0:00

Today on the A.I. Daily A.I.

0:02

Brief, X.A. has merged with X,

0:04

and here's what it means for

0:06

the broader frontier lab

0:08

battle. Before that in the

0:11

headlines, A.I. sees its first big

0:13

IPO this year, and we talk

0:15

about how it went. The A.I.

0:18

Daily Brief is a daily podcast

0:20

and video about the

0:22

most important news and

0:24

discussions in A.I.I. News You Need

0:27

in around five minutes. We kick

0:29

off today with another big story

0:31

from Friday. Our main story is

0:33

about XAI combining with X. This

0:36

one is about Corweave's IPO, which unfortunately

0:38

turned into a bit of

0:40

an anti-climax. The company stock

0:42

dropped by 2.5% at launch on

0:44

Friday after fundraising targets were heavily

0:47

downsized. Corweve raised 1.5 billion from

0:49

the share sale, but had initially priced

0:51

the IPO to raise over 2.7 billion.

0:53

At one point in the planning phase, the stock

0:55

was to be priced at $55 per share, but

0:58

ended up going out at just $40. The stock

1:00

ended the day flat, closing at $40

1:02

and receiving no IPO pop. Bloomberg sources

1:04

said that half of the shares went

1:06

to the three largest investors in the deal,

1:09

with 90% going to the top 15. One of

1:11

those large investors was invidia who

1:13

took a $250 million allocation to add

1:15

to their pre-existing 6% stake in the

1:17

company. CEO Michael and Trader said that

1:20

without invidia, the IPO quote wouldn't have

1:22

closed. He also added if 27 others

1:24

didn't show up it wouldn't have closed.

1:26

And while the headlines are calling this

1:28

a dud of an IPO and pointing to

1:30

Coreweave specifically in AI more generally, what

1:33

in traders pointing out is that this is

1:35

a terrible moment to go public. The

1:37

NASDAQ index also had a 2.5% overall

1:39

drop on Friday contributing to a 13%

1:41

decline this month. Risk assets have been

1:44

struggling mightily in a market that

1:46

is characterized by insecurity and volatility

1:48

and hyped up tech IPOs squarely

1:50

in the risk asset category. Now to

1:52

the extent that people are looking at what

1:55

wasn't working about the IPO in the context

1:57

of Corweaven AI rather than larger market volatility,

1:59

people are reading this either as a

2:01

potential signal for an AI infrastructure bubble

2:03

or they're pointing to some idiosyncratic warning

2:05

signs in Corweave specifically. On the AI

2:08

bubble conversation Corweave co-founder Brandon McBee is

2:10

dismissive saying this conversation around an AI

2:12

bubble seems to come up every three

2:14

to six months or so and then

2:16

it drops away. What we see on

2:18

the ground and what I'm sure you're

2:20

hearing in Silicon Valley is just consistent

2:23

growing demand. Obviously if you are a

2:25

regular listener to this show you'll know

2:27

that that opinion is much closer to

2:29

my view than the idea that there's

2:31

some massive bubble just waiting to deflate.

2:33

The core weave specific problems might be

2:35

a little bit more difficult to brush

2:38

off. Core weave already has quite a

2:40

bit of debt and may need to

2:42

raise more to make up for the

2:44

shortfall in the IPO, that is of

2:46

course if they actually needed that capital.

2:48

The company faces repayments of $7.5 billion

2:50

by the end of next year, although

2:53

they could also be able to refinance.

2:55

Like many companies in this AI infrastructure

2:57

space, they also have a highly concentrated

2:59

customer base. Microsoft represented 62% of their

3:01

revenue last year, with a further 15%

3:03

coming from an unknown single large customer.

3:05

Microsoft has already walked away from their

3:08

option to extend leases with the company.

3:10

However, Corweve did seal a big deal

3:12

with Open AI in the lead-up to

3:14

the IPO. Bloomberg's Dave Lee commented that

3:16

unlike other big cloud providers, Corweve really

3:18

doesn't have anywhere to hide. He wrote,

3:20

While Corweave has some unique vulnerabilities, the

3:23

bigger picture here is that its accounts

3:25

will finally lay bare in quarterly reports,

3:27

the brutal economics of an industry that

3:29

is burning through an unprecedented amount of

3:31

cash in pursuit of some kind of

3:33

lucrative application, nobody has quite figured out

3:35

yet figured out yet, nobody has quite

3:38

figured out yet yet. Corweave can't obfuscate

3:40

growth in AI services, or can't obfuscate

3:42

growth in AI services, or can it

3:44

hide the interconnectedness of the industry, or

3:46

if the interconnected Joe Sai worn this

3:48

week. It's on the balance sheet of

3:50

core weave where the clues might emerge,

3:53

written, for the first time, in plain

3:55

black and white for all to see.

3:57

I continue to be skeptical of this

3:59

type of analysis, but if for no

4:01

other reason that as a meta understanding

4:03

of where the market is, it's worth

4:05

noting this is a fairly common opinion.

4:08

Speaking of... of IPOs, Proplexity CEO Arvan

4:10

Shritovas is denying that the company is

4:12

under financial pressure and needs to rush

4:14

to an IPO. A few days ago,

4:16

a Reddit user called Nothing Ever Happened,

4:18

aired out a theory on Proplexity's subreddit,

4:20

writing, I've recently noticed Proplexity making lots

4:23

of changes to cut costs. My theory

4:25

is that they're doing horribly financially. Those

4:27

charges included an insider telling them that

4:29

all funding for marketing and partnerships has

4:31

been paused. and layoffs which the editor

4:33

discovered, quote unquote, by digging into LinkedIn

4:35

profiles and finding a lot of former

4:38

employees. The key complaint were changes to

4:40

how the services uses auto mode, which

4:42

now removes model selection from the user

4:44

during follow-up questions. The editor claimed that

4:46

their follow-up questions were always answered by

4:48

the default cheaper model, rather than a

4:50

high-end reasoning model like OpenAIs01. Unless we

4:53

think that this is just one complaining

4:55

user, perplexity CEO are of Unshrinivos took

4:57

to post a response, which he also

4:59

copied over onto X as well. Now

5:01

he didn't reference the original post, but

5:03

did give some plausible explanations for each

5:05

of the points, and included several others

5:08

addressing complaints about degradation of service. Regarding

5:10

auto mode, Trinovas claimed that it was

5:12

a UX improvement to remove the model

5:14

selection and follow-up questions. He wrote that

5:16

the goal is to quote, let the

5:18

AI decide for the quick fast-answer query,

5:20

or a slightly slower multi-step pro search

5:23

query, or slow reasoning mode query, or

5:25

a really slow deep research query. The

5:27

long-term future is that decides the amount

5:29

of compute to apply to a question.

5:31

and maybe clarify with the user when

5:33

not super sure. Our goal isn't to

5:35

save money and scam you in any

5:38

way. It's genuinely to build a better

5:40

product with less clutter and simple selector

5:42

for customization options for the technically adept

5:44

and well-informed users. This is the right

5:46

long-term convergence point. And by the way,

5:48

I will say at this point that

5:50

one of the big UI-UX complaints around

5:53

these services has been model selector type

5:55

of issues. This is something that Tim

5:57

Altman and Chatmen and ChatGPT have discussed

5:59

extensively extensively extensively as well. Users hate

6:01

the fact that they have to look

6:03

through and understand which of the models

6:05

is good for different things. So I

6:08

don't think it's some conspiracy theory to

6:10

think that perplexity... which is an extremely

6:12

UI-U-X focused company, is just trying to

6:14

improve that part of the experience. Now

6:16

maybe more pointedly, was this paragraph from

6:18

Shinavas who wrote, are we running out

6:20

of funding and facing market pressure to

6:23

IPO? No, we have all the funding

6:25

we've raised and our revenue is only

6:27

growing. The objective behind auto mode is

6:29

to make the product better, not to

6:31

save costs. I've learned it's better to

6:33

communicate more transparently to avoid any incorrect

6:35

conclusions. Re IPO, we have no plans

6:38

of IPO before 2028. Ultimately, when it

6:40

comes to perplexity, I think that their

6:42

bigger challenge is that every other frontier

6:44

lab also wants to be the gateway

6:46

to search. They're all working to improve

6:48

not only their underlying models, but also

6:50

their search experience. That, more than any

6:53

potential cost savings from auto model selection,

6:55

that more than any potential cost savings

6:57

from auto model selection, is going to

6:59

be the challenge that perplexity has to

7:01

overcome. As of recent notes, perplexity claims

7:03

that they've crossed $100 million in annualized

7:05

and annualized revenue. For now that that

7:08

is going to do it for today's

7:10

AIDily Brief Headlines edition, next up the

7:12

main episode. Today's episode is brought to

7:14

you by Vanta. Trust isn't just earned,

7:16

it's demanded. Whether you're a startup founder

7:18

navigating your first audit or a seasoned

7:20

security professional scaling your GRC program, proving

7:23

your commitment to security has never been

7:25

more critical or more complex. That's where

7:27

Vanta comes in. Businesses use Vanta to

7:29

establish trust by automating compliance needs across

7:31

over 35 frameworks like Sock 2. and

7:33

ISO 2701. Centralized security workflows complete questionnaires

7:35

up to 5X faster and proactively manage

7:38

vendor risk. Vanta can help you start

7:40

or scale up your security program by

7:42

connecting you with auditors and experts to

7:44

conduct your audit and set up your

7:46

security program quickly. Plus with automation and

7:48

AI throughout the platform Vanta gives you

7:50

time back so you can focus on

7:53

building your company. Join over 9,000 global

7:55

companies like Atlacian, Cora and Factory who

7:57

use Vanta to manage risk and improve

7:59

security in real time. For a limited

8:01

time, this audience gets $1,000 off Vanta

8:03

at vanta.com. That's vanta.com/N-L-W for $1,000 off.

8:05

All right, a ideally brief listeners. Today

8:08

I'm excited to tell you about the

8:10

disruption incubator. One of the things that

8:12

our team sees all the time is

8:14

a lot of frustration from enterprises. There's

8:16

a fatigue around small incremental solutions, a

8:18

concern around not thinking big enough, tons

8:20

of bureaucratic challenges, of course, inside big

8:23

companies. And frankly, we just hear all

8:25

the time from CEOs, CTOs, other types

8:27

of leaders that they want to ship

8:29

some groundbreaking AI agent or product or

8:31

feature. In many cases, they even have

8:33

a pretty well thought out vision for

8:35

what this could be. Their teams are

8:38

just not in an environment conducive to

8:40

that type of ambition. Well, it turns

8:42

out our friends at Fractional have experienced

8:44

the exact same thing. Fractional are the

8:46

top AI engineers specializing in transformative AI

8:48

product product development and to answer this

8:50

particular challenge. They have, with perhaps a

8:53

little bit of help from super intelligent,

8:55

set up what they're calling the disruption

8:57

incubator for exactly this type of situation.

8:59

The idea of the disruption incubator is

9:01

to give a small group of your

9:03

most talented people an overly ambitious mandate,

9:05

something that might have taken one to

9:08

two years within their current construct. Send

9:10

them to San Francisco to work with

9:12

the team at fractional, and within two

9:14

to three months, ship something that would

9:16

have previously been impossible. The idea here

9:18

is that you are not just building

9:20

some powerful new agent or AI feature,

9:23

but you're actually investing in your AI

9:25

leadership at the same time. If this

9:27

is something interesting to you, send us

9:29

a note at Agent at B super

9:31

dot AI with the word disruption in

9:33

the title and we will get right

9:35

back to you with more information. Again

9:38

that's Agent at B super dot AI

9:40

with disruption in the subject line. Welcome

9:42

back to the AI Daily Brief. Today

9:44

we are discussing the state of the

9:46

battle among the frontier labs for AI

9:48

supremacy and the specific context for the

9:50

conversation is... Late on Friday, we got

9:53

news, Elon Musk's X-A-I, which is of

9:55

course the parent company of Groc, and

9:57

the home of all his generative AI

9:59

adventures. had acquired X, which is of

10:01

course the former Twitter. The announcement

10:03

post which came out at 5.20

10:05

p.m. Eastern Time on Friday read,

10:07

XAI has acquired X in an

10:09

all-stock transaction. The combination values XAI

10:12

at 80 billion and X at 33 billion,

10:14

which is 45 billion less 12 billion in

10:16

debt. Since its founding two years ago,

10:18

XAI has rapidly become one of

10:20

the leading AI labs in the

10:22

world, building models and data centers

10:24

at unprecedented speed and scale. X is

10:26

the digital town square where more than 600

10:29

million active users go to find the real-time

10:31

source of ground truth and in the last

10:33

two years has been transformed into one of

10:35

the most efficient companies in the world positioning

10:37

it to deliver scalable future growth. X AI

10:40

and X's futures are intertwined. Today we

10:42

officially take the step to combine the

10:44

data models compute distribution and talent. This

10:46

combination will unlock immense potential by blending

10:48

X AI's advanced AI capability and expertise

10:50

with X's massive reach. The combined company

10:53

will deliver smarter, more meaningful experiences to

10:55

billions of people while staying true to

10:57

our core mission of seeking truth and

10:59

advancing knowledge. This will allow us to

11:01

build a platform that doesn't just reflect the

11:04

world, but actively accelerates human progress.

11:06

I would like to recognize the hardcore

11:08

dedication of everyone at XAI and X that has

11:10

brought us to this point. This is just the

11:12

beginning. Now on the one hand, the companies

11:15

are both private, and Elon presumably has the

11:17

support of investors, so we can pretty much

11:19

do hear what he wants. Still the deal is

11:21

far from normal. The Wall Street Journal reports,

11:23

the new valuations were determined during

11:25

negotiations between the two Musk arms,

11:27

which both had the same advisors,

11:29

people familiar with the matter said. The

11:31

last time X AI raised money was in December,

11:33

and it was thought to be valued at

11:35

around $40 billion, so this deal implies a

11:37

doubling in three months. That's obviously quite

11:40

an acceleration, but not necessarily totally out

11:42

of sync with the world of AI.

11:44

The journal points out that this isn't the first

11:46

time Elon has done something like this. Back in

11:48

2016, Elon used Tesla stock to buy

11:50

his solar energy company, Solar City. Musk

11:52

is apparently a great dealmaker when

11:55

he's negotiating with himself. Still, if

11:57

you hold aside the mechanics and the

11:59

valuations, it's... very clear why this deal

12:01

makes sense for the two companies. Musk

12:03

has already been open that the Grock

12:05

model was trained on X data, and

12:07

the chatbot is now embedded in the

12:09

platform as a native assistant. The two

12:11

platforms are already deeply entwined in their

12:13

user experience, their resources, and even some

12:15

of their personnel. For X, the merger

12:17

takes the pressure off of it to

12:20

thrive exclusively as an independent social media

12:22

platform. Advertising revenue has not been without

12:24

challenges since Musk took over in 2022

12:26

over in 2022. And while there have

12:28

been signs that the numbers had recovered

12:30

had recovered in the past few months

12:32

had recovered in the past few months,

12:34

X now has the additional economic value

12:36

of simply becoming a data repository and

12:38

a portal for X-A-I. Opinions on this

12:40

deal, as with so much in Elon

12:42

Musk world, basically come down to what

12:44

you think of Elon Musk. The pro-

12:46

Musk side is represented by posts like

12:49

this one from Fernando Cow, which writes,

12:51

When Musk bought Twitter, everyone was confused.

12:53

Why would a man focused on electric

12:55

vehicles, space travel, and neural interfaces want

12:57

a social media platform? But, a social

12:59

media platform. X had something incredibly valuable

13:01

that most AI companies desperately need. Real-time,

13:03

human-generated diverse data from 600 million active

13:05

users. This is the perfect fuel for

13:07

AI models, and it's exactly what XAI

13:09

needs to compete with Open AI and

13:11

Enthropic. Investor Chamath Palahapatiya writes, the currently

13:13

best-ranked consumer AI model has just acquired

13:15

the most complete corpus of scaled real-time

13:18

information on the internet. The data will

13:20

be a part of the pre-training to

13:22

make the models XAI makes even more

13:24

differentiated more differentiated. This is a smart

13:26

move in a moment when other model

13:28

makers are caught up and slowed down

13:30

in copyright lawsuits, like open AI for

13:32

training data or pre-training quality like meta.

13:34

On the flip side is the common

13:36

take represented by this one from Compound

13:38

248. They write, it's hard to know

13:40

what to make of XAI buying X.

13:42

My gut is its max of desperation.

13:44

On the surface, the deal values X

13:46

flat to Twitter's 2022 takeout value, despite

13:49

massive underperformance on financial metrics. Sounds like

13:51

a win for X shareholders. But it

13:53

is a stock deal and ex-owners will

13:55

now own 29% of combined company shifting

13:57

from a near pureplay social media bet

13:59

to an AI bet that's very much

14:01

on the complex a deluded share in

14:03

Twitter. Yes, XAI is a powerful model,

14:05

but not unusually so. XAI has de

14:07

minimis revenue, is hemorrhaging cash, and its

14:09

prospective business opportunity seems very difficult given

14:11

the relevant competition. A has a head

14:13

start, B is a murderer's row, and

14:15

C has existing business and go-to-market strategies

14:18

to build on. X's $12 billion of

14:20

very high-cost debt isn't going away. It

14:22

will be in perpetual cash raising mode

14:24

until that changes, which leaves it risk

14:26

to the whimss of the fund- I

14:28

wouldn't bet against Elon, but I'd be

14:30

very nervous as a combined company owner.

14:32

Now honestly, this is actually fairly middle

14:34

of the road. A more reflective antagonistic

14:36

take comes from Adam Cochran who wrote,

14:38

in other words, Musk used his pumped-up

14:40

XAI stock to pay multiple times over

14:42

value for XAI stock to pay multiple

14:44

times over value for X, but still

14:46

taking 11 billion loss on the transaction

14:49

while screwing over XAI investors and X

14:51

investors and to sell your data to

14:53

his own AI company. But otherwise it's

14:55

not a breakthrough model and it's terribly

14:57

monetized. Again, the middle of the road

14:59

take really focuses on data. Ralph Paul

15:01

from Real Vision writes, it was always

15:03

about the data for the AI. I

15:05

talked about this when he first bought

15:07

X and said it was a bargain

15:09

back then due to the AI training

15:11

data. And the AI is all about

15:13

the robots and the robots are all

15:15

about Mars, as is everything else. I

15:18

think that while the focus on data

15:20

makes sense, people might be underestimating the

15:22

value of the integrated experience with Twitter

15:24

content. To the extent that these companies

15:26

are all competing to be the next

15:28

generation search portal where people begin their

15:30

internet experience, Grock offers something fundamentally different

15:32

that none of the competitors, Google, Open

15:34

AI, Antropic, Proplexity, etc., offer, which is

15:36

the ability to integrate the meta-conversation into

15:38

deep research. I think I have a

15:40

particular point of view on this given

15:42

that I've now built two podcasts, for

15:44

both of which a major value proposition

15:46

is the fact that we don't just

15:49

talk about the discussion around the news.

15:51

The thing that takes longest in producing

15:53

both the breakdown in the AI Daily

15:55

Brief is going through and understanding dozens

15:57

if not hundreds of different opinions around

15:59

anything that's happening. in order to be

16:01

able to synthesize that into a coherent

16:03

view not just of what actually happened,

16:05

but what's likely to happen next based

16:07

on how people are receiving that news.

16:09

This is again not commenting on any

16:11

of the questions of self-dealing or valuations

16:13

or anything like that. I just think

16:15

that the Grock Twitter merger has value

16:18

beyond just a pre-training data play. So

16:20

what's happening though beyond Grock if we're

16:22

using this as a way to catch

16:24

up on the state of the AI

16:26

frontier lab battle? While elsewhere in the

16:28

AI space, the biggest players seem to

16:30

be ducking it out for leadership in

16:32

the major verticals. We have, of course,

16:34

discussed extensively open AI's new image generator,

16:36

which has been just absolutely sucking all

16:38

of the oxygen out of the room.

16:40

But because it was so dominant last

16:42

week, a lot of people missed the

16:44

fact that anthropics dominance in coding seems

16:46

to be contested for the first time

16:49

in months. The same GPT40 update that

16:51

made chatGPT so much better at image

16:53

coding. According to Artificial Analysis's Intelligence Index,

16:55

GPT-40 is now the top non-reasoning model

16:57

overtaking Entropics Clause 3.7 sonnet. Now, their

16:59

index combines a range of different coding

17:01

and knowledge benchmarks to come up with

17:03

a blended intelligence score, but digging into

17:05

the coding-specific scores, GPT-40 is now at

17:07

the top of the leaderboard. At the

17:09

same time though, where there's also a

17:11

ton of chatter that actually, in fact,

17:13

Google's Gemini 2.5, isn't even better coding

17:15

standard. During last week's release, the reasoning

17:18

model was clearly a high-performing coding assistant

17:20

based on the benchmarks, but having tested

17:22

it now for a few days, Venture

17:24

Beat broke down why this model could

17:26

be a big step up for programmers.

17:28

They noted that, like Open AI's models,

17:30

Google provides full access to chain of

17:32

thought reasoning. For programmers, that means you

17:34

can follow the model along precisely and

17:36

audit their results, picking up and correcting

17:38

errors along the way. Venture Beat wrote,

17:40

In practical terms, this is a breakthrough

17:42

for trust and steerability. Enterprise users evaluating

17:44

output for critical tasks like reviewing policy

17:47

implications, coding logic, or summarizing complex research

17:49

can now see how the model arrived

17:51

at an answer. That means they can

17:53

validate, correct, or redirect it with more

17:55

confidence. It's a major evolution from the

17:57

black box field that still plagues many

17:59

LLLM outputs. Many coders have also discovered the

18:01

Gemini 2.5 pro is much better at succeeding

18:03

at one-shot tasks. The strong reasoning is

18:05

a possible explanation. The model lays out its

18:07

design and code structure before writing a single

18:09

line of code. Now this could also be

18:12

just an artifact of the observability, allowing

18:14

programmers to see exactly what the model

18:16

is doing throughout. Another benefit that could

18:18

help is geminize 1 million token context

18:20

window. Anthropic is only now preparing to

18:22

release a 500,000 token context window for

18:24

Claude, an upgrade from the 200,000 tokens

18:26

they offer currently. Large Context Windows allow

18:28

bigger code bases to be uploaded, and

18:30

more importantly, to be understood by the

18:32

model while working on coding problems. One

18:35

feature that also feels like it's under-explored

18:37

at the moment, is the new workflows

18:39

that are opened up by the multimodal

18:41

reasoning capabilities present in Gemini 2.5 Pro.

18:43

Like the new version of GPT-4O, Gemini 2.5

18:45

can apply native reasoning to image inputs. This

18:47

is valuable for more than allowing these models

18:50

to easily edit images to easily edit

18:52

images. Developers are starting to realize there.

18:54

Yancey Min and A.I. Figma Plug-in designer

18:56

walked through a tinkering session with GBT-40.

18:58

First, he discovered that the model can

19:00

take interface code and generate an image

19:02

of the interface. Then he found the

19:05

model can modify the code based on

19:07

visual alterations to the interface. In

19:09

this case, Min-brushed over a tab,

19:11

in visual alterations to the interface. In

19:13

this case, Min, brushed over a tab in

19:15

the visual alterations to the interface. In this

19:18

case, Min, Min brushed over barely to a

19:20

gantic support, tweeting, tweeting, tweeting, tweeting,

19:22

tweeting, tweeting, to MCP or not

19:24

to MCP, that's the question. Let

19:27

me know in the comments. As you might imagine,

19:29

the replies were strongly in favor

19:31

of Google supporting the universal protocol

19:33

for agentic tooling. And I actually think

19:35

maybe to take a step back and sum

19:37

this all up, this is a really reflective

19:39

evolution of the frontier model battle. Increasingly,

19:42

the competition is going to be

19:44

less about general performance and more

19:46

about specific use cases. We're evolving

19:48

in such a way that people are actually

19:50

integrating these tools at the core of new

19:52

and existing workflows. and they're picking the best

19:54

models and the best interfaces and the best

19:56

experiences and the best products for whatever it

19:59

is they're trying to do. Coding is clearly

20:01

one of the breakout use cases, which

20:03

is why there's so much competition

20:05

around it. around it. Deep -style searching is also

20:07

clearly going to be a core

20:09

experience, and that's why the XAIX merger is more

20:11

than just about more than just about Elon's

20:13

financial engineering. Ultimately, I expect the

20:15

the course of the next year not

20:17

only continuous upgrades to the underlying models,

20:19

but more and more focus on these

20:21

specific domain areas and specific use cases

20:23

rubber the rubber is actually hitting the

20:25

road when it comes to the business

20:27

applications of the underlying technology. Anyways

20:29

interesting stuff to kick to our week, our

20:31

but that's going to do it for

20:33

today's do Brief. today's AI Daily you listening or

20:35

watching you always, or and until next time, until

20:37

next time, peace.