0:00

Hello and welcome to

0:02

the last week in

0:04

a podcast where you

0:06

can hear a chat

0:08

about what's going on

0:10

with AI. As usual

0:12

in this episode we

0:14

will be summarizing and

0:16

discussing some of last

0:18

week's most interesting AI

0:20

news. You can go

0:22

to the episode description

0:24

for all the links

0:26

and timestamps and also

0:28

to last week in

0:30

ai.com on your laptop to be

0:33

able to read those articles yourself

0:35

as well. As always, I'm one of

0:37

your hosts Andre Karenkov. I studied

0:39

AI in grad school and I

0:41

now work at the Genitive AI

0:44

startup Astrocade. And I'm your other host,

0:46

Jeremy Harris. I'm with Pladstone A.I. National

0:48

Security Company, which you know about if

0:50

you listen to the podcast, you also

0:53

know about Astricade now, a bunch of

0:55

you listen to, you know about all of this,

0:57

you know, what you don't know, though, is that this

0:59

morning, at the early hour, I think it

1:01

was like three or something in the morning,

1:03

I discovered that I have bats in

1:05

my house, which is fun, which is really

1:07

fun, especially when you have like a six-month

1:09

old, you have bats. and then you start

1:12

googling things. So anyway, you had

1:14

pest control come in. That's why, wow,

1:16

my hair looks like Cosmo Kramer

1:18

right now. I've just been running my

1:21

fingers through it for quite a bit.

1:23

So anyway, we got everything on for

1:25

showtime though, because the show will go

1:27

on. Yeah, but if you forget any

1:29

details wrong, you know, it's, it's the

1:31

shock residual shock, the bats, you know, maybe

1:34

that. It's just, I'll be on the lookout.

1:36

Well, let's do a quick preview

1:38

of what we'll be talking about

1:40

in this episode. It's going to

1:42

be a bit of a relaxed

1:45

one. There's nothing to sort of

1:47

world shattering, but a variety of

1:49

pretty interesting stories, tools and apps.

1:51

We have some new impressive

1:53

models out of China, some new

1:55

stuff from Open AI as well,

1:58

Google and FARPIC, everyone launched. stuff,

2:00

applications and business, as we often

2:02

do, going to be talking a

2:05

lot about hardware and GPUs, a

2:07

little bit about fundraising as well,

2:09

projects and open source, we'll be

2:11

talking about the model context protocol,

2:13

which has been all the rage

2:16

in the community recently, and a

2:18

couple new models. As usual, researching

2:20

investments, we got to talk about

2:22

reasoning techniques, inference time scaling techniques.

2:24

but also some new kind of

2:27

developments in the space of how

2:29

you implement your models. Policy and

2:31

safety, we have some more analysis

2:33

of what's going out of China,

2:35

US national security, things like that.

2:38

And finally, we will actually talk

2:40

a little bit about the world

2:42

of art and entertainment with some

2:44

news about copyright. So let's just

2:47

get straight into it. in tools

2:49

and apps. The first story is

2:51

about Baidu launching two new versions

2:53

of the Ernie model, Ernie 4.5

2:55

and Ernie X1. So Ernie initially

2:58

released two years ago and now

3:00

we have Ernie 4.5, presumably, I

3:02

don't know, it sounds like kind

3:04

of two coincided for, to PD

3:06

4. And then Ernie X1 is

3:09

the reasoning variant of Ernie that

3:11

Baidu says is on par with

3:13

Deep Seek R1, but at half

3:15

the price, and both of these

3:17

models are multimodal. They can process

3:20

videos, images, and audio as well.

3:22

They also say Ernie 4.5 is

3:24

kind of emotionally intelligent. They can

3:26

understand memes and satire, which is

3:28

interesting. So... I think we don't

3:31

have like a great sense of

3:33

the tool landscape in China is

3:35

my impression. I really wish I

3:37

would know like if you are

3:39

a user of a chat bot.

3:42

we got to charge BT or

3:44

Claude to give up queries, I

3:46

think it seems likely that Ernie

3:48

is sort of filling that role

3:50

and the fact that there's new

3:53

models and the fact that they're

3:55

really competitive price-wise is a big

3:57

deal. The like number one downloaded

3:59

app in China just switched to

4:02

a new AI chatbot that is

4:04

not Deep Seek. So things are

4:06

definitely moving. The big advantage here

4:08

with this launch seems to be

4:10

cost. At least that's what they're

4:13

leaning into with a lot of

4:15

the discussion around this. So the

4:17

goal that Baidu has, which Baidu

4:19

of course is China's, roughly China's

4:21

Google, right? They own search there.

4:24

Their goal is to progressively integrate

4:26

Ernie 4.5 and their X1 reasoning

4:28

model into all their product ecosystem,

4:30

including Baidu search, which is sort

4:32

of interesting. So we'll see a

4:35

rollout of the generative AI. capabilities

4:37

in that context. Yeah, so ultimately

4:39

it does come down to price

4:41

a lot of it. So for

4:43

context, there's really handy table in

4:46

one of the articles that looked

4:48

at this comparing GPT 4.5 per

4:50

token cost to Deep Seek v.

4:52

3 to Bernie, to Ernie 4.5,

4:54

it's quite interesting, right? So input

4:57

costs for input tokens, 75 bucks

4:59

for a million tokens. This is

5:01

for GPT 4.4. Deep Seek v3

5:03

that drops to basically 30 cents

5:06

earning 4.5 is about 60 cents

5:08

or so per 1 million tokens

5:10

So you know you're talking orders

5:12

of magnitude less Also the case

5:14

that these models are less performant

5:17

so that's sort of the tradeoff

5:19

there, but where things yeah, I

5:21

think just to give a bit

5:23

of a perspective Deep seek v

5:25

free is more comparable to something

5:28

like GPT for O in open

5:30

eyes, say models or free meaty

5:32

for instance where aggression isn't that

5:34

crazy it's maybe I forget one

5:36

dollar ish per million tokens so

5:39

they're comparable. GPD 4.5 is just

5:41

crazy crazy pricing compared to everything

5:43

else and that's a It's the

5:45

way to think about 4.5, I

5:47

think we touched on this a

5:50

couple episodes ago, but it's a

5:52

base model, but it's not a

5:54

base model for, let's say, mass

5:56

production, right? These are high, high

5:58

quality tokens, probably best used to

6:01

create things like synthetic data sets

6:03

or to answer very specific kinds

6:05

of questions, but you're not looking

6:07

at this as something that you

6:10

want to productize, just because you're

6:12

right. I mean, it's two orders

6:14

of magnitude more expensive than other

6:16

base models. Where you actually see

6:18

the lift here. especially for Ernie

6:21

X1, right, with this is the

6:23

reasoning model, is on the reasoning

6:25

side, right? So opening eyes is

6:27

roughly 50 times more expensive than

6:29

Ernie X1. Ernie X1 is about

6:32

half the cost of R1 for

6:34

input tokens, and actually that's also

6:36

true for output tokens. So it's

6:38

quite significant, especially again relative to

6:40

O1, and shows you. One of

6:43

two things, either Chinese engineering is

6:45

actually really really really that good,

6:47

or there's some state subsidy thing

6:49

going on in the background. I

6:51

think the latter is somewhat less

6:54

plausible at this point, though I

6:56

wouldn't rule it out. certainly there's

6:58

some amazing engineering making these these

7:00

margins possible and it's that's a

7:02

pretty remarkable thing here right I

7:05

mean the cost just collapsing for

7:07

reasoning this implies that there's some

7:09

reasoning specific engineering going on in

7:11

the background and you know you

7:14

should expect that to apply to

7:16

training as well as inference going

7:18

forward yeah and it's kind of

7:20

funny in a way there is

7:22

a parallel here between Baido and

7:25

Google where Google likewise has quite

7:27

competitive pricing, especially for Gemini to

7:29

flash thinking. So I could also

7:31

see it being, you know, just

7:33

a company strategy kind of thing.

7:36

Baidu is gigantic, they're printing money

7:38

with Search, so they could also

7:40

kind of eat the additional costs

7:42

to undermine something like Deep Seek,

7:44

which is a startup, right, too.

7:47

Lock in the market. But either

7:49

way, exciting news and I guess

7:51

if you're in China, I don't

7:53

believe you can use charge property.

7:55

So if nothing else, it's good

7:58

that they are comparable tools for

8:00

people to use. but not miss

8:02

out on the fun of these

8:04

advanced elements. I will say I

8:06

don't know that Baidu would be

8:09

subsidizing at the level of at

8:11

least their base model because they

8:13

are actually more expensive than Deep

8:15

Seek V3 on Ernie 4.5. Where

8:18

you see that flip is with

8:20

the reasoning models, which itself is,

8:22

yeah, that's kind of interesting, right?

8:24

I mean, to me at least,

8:26

that seems to imply something about

8:29

reasoning, like engineering for the like

8:31

computer architecture behind reasoning, or more

8:33

token efficiency and therefore compute efficiency

8:35

at the at the reason I

8:37

shouldn't say therefore maybe alternatively compute

8:40

efficiency at the reasoning stage but

8:42

you're right there's all kinds of

8:44

things start to muddy the water

8:46

is when you start thinking about

8:48

the economics of these things as

8:51

they represent a larger and larger

8:53

fraction of the corporate bottom line

8:55

even for big companies like I

8:57

do like Google these companies are

8:59

gonna be forced to show us

9:02

their hand in a sense right

9:04

they're gonna have to sell these

9:06

tokens for profit and we will

9:08

eventually learn what their actual margins

9:10

are what their actual margins are

9:13

It's debatable whether we're learning that

9:15

just yet. Yeah, I don't think

9:17

we are. It's very much unknown

9:19

and I haven't seen any kind

9:22

of strong analysis to explain it.

9:24

There's you know, yeah, it's just

9:26

a mystery what kind of tricks

9:28

people are pulling but I would

9:30

also kind of bet that the

9:33

margins aren't great. The one thing

9:35

we do know Deep Seek claimed

9:37

at least that they were making

9:39

a profit and had a positive

9:41

margin on their models and I

9:44

could see that not be the

9:46

case for, you know, for instance,

9:48

Open AI where their revenue is

9:50

in the billions, but the real

9:52

question is, are they actually making

9:55

a profit? Last thought on this

9:57

too, on the economic side, like

9:59

when we think about what it

10:01

means for Deep Sea to claim

10:03

that they're generating positive returns, I

10:06

think there's a... an important question

10:08

here about whether that's operating expenses

10:10

or CAPX factored in, right? So

10:12

we saw in their paper that

10:14

they famously talked about how they

10:17

trained V3 on six million dollars

10:19

of compute infrastructure. Now, or sorry,

10:21

on a six million dollar compute

10:23

budget, that was, it seems in

10:26

retrospect, the actual operating expenses of

10:28

running that compute, not the capital

10:30

expenses associated with. the tens of

10:32

millions of dollars as it would

10:34

have been of compute hardware. So

10:37

it's always hard to know like

10:39

what do you amortize? How do

10:41

you factor in what's apples to

10:43

apples? Yeah, it's hard to say

10:45

like deep seek is profitable, but

10:48

on a per token basis, just

10:50

for inference, I believe the claim

10:52

is they're making money which yeah,

10:54

on an all back spaces. Yeah,

10:56

interesting. Yeah. Moving right along, next

10:59

we have Open AI and they

11:01

are releasing new two new speech-to-tech

11:03

models, G.P.40 transcribed and G.P.40 mini

11:05

transcribe, which are basically replacing their

11:07

whisper models. Open has already had

11:10

this as a service for quite

11:12

a while. The exciting new thing

11:14

here is the text-to-speech model, G.P.40

11:16

mini dash-T-T-S, which is more long

11:18

lines of 11 labs where you

11:21

can produce very natural human sounding

11:23

speech and along with announcement of

11:25

the models open AI has also

11:27

launched and you cite open AI

11:29

dot f m which is a

11:32

demo site where you can go

11:34

and mess around and kind of

11:36

hear of the outputs and this

11:38

is kind of a fun trend

11:41

I gotta say where these companies

11:43

increasingly are launching these little fun

11:45

toys to get a sense for

11:47

what these models are capable of.

11:49

One last thing, again, probably should

11:52

comment on pricing. The pricing is

11:54

very competitive, the transcription. for GP40,

11:56

it's 0.6 cents per minute, 0.6

11:58

cents, so like 0.006 dollars, I

12:00

guess. And GD40, mini-TTS, is 1.5

12:03

cents per minute, which is much

12:05

lower than a competitor, like 11

12:07

Labs, for instance. So, yeah, I

12:09

think it's interesting to see opening

12:11

expanding their model suite to these

12:14

new domains where they're there sort

12:16

of less. focused. We've seen them

12:18

kind of move away from text

12:20

to image for instance, Delhi hasn't

12:22

had an update in forever. And

12:25

so I guess this makes a

12:27

lot of sense that they have

12:29

very competitive things to offer given

12:31

their investment in the advanced voice

12:33

mode in chatGPT. It's sort of

12:36

reminiscent of the problem that meta

12:38

faces, right, where they're, you know,

12:40

they reach like, whatever, 3 billion

12:42

people around the world, at a

12:45

certain point, when your market penetration

12:47

is so deep, one of the

12:49

only things you can do to

12:51

keep growing is to grow the

12:53

market. And so meta invests, for

12:56

example, in getting more people on

12:58

the internet, in other countries, like

13:00

in countries that don't have internet

13:02

access typically or have less of

13:04

it. And so they're literally just

13:07

trying to like grow the... pool

13:09

of people like they can they

13:11

can tap for this in the

13:13

same way I think there's a

13:15

lens on this that's similar right

13:18

so you're only able to interact

13:20

with chat gBT through certain modalities

13:22

or with open AI products for

13:24

certain modalities and by achieving greater

13:26

ubiquity by reaching into your life

13:29

more and making more of the

13:31

conversational tooling available to you that

13:33

that really does effectively increase their

13:35

their market right like you don't

13:37

have to be in front of

13:40

a computer necessarily or in the

13:42

same way or engaged in the

13:44

same way to use the product.

13:46

And obviously they've had other other

13:49

voice products before, but it's sort

13:51

of part of if I'm open

13:53

AI, I'm really thinking about multimodality,

13:55

but from the standpoint of increasing

13:57

the number of context, life context

14:00

in which I can reach you

14:02

and text to image still requires

14:04

you. to be in front of

14:06

a screen, same as writing text

14:08

on chat GPT, whereas audio is

14:11

just this like, you know, greater

14:13

reach for modality wise. So I

14:15

think strategically it's an interesting play

14:17

for them. ethically all kinds of

14:19

issues. I mean, you know, you

14:22

think about the modality of audio

14:24

as being one that is much

14:26

more intimate to humans and an

14:28

easier way to plug into your

14:30

inner world. And that's, I think,

14:33

something, you know, we look at,

14:35

like, what record did to people

14:37

just through text, right? The suicidal

14:39

ideation, the actual suicides, if nothing

14:41

else for open AI. There is

14:44

one figure by the way in

14:46

the article at least we're linking

14:48

to here and it's just a

14:50

piece of research looking at the

14:53

the word error rate comparisons across

14:55

leading models for different languages as

14:57

part of this kind of tooling

14:59

I just I find it really

15:01

interesting like Arabic and Hindi there's

15:04

a lot of struggle there those

15:06

are some of the worst performing

15:08

languages obviously English one of the

15:10

better performing ones I'd love to

15:12

see an overlay of this relative

15:15

to the amount of data that

15:17

was used to train the model,

15:19

so you can see in relative

15:21

terms, like, which languages are, in

15:23

a sense, like, harder for AI

15:26

to pronounce, to kind of speak.

15:28

I think there's something, anyway, linguistically

15:30

just fascinating about that, if nothing

15:32

else. So, anyway, overall, interesting launch,

15:34

and I think we're gonna see

15:37

more and more of this, right?

15:39

It's gonna be more expected to

15:41

have very high quality audio models

15:43

and linking them specifically to agents,

15:45

sort of Star Trek computer style.

15:48

Yeah, I guess one thing worth

15:50

noting on the kind of ethics

15:52

side is I don't believe they're

15:54

offering voice cloning technology, which is

15:57

where you can really get into

15:59

trouble very easily. So I think

16:01

opening up is being a little

16:03

careful these days in general to

16:05

not court controversy. That's part of

16:08

why it took them forever to

16:10

release SORA potentially. And in this.

16:12

API, this demo, they are releasing

16:14

something like a dozen voices you

16:16

can use with names like alloy,

16:19

ash, echo, fable, onyx, nova, kind

16:21

of, I don't know, human, I

16:23

guess we are not even trying

16:25

to make human sounding, and you

16:27

can also assign them a vibe

16:30

in this demo like cowboy, auctioneer,

16:32

all-timey, serene, with a lot of

16:34

this kind of steering that you

16:36

can do as well. So. Yeah,

16:38

I think it's pretty exciting. And

16:41

as ever will release some new

16:43

APIs, which really enables downstream of

16:45

Open AI and these other companies

16:47

for others to build exciting new

16:49

applications of AI. And onto a

16:52

few more quick stories. Next up

16:54

also Open AI, they have released

16:56

01-1-Pro into their developer API. it's

16:58

actually limited to developers who spent

17:01

of these five dollars on this

17:03

and it cost a hundred fifty

17:05

per million tokens for input and

17:07

six hundred dollars per million tokens

17:09

generated so that's very very high

17:12

prices obviously that's as we've said

17:14

GP 4.5 was seventy five dollars

17:16

for one million output tokens and

17:18

that's yeah two two orders of

17:20

magnitude easily above what you would

17:23

typically charge. Yeah, I'm trying to

17:25

think if it's two or three

17:27

orders of magnitude. It might be

17:29

approaching three orders of market here

17:31

actually. So yeah, interesting strategy here

17:34

from Open Eye where you haven't

17:36

seen any other companies release these

17:38

very expensive products yet and Open

17:40

Eye is increasingly doing that with

17:42

Chesapeake Pro, they're $200 per month

17:45

subscription with QB4. It makes me

17:47

wonder if this is an attempt

17:49

to become more profitable or if

17:51

this is them sort of testing

17:53

waters There could be various readings

17:56

I suppose. Yeah, it's also I

17:58

mean, it's interesting to note this

18:00

is not an order of magnitude

18:02

larger than what GPT-3's original pricing

18:05

was. I was just looking it

18:07

up in the background here to

18:09

check because I seem to remember

18:11

it being you know on it

18:13

back then it was priced per

18:16

per a thousand tokens with reasoning

18:18

models you tend to see more

18:20

per million tokens just because of

18:22

the number of tokens generated but

18:24

sort of reminds me in the

18:27

military or the history of the

18:29

military there is often this this

18:31

restriction where it's like people can

18:33

only carry I forget what it

18:35

is, 60 pounds or something of

18:38

equipment. And so over time, you

18:40

tend to see like the amount

18:42

of equipment that a soldier carries

18:44

doesn't tend to change or the

18:46

weight of it, but of course

18:49

the kind of equipment they carry

18:51

just changes to reflect technology. This

18:53

sort of seems similar, right? There's

18:55

like almost a perito frontier of

18:57

pricing, at least for the people

19:00

who are willing to reach for

19:02

the most intelligent products, you know,

19:04

and you're constantly reaching for it.

19:06

This is a push-forative. Even relative.

19:08

There's all kinds of feedback people

19:11

have been getting. There's complaints about,

19:13

though, this model struggled with Sudoku

19:15

puzzles, apparently, and optical illusions and

19:17

things like that. People say, you

19:20

know, at a certain point, anything

19:22

you launch at a high-priced point,

19:24

especially if you're opening eye, people

19:26

will complain that it's not like

19:28

super intelligence. And so... Yeah, but

19:31

there's also an interesting parallel here

19:33

where O1-Pro... just in terms of

19:35

benchmarks and I think in general

19:37

in terms of the vibe of

19:39

what people think is that it's

19:42

not sort of significantly better than

19:44

01 and that parallels duty 4.5

19:46

you know it's better but it's

19:48

not sort of a huge leap

19:50

so there is an interesting kind

19:53

of a demonstration of probably it's

19:55

harder to get you know huge

19:57

leaps and performance and people are

19:59

going to be more critical now

20:01

of if you were not offering

20:04

something that's like you know really

20:06

between Jupiter 4 3.5 and 4

20:08

for instance Yeah, I think it's

20:10

quite use case specific too, right?

20:12

So as we've seen, you know,

20:15

the kinds of issues people are

20:17

running into optical illusions, you know,

20:19

Sudoku puzzles, this sort of thing,

20:21

are pretty far from the standard,

20:24

you know, the actual workloads that

20:26

Open AI is targeting, right? Their

20:28

focus is, can we build something

20:30

that helps us automate AI research

20:32

as quickly as possible? Those sorts

20:35

of benchmarks, yeah, we are seeing

20:37

needle moving there. There's also some

20:39

interesting stuff that we'll talk about

20:41

for meter, suggesting that in fact

20:43

that is what's happening here. That

20:46

on those particular kinds of tasks,

20:48

we're seeing pretty significant. acceleration with

20:50

scale, but you're right, right? It's

20:52

this funny uneven surface, just like

20:54

how humans are funny and uneven,

20:57

right? Like you have a really

20:59

talented artist who can't write a

21:01

line of code to save their

21:03

lives, right? And vice versa. So

21:05

another instance of the paradox of

21:08

what's hard for AI isn't necessarily

21:10

hard for humans. And moving away

21:12

from Open AI to Google, we

21:14

now have another feature. Another instance

21:16

of canvas, this time on Gemini,

21:19

and they're also adding audio overview.

21:21

So I don't know why they

21:23

do this, why these ALMs just

21:25

copy each other's names. We had

21:28

deep research showing up in multiple

21:30

variants. Now we have a canvas,

21:32

which is also on ChadGPT, and

21:34

I think on the topic, it's

21:36

called Artifacts. Basically the same idea,

21:39

where now as you're working on

21:41

something like code, for instance. or

21:43

like a web app, for instance,

21:45

you can have a site panel

21:47

showing this living document rendering of

21:50

it with a chatbot to the

21:52

left, so you can essentially interactively

21:54

work and see a preview of

21:56

what you're getting. And you also

21:58

have audio overviews, which is pretty

22:01

much something like notebook, LM, that

22:03

you can upload documents and get

22:05

this podcast. style conversation going on.

22:07

So nothing sort of conceptually new

22:09

going on here, but I think

22:12

an interesting convergence across the board

22:14

of all these tools, everyone has

22:16

canvas, everyone has deep research, everyone

22:18

seems to have kind of the

22:20

same approach to implementing LM interfaces.

22:23

Speaking of that, in fact, the

22:25

next story is about Enfropic and

22:27

them adding web search capabilities to

22:29

Claude. That is now in preview

22:32

for paid users in the US

22:34

and that will basically work the

22:36

same as it does in Chajabiki

22:38

and other models. You can enable

22:40

it to work with cloud 3.7

22:43

and then it will be able

22:45

to provide direct citations from web-source

22:47

information. So yeah, not much else

22:49

to say we're getting a website

22:51

for cloud which will enable it

22:54

to be more useful. It's interesting

22:56

because it's like the tee up

22:58

to this is anthropic being a

23:00

little bit more shy than other

23:02

companies to roll up the the

23:05

web search product into their Into

23:07

their agents and I mean this

23:09

is consistent with the threat models

23:11

that they take seriously right things

23:13

like loss of control right which

23:16

typically involve you know an AI

23:18

model going on to the internet

23:20

maybe replicating its weight somehow and

23:22

internet access is kind of central

23:24

to a lot of these things

23:27

I don't know if that was

23:29

part of this, it at least

23:31

is consistent with it. So the

23:33

result is that there may be

23:36

a little bit later the party

23:38

than others. Apparently, according to these

23:40

initial tests, you don't always see

23:42

web search used for current events

23:44

related questions. But when that happens,

23:47

you do get these nice in

23:49

line citations, pull from sources, it

23:51

does look at social media, and

23:53

then of course news sources. like

23:55

NPR, like Reuters, they cite in

23:58

the examples they show. So, you

24:00

know, pretty standard product in the

24:02

inline citation approach that you see

24:04

with deep research, for example, certainly

24:06

making an appearance here. And last

24:09

up, again, along the lines of

24:11

these stories. we have XAI launching

24:13

a new API, this one for

24:15

generating images. So we have a

24:17

new model called Rock 2 Image

24:20

12, and you can now query

24:22

it. For now it's quite limited.

24:24

You can only generate 10 images

24:26

per request and you are limited

24:28

to 5 requests per second. The

24:31

cost there is 7 cents per

24:33

image. which is slightly above what,

24:35

for instance, Blackforest Labs charges. They

24:37

are the developers of Flux and

24:40

competitive with another offering for my

24:42

diagram. So I think, yeah, interesting

24:44

to see XAI expanding their APIs

24:46

once again, they released their own

24:48

image generation back in December and

24:51

it's kind of looked competitive with

24:53

something like Google's latest generation of

24:55

where we focus is really shifted

24:57

towards careful instruction following in your

24:59

image generation. So yeah, X AI

25:02

is as ever trying to catch

25:04

up or moving quite rapidly to

25:06

expand their offerings. Yeah, they really

25:08

are. And I think when we

25:10

first covered Blackforest Labs partnership with

25:13

X AI, one of the first

25:15

things that we said was like,

25:17

hey, this is. because I think

25:19

they raised a big round, right,

25:21

on the back of the incredible

25:24

distribution that they were going to

25:26

get through X AI and the

25:28

kind of vote of confidence that

25:30

reflects it from Elon. But at

25:32

the time we were talking about,

25:35

hey, you know, this is a

25:37

pretty strategically dicey position for Blackforth

25:39

Labs because the one thing we've

25:41

consistently seen from from all the

25:44

AI companies is Once they start

25:46

getting you in for chat, eventually

25:48

they start rolling out multimodal features,

25:50

and it's not clear that those

25:52

aren't best built in house for

25:55

any number of reasons, not just

25:57

including the fact that you want

25:59

to kind of internalize all the

26:01

revenues you can from the whole

26:03

from the whole stack, but also

26:06

once you have a good reasoning

26:08

model or rather a good foundation

26:10

model, that foundation model can be

26:12

mine for multimodality. post hoc and

26:14

you just kind of get to

26:17

amortize your investment across more modalities

26:19

and so it's just this natural

26:21

move to kind of keep crawling

26:23

into you're creeping into adjacent markets

26:25

like image generation video generation which

26:28

is also something that X AI

26:30

is looking at so yeah I

26:32

mean it's kind of interesting for

26:34

Blackforce Labs this probably is going

26:36

to be a big challenge for

26:39

them I don't know how extensive

26:41

their partnership continues to be at

26:43

this point but it's a dicey

26:45

time to be to be one

26:47

of these companies. And on two

26:50

applications and business we begin with

26:52

some announcements from invidia. There's a

26:54

preview of their plans in 2026

26:56

and 2027. They have the Ruben

26:59

family of GPUs coming in 2026

27:01

and then Ruben Ultra in 2027.

27:03

So that will also come along

27:05

with a new, I guess. server

27:07

layout with ability to combine 576

27:10

GPUs per rack, which, you know,

27:12

I guess it's very much following,

27:14

you know, the tracks of very,

27:16

very crazy enhancement to computing that

27:18

invidious been able to continue creating

27:21

with, you know, B200, I believe

27:23

it's now, and now this is

27:25

their plans for the next couple

27:27

years. Yeah, there's a lot going

27:29

on with this update. It's actually

27:32

pretty interesting and quite significant, especially

27:34

on the data center side, in

27:36

terms of the infrastructure that will

27:38

be required to accommodate these new

27:40

chips. A couple things here, right?

27:43

So there is this configuration of

27:45

the Blackwell called the NBL 72,

27:47

is the certain name of this

27:49

configuration. This is where you have,

27:51

so, okay, imagine a tray that

27:54

you're in a slot into. a

27:56

rack, a server rack, right? So

27:58

on that tray, you're gonna have

28:00

four GPUs. All right, so each

28:03

trache. contains four GPUs, and in

28:05

total, in that whole rack, you're

28:07

gonna have 72, I'm sorry, you're

28:09

actually gonna have 144 GPUs total,

28:11

but because two of those GPUs

28:14

show up on the same motherboard,

28:16

God. So each freaking, each freaking

28:18

tray that you slot into the

28:20

rack has two motherboards on it,

28:22

each of those motherboards has two

28:25

GPUs, two B200 GPUs. So in

28:27

total, you're putting in four GPUs

28:29

per tray. But they're kind of

28:31

divided into two motherboards, each with

28:33

two GPUs. Anyway, this led to

28:36

the thing being called the NFL

28:38

72, when in reality there's 144

28:40

GPUs on there, at least Jensen-Hwang

28:42

says it would have been more

28:44

appropriate to call it the NFL-144-L.

28:47

Okay. What's actually interesting in this

28:49

setup, they're calling the Ruben-Nvil-L-144 rack.

28:51

There's not more GPUs there. It's

28:53

not that there's twice as many

28:55

GPUs as the Nvil-72 with the

28:58

black wells. It's just that they're

29:00

counting them differently now. So they're

29:02

saying, actually, we're going to count

29:04

all the GPUs. So if, I

29:07

think back in the day we

29:09

did talk about the Nvil-72 setup,

29:11

this is basically just the same

29:13

number of GPUs. If that didn't

29:15

make any sense, just delete it

29:18

from your mind. It's both some

29:20

of the things that are actually

29:22

interesting. The story is it's comparable

29:24

in the number of GPUs to

29:26

the current set of top line

29:29

GPUs. So they're kind of pitching

29:31

it as you can spot it

29:33

into your existing infrastructure more or

29:35

less and just to jump into

29:37

numbers a little bit, you're getting

29:40

roughly three times the inference and

29:42

training performance in terms of just

29:44

rock compute memory is faster by

29:46

close to two-ish or multiplier of

29:48

two kind of like yeah you're

29:51

seeing multipliers on top of the

29:53

current one so quite significant change

29:55

in performance if you do upgrade

29:57

yeah so so when it comes

29:59

to Rubin right which is the

30:02

the sort of next generation coming

30:04

online at FP4, you're seeing, yeah,

30:06

3X more flops, right, three times

30:08

more, more logic capacity. Now the,

30:11

on the memory side, things actually

30:13

do get somewhat interesting. The memory

30:15

capacity is going to be 288

30:17

gigabytes per GPU, right? That is

30:19

the same as the B300. So

30:22

no actual change in terms of

30:24

the, like, per GPU. memory capacity.

30:26

We'll get back to why that

30:28

matters a bit less in a

30:30

second, but that's kind of part

30:33

of the idea. The memory bandwidth

30:35

is improving. It's almost doubling. Maybe,

30:37

yeah, it's short of doubling. So

30:39

the memory bandwidth is really, really

30:41

key, especially when you look at

30:44

inference. So that's one of the

30:46

reasons why this is really being

30:48

focused on. But there's also a

30:50

bunch of things like the, so

30:52

the cables that connect GPUs together.

30:55

on roughly speaking on one rack,

30:57

if you want to imagine it

30:59

that way. Those are called envy

31:01

link cables, super super high bandwidth.

31:03

Those are doubling in throughput. So

31:06

that's really a really big advance.

31:08

There's also stuff happening on the

31:10

networking side, but we don't need

31:12

to touch that. Bottom line is

31:15

envy link cables used to be

31:17

the way you connected GPUs across

31:19

different. trays in the same rack

31:21

and maybe adjacent racks depending on

31:23

the configuration. But it's very local,

31:26

very very tight, very high bandwidth

31:28

communication. What's happening here is each

31:30

of these motherboards, you know, that

31:32

you're slotting into your rack, they

31:34

have a CPU and two GPUs.

31:37

And we talked about this in

31:39

the hardware episode as to why

31:41

that is. The CPUs like the

31:43

orchestra conductor, the GPUs are like

31:45

the instruments that, you know, they're

31:48

actually doing the hard work in

31:50

the heavy lifting. Typically the CPU

31:52

would be connected to the GPUs

31:54

through a PCIE connection. So this

31:56

is a relatively low bandwidth compared

31:59

to NVLink. Now they're moving over

32:01

to NVLink as well for the

32:03

CPU 2 GPU connection. That's actually

32:05

a real... really big deal. It

32:07

comes with a core to core

32:10

interface. So right now, the GPUs

32:12

and CPUs are going to share

32:14

a common memory space. So essentially

32:16

directly accessing each other's memory. Whatever

32:19

is in memory on the CPU,

32:21

the GPU can access right away

32:23

and vice versa. That's a really,

32:25

really big change. It used to

32:27

not be the case. You used

32:30

to have independent CPU and GPU

32:32

memory. the GPUs themselves would share

32:34

a common memory space if they

32:36

were connected via NVLink and in

32:38

fact that's kind of that's part

32:41

of the idea here that's what

32:43

makes them a coherent wad of

32:45

compute and it's also part of

32:47

the reason why the memory capacity

32:49

on those GPUs matters a bit

32:52

a bit less because you're adding

32:54

you're kind of combining all your

32:56

GPUs together and they have a

32:58

shared memory space so if you

33:00

can just add to the number

33:03

of GPUs, you have, you're effectively

33:05

adding true memory capacity. So that's

33:07

kind of an important difference there.

33:09

So anyway, last thing I'll mention,

33:11

they say that apparently Ruben Ultra

33:14

is going to come out. This

33:16

is, so there's going to be

33:18

Ruben and then Ruben Ultra. Ruben

33:20

Ultra is coming out the second

33:23

half of 2027. It'll come with

33:25

a Ruben GPU and a Vera

33:27

CPU, like invidia tends to do,

33:29

right? And so Vera is the

33:31

CPU, Reuben is the GPU. Apparently,

33:34

the full rack is going to

33:36

be replaced by this 576 GPU

33:38

setup, a massive number. That is

33:40

essentially, so they don't specify the

33:42

power consumption, but it's clear from

33:45

other kind of industry products that

33:47

are coming out. We're tracking for

33:49

one megawatt per rack. And just

33:51

worth emphasizing. That's a thousand kilowatts.

33:53

That is a thousand homes worth

33:56

of power going to a single

33:58

rack in a server, in a

34:00

data center. That's insane, right? So

34:02

the power density required for this

34:04

is going through the roof, the

34:07

cooling requirements, all this stuff. It's

34:09

all really cool. And anyway, this

34:11

is a very, very big motion.

34:13

Just to dive in a little.

34:15

it into the numbers, just fun,

34:18

right? So the compute numbers are

34:20

in terms of flops, which is

34:22

floating point operations per second, basically

34:24

multiplications per second or additions per

34:26

second. And the numbers we get

34:29

with these announced upcoming things like

34:31

their Rubin are now for inference,

34:33

3.6 XA flops of inference. So

34:35

XA is... is Quintilian. It's 10

34:38

to the 18. Quintilian is the

34:40

one after Quintilian. So I can't

34:42

even imagine how many, I mean,

34:44

I guess I know how many

34:46

zeros it is, but it's very

34:49

hard to imagine a number that

34:51

long. And that's just where we

34:53

are at. Also worth mentioning, so

34:55

this is the plans for 2026,

34:57

2027. They also did announce for

35:00

late of a year the coming

35:02

of B300, which is B300. Also,

35:04

you know, is an improvement in

35:06

performance of about 1.5. They also

35:08

didn't announce the ultra-bariance of Blackwell,

35:11

both 200 and 300. And the

35:13

emphasis they are starting to add,

35:15

I think, is more on the

35:17

inference side. They definitely are saying

35:19

that these are good models for

35:22

the age of reasoning. So they're

35:24

capable of outputting things fast in

35:26

addition to training well and that's

35:28

very important for reasoning of course

35:30

because the whole idea is you're

35:33

using up more tokens to get

35:35

better performance. So they're giving some

35:37

numbers like for instance Blackwell Ultra

35:39

will be able to deliver up

35:42

to 1,000 tokens per second on

35:44

Deep Seek R1 and that's you

35:46

know comparable usually you would be

35:48

seeing something like 100-200 tokens per

35:50

second a thousand dollars per second

35:53

is very fast. Yeah, and then

35:55

the inference focus is reflected too.

35:57

in the fact that they're looking

35:59

at you know FP4 flops denominated

36:01

performance right so when you go

36:04

to inference often you're your inferencing

36:06

quantized models inferencing at FP4 so

36:08

lower resolution and also this the

36:10

the memory bandwidth side becomes really

36:12

important for inference disproportionately relative to

36:15

training at least on the current

36:17

paradigm so that's kind of you

36:19

know part of the reason that

36:21

you're seeing those big big lifts

36:23

at that that end of things

36:26

is because of the inference. And

36:28

next story is also about some

36:30

absurd sounding numbers of hardware. This

36:32

one is from Apple. They have

36:34

launched a Mac studio offering and

36:37

the top line configuration where you

36:39

can use the entry ultra chip

36:41

with 32 CPU's 80 core GPUs

36:43

that can even run the Deep

36:46

Seek R1 model. That's the 671

36:48

billion parameter AI model. Fewer at

36:50

inference, you're using about 37 billion

36:52

per output I believe, but still,

36:54

this is hundreds of gigabytes of

36:57

memory necessary to be able to

36:59

run it and just fit it

37:01

in there. Yeah, Apple's also doing

37:03

this weird thing where they're not

37:05

designing like GPUs for their data

37:08

centers, including for AI workloads. They

37:10

seem to be basically like doing

37:12

souped up CPUs, kind of like

37:14

this, with just like gargantuan amounts

37:16

of VRAM, that again, have this

37:19

very large kind of shared pool

37:21

of memory, right? We talked about

37:23

like coherent memory on the Blackwell

37:25

side, right, and on the Rubin

37:27

side, just the idea that. if

37:30

you have a shared memory space,

37:32

you can pool these things the

37:34

other. Well, they're not as good

37:36

at the shared memory space between

37:38

CPUs. What they do is they

37:41

have disgusting amounts of RAM, one

37:43

GPU, right? So like, 512 gigs

37:45

is, it's just wild. Like, it's,

37:47

anyway. for a CPU at least.

37:50

And we're talking here about, when

37:52

you say memory, we mean really

37:54

something like RAM, right? And so

37:56

if you have a laptop, right,

37:58

if you buy a Mac, for

38:01

instance, typically you're getting eight gigabytes,

38:03

maybe 16 gigabytes of RAM, the

38:05

fast type of memory, read, what

38:07

does it, read something memory? It's

38:09

random access memory, right. As opposed

38:12

to the slower memory of. Let's

38:14

say an SSD or things where

38:16

you can easily get terabytes to

38:18

get that crazy amount of Random

38:20

access memory is insane when you

38:23

consider that typically it's like eight

38:25

16 gigabytes and you know This

38:27

is expensive memory. It's it's stupid

38:29

expense. It's also like Yeah, there's

38:31

different kinds of RAM and we

38:34

talked about that in our hardware

38:36

episode This is a combined CPU

38:38

GPU set up by the way

38:40

so 32 core CPU 80 core

38:42

GPU but shared memory across the

38:45

board. So V-Ram is like really

38:47

close to the logic, right? So

38:49

this is like the most, as

38:51

you said, exquisitely expensive kind of

38:54

memory you can put on these

38:56

things. They're opting to go in

38:58

this direction for very interesting reasons,

39:00

I guess. I mean, it does

39:02

mean that they're disadvantaged in terms

39:05

of being able to scale their

39:07

data center, you know, infrastructure, because

39:09

of networking, at least as far

39:11

as I can tell. It's a

39:13

very interesting standalone standalone machine. I

39:16

mean, it's a pretty wild specs.

39:18

Right. Yeah, exactly. If you go

39:20

to the top line offerings, and

39:22

this is, you know, a physical

39:24

product you can buy as a...

39:27

Yeah, it's a Mac, right? Yeah,

39:29

it's a Mac, yeah, it's a

39:31

Mac, it's like a big, kind

39:33

of cube-ish thing. And if you

39:35

go to the top line of

39:38

recreation, it's something like $10,000, don't

39:40

quote me on that, but it's,

39:42

you know, you know, expensive, expensive,

39:44

expensive, expensive, expensive, as well. It

39:46

does come with other options for

39:49

a reason M4 Max CPU and

39:51

GPUs less powerful than an free

39:53

ultra, but anyway. Very kind of

39:55

beefy offering now from Apple. Next

39:58

we have something a bit more

40:00

forward-looking. Intel is apparently reaching an

40:02

exciting milestone for the 18A1.8 nanometer

40:04

class wafers with a first run

40:06

at the Arizona Fab. So this

40:09

is apparently ahead of schedule. They

40:11

have these Arizona Fabs. Fab 52

40:13

and Fab 62, Fab as we've

40:15

covered before is where you try

40:17

to make your chips and 1.8

40:20

nanometer is the next kind of

40:22

frontier in terms of scaling down

40:24

the resolution. the density of logic

40:26

you can get on a chip.

40:28

So the fact that they're running

40:31

these test wafers, they're ensuring that

40:33

you can transfer the fabrication process

40:35

to these new Arizona facilities, I

40:37

guess the big deal there is

40:39

partially that these are located within

40:42

the US, within Arizona, and they

40:44

are seemingly getting some success and

40:46

are ahead of schedule, as you

40:48

said, and that's impressive because fabs

40:50

are an absurdly complex engineering project.

40:53

Yeah Intel is in just this

40:55

incredibly fragile space right now as

40:57

has been widely reported and we've

40:59

talked about that a fair bit

41:01

I mean they need to blow

41:04

it out of the water with

41:06

with 18a in their future notes

41:08

I mean this is like a

41:10

make-or-break stuff so forward progress yeah

41:13

they had their their test facility

41:15

in Hillsborough Oregon who that was

41:17

doing 18a production as you said

41:19

on a test basis and they're

41:21

now successfully getting the first test

41:24

wafers in their new Arizona Fab

41:26

out so that's great But yeah,

41:28

it'll eventually have to start running

41:30

actual chips for commercial products. The

41:32

big kind of distinction here is

41:35

they're actually manufacturing with 18A, these

41:37

gate all around transistors. I think

41:39

we talked about this in the

41:41

hardware episode. We'll go into too

41:43

much detail. This is a specific

41:46

geometry of transistor that allows you

41:48

to have better control over the

41:50

flow of electrons through your transistor,

41:52

essentially. It's a big, big challenge

41:54

people have had in making transistors

41:57

small and smaller. You get all

41:59

kinds of current leakage. The current,

42:01

by the way, is sort of

42:03

like the thing that carries information

42:05

in your computer. And so you

42:08

want to make sure that you

42:10

don't have current leakage to kind

42:12

of have ones become zeros or

42:14

let's say operation, like, you know,

42:17

a certain kind of gate, turn

42:19

into the wrong kind of gate.

42:21

That's the idea here. this gate-all-round

42:23

transistor based on a ribbon-fed design.

42:25

And yeah, so we're seeing that

42:28

come-to-market gate-all-round is is something that

42:30

TSMC is moving towards as well.

42:32

And you know, it's just going

42:34

to be the next, essentially the

42:36

next beat of production. So here

42:39

we have 18A, kind of early

42:41

signs of progress. And now moving

42:43

away from hardware more to business-y

42:45

stuff, XAI has acquired a genative

42:47

AI startup. which is focused on

42:50

text to video similar to SORA.

42:52

They also have AI powered, or

42:54

initially they're worked on AI powered

42:56

for the tools and then pivoted.

42:58

So I suppose on surprising in

43:01

a way that they are working

43:03

on text to video as well.

43:05

They just want to have all

43:07

the capabilities at XAI, and this

43:09

presumably will make that easier to

43:12

do. Yeah, one of the founders

43:14

had some quote, I think it

43:16

might have been on X, I'm

43:18

not sure, but he said, we're

43:21

excited to continue scaling these efforts

43:23

on the largest cluster in the

43:25

world, Colossus, as part of XAI,

43:27

so it seems like they'll be

43:29

given access to Colossus as part

43:32

of this, maybe not shocking, but

43:34

kind of an interesting subnote. They

43:36

were backed by some really impressive

43:38

VCs as well, so Alexis Zahanian,

43:40

as were like famous for like

43:43

being the co-founder, read it of

43:45

course, and doing his own VC

43:47

stuff, and SV Angel too. So

43:49

pretty interesting acquisition and a nice

43:51

soft landing too for folks in

43:54

a space that otherwise, you know,

43:56

I mean, they're either going to

43:58

acquire you or they're going to

44:00

eat your lunch. So I think

44:02

that's probably the best outcome for

44:05

people working on the different modalities,

44:07

at least. at least on my

44:09

view of the market. Yeah, and

44:11

I guess acquisition makes sense. The

44:13

startup has been around for over

44:16

two years and they have already

44:18

trained multiple video models. Hotshot Excel

44:20

and hotshot, and they do produce

44:22

quite good looking videos, so. make

44:25

some sense for opening eye to

44:27

acquire them, if only for the

44:29

kind of brain power and expertise

44:31

in that space. Yeah, man, they're

44:33

old. They've been around for like

44:36

two years, right? Like, yeah, that

44:38

was what, free SORA or like

44:40

a round-of-time SORA. Yeah, yeah, it's

44:42

funny. It's just funny how the

44:44

AI business cycle is so short,

44:47

like, like, these guys have been

44:49

around for all the 24 months.

44:51

Very experts, very veterans. And onto

44:53

the last story, Tencent is reportedly

44:55

making massive and video H20 chip

44:58

purchases. So they are supposedly meant

45:00

to support the integration of Deep

45:02

Seek into We Chat, which kind

45:04

of reminds me of meta where

45:06

meta has this somewhat interesting drive

45:09

to... Let you use a llama

45:11

everywhere and Instagram and all their

45:13

messaging tools. So this would seem

45:15

to be in a way similar

45:17

where Tencent would allow you to

45:20

use deep seek within we chat.

45:22

Yeah, part of what's going on

45:24

here too is the standard stockpiling

45:26

that you see China do and

45:29

Chinese companies do ahead of the

45:31

anticipated crackdown from the United States

45:33

on export controls. And in this

45:35

case, the age 20 has been

45:37

kind of identified as one of

45:40

those chips that's likely to likely

45:42

to be... shut down for the

45:44

Chinese market in the relatively near

45:46

term. So it makes all the

45:48

sense in the world that they

45:51

would be stockpiling for that purpose.

45:53

But it is also the case

45:55

that you got R1 that has

45:57

increased dramatically the demand for access

45:59

to hardware. It's sort of funny

46:02

how quickly we pivoted from, oh

46:04

no, R1 came out and so

46:06

invidious stock crashes to, oh, actually

46:08

R1. great news for invidia. Anyway,

46:10

I think it's the turnaround that

46:13

we sort of expected. We talked

46:15

about this earlier and there has

46:17

been apparently a short-term supply shortage

46:19

in China regarding these age 20

46:21

chips. So like there's so much

46:24

demand coming in from 10 cent

46:26

that it's this sort of like

46:28

rate limiting for invidia to get

46:30

age 20s into the market there.

46:33

So kind of interesting, they've previously

46:35

placed orders on the order of

46:37

hundreds of thousands between them and

46:39

bite dance. back I think last

46:41

year was almost a quarter million

46:44

of these these Japanese so yeah

46:46

pretty big customers and onto projects

46:48

and open source we begin with

46:50

a story from the information titled

46:52

on tropics not so secret weapon

46:55

that's giving agents a boost I

46:57

will say kind of a weird

46:59

spin on this whole story but

47:01

anyway that's the one we're linking

47:03

to and it covers the notion

47:06

of MCP model context protocol which

47:08

Anthropic released all the way back

47:10

in November. We hopefully covered it.

47:12

I guess we don't know. I

47:14

was trying to remember, yeah. Yeah,

47:17

I think we did. And the

47:19

reason we're coming in now is

47:21

that it sort of blew up

47:23

over the last couple of weeks

47:25

if you're in the AI developer

47:28

space or you see people hacking

47:30

on AI, that it has been

47:32

to talk of a town, so

47:34

to speak. So Molokontax protocol, broadly

47:37

speaking is something... like an API,

47:39

like a standardized way to build

47:41

ports or mechanisms for AI agents

47:43

or AI, I guess, models to

47:45

call on services. So it standardizes

47:48

the way you can provide things

47:50

like tools. So there's already many,

47:52

many integrations following the standard for

47:54

things like slack, perplexity, notion, etc.

47:56

where if you adopt a particle

47:59

and you provide an FCP compatible

48:01

kind of opening, you can then

48:03

have an MCP client, which is

48:05

your AI model, call upon this

48:07

service. And it's very much like

48:10

an API for a website, where

48:12

you can have a particular URL

48:14

to go to, particular kind of

48:16

parameters, and you get something back

48:18

in some format. Here, the difference

48:21

is that, of course, this is

48:23

more specialized for. AI models in

48:25

particular, so it provides tools, it

48:27

provides like a prompt to explain

48:29

the situation, things like that. Personally,

48:32

I'm in the camp of people

48:34

who are a bit confused and

48:36

kind of think that this is

48:38

an API for an API kind

48:40

of situation, but either way, it

48:43

has gotten very popular. That is

48:45

exactly what it is. Yeah, it's

48:47

an API for an API. It's

48:49

also, I guess, a transition point.

48:52

or could be viewed that way,

48:54

you know, in the sense that

48:56

eventually you would expect models to

48:58

just kind of like figure it

49:00

out, you know, and have enough

49:03

context and ability to uncover whatever

49:05

information is already on the website.

49:07

to be able to use tools

49:09

appropriately, but there are edge cases

49:11

where you expect this to be

49:14

worthwhile. Still, this is going to

49:16

reduce things like hallucination of tools

49:18

and all kinds of issues that

49:20

when you talk about agents, like

49:22

one failure anywhere in a reasoning

49:25

chain or in an execution chain,

49:27

it can cause you to fumble.

49:29

And so, you know, this is

49:31

structurally a way to address that

49:33

and quite important in that sense.

49:36

It is also distinct from a

49:38

lot of the tooling that opening

49:40

eyes come out with, but it

49:42

sounds similar like the agent's API.

49:44

where they're focused more on chaining

49:47

tool uses together, whereas MCP, as

49:49

you said, is more about helping

49:51

make sure that each individual instance

49:53

of tool use goes well, right?

49:56

That the agent has what it

49:58

needs to kind of ping the

50:00

tool properly interact with it and

50:02

find the right tool rather than

50:04

necessarily chaining them together. So there

50:07

you go. MCP. is a nice

50:09

kind of clean open source play

50:11

for anthropic too. They are going

50:13

after that kind of more startup

50:15

founder and business ecosystem. So pretty

50:18

important from a marketing standpoint for

50:20

them too. Right, yeah, exactly. So

50:22

back in November, they announced this,

50:24

they introduced us as an open

50:26

standard, and they also released open

50:29

source repositories with some example of

50:31

model context product called servers. and

50:33

as well the specification and like

50:35

a development toolkit. So I honestly

50:37

haven't been able to track exactly

50:40

how this blew up. I believe

50:42

there was some sort of tutorial

50:44

given at some sort of convention

50:46

like the AI engineer convention or

50:48

something and then it kind of

50:51

took off and everyone was very

50:53

excited about the idea of model

50:55

context protocols right now. Moving on

50:57

to new models, we have Mistral

51:00

dropping a new open source model

51:02

that is comparable to 2P400 mini

51:04

and is smaller. So they have

51:06

Mistral small 3.1, which is seemingly

51:08

better than similar models, but only

51:11

has 24 billion parameters. Also can...

51:13

take on more input tokens, 120,000

51:15

tokens, and is fairly speedy at

51:17

150 tokens per second. And this

51:19

is being released under the Apache

51:22

2 license, meaning that you can

51:24

use it for whatever you want,

51:26

business implications, etc. I don't think

51:28

there's too much to say here,

51:30

other than like kind of a

51:33

nitpick here, but they say it

51:35

outperforms copper models like Gemini 3.

51:37

GD40 mini while delivering infant speeds

51:39

as you said of 150 tokens

51:41

per second but like you can't

51:44

just say that shit like it

51:46

doesn't mean anything to say yeah

51:48

it depends on what infrastructure you're

51:50

using yeah what's the stack dude

51:52

like like you know like I

51:55

can move at a hundred like

51:57

at a hundred miles an hour

51:59

if I'm in a Tesla that's

52:01

what makes me anyway so they

52:04

do give that information but it's like

52:06

buried like in the little like great

52:08

this is from their blog post where

52:10

we get these numbers so I guess

52:12

as with any of these model

52:15

announcements you go to the company

52:17

log you would get a bunch

52:19

of numbers on benchmarks showing

52:21

that it's the best You

52:23

have comparisons to Gemma free

52:26

from Google to coherent, AIIA,

52:28

Jupiter for a mini, cloud

52:30

3.5 haiku, and on all

52:32

of these things like MMLU,

52:34

human evil, math, it typically is

52:36

better, although, you know, I would

52:38

say it doesn't seem to be

52:40

that much better than Gemma at

52:43

least, and in many cases

52:45

is not better than 3.5 haiku

52:47

and Jupiter for a mini, but

52:49

yeah, still quite good. the 150 tokens

52:52

per second to for context is a batch

52:54

size 16 on 4 H 100s. They actually

52:56

like even in the technical post they write

52:58

while delivering inference speeds of 150 tokens

53:00

per second without further qualifying but it's

53:02

in like it's in this like small

53:05

gray text underneath an image that you

53:07

have to look for where you actually

53:09

find that context so don't expect this

53:11

to run at 150 tokens per second

53:13

on your laptop right that's that's just

53:15

not going to happen because you know

53:17

for H 100s is like that's quite

53:19

a lot of horsepower horsepower. still yeah

53:21

it's incremental improvement more

53:24

more open source coming

53:26

from from Israel and patchy

53:29

2.0 license so you know highly

53:31

permissive and one more model

53:33

you have exa on deep

53:35

reasoning enhanced language models coming

53:38

from LG AI research so

53:40

these are new models new

53:42

family models 2.4 billion seven

53:45

point eight billion and 32

53:47

billion parameters These are

53:49

optimized for reasoning

53:51

tasks and seemingly

53:54

are on par or

53:56

out of our performing

53:58

variations of R1. are one is

54:01

the giant one that's 671

54:03

billion. There's distilled

54:05

versions of those models

54:08

at comparable sizes and

54:10

in the short technical

54:12

report that they provide they

54:14

are showing that it seems

54:16

to be kind of along the lines

54:18

of what you can get with

54:21

those distilled R1 models

54:23

and similar or better

54:25

than Open AI01 mini.

54:27

Yeah, it's also, it's kind of interesting,

54:29

it seems to, I described, again, not

54:32

a lot of detail in the paper,

54:34

so it makes it hard to reconstruct,

54:36

but it does seem to be at

54:38

odds with some of the things that

54:40

learn in the Deep Seek R1 paper,

54:42

for example. So there's, they start

54:45

with, it seems, an instruction tuned

54:47

model, base model, the XA1,

54:49

3.5 instruct models, and then they

54:51

add onto that. a bunch of fine tuning,

54:53

they do supervised fine tuning, presumably this

54:56

is for like the reasoning structure, and

54:58

then DPO, so standards, or like, RL

55:00

stuff, and online RL. So, you know, this

55:02

is, you know, quite a bit of supervised

55:04

fine tuning of trying to teach the model

55:07

how to solve problems in the way you

55:09

wanted to solve them, rather than just like

55:11

giving it a reinforcement learning signal

55:13

and reward signal and like kind

55:15

of have at it like R10 did.

55:18

So yeah, kind of an interesting

55:20

alternative more, let's say. more

55:22

inductive prior laden approach and

55:24

first time as well that I

55:26

think we've covered anything from

55:28

LGA high risk so yeah

55:30

Exxon these models appear to

55:33

have already been existing and

55:35

being released Exxon 3.5 we're

55:37

back from December which be

55:39

somehow missed at the time

55:41

yeah fun fact Exxon stands

55:43

for expert AI for everyone

55:45

you got a lot when people

55:48

come up these acronyms in a

55:50

very creative way And they are

55:52

open sourcing it on hugging face

55:54

with some restrictions, this is primarily

55:56

for research usage. Odd to research and

55:59

investment. we begin the paper

56:01

sample scrutinize and scale effective inference

56:03

time search by scaling verification it

56:05

is coming from Google and let

56:07

me just double check also you

56:09

see Berkeley and it is quite

56:12

exciting I think as a paper

56:14

it basically is making the case

56:16

or presenting the idea of that

56:18

to do scaling inference time scaling

56:20

via reasoning where you have the

56:23

model you know once you already

56:25

train your model once you stop

56:27

updating your weights can you use

56:29

your model more to get smarter

56:31

if you just kind of do

56:33

more outputs in some way and

56:36

what you've seen in recent months

56:38

is inference time scaling via reasoning

56:40

where you have the model you

56:42

know output a long chain of

56:44

tokens where it does various kinds

56:46

of strategies to do better at

56:49

a given complicated task like planning

56:51

sub steps like verification backtracking these

56:53

things we've already covered. Well this

56:55

paper is saying instead of that

56:57

sort of scaling the output in

56:59

terms of a chain an extended

57:02

chain of tokens you can instead

57:04

sample many potential outputs like just

57:06

do a bunch of outputs from

57:08

scratch. And I kind of varied

57:10

up so you get many possible

57:12

solutions. And then if you have

57:15

a verifier and you can kind

57:17

of compare and combine these different

57:19

outputs, you can be as effective

57:21

or even in some cases more

57:23

effective than the kind of traditional

57:25

reasoning, inference time scaling paradigm. Again,

57:28

yeah, quite interesting in the paper.

57:30

They have a table one where

57:32

they are giving this example of

57:34

if you sample a bunch of

57:36

outcomes and you have their... that

57:38

is good, you can actually outperform

57:41

01 preview. And many other techniques,

57:43

you can get better numbers on

57:45

the hard reasoning benchmarks like Amy,

57:47

where you can solve eight out

57:49

of the 15 problems now, which

57:51

is insane, but that's where we

57:54

are. And on math and live

57:56

bench math and live bench reasoning.

57:58

So yeah, very interesting idea to

58:00

sample a bunch of solutions and

58:02

then just. compare them and can

58:04

combine them into one final output.

58:07

Yeah, I think this is one

58:09

of those key papers that, again,

58:11

I mean, we see this happen

58:13

over and over. Scaling is often

58:15

a bit more complex than people

58:18

assume, right? So at first, famously

58:20

we had pre-training, scaling, pre-training, compute,

58:22

that brought us from, you know,

58:24

GPD2, G3. to GPD4, now we're

58:26

in the inference time compute paradigm

58:28

where, you know, this analogy that

58:31

I like to use, like, you

58:33

have 30 hours to dedicate to

58:35

doing well on a test, you

58:37

get to choose how much of

58:39

that time you dedicate to studying,

58:41

how much you dedicate to actually

58:44

spending writing the test. And what

58:46

we've been learning is, you know,

58:48

scaling, pre-training computers, basically all study

58:50

time. And so if you just

58:52

do that and you effectively give

58:54

yourself like one second to write

58:57

the test, well, there's only so

58:59

well you can do, right? You

59:01

eventually start to saturate if you

59:03

just keep growing pre-training and you

59:05

don't grow inference time compute. You

59:07

don't invest more time at test

59:10

time. So essentially you have two-dimensional

59:12

scaling. If you want to get

59:14

the sort of indefinite returns, if

59:16

you want the curves to just

59:18

keep going up and not saturate,

59:20

you have to scale two things

59:23

at the same time. This is...

59:25

a case like that, right? This

59:27

is a case of a scaling

59:29

law that would be hidden to

59:31

a naive observer just looking at

59:33

this as a kind of a

59:36

one variable problem when in reality

59:38

it's a multivariate problem and suddenly

59:40

when you account for that you

59:42

go, oh wow, there's a pretty

59:44

robust scaling trend here. And so

59:46

what are these two variables? Well,

59:49

the first is scaling the number

59:51

of sampled responses, so just the

59:53

number of shots on goal, the

59:55

number of attempts that your model

59:57

is going to make at solving

59:59

a given problem, but you have

1:00:02

to improve verification capabilities at the

1:00:04

same time, right? So they're asking

1:00:06

the question, what test time scaling

1:00:08

trends come up as you scale

1:00:10

both the number of sampled responses

1:00:12

and your verification capabilities? One of

1:00:15

the things crucially that they find

1:00:17

though, is you might not evenly

1:00:19

think if you have a... a

1:00:21

verifier, right? And I want to

1:00:23

emphasize this is not a verifier

1:00:26

that has access to ground truth,

1:00:28

right? This is like, and I

1:00:30

think they use like Claude 3.7

1:00:32

for this, 3.7, so on and

1:00:34

so on and for this, but

1:00:36

basically there's a model that's going

1:00:39

to look at, say the 20

1:00:41

different possible samples in other words

1:00:43

that you get from your model

1:00:45

to try to... solve a problem.

1:00:47

And this verifier model is going

1:00:49

to contrast them and just determine,

1:00:52

based on what it knows, not

1:00:54

based on access to actual ground

1:00:56

truth or any kind of symbolic

1:00:58

system like a calculator, it's just

1:01:00

going to use its own knowledge

1:01:02

in the moment to determine which

1:01:05

of those 20 different possible answers

1:01:07

is the right one. And what

1:01:09

you find is, as you scale

1:01:11

the number of possible answers that

1:01:13

you have your verifier look at,

1:01:15

you might not even think well...

1:01:18

with just so much gunk in

1:01:20

the system, probably the verifiers performance

1:01:22

is going to start to drop

1:01:24

over time. It's just like it's

1:01:26

so much harder for it to

1:01:28

kind of remember which of these

1:01:31

was that good and which was

1:01:33

bad and all that stuff. So

1:01:35

eventually you would expect the kind

1:01:37

of trend to be that your

1:01:39

performance would saturate maybe even drop.

1:01:41

But what they find is the

1:01:44

opposite. The performance actually keeps improving

1:01:46

and improving. And the reason for

1:01:48

that seems to be that as

1:01:50

you increase the number of. samples,

1:01:52

the number of attempts to solve

1:01:54

a problem, the probability that you

1:01:57

get a truly exquisite answer that

1:01:59

is so much better than the

1:02:01

others, like that contrasts so much

1:02:03

with the median other answer, that

1:02:05

it's really easy to pick out,

1:02:07

increases. And that actually makes the

1:02:10

verifiers... easier. And so they refer

1:02:12

to this as an instance of

1:02:14

what they call implicit, I think

1:02:16

it was implicit scaling, was the

1:02:18

term, yeah, implicit scaling, right? So

1:02:20

essentially the, yeah, this idea that

1:02:23

you're more likely to get one

1:02:25

like exquisite outlier that favorably contrasts

1:02:27

with the crappy median samples. And

1:02:29

so in this sense, I mean,

1:02:31

I feel like the term verifier

1:02:34

is maybe not the best one.

1:02:36

Really what they have here is

1:02:38

a contraster. When I hear verifier,

1:02:40

I tend to think of ground

1:02:42

truth. I tend to think of

1:02:44

something that is actually kind of

1:02:47

like, you know, giving us, you

1:02:49

know, checking, let's say code and

1:02:51

seeing if it compiles properly. Yeah.

1:02:53

compiler, you could say, where it

1:02:55

takes a bunch of possible outputs

1:02:57

and then from all of that

1:03:00

picks out the best. kind of

1:03:02

guess to answer. Exactly, yeah, and

1:03:04

this is why I don't know

1:03:06

if it's not a term that

1:03:08

people use, but if it was

1:03:10

like contrast or is it really

1:03:13

what you're doing here, right? You're

1:03:15

like, you're sort of doing, in

1:03:17

a way, a kind of contrast

1:03:19

of learning, well, not learning necessarily,

1:03:21

because it's all happening at inference

1:03:23

time, but yeah. So the performance

1:03:26

is really impressive. There are implications

1:03:28

of this for the design of

1:03:30

these systems as well, so you're

1:03:32

trying to find ways to wrangle

1:03:34

problems to wrangle problems into a

1:03:36

shape where you can take advantage

1:03:39

of implicit scaling. where you can

1:03:41

have your model pump out a

1:03:43

bunch of responses in the hopes

1:03:45

that you're going to get, you

1:03:47

know, one crazy outlier that makes

1:03:49

it easier for the verifier to

1:03:52

do its job. So yeah, again,

1:03:54

you know, I think a really

1:03:56

interesting case of multi-dimensional scaling laws,

1:03:58

essentially, that are otherwise easily missed

1:04:00

if you don't invest in both

1:04:02

verifier performance and sampling at the

1:04:05

same time. Exactly, and this is,

1:04:07

I think, important context to provide.

1:04:09

The idea of sampling many answers

1:04:11

and just kind of picking out

1:04:13

the answer that occurred the most

1:04:15

times in all these samples is

1:04:18

a well-known idea. There's also, I

1:04:20

mean... Self-consistency. Self-consiciency, exactly. a majority

1:04:22

vote essentially is one well-established technique

1:04:24

to get better performance and there's

1:04:26

even more complex things you could

1:04:28

do also with like a mixture

1:04:31

of, I forget the term that

1:04:33

exists, but the general idea of

1:04:35

paralyzed generation of outputs and generating

1:04:37

multiple outputs potentially of multiple models

1:04:39

is well known and the real

1:04:42

insight as you said here is

1:04:44

that you need a strong verifier

1:04:46

to be able to really leverage

1:04:48

it. in the stable one they

1:04:50

show that if you compare like

1:04:52

you do get better performance quite

1:04:55

a bit if you just do

1:04:57

consistency if you just sample 200

1:04:59

responses and pick out the majority

1:05:01

compared to the thing where you

1:05:03

don't do any scaling you're getting

1:05:05

four problems out of 15 on

1:05:08

Amy as opposed to one you're

1:05:10

getting you know a significant jump

1:05:12

in performance but after 200 if

1:05:14

you go to 1, you basically

1:05:16

stop getting better. But if you

1:05:18

have a strong verifier, basically a

1:05:21

strong selector among the things, instead

1:05:23

of majority voting, you have some

1:05:25

sort of intelligent way to combine

1:05:27

the answers, you get a huge

1:05:29

jump, a huge difference between just

1:05:31

consistency at 100 and verification at

1:05:34

200. And one reason this is

1:05:36

very important is, first, they also

1:05:38

highlight the fact that verification is

1:05:40

a bit understudy, I think, in

1:05:42

the space of LLLMs, even introduce

1:05:44

a benchmark specifically for a verification.

1:05:47

The other reason that this is

1:05:49

very notable is that if you're

1:05:51

doing sampling based techniques, you can

1:05:53

paralyze the sampling and that is

1:05:55

different from extending your reasoning or

1:05:57

your search because reasoning via more

1:06:00

tokens is sequential, right? It's going

1:06:02

to take more time. Scaling via

1:06:04

sampling you can paralyze all the

1:06:06

samples. then just combine them, which

1:06:08

means that you can get very

1:06:10

strong reasoning at comparable timescales to

1:06:13

if you just take one output,

1:06:15

for instance. So that's a very

1:06:17

big deal. Yeah. Another key reason

1:06:19

why this works too, we covered

1:06:21

a paper, I think it was

1:06:23

months ago, that was pointing out

1:06:26

that if you look at all

1:06:28

the alignment techniques that people used

1:06:30

to get more value out of

1:06:32

their their models, right? They kind

1:06:34

of assumed this one. query one

1:06:37

output picture. Like let's align within

1:06:39

that context. Whereas in reality, what

1:06:41

you're often doing, especially with agents

1:06:43

and inference time compute, is you

1:06:45

actually are sampling like a large

1:06:47

number of outputs and you don't

1:06:50

care about how shitty the average.

1:06:52

generated sample is generated solutions. What

1:06:54

you care about is, is there

1:06:56

one exquisite one in this batch?

1:06:58

And what they did in that

1:07:00

alignment paper is they found a

1:07:03

way to kind of like upweight

1:07:05

just the most successful outcomes and

1:07:07

use that for a reward signal

1:07:09

or some kind of gradient update

1:07:11

signal. But this is sort of

1:07:13

philosophically aligned with that, right? It's

1:07:16

saying. you know, like you said,

1:07:18

self consistency is the view that

1:07:20

says, well, let's just do wisdom

1:07:22

in numbers and say we generated,

1:07:24

you know, I don't know, like

1:07:26

a hundred of these outputs. And

1:07:29

the most common response, most consistent

1:07:31

answer was this one. So let's

1:07:33

call this the one that we're

1:07:35

going to cite as our output.

1:07:37

But of course, if your model

1:07:39

has a consistent failure mode, that

1:07:42

can also cause you to true

1:07:44

self-consistency, identify that, you know, kind

1:07:46

of settle on that failureor mode.

1:07:48

in this big pot and that's

1:07:50

really what this is after. So

1:07:52

a lot of interesting times to

1:07:55

other lines of research as you

1:07:57

said and I think a really

1:07:59

interesting and important paper. And next

1:08:01

up we have the paper block

1:08:03

diffusion interpreting between outer aggressive and

1:08:05

diffusion language models and also I

1:08:08

think quite an interesting one. So

1:08:10

typically when you're using an OLM

1:08:12

you're using outer aggression which is

1:08:14

just saying that you're computing one

1:08:16

talk at a time, right? You

1:08:18

start with one word, then you

1:08:21

select the next word, then you

1:08:23

select the next word. You have

1:08:25

this iterative process, and that's a

1:08:27

limitation of traditional alms, because you

1:08:29

need to sequentially do this one

1:08:31

step at a time. You can't

1:08:34

generate an entire output sequence all

1:08:36

at once, as opposed to diffusion,

1:08:38

which is a generation mechanism, a

1:08:40

way to the generation that is

1:08:42

kind of giving you an entire

1:08:45

answer at once. And diffusion is

1:08:47

the thing that's typically used for

1:08:49

image generation where you start with

1:08:51

just a noisy image, a bunch

1:08:53

of noise, you iteratively update the

1:08:55

entire image all at once until

1:08:58

you get to a good solution.

1:09:00

And we covered, I believe, maybe

1:09:02

a week ago or two weeks

1:09:04

ago, a story of a diffusion-based

1:09:06

alum that was seemingly performed pretty

1:09:08

well. There was a company that

1:09:11

was a company that was a

1:09:13

company that was a company that

1:09:15

was a company that made the

1:09:17

claim, although we didn't provide too

1:09:19

much research on it. Well, this

1:09:21

paper is talking about, well, how

1:09:24

can we combine the strengths of

1:09:26

both approaches? The weakness of diffusion

1:09:28

is typically just doesn't work as

1:09:30

well for LLLMs, and there's kind

1:09:32

of various hypotheses, it's an interesting

1:09:34

question of why it doesn't work,

1:09:37

but it also doesn't work for

1:09:39

arbitrary length. You can only generate

1:09:41

a specific kind of horizon. and

1:09:43

some other technical limitations. So the

1:09:45

basic proposal in the paper is,

1:09:47

well, you can have this idea

1:09:50

of block diffusion where you still

1:09:52

sample aggressively, like sample one step

1:09:54

at a time, but instead of

1:09:56

sampling just one word or one

1:09:58

token, you use diffusion to generate

1:10:00

a chunk of stuff. So you

1:10:03

generate several tokens all at once

1:10:05

of diffusion in parallel. and then

1:10:07

you ought to aggressively keep doing

1:10:09

that and you get to be,

1:10:11

you know, the best of both

1:10:13

worlds. so to speak. So an

1:10:16

interesting idea, kind of architecture you

1:10:18

haven't seen before and potentially could

1:10:20

lead to stronger or faster models.

1:10:22

Yeah, it's also more paralyzable, right?

1:10:24

So the big advantage because these

1:10:26

within these blocks you're able to

1:10:29

just like denoise in one shot

1:10:31

with all the text in there.

1:10:33

It means you can parallelize more.

1:10:35

They try, I think, with blocks

1:10:37

of various sizes, like sizes of

1:10:39

four tokens, for example, right? So,

1:10:42

like, what's denoise four tokens at

1:10:44

a time? They do it with

1:10:46

this interesting kind of like, well,

1:10:48

it's essentially, they put masks on

1:10:50

and kind of gradually remove the

1:10:53

masks on the tokens as they

1:10:55

denoise. That's their interpretation of how

1:10:57

denoising would work. The performance is

1:10:59

lower, obviously, than state of the

1:11:01

art for. autoregressive models. Think of

1:11:03

this more as a proof of

1:11:06

principle that there are favorable, favorable

1:11:08

scaling characteristics. There's some promise here.

1:11:10

In my mind, this fits into

1:11:12

the kind of some of the

1:11:14

mamba stuff where, you know, the

1:11:16

next logical question is, okay, like,

1:11:19

how much scale will this work

1:11:21

at? And then do we see

1:11:23

those lost curves eventually with some

1:11:25

updates, some some hardware, jiggery, and

1:11:27

then crossing the loss curves that

1:11:29

we see for traditional auto-regressive modeling.

1:11:32

I mean, it is interesting either

1:11:34

way. They found a great way

1:11:36

to kind of break up this

1:11:38

problem. And one reason speculatively that

1:11:40

diffusion does not work as well

1:11:42

with text is partly that like

1:11:45

you tend to think as you

1:11:47

write. And so your previous words

1:11:49

really will affect in a causal

1:11:51

way where you're going and trying

1:11:53

to do diffusion. at the, you

1:11:55

know, in parallel across a body

1:11:58

of text like that, from an

1:12:00

inductive prior standpoint, doesn't quite match

1:12:02

that intuition, but could very well

1:12:04

be wrong. It's sort of like

1:12:06

top of mind thought, but anyway,

1:12:08

it's a good paper. The question

1:12:11

is always, will it scale, right?

1:12:13

And there are lots of good

1:12:15

proofs of principle out there for

1:12:17

all kinds of things, whether they

1:12:19

end up getting reflected in in

1:12:21

scaled training runs is the big

1:12:24

question. Exactly, and this does require

1:12:26

quite different training and you know

1:12:28

models from what all these other

1:12:30

alums are so decent chances won't

1:12:32

have a huge impact just because

1:12:34

you have to change up and

1:12:37

all these train models already are

1:12:39

alums in the out of aggressive

1:12:41

sense diffusion is the whole new

1:12:43

kind of piece of a puzzle

1:12:45

that isn't typically worked on but

1:12:47

nevertheless interesting as you said similar

1:12:50

to Mombop. Next, we have communication

1:12:52

efficient language model training, scales reliably,

1:12:54

and robustly scaling laws for deloco,

1:12:56

distributed law computation training. And I

1:12:58

think I'll just let you take

1:13:01

this one, Jeremy, since I'm sure

1:13:03

you dove deep into this. Oh,

1:13:05

yeah. I mean, so I just

1:13:07

think deloco, if you're interested in

1:13:09

anything from like national security to

1:13:11

the future data center design, is.

1:13:14

And you have China competition generally

1:13:16

is just so important and in

1:13:18

this general thread, right, well, like

1:13:20

together AI, distributed training, all that

1:13:22

stuff. So as a reminder, when

1:13:24

you're thinking about deloco, like how

1:13:27

does it work? So this basically

1:13:29

is the answer to a problem,

1:13:31

which is that traditional training runs

1:13:33

in like data peril training, like

1:13:35

in data centers today, happens in

1:13:37

this way where you have a

1:13:40

bunch of number crunching and a

1:13:42

bunch of communication, like a burst

1:13:44

of communication that has to happen.

1:13:46

at every time step you like

1:13:48

share gradients you update like across

1:13:50

all your GPUs you update your

1:13:53

model weights and then you go

1:13:55

on to the next the next

1:13:57

mini batch right or the next

1:13:59

part of your data set and

1:14:01

you repeat right run your computations

1:14:03

calculate the gradients update the model

1:14:06

weights and so on and there's

1:14:08

this bottleneck communication bottleneck that comes

1:14:10

up when you do this at

1:14:12

scale where you're like just waiting

1:14:14

for the communication to happen and

1:14:16

so the question then is going

1:14:19

to be okay well what if

1:14:21

we can set up Sort of

1:14:23

smaller pockets, because you're basically waiting

1:14:25

for the stragglers, right? The slowest

1:14:27

GPUs are going to dictate when

1:14:29

you find... sync up everything and

1:14:32

you can move on to the

1:14:34

next stage of training. So what

1:14:36

if we could set up a

1:14:38

situation where we have a really

1:14:40

small pocket of like a mini

1:14:42

data center in one corner working

1:14:45

on its own independent copy of

1:14:47

the model that's being trained? And

1:14:49

then another mini data center doing

1:14:51

the same and another and another

1:14:53

and very rarely we have an

1:14:56

outer loop where we do a

1:14:58

general upgrade of the whole thing

1:15:00

so that we're not constrained by

1:15:02

the lowest common denominator straggler in

1:15:04

that group. So this is going

1:15:06

to be the kind of philosophy

1:15:09

behind Du Loco. You have an

1:15:11

outer loop that essentially is going

1:15:13

to, in a more, think of

1:15:15

it as like this like wise

1:15:17

and slow, slow loop that updates

1:15:19

based on what it's learned from

1:15:22

all the local data centers that

1:15:24

are running their own training runs.

1:15:26

And then within each local data

1:15:28

center, you have this much more

1:15:30

radical, aggressive loop, more akin to

1:15:32

what we see in traditional data

1:15:35

parallel training in data centers, which,

1:15:37

you know, they're running. Anyway, we

1:15:39

have a whole episode on Deloco.

1:15:41

Check it out. I think we

1:15:43

talk about the atom optimizer or

1:15:45

atom W optimizer that runs at

1:15:48

the local level, and then the

1:15:50

sort of like more gradient descent

1:15:52

nester off momentum. Optimizer on the

1:15:54

outer loop. The details there don't

1:15:56

matter too much. This is a

1:15:58

scaling law paper basically. What they're

1:16:01

trying to figure out is how

1:16:03

can we study the scaling laws

1:16:05

that predict the performance, let's say,

1:16:07

of these models based on how

1:16:09

many model copies, how many many

1:16:11

data centers we have running at

1:16:14

the same time, and the size

1:16:16

of the models that we're training.

1:16:18

And they test at meaningful scales.

1:16:20

They go all the way up

1:16:22

to 10 billion parameters. And essentially

1:16:24

they were able through their their

1:16:27

scheme through their hyper parameter optimization

1:16:29

to reduce the total communication required

1:16:31

by a factor of over a

1:16:33

hundred in their scheme. It's a

1:16:35

great great paper. I think one

1:16:37

of the Wilder things about it

1:16:40

is that they find that even

1:16:42

when you just have a single

1:16:44

replica or let's say a single

1:16:46

mini data center, you still get

1:16:48

a performance lift. This is pretty

1:16:50

wild, like relative to the current

1:16:53

like purely data parallel scheme. So

1:16:55

it benefits you to have this

1:16:57

kind of radical quick updating interloop

1:16:59

and then to add on top

1:17:01

of that this slow wise outer

1:17:04

loop, even if you only have

1:17:06

a single data center that's like

1:17:08

you could you could just do

1:17:10

just the radical interloop and that

1:17:12

would be enough. But by adding

1:17:14

this like more strategic level of

1:17:17

optimization that comes from the Nestorov

1:17:19

momentum, the sort of slower gradient

1:17:21

updates for the outer loop, you

1:17:23

get better performance. That's highly counterintuitive,

1:17:25

at least to me, and it

1:17:27

does suggest that there's a kind

1:17:30

of stabilizing influence that you're getting

1:17:32

from just that new outer loop.

1:17:34

Last thing I'll mention is from

1:17:36

a sort of national security standpoint,

1:17:38

one important question here is how

1:17:40

fine-grained can this get? Can Deloco

1:17:43

continue to scale successfully if we

1:17:45

have not... one, not three, not

1:17:47

eight, but like a thousand of

1:17:49

these many data centers, right? If

1:17:51

that happens, then we live in

1:17:53

a world where essentially we're doing

1:17:56

something more like bit torrent, right,

1:17:58

for training models at massive scale,

1:18:00

and we live in a world

1:18:02

where it becomes a lot harder

1:18:04

to oversee training runs, right? If

1:18:06

we do decide that training models

1:18:09

at scale introduces WMP level capabilities

1:18:11

through cyber risk, through biarisk, whatever,

1:18:13

It actually gets to the point

1:18:15

where if there's a mini data

1:18:17

center on every laptop and every

1:18:19

GPU, if Deloca, that is the

1:18:22

promise of Deloco in the long

1:18:24

run, then yeah, what is the

1:18:26

meaningful tool set that policy has,

1:18:28

that government has, to make sure

1:18:30

that things don't get misused, that

1:18:32

you don't get the proliferation of

1:18:35

WMD level capabilities in these systems.

1:18:37

So I think it's a really,

1:18:39

really interesting question, and this paper

1:18:41

is a step in that direction.

1:18:43

like eight essentially of these many

1:18:45

data centers, but I think we're

1:18:48

going to see a lot more

1:18:50

experiments in that direction in the

1:18:52

future. Right, and this is following

1:18:54

up on. their initial paper, Deloco,

1:18:56

back in September of 2024, I

1:18:58

think we mentioned as a time

1:19:01

that it's quite interesting to see

1:19:03

Google, Google, deep mind publishing this

1:19:05

work because it does seem like

1:19:07

something you might keep secret. It's

1:19:09

actually quite impactful for how you

1:19:12

build your data centers, right? And

1:19:14

once again, you know. They do

1:19:16

very expensive experiments to train you

1:19:18

know billion parameter models to verify

1:19:20

that compared to the usual way

1:19:22

of doing things This is Comparable

1:19:25

let's say and can achieve similar

1:19:27

performance. So you know a big

1:19:29

deal if you're a company that

1:19:31

is building out data centers and

1:19:33

spending billions of dollars onto a

1:19:35

couple quicker stories because we are

1:19:38

as always starting to run out

1:19:40

of time. The first one is

1:19:42

Transformers without normalization, and that I

1:19:44

believe is from meta. They're introducing

1:19:46

a new idea called Dynamic 10H,

1:19:48

which is a simple alternative to

1:19:51

traditional normalization. So just keep a

1:19:53

simple, you have normalization, which is

1:19:55

when you make everything sum up

1:19:57

to one as a typical thing,

1:19:59

typical step in transformer architectures. And

1:20:01

what they found in this paper

1:20:04

is. you can get rid of

1:20:06

that if you add this little

1:20:08

computational step of 10H, basically a

1:20:10

little function that kind of flattens

1:20:12

things out. It ends up looking

1:20:14

similar to normalization. And that's quite

1:20:17

significant because normalization requires you to

1:20:19

do a computation over, you know,

1:20:21

a bunch of outputs all at

1:20:23

once. This is a per output

1:20:25

computation that... could have a meaningful

1:20:27

impact on the total computation kind

1:20:30

of requirements of the transformer. Next

1:20:32

up, I think, Jeremy, you mentioned

1:20:34

this one, we have an interesting

1:20:36

analysis measuring AI abdish. to complete

1:20:38

long tasks. And it is looking

1:20:40

at how 13 frontier AI models,

1:20:43

going from 2019 to 2025, how

1:20:45

long a time horizon they can

1:20:47

handle, and they found that the

1:20:49

ability to get to 50% task

1:20:51

completion time horizon, so on various

1:20:53

tasks that require different amounts of

1:20:56

work, That has been doubling approximately

1:20:58

every seven month and they have

1:21:00

you know a curve fit That's

1:21:02

kind of like a more's law

1:21:04

basically kind of introducing the idea

1:21:06

of a very strong trend towards

1:21:09

the models improving on this particular

1:21:11

measure Yeah, this is a really

1:21:13

interesting paper generated a lot of

1:21:15

a lot of discussion including by

1:21:17

the way a tweet from Andrew

1:21:20

Yang who treated this paper. He

1:21:22

said, guys, AI is going to

1:21:24

eat a shit ton of jobs.

1:21:26

I don't see anyone really talking

1:21:28

about this meaningfully in terms of

1:21:30

what to do about it for

1:21:33

people. What's the plan? Kind of

1:21:35

interesting, because like, I don't know,

1:21:37

it's, it's worlds colliding here a

1:21:39

little bit, right, that the political

1:21:41

and the, the like deep AGI

1:21:43

stuff, but yeah, this out of

1:21:46

meter, I think one of the

1:21:48

important caveats that's been thrown around

1:21:50

and fairly so is How long

1:21:52

does a task have to be

1:21:54

before an AI agent fails at

1:21:56

about 50% of the time? Right?

1:21:59

They call that the 50% time

1:22:01

horizon. And the observation is that,

1:22:03

yeah, this, as you said, like,

1:22:05

that time horizon has been increasing

1:22:07

exponentially quite quickly. Like, it's been

1:22:09

increasing doubling every seven months, as

1:22:12

they put it, which itself, I

1:22:14

mean. kind of worth flying. Doubling

1:22:16

every seven months, training compute for

1:22:18

frontier AI models grows, doubles every

1:22:20

six months or so. So it's

1:22:22

actually increasing at about the same

1:22:25

rate as training compute that we.

1:22:27

Now, that's not fully causal. There's

1:22:29

other stuff besides training compute that's

1:22:31

increasing the actual performance of these

1:22:33

models, including algorithmic improvements. But still,

1:22:35

it does mean we should expect

1:22:38

kind of an exponential course of

1:22:40

progress towards ASI if our benchmark

1:22:42

of progress towards ASI is this

1:22:44

kind of like 50% performance threshold,

1:22:46

which I don't think is unreasonable.

1:22:48

It is true that it depends

1:22:51

on the task, right? So not

1:22:53

all tasks show the same rate

1:22:55

of improvement. not all tasks show

1:22:57

the same performance, but what they

1:22:59

do find is that all tasks

1:23:01

they've tested basically show an exponential

1:23:04

trend. That itself is a really

1:23:06

important kind of detail. The tasks

1:23:08

they're focusing on here though, I

1:23:10

would argue, are actually the most

1:23:12

relevant. These are tasks associated with

1:23:15

automation of machine learning engineering tasks,

1:23:17

machine learning research, which is explicitly

1:23:19

the strategy that open AI, that

1:23:21

anthropic, that Google, are all gunning

1:23:23

for, Can we make AI systems

1:23:25

that automate AI research so that

1:23:28

they can rapidly get better at

1:23:30

improving themselves or at improving AI

1:23:32

systems and you close this loop

1:23:34

and you know we get recurs

1:23:36

of self-improvement essentially and then take

1:23:38

off to super intelligence? I actually

1:23:41

think this is quite relevant. One

1:23:43

criticism that I would have of

1:23:45

the curve that they show that

1:23:47

shows doubling time being seven months

1:23:49

and by the way they extrapolate

1:23:51

that to say, okay so then

1:23:54

we should assume that based on

1:23:56

this, AI systems will be able

1:23:58

to automate many software tasks that

1:24:00

currently take humans a month sometime

1:24:02

between late 2028 and early 2031.

1:24:04

So those are actually quite long

1:24:07

ASI timelines if you think ASI

1:24:09

has achieved once AI can do

1:24:11

tasks that it takes humans about

1:24:13

a month to do, which I

1:24:15

don't know, maybe fair. If you

1:24:17

actually look at the curve though,

1:24:20

it does noticeably steepen much more

1:24:22

recently and in precisely kind of

1:24:24

the point where synthetic data self-improvement

1:24:26

like reinforcement learning on chain of

1:24:28

thought with verifiable rewards. So basically

1:24:30

the kind of strawberry concept started

1:24:33

to take off. And so I

1:24:35

think that's pretty clearly its own

1:24:37

distinct regime. Davidad had a great

1:24:39

set of tweets about this. But

1:24:41

fundamentally, I think that there is

1:24:43

maybe an error being made here

1:24:46

and not recognizing a new regime.

1:24:48

And it's always going to be

1:24:50

debatable because the sample sizes are

1:24:52

so small. But we're taking a

1:24:54

look at that plot, seeing if

1:24:56

you agree for yourself, that the

1:24:59

sort of last, I guess, six

1:25:01

or so entries. in that plot

1:25:03

actually do seem to chart out

1:25:05

a deeper slope and a meaningfully

1:25:07

steeper one that could have, I

1:25:09

mean, that same one month R&D

1:25:12

benchmark being hit more like, you

1:25:14

know, even 2026, even 2025? Pretty

1:25:16

interesting. Yeah, and I think it's

1:25:18

important to know that this is

1:25:20

kind of a general idea. You

1:25:23

shouldn't kind of take this too

1:25:25

literally. Obviously, it's a bit subjective

1:25:27

and very much depends on the

1:25:29

tasks. Here is a relatively small.

1:25:31

data set they're working with. So

1:25:33

it's mainly software engineering tasks. They

1:25:36

have three sources of tasks, Hcast,

1:25:38

which is 97 software tasks that

1:25:40

range from one minute to 30

1:25:42

hours. They have seven difficult machine

1:25:44

learning research engineering tasks that take

1:25:46

eight hours each. And then these

1:25:49

software atomic actions that take one

1:25:51

second to 30 seconds for software

1:25:53

engineers. pretty limited variety of tasks

1:25:55

in general and the length of

1:25:57

tasks you by a way is

1:25:59

meant to be measured in terms

1:26:02

of how long they would take

1:26:04

to a human professional which of

1:26:06

course there is also variance there.

1:26:08

So the idea is I think

1:26:10

more so the general notion of

1:26:12

X percent task completion time horizon

1:26:15

which is defined as For a

1:26:17

task that it takes X amount

1:26:19

of time for human to complete

1:26:21

can an AI complete it successfully

1:26:23

You know 50% of time 80%

1:26:25

of time. So right now we

1:26:28

are going up to eight hours.

1:26:30

Again, it's not talking about how

1:26:32

fast the AI is itself, by

1:26:34

the way. It was just talking

1:26:36

about success rate. So interesting ideas

1:26:38

here on tracking the future in

1:26:41

the past. And just one more

1:26:43

thing to cover. And it is

1:26:45

going to be H-cast, human-collibrated Tommy

1:26:47

software tasks. So this is the

1:26:49

benchmark of 189 machine learning, cyber

1:26:51

security, software engineering, and also general

1:26:54

reasoning tasks. use a subset of

1:26:56

that I think in the previous

1:26:58

analysis. So they collected 563 human

1:27:00

baselines, that's over 1,500 hours from

1:27:02

various people, and that collected a

1:27:04

data set of tasks to then

1:27:07

measure AI on. All right, moving

1:27:09

on to policy and safety. We

1:27:11

start with Zochi, an ontology project.

1:27:13

This is an applied research lab,

1:27:15

and they have announced. This project,

1:27:17

Sochi, which they claim is the

1:27:20

world's first artificial scientist. I know,

1:27:22

right? The difference, I suppose, here

1:27:24

is that they got Sochi to

1:27:26

publish a peer-reviewed paper at Eichler,

1:27:28

at Eichler workshops, I should note.

1:27:31

This AI wrote a paper submitted

1:27:33

it, the reviewers, human reviewers, of

1:27:35

this prestigious conference Eichler, then reviewed

1:27:37

it and let it, thought it

1:27:39

was worthy of publication. So that

1:27:41

seems to be a proof of

1:27:44

concept that we are getting to

1:27:46

a point where you can make

1:27:48

AI do AI research, AI research

1:27:50

conference, which Jeremy is, as you

1:27:52

noted, of a very important detail

1:27:54

in tracking where we are with

1:27:57

the ability of AI to improve

1:27:59

itself and kind of potentially reaching

1:28:01

super intelligence. Yeah, and then we've

1:28:03

got... So I think we've covered

1:28:05

a lot of these, but there

1:28:07

are quite a few labs that

1:28:10

have claimed to be the first

1:28:12

AI scientist, you know, Sakana famously,

1:28:14

the AI scientist, you know, Google

1:28:16

has its own sort of research

1:28:18

product and then there's auto science.

1:28:20

But this one is, I will

1:28:23

say, quite impressive, how to look

1:28:25

at some of the papers and

1:28:27

they're good. The authors, the company.

1:28:29

If that's the right term for

1:28:31

ontology guy, I couldn't find any

1:28:33

information about them I looked on

1:28:36

crunch base I looked on pitch

1:28:38

book I tried you know, try

1:28:40

a lot of stuff I think

1:28:42

that this is kind of their

1:28:44

coming up party, but they don't

1:28:46

have that I could see any

1:28:49

information about like where they come

1:28:51

from what their funding is and

1:28:53

all that so hard to know

1:28:55

we do know based on the

1:28:57

the papers they put out that

1:28:59

their co-founders are Andy Zoo and

1:29:02

Ron Ariel. Ron Ariel previously was

1:29:04

at Intel Labs under Joshua Bach.

1:29:06

So if you're a fan of

1:29:08

his work, you know, it might

1:29:10

kind of ring a bell and

1:29:12

then a couple other folks. So

1:29:15

I think maybe most useful is

1:29:17

there's fairly limited information about the

1:29:19

actual model itself and how it's

1:29:21

set up, but it is a

1:29:23

multi-agent setup that we do know.

1:29:25

It's produced a whole bunch of

1:29:28

interesting papers. So I'm just gonna

1:29:30

mention one. It's called CS-Reft, I

1:29:32

guess it's how you'd pronounce a

1:29:34

compositional subspace representation fine tuning. And

1:29:36

just to give you an idea

1:29:39

of how creative this is, so

1:29:41

you have a model, if you

1:29:43

try to retrain the model to

1:29:45

perform a specific task, it will

1:29:47

forget how to do other things.

1:29:49

And so what they do is

1:29:52

essentially identify a subspace of the

1:29:54

parameters in the model to focus

1:29:56

on one task, and then a

1:29:58

different. subspace of those parameters to

1:30:00

or sorry I should say not

1:30:02

even the parameters of the activations

1:30:05

to apply a transformation in order

1:30:07

to have the model perform a

1:30:09

different task and so anyway just

1:30:11

cognizant of our time here I

1:30:13

don't think we have time to

1:30:15

go into detail but this like

1:30:18

paper I'm frustrated that I discovered

1:30:20

it through the Zocte paper because

1:30:22

I will have wanted to cover

1:30:24

this one, like for our full

1:30:26

time block here, it is fascinating.

1:30:28

It's actually really, really clever. They

1:30:31

also did, anyway, some other kind

1:30:33

of AI safety, red teaming, vulnerability

1:30:35

detection stuff that is also independently

1:30:37

really cool. But bottom line is,

1:30:39

and this is an important caveat,

1:30:41

so this is not 100% automated.

1:30:44

They have a human in the

1:30:46

loop. reviewing stuff. So they take

1:30:48

a look at the intermediate work

1:30:50

that Zochi puts out before allowing

1:30:52

it to do further progress. Apparently

1:30:54

this happens at three different key

1:30:57

stages. The first is before sensitive,

1:30:59

sorry, before extensive experimentation starts, so

1:31:01

before a lot of, say, computing

1:31:03

resources report in, after the results

1:31:05

are kind of solidified, they say,

1:31:07

so somewhat grounded, but before the

1:31:10

manuscript is written. And then again,

1:31:12

after the manuscript is written. So

1:31:14

you still do have a lot

1:31:16

of a lot of humans in

1:31:18

the loop acting to some extent

1:31:20

as a hallucination filter among other

1:31:23

things. There is by the way

1:31:25

a section in this paper on

1:31:27

recursive self-improvement. That I thought was

1:31:29

interesting. They say during development we

1:31:31

observed early indications of this recursive

1:31:34

advantage, they're referring here to recursive

1:31:36

self-improvement, when Zoshi designed novel algorithmic

1:31:38

components to enhance the quality of

1:31:40

its generated research hypotheses. These components

1:31:42

were subsequently incorporated into later versions

1:31:44

of the system architecture, improving overall

1:31:47

research quality. So this isn't all

1:31:49

going to happen at once necessarily

1:31:51

with one generation of model, but

1:31:53

it could happen pretty quick. And

1:31:55

I think this is kind of

1:31:57

an interesting canarian of coal mine

1:32:00

for recursive self-improvement. Right, exactly. And

1:32:02

yeah, as you said, I think,

1:32:04

looking at the papers, they seem...

1:32:06

significantly more creative and interesting than

1:32:08

what we've seen, for instance, from

1:32:10

Sakana. It's the Kana had a

1:32:13

lot of built-in structure as well.

1:32:15

It was pretty easy to criticize.

1:32:17

Worth noting, also, this takes a

1:32:19

high-level research direction as an input,

1:32:21

and the human can also provide

1:32:23

input at any time. high level

1:32:26

feedback at any time, which apparently

1:32:28

is also used during paper writing.

1:32:30

So a lot of caveats to

1:32:32

the notion that this is an

1:32:34

AI scientist, right? It's an AI

1:32:36

scientist over human advisor slash supervisor,

1:32:39

but the outputs and the fact

1:32:41

that they got published is pretty

1:32:43

impressive. With AI doing the majority

1:32:45

of work and no time to

1:32:47

get into the details, but this

1:32:49

is similar to previous work in

1:32:52

AI research where they give it

1:32:54

a structured kind of plan where,

1:32:56

you know, they give it a

1:32:58

high level of direction, it does

1:33:00

ideation, it makes a plan, hypothesis

1:33:02

generation, it goes off and does

1:33:05

experiments and eventually it starts writing