0:00

Today on the AI Daily Brief,

0:02

Entropic introduced

0:04

their power user tier. Before

0:06

that in the headlines, Google

0:09

Next is all about agents.

0:11

The AI Daily Brief is

0:14

a daily podcast and video

0:16

about the most important

0:18

news and discussions in

0:20

AI. To join the

0:23

conversation, follow the Discord link

0:25

in our show notes. First

0:27

up, as I've been mentioning a couple times,

0:29

for those of you who are looking for

0:32

an ad-free version of the A.I. Daily Brief,

0:34

you can now head on over to patron.com/A.I.

0:36

Daily Brief to find that. Second, next week

0:38

is spring break, so you will not not

0:40

have shows, but they will be a little

0:43

bit different than normal. Lastly, something I

0:45

want to gauge people's perspective on. The

0:47

A.I. Daily Brief community has been hugely

0:49

supportive of and important in the superintelligent

0:51

story. We're. However, I'm trying to gauge

0:53

interest. If this is something you think

0:55

we should explore, send me a note

0:58

at NLW at B super dot AI

1:00

with super in the title. Thanks in

1:02

advance for your perspective. And with that,

1:04

let's get into today's show. Welcome back

1:06

to the AI Daily Brief Headlines edition, all the

1:08

Daily AI News you need in around five minutes.

1:11

We kick off today with Antropic getting themselves

1:13

a power user tier. Now, you might

1:15

remember when Open Open AI introduced their

1:17

200 dollar a month tier back in

1:19

December, many believe that the demand just

1:21

wouldn't be there. However, while we don't know how

1:24

popular exactly that subscription is, it's generated enough

1:26

usage that Sam Altman said that the company

1:28

was actually losing money on the deal. Entropic

1:30

is calling their version of the premium subscription Claude

1:32

Max. The product allows users to pay to

1:34

get around Entropic's to pay to get around

1:36

Entropic's notoriously troublesome rate limits, gets priority responses

1:38

during heavier traffic periods, and early access to

1:41

new features thrown in as a sweetener. If

1:43

you've spent any time on AI Twitter, you'll

1:45

know that people have been constantly asking Entropic for the

1:47

ability to pay more for more for more for more

1:49

for more for more for more for more for more

1:51

service. There's actually two different levels of

1:53

Claude Max. For $100, users get five times

1:56

the rate limits of the $20 a month

1:58

pro tier. That math works out. But for

2:00

those users who really burn through tokens, they'll

2:02

want the $200 level that allows for 20

2:04

times the rate limit. What's not on

2:07

offer, unfortunately, for some is unlimited

2:09

usage. So is this about boosting revenue or

2:11

simply recognizing that power users are underserved?

2:13

As they say in the business, why

2:15

not both? Anthropics product lead Scott White

2:18

said that the company isn't ruling out

2:20

adding even more pricey subscriptions, saying we

2:22

could see a $500 a month level.

2:24

Ultimately, he said that the product roadmap

2:26

is guided by user feedback. And the

2:28

one loud, consistent piece of user feedback we've

2:30

seen over the past year is that power

2:32

users want to pay for more usage. Next

2:34

up, a little more on the terror fallout, Invidia

2:36

has secured a carve-out for their

2:39

China-based chips. Industry insiders had widely

2:41

expected the administration to clamp down on

2:43

export of H20 chips, which is the

2:45

downgraded GPUs that are designed to get

2:48

around export controls. NPR is reporting that

2:50

the additional restrictions won't go ahead after Jensen

2:52

Huang attended a dinner at Maralaga last week.

2:54

Sources said that restrictions had been in the

2:57

works for months and were ready to be

2:59

implemented as soon as this week. They said that the

3:01

president changed his mind after Huang promised new

3:03

data center investment in the US. Chris Miller

3:05

at Tufts University History Professor and semiconductor

3:07

expert commented, even though these chips are

3:10

specifically modified to produce their performance, thus

3:12

making them legal to sell to China,

3:14

they are better than many perhaps most

3:16

of China's homegrown chips. China still can't produce

3:18

the volume of chips it needs domestically, so

3:20

it is critically reliant on imports of invidious

3:23

chips. Then again, that view is a little up

3:25

in the air after recent reports of efficient training

3:27

runs on new Huawei chips. However, even if

3:29

the Chinese AI industry is trying to wean

3:31

themselves off of invidia, the market is still

3:33

critical for the dominant chipmaker. Thirteen percent of

3:36

invidious official demand comes from China, and that

3:38

figure could be much higher if you account

3:40

for evasion of export controls through Southeast Asia.

3:42

All in all, it adds up to us

3:44

continuing to not have a coherent picture of

3:46

the administration's strategy when it comes to chip

3:48

controls. Following inauguration, Trump pledged to wind

3:50

back many of the restrictions on the AI

3:53

industry. However, the enhanced export controls introduced in

3:55

the final weeks of the Biden presidency are

3:57

still in place. These regulations put limits...

3:59

on a huge portion of the world

4:02

including friendly countries like Israel and India.

4:04

Then again this does also seem like reinforcement

4:06

of the idea that there's always a deal

4:08

to be made when it comes to Trump.

4:10

Now over the last couple of days you

4:12

may have heard me wax poetic around what

4:14

I think the implications of some of this

4:16

tariff stuff are likely to be on venture

4:19

capital. Nominally I think that it's going to

4:21

be a harder fundraising environment not only for

4:23

startups but also for VCs themselves, but

4:25

countering that point appears to be Andrei

4:28

and Horowitz and Horowitz. Sources said

4:30

the firm is looking to capitalize

4:32

on high international demand for investments

4:34

in American companies. They added that international

4:36

LPs view the fund as a way to

4:38

more easily invest money in the USAI sector

4:41

without the restrictions. So it sounds like this

4:43

actually might be playing into and taking advantage

4:45

of some of the tariffs. Last year A16Z

4:47

raised 7.2 billion scattered across the

4:49

themes of American dynamism, apps, games,

4:51

infrastructure and growth. This fund then is

4:54

both significantly larger and more focused

4:56

than previous efforts. The gigantic size

4:58

brings up questions about whether venture capital

5:00

can scale up in this rarefied air.

5:02

Softbank is perhaps the obvious comparison. They

5:05

raised their $100 billion vision fund in

5:07

2017 with very mixed results. The second

5:09

vision fund raised in 2019 is a

5:11

relatively more modest but still massive 56

5:13

billion. The other comp that springs to mind

5:16

is Sequoio, who currently manage over 56

5:18

billion in assets overall. Still there hasn't

5:20

been a venture strategy as capital intensive as

5:22

AI in the past. In fact, part of

5:24

the reason that companies like Open AI had

5:26

to turn to big tech partners like Microsoft

5:28

is that there simply wasn't enough dry powder

5:30

in the venture capital coffers for them to

5:33

get what they needed. Reuter sources said that

5:35

a significant portion of the fund would

5:37

be set aside for follow-on investments in

5:39

companies already in A16Z's portfolio. And

5:42

with portcos like Mistral and Safe Super Intelligence

5:44

and data bricks, there is a lot of

5:46

money to be spent. There is certainly a lot

5:48

of capital need. Lastly today, a little

5:50

bit of institutional psychodrama. Open AI has countersued

5:52

Elon Musk asking the court to bring an

5:55

end to the billionaire's legal challenge. The court

5:57

filing called for Musk to be prohibited from

5:59

taking... quote, further unlawful and unfair action

6:01

and held responsible for the damage he

6:04

has already caused. It stated, Open AI

6:06

is resilient, but Musk's actions have taken a

6:08

toll. Should his campaign persist, greater harm is

6:10

threatened to Open AI's ability to govern in

6:13

service of its mission to the relationships that

6:15

are essential to furthering that mission and the

6:17

public interest. Musk's continued attacks

6:19

on Open AI culminating most recently in

6:21

a fake takeover bid designed to disrupt

6:23

Open AI's disrupt Open AI's future must

6:25

cease. Musk's attorney immediately fired back in

6:28

a press statement he said. Had Open AI's

6:30

board genuinely considered Musk's bid as they were

6:32

obligated to do, they would have seen how

6:34

serious it was. It's telling that having to

6:36

pay fair market value for Open AI's assets

6:38

allegedly interferes with their business plans. The case

6:40

is currently at a slow point as the parties

6:43

await a jury trial next spring. Musk's attempt

6:45

to seek an injunction to stop Open AI

6:47

from converting to a non-profit was rejected in

6:49

March. So technically there isn't anything

6:51

stopping California Attorney General Rob Bonsa

6:53

from making a decision on the

6:55

conversion, however complaints continue to roll

6:57

in and the active litigation

7:00

gives him a good excuse to delay.

7:02

Meanwhile, of course, Open AI has a

7:04

huge financial incentive to get this wrapped

7:06

up quickly. The company's latest fundraising round

7:08

featured $10 billion from Soft Bank that

7:10

is contingent on the conversion being completed

7:12

by the end of the year. That friends

7:14

is going to do it for today's Aye

7:17

Daily Brief Headlines Edition, next up the main

7:19

episode. In our fast-paced world, having

7:21

a solid AI plan can

7:23

make all the difference. Enabling

7:25

organizations to create new value,

7:28

grow, and stay ahead of

7:30

the competition is what it's

7:32

all about. KPMG is here to

7:34

help you create an AI strategy

7:37

that really works. Don't wait. Now's

7:39

the time to get ahead. Check

7:41

out real stories from KPMG of

7:44

how AI is driving success

7:46

with its clients at KPMG.US.

7:48

US slash AI. Today's episode is brought

7:50

to you by Vanta. Vanta is a trust management

7:52

platform that helps businesses automate

7:54

security and compliance, enabling them

7:57

to demonstrate strong security practices

7:59

and scale. In today's business landscape,

8:01

businesses can't just claim security. They have

8:03

to prove it. Achieving compliance with a

8:05

framework like SOC 2, ISO 2701, HIPAA,

8:07

GDPR, etc. is how businesses can demonstrate

8:10

strong security practices. And we see how

8:12

much this matters every time we connect

8:14

enterprises with agent services providers

8:16

at Superintelligent. Many of these

8:18

compliance frameworks are simply not

8:21

negotiable for enterprises. The problem

8:23

is that navigating security and compliance

8:25

is time-consuming and complicated. It can take

8:27

months of work and use up valuable

8:29

time and resources. Vanta makes it easy

8:31

and faster by automating compliance across 35

8:33

plus frameworks. It gets you audit ready

8:36

in weeks instead of months and saves

8:38

you up to 85% of associated

8:40

costs. In fact, a recent IDC white

8:42

paper found that Vanta customers achieve $535,000

8:44

per year in benefits, and the platform

8:46

pays for itself in just three months.

8:48

The proof is in the numbers. More

8:51

than 10,000 global companies trust Vanta,

8:53

including Atlian, Cora, Cora, and more. For

8:55

a limited time, listeners get

8:57

$1,000 off at vanta.com/

8:59

NLW. That's va-n-t-a.com/NLW for $1,000 off.

9:02

Today's episode is brought to you

9:04

by Super Intelligent and more specifically

9:06

Super's Agent Readiness Audits. If you've

9:09

been listening for a while, you

9:11

have probably heard me talk about

9:13

this. But basically the idea of

9:15

the agent readiness audit is that

9:18

this is a system that we've

9:20

created to help you benchmark and

9:22

map opportunities in your organizations where

9:25

agents could specifically help you solve

9:27

your problems, create new opportunities in

9:29

a way that again is completely

9:31

customized to you. When you do one

9:33

of these audits, what you're going to

9:35

do is a voice-based agent interview where

9:37

we work with some number of your

9:39

leadership and employees to map what's going

9:41

on inside the organization and to figure

9:44

out where you are in your agent

9:46

journey. That's going to produce an agent

9:48

readiness score that comes with a

9:50

deep set of explanations, strength, weaknesses,

9:52

key findings, and of course, a

9:54

set of very specific recommendations that then

9:56

we have the ability to help you

9:58

go find the right part. to actually

10:00

fulfill. So if you were looking for

10:02

a way to jumpstart your agent strategy,

10:04

send us an email at agent at

10:06

b-super dot AI, and let's get you

10:08

plugged into the agentic era. Welcome back

10:10

to the AI Daily Brief. In a

10:12

massive shock to literally no one paying

10:15

any attention to this space, Google Next

10:17

is all about agents. Yes, friends, this

10:19

week we are headed into the next

10:21

round of big tech conference season. And

10:23

unsurprisingly, Google's Cloud Next conference featured a

10:25

huge lineup of AI announcements designed to

10:27

all in some, I think, make the

10:29

technology feel and be more useful. The

10:31

company's annual cloud conference was held in

10:33

Las Vegas this week and was absolutely

10:35

squarely focused on taking the next step

10:37

in the AI race. We've got agentic

10:39

infrastructure, new models, a new AI chip,

10:42

and much, much more. A pair of

10:44

announcements about agentic infrastructure could end up

10:46

having the most impact of all. Back

10:48

a couple of weeks ago on March

10:50

30th, in the wake of Open AI

10:52

announcing that they were going to support

10:54

MCP or the Model Context Protocol, Google

10:56

CEO Sundar Pichai tweeted to MCP or

10:58

not to MCP, that's the question. Let

11:00

me know in the comments. One point

11:02

eight million views and a thousand comments

11:04

later, we got the answer and it

11:06

was a yes. Deep Mine CEO Demiss

11:08

Asavis tweeted, MCP is a good protocol

11:11

and it's rapidly and it's rapidly becoming

11:13

and SDK. look forward to developing it

11:15

further with the MCP team and others

11:17

in the industry. And to reiterate, this

11:19

means that all three of the leading

11:21

US labs are now supporting MCP. Now

11:23

for those of you who don't know

11:25

what the heck I'm talking about, we

11:27

covered MCP in depth a couple weeks

11:29

back, but in short, if you're just

11:31

trying to understand the implications, it means

11:33

a significant boost for interoperability and compatibility

11:35

across the rapidly developing agentic infrastructure layer.

11:38

The second big agentic announcement was the

11:40

unveiling of Google's agent development kit and

11:42

a new interoperability standard called agent to

11:44

agent. As the name suggests... the standard

11:46

is seeking to harmonize the way agents

11:48

communicate with each other rather than how

11:50

they interact with tools. Rouse, who are

11:52

up in any VP of Google's cloud

11:54

business application platform, insisted that A2A is

11:56

not competing with MCP, which of course

11:58

is more about tool use. They said,

12:00

we see MCP and A2A as complementary

12:02

capabilities. The way we are looking at

12:05

agent to agent is at a higher

12:07

layer of abstraction to enable applications in

12:09

agents to talk to each other. So

12:11

think of it as a layered stack

12:13

where MCP operates for tools and data.

12:15

Words, words, words, but basically the point

12:17

is that they're trying explicitly to not

12:19

compete with MCP here, but instead be

12:21

a standard for something different. H2A has

12:23

50 companies on board to support the

12:25

open standard, including Salesforce, service now, and

12:27

workday. Sir Apanini said that Google isn't

12:29

necessarily looking to compete with other consortiums

12:32

working on their own solutions, saying, we

12:34

will look at how to align with

12:36

all of the protocols. There will always

12:38

be some protocol with a good idea,

12:40

and we want to figure out how

12:42

to compete with a good idea, and

12:44

a good idea, and a good ideas,

12:46

Standard ways for agents to coordinate could

12:48

reduce the amount of complexity, for example,

12:50

in multi-agent systems, could mean that agents

12:52

talk to their counterparts and other companies,

12:54

making them more capable of getting work

12:56

done without getting humans involved. So what

12:59

is it all amount to? Well, MIT

13:01

PhD Tobin-South says, my take on Ado

13:03

away from Google and friends is that

13:05

they're trying to create a communications hierarchy

13:07

with MCP as tool use and A-to-A

13:09

as coordination and communication. Hubspot founder and

13:11

now agent.a creator Darmes Shah writes, shockingly

13:13

this doesn't change everything. It's very very

13:15

early, but here are my initial thoughts.

13:17

I'm a big believer in multi-agent networks

13:19

in agent-to-agent communication. It's good that there's

13:21

now an open standard out there for

13:23

it. They cover some very key needs.

13:26

Capability discovery, agents being able to send

13:28

messages to each other, being able to

13:30

work on tasks that are long-lived, to

13:32

a-sink, weaving human UX into the agentic

13:34

flow, etc, etc. This is not a

13:36

replacement for MCP. In fact, during their

13:38

announcement post, Google included a helpful diagram

13:40

that illustrates how EDA and MCP fit

13:42

together. Still, this feels a bit... heavy

13:44

to me. It's trying to do a

13:46

lot. In a way that's good because

13:48

you get a bunch of capabilities out

13:50

of the box like Asing tasks and

13:53

user experience negotiation, but the trade-off is

13:55

that heavier protocols are harder to implement

13:57

and as such you don't get the

13:59

quick adoption you see with lighter weight

14:01

things, so I don't anticipate MCP-style adoption.

14:03

Reading between the lines, this feels like

14:05

they're solving for a lot of mega

14:07

enterprises and big consulting firms looking to

14:09

build multi-agent systems inside the corporation. It's

14:11

less about connecting agents' will be interesting

14:13

to see actual usable implementations of this

14:15

outside the Fortune 1000 companies. Overall, this

14:17

is good news though, moves us further

14:20

down the multi-agent systems road. Other random

14:22

little agentic notes, Gemini Code Assist is

14:24

getting an agentic upgrade. This is their

14:26

cursor competitor nominally and it can now

14:28

deploy agents to complete complex programming tasks

14:30

across multiple steps. This kind of agentic

14:32

feature has been a game changer on

14:34

other platforms with programmers using it to

14:36

automate repetitive tasks like code migration. Google

14:38

has also released a security agent as

14:40

part of their new unified security platform.

14:42

The goal is to have an agent

14:44

on the beat that can recognize and

14:46

remediate threats before they become major problems.

14:49

The Chief Information Security Officer at Charles

14:51

Schwab, Bashar Abbasido said, Google is transforming

14:53

security operations and enabling our vision to

14:55

stay proactive in responding to cyber threats.

14:57

The platform has empowered our team to

14:59

focus on strategic initiatives and high-value work.

15:01

We didn't get a massive new model,

15:03

but what we did get was Gemini

15:05

2.2. Like its predecessor, this is a

15:07

smaller model designed to deliver efficient performance

15:09

with extremely low latency. The big change

15:11

is a lot more customization. Google wrote,

15:13

you can do the speed, accuracy, and

15:16

cost balancing for your specific needs. This

15:18

flexibility is key to optimizing flash performance

15:20

and high volume cost sensitive applications. Gemini

15:22

2.5 flash is a reasoning model at

15:24

launch and will likely end up being

15:26

the cheapest on the market. The model

15:28

is designed to adjust the depth of

15:30

reasoning based on the complexity of the

15:32

prompt, which is similar it seems to

15:34

the approach that Open AI is pursuing.

15:36

Google is aiming to provide a model

15:38

that hits the sweet spot between performance

15:40

and cost writing. It's the ideal engine

15:43

for responsive virtual assistance and real-time... summerization

15:45

tools where efficiency at scale is key.

15:47

Google also announced that they plan to

15:49

bring Gemini models to on-premise deployments starting

15:51

in Q3. On the chip side, Google

15:53

has announced the seventh generation of their

15:55

tensor processing unit, which they're calling Ironwood.

15:57

Now, unlike GPUs, are specifically designed for

15:59

AI computing tasks. GPUs, meanwhile, are more

16:01

generalized across all mathematical functions. In fact,

16:03

in the early days, it was a

16:05

slight coincidence that the chip architecture that

16:07

powers 3D gaming was also highly performed

16:10

for AI. The core bet with developing

16:12

TPUs instead of GPUs was that specially

16:14

designed architecture would be more efficient. So

16:16

far that theory hasn't totally played out

16:18

with invidious GPU architectures still at the

16:20

top of the pack. However, Google is

16:22

hoping that Ironwood might finally validate that

16:24

thesis. The company claims that their new

16:26

processor can deliver 24 times the computing

16:28

power of the world's fastest supercomputer when

16:30

deployed at scale. Previous generations of the

16:32

hardware were designed for inference. I'm in

16:34

Vadat, Google's vice president of machine learning

16:37

said, Ironwood is built to support this

16:39

next phase of generative AI and its

16:41

tremendous computational and communication requirements. This is

16:43

what we call the age of inference,

16:45

where AI agents will proactively retrieve and

16:47

generate data to collaboratively deliver insights and

16:49

answers, not just data. Now there are

16:51

a bunch of numbers and specs that

16:53

come with this thing, but rather than

16:55

try to explain an exa-flop, the promise

16:57

here is that Ironwood is around twice

16:59

as in invidious age 100. Still the

17:01

biggest difference is actually scale. A maximum

17:04

size pod of Blackwell B200 chips is

17:06

576 chips before they require outside networking,

17:08

while Ironwood claims to be capable of

17:10

being deployed in a 9, 216 chip

17:12

pod. This is also a massive jump

17:14

from Google's previous generation of TPUs, which

17:16

were called Trillium. The company has achieved

17:18

a fourfold increase in computing power compared

17:20

to the 2024 model. Ironwood is also

17:22

more efficient, delivering twice the performance per

17:24

watt compared to Trillium. And this TPU

17:26

and the focus on efficient inference definitely

17:28

suggests that Google is scaling up to

17:31

service compute-hungry reasoning models and the... agents

17:33

by them. Now there is

17:35

more going is more

17:37

going on In a well.

17:39

of In a big

17:41

vote of confidence

17:43

for Google's Silicon, Ilia announced

17:45

that his startup

17:47

will use Google will use

17:49

Google Samsung announced that

17:51

Gemini would be

17:53

added to their be added

17:55

to their new Google's Google's

17:58

platform now features a

18:00

music generation model.

18:02

Ultimately, while there wasn't

18:04

one big huge

18:06

thing, at least not

18:08

a hitch you

18:10

over the head, over the

18:12

head. a pretty remarkable

18:14

shift from from black

18:16

glue on pizza on

18:18

all of this in

18:20

just about a

18:22

year. a year. AI for success writes,

18:25

Google Deep Mind is destined

18:27

to win the win

18:29

the AGI why, they

18:31

have the data advantage,

18:33

they own the they

18:35

own distribution channel, have all the

18:37

best models right now, they have

18:39

everything and they dominate all four

18:41

key areas, and they foundation models, key accelerator,

18:43

hardware. I'm not sure how

18:46

this all shakes out, but if you're

18:48

Google, you gotta be happy that that is

18:50

the narrative among many these days. Google, That's

18:52

gonna do it for today's AI Daily Brief. the

18:54

Appreciate you listening or watching as always That's

18:56

until next time, peace. AI Daily Brief. Thank

19:02

you.