
These are true. What would life on Mars
look like?

01:39:57.035 --> 01:39:59.735
Both not. All right, let's see how daily
life works.

01:40:00.075 --> 01:40:02.795
Life would happen mostly inside these
habitats.

01:40:03.475 --> 01:40:06.395
How does that work? Do people get upset at
you if you do certain things?

01:40:07.675 --> 01:40:09.034
Uh, don't smoke that.

01:40:09.475 --> 01:40:10.835
You do not want to smoke that.

01:40:11.835 --> 01:40:12.795
Uh,

01:40:12.795 --> 01:40:16.015
yeah, going for the viral moment, there's
a video, uh, there's a a famous thing

01:40:16.035 --> 01:40:19.315
where Elon Musk smokes some weed on Joe
Rogan, and she's like, oh, no, don't smoke

01:40:19.355 --> 01:40:21.755
this, and the result is really hilarious.

01:40:26.635 --> 01:40:27.794
Okay, I won't.

01:40:29.835 --> 01:40:30.835
You're welcome.

01:40:31.595 --> 01:40:32.595
I...

01:40:33.475 --> 01:40:35.835
I think with no reactions, folks, this is
the world we're in.

01:40:36.075 --> 01:40:39.875
You can now talk to videos and have them
just do whatever you want.

01:40:40.715 --> 01:40:44.215
If you enjoyed the show, first of all,
thank you so much for our live producer

01:40:44.275 --> 01:40:48.375
assistant that did a lot of work, a lot of
lifts, so all the notifications and stuff.

01:40:48.675 --> 01:40:52.794
Uh, second of all, huge thanks for, uh,
guest, uh, impromptu guest Florian S.

01:40:53.035 --> 01:40:55.515
for jumping on with, uh, Benchmark Heaven.

01:40:55.755 --> 01:40:57.015
Uh, thank you, Nisten.

01:40:57.195 --> 01:40:58.975
Uh, thank you, Peter. Thank you, Yam.

01:40:59.015 --> 01:41:03.474
Thank you, Wolfram, who joined us, uh, and
from and Mazir Penahi from AI Engineer at

01:41:03.675 --> 01:41:08.434
Paris. And folks, if you missed any part
of the show, you can rewatch the whole

01:41:08.515 --> 01:41:13.154
show as it was live on ThursdAI.live, or
you can subscribe to a newsletter

01:41:13.835 --> 01:41:17.135
or our YouTube, which, if you aren't
subscribed to our YouTube, we're doing a

01:41:17.195 --> 01:41:19.915
lot of work. It will really help us if you
hit that sub button.

01:41:20.035 --> 01:41:24.005
It's completely free for you and costs you
nothing, and you can, see notifications

01:41:24.085 --> 01:41:25.164
when we post new clips.

01:41:25.565 --> 01:41:26.565
And,

01:41:27.050 --> 01:41:30.909
next week show is going to be live from
Mosconi's house, a little different time.

01:41:31.010 --> 01:41:32.629
So we're gonna go 11 a.m.

01:41:32.770 --> 01:41:33.770
Pacific,

01:41:33.810 --> 01:41:37.699
so a little like, 3-ish hours later than
our usual, and go for around 2 hours.

01:41:37.860 --> 01:41:40.505
First hour is gonna be all about news, the
second hour is gonna be all about, uh,

01:41:40.505 --> 01:41:41.505
CoreWeave stuff.

01:41:41.690 --> 01:41:42.810
It was a pleasure of mine.

01:41:43.290 --> 01:41:44.350
What a crazy week

01:41:45.090 --> 01:41:46.909
to have a show that talks about AI, folks.

01:41:47.010 --> 01:41:48.329
Just an incredible, crazy week.

01:41:48.890 --> 01:41:49.979
And we'll see you here next week.