0:00

Today on the A.I. Daily

0:02

Brief, former open A.I. co-founder

0:04

Andre Carpathy, on how LLM's

0:07

flipped the script on technological

0:09

diffusion, or, as I'm framing

0:11

it, A.I. Empowerment. The A.I.

0:14

Daily Brief is a daily

0:16

podcast and video about the

0:18

most important news and

0:20

discussions in A.I. To

0:22

join the conversation, follow the

0:25

discord link in our show. First up,

0:27

as I've been mentioning a couple times,

0:29

for those of you who are looking

0:31

for an ad-free version of the AI

0:33

Daily Brief, you can now head on

0:35

over to patreon.com slash AI Daily

0:37

Brief to find that. Lastly, something

0:40

I want to gauge people's perspective

0:42

on. The AI Daily Brief community

0:44

has been hugely supportive of and

0:47

important in the super intelligent story.

0:49

We're considering reserving part of our

0:51

current round for investors from this

0:54

community. However I'm trying to gauge

0:56

interest. Thanks in advance for your perspective.

0:58

And with that, let's get into today's

1:00

show. Hello friends, welcome back to another long

1:02

reads episode of the AI Daily Brief. One

1:05

more quick reminder before we get into it.

1:07

As you are listening to this, I am

1:09

in Florida, probably just about to surprise our

1:12

kids with their first Disney World trip, which

1:14

should be great. As I mentioned, though, this

1:16

means that next week's episodes are a little

1:18

bit different. We have a slate of really

1:21

interesting interviews, lots of talk about agents and

1:23

vibe coding and big technological changes, stuff that

1:25

I think will be a really interesting and

1:27

enlightening change of pace. If some crazy thing

1:29

happens, you can be assured that I will

1:32

be a really interesting and enlightening change of

1:34

pace. If some crazy thing happens, you can

1:36

be sure that I will be back at

1:38

the end of next week with an Easter

1:40

Long Read Sunday and then normal episodes to

1:42

follow that. For now though, let's jump over

1:44

into this recent post from Andre Carpathy.

1:47

It's basically a blog post but shared

1:49

to X. And it's always a real treat

1:51

when we get to read a piece from

1:53

one of the big thinkers in the space.

1:55

Especially because these types of pieces tend to

1:57

be more conversational, let's say, then the essays...

1:59

that get published in an op-ed section

2:02

in Bloomberg or something. By choosing to

2:04

put this on X, Andre's inviting discourse

2:06

in conversation, and we've seen in the

2:09

past what happens when Andre invites discourse

2:11

in conversation. When the man introduced the

2:13

term vibe coding, which although it's kind

2:16

of been warped from how he was

2:18

originally using it, is obviously one of

2:20

the most influential concepts of the year.

2:22

In any case, this piece is called

2:25

Power to the People, how LLLM's flipped

2:27

flip the script on technology diffusion technology

2:29

diffusion, and we're going to read it.

2:32

Once again, you guys are on a

2:34

roll, this is actually me reading it,

2:36

rather than AI. Andre writes, transformative technologies

2:39

usually follow a top-down diffusion path, originating

2:41

in government or military contexts, passing through

2:43

corporations, and eventually reaching individuals. Think electricity,

2:46

cryptography, computers, flight, the internet, or GPS.

2:48

This progression feels intuitive. New and powerful

2:50

technologies are usually scarce, capital intensive, and

2:52

their use requires specialized technical expertise in

2:55

the early stages. So it strikes me

2:57

as quite unique and remarkable that LLLM's

2:59

display a dramatic reversal of this pattern.

3:02

They generate disproportionate benefit for regular people,

3:04

while their impact is a lot more

3:06

muted and lagging in corporations and governments.

3:09

Chat GPT is the fastest growing consumer

3:11

application in history, with 400 million weekly

3:13

active users, who use it for writing,

3:16

coding, translation, deep research, brainstorming, etc. This

3:18

is not a minor upgrade to what

3:20

existed before. It's a major multiplier to

3:23

an individual's power level across a broad

3:25

range of capabilities, and the barrier to

3:27

use is incredibly low. The models are

3:29

cheap, free even, fast, available to anyone

3:32

on demand behind a URL or even

3:34

local machine, and they speak anyone's native

3:36

language, including tone slang or emoji. This

3:39

is insane. As far as I can

3:41

tell, the average person has never experienced

3:43

the technological unlock this dramatic this fast.

3:46

Why then are the benefits a lot

3:48

more muted in the corporate and government

3:50

realms? I think the first reason is

3:53

that L alarms offer a very specific

3:55

profile of capability, that of merely quasi-expert

3:57

knowledge and performance, but simultaneously across... a

3:59

very wide variety of domains. In other

4:02

words, they are simultaneously versatile but also

4:04

shallow and fallible. Meanwhile, an organization's unique

4:06

superpower is the ability to concentrate diverse

4:09

expertise into a single entity by employing

4:11

engineers, researchers, analysts, lawyers, lawyers, marketers, etc.

4:13

While LLMs can certainly make these experts

4:16

more efficient individually, e.g. drafting initial legal

4:18

clauses, generating boilerplate code, etc. The improvement

4:20

to the organization takes the form of

4:23

becoming a bit better at the things

4:25

it could already do. In contrast, an

4:27

individual will usually only be an expert

4:29

in at most one thing. So the

4:32

broad quasi-experties offered by the LLM fundamentally

4:34

allows them to do things they couldn't

4:36

do before. People can now vibe code

4:39

apps. They can approach legal documents. They

4:41

can grogesoteric research papers. They can do

4:43

data analytics. They can do data analytics.

4:46

They can generate multimodal content for branding

4:48

and marketing. They can do all of

4:50

this at an adequate capability without involving

4:53

an additional expert. Think, various integrations, legacy

4:55

systems, corporate brand or style guides, stringent

4:57

security protocols, privacy considerations, internationalization, regulatory compliance,

5:00

and legal risk. There are a lot

5:02

more variables, a lot more constraints, a

5:04

lot more considerations, and a lot lower

5:06

margin for error. It's not so easy

5:09

to pull all of it into a

5:11

context window. You can't just vibe code

5:13

something. You might be one disastrous hallucination

5:16

away from losing your job. And third,

5:18

there is the well-documented inertia of a

5:20

larger organization. featuring culture, historical precedence, political

5:23

turf that escalate in periods of rapid

5:25

change, communication overhead, retraining challenges of a

5:27

distributed workforce, and good old-fashioned bureaucracy. These

5:30

are major headwinds when it comes to

5:32

rapid adoption of a sparkling new versatile

5:34

but shallow and fallible tool. I don't

5:36

wish to downplay the impacts of LLLMs

5:39

and corporations or governments, but at least

5:41

for the moment and in aggregate across

5:43

society, they have been significantly more life-altering

5:46

for individuals than they have been for

5:48

organizations. Mary Jim and Joe are experiencing

5:50

the majority of the benefit, not Google

5:53

or the government of the United States.

5:55

Looking forward, the continued diffusion of LLLM's

5:57

of course depends on continued performance improvement

6:00

and its capability profile. The benefit distribution

6:02

overall is particularly interesting to chart and

6:04

depends heavily on the dynamic range of

6:06

the performance as a function of capital

6:09

expenditure. Today, frontier-grade LLLM performance is very

6:11

accessible and cheap. Beyond this point, you

6:13

cannot spend a marginal dollar to get

6:16

better performance, reliability or autonomy. Money can't

6:18

buy better chat-GPT. Bill Gates talks to

6:20

GPT-40 just like you do. But can

6:23

this be expected to last? Train time

6:25

scaling, increasing parameters and data, test time

6:27

scaling, increased time, and model ensembles increased

6:30

batch are forces increasing the dynamic range.

6:32

On the other hand, model distillation, the

6:34

ability to train disproportionately powerful small models

6:36

by training to mimic the big model

6:39

has been a forced decreasing dynamic range.

6:41

Certainly the moment money can buy dramatically

6:43

better chat GTT, things change. Large organizations

6:46

get to concentrate their vast resources to

6:48

buy more intelligence, and within the category

6:50

of individual too, the elite may once

6:53

again split away from the rest of

6:55

society. Their child will be tutored by

6:57

GPT-8 Pro Max High, yours by GPT-6

7:00

mini. But at least at this moment

7:02

in time, we find ourselves in a

7:04

unique and unprecedented situation in the history

7:07

of technology. If you go back through

7:09

various sci-fi, you'll see that very few

7:11

would have predicted the AI-revolution would feature

7:13

this progression. It was supposed to be

7:16

a top-secret government mega-brain project wielded by

7:18

the generals, not chat-GPT, appearing basically overnight

7:20

and for free on a device already

7:23

in for free on a device already

7:25

in, Surprise, the future is already here

7:27

and it is shockingly distributed. Power to

7:30

the people. Personally, I love it. Today's

7:32

episode is brought to you by Vanta.

7:34

Vanta is a trust management platform that

7:37

helps businesses automate security and compliance, enabling

7:39

them to demonstrate strong security practices and

7:41

scale. In today's business landscape, businesses can't

7:43

just claim security, they have to prove

7:46

it. Achieving compliance with a framework like

7:48

SOC 2, ISO 2701, HIPPA, GDPR, and

7:50

more. is how businesses condemn it. demonstrate

7:53

strong security practices. And we see how

7:55

much this matters every time we connect

7:57

enterprises with agent services providers at super

8:00

intelligent. Many of these compliance frameworks are

8:02

simply not negotiable for enterprises. The problem

8:04

is that navigating security in compliance is

8:07

time-consuming and complicated. It can take months

8:09

of work and use up valuable time

8:11

and resources. VANTA makes it easy and

8:13

faster by automating compliance across 35 plus

8:16

frameworks. It gets you audit ready in

8:18

weeks instead of months and saves you

8:20

up to 85% of associated costs. In

8:23

fact, recent IDC White Paper found that

8:25

Vanta customers achieve $535,000 per year in

8:27

benefits, and the platform pays for itself

8:30

in just three months. The proof is

8:32

in the numbers. More than 10,000 global

8:34

companies trust Vanta, including Atlasian, Cora, and

8:37

more. For a limited time, listeners get

8:39

$1,000 at vanta.com/Nelw for $1,000 dollars off.

8:41

Today's episode is brought to you by

8:44

Super Intelligent and more specifically Super's Agent

8:46

Readiness Audits. If you've been listening for

8:48

a while you have probably heard me

8:50

talk about this, but basically the idea

8:53

of the Agent Readiness Audit is that

8:55

this is a system that we've created

8:57

to help you benchmark and map opportunities

9:00

in your organizations where agents could specifically

9:02

help you solve your problems, create new

9:04

opportunities in a way that again is

9:07

completely customized to you. When you do

9:09

one of these audits, what you're going

9:11

to do is a voice-based agent interview

9:14

where we work with some number of

9:16

your leadership and employees to map what's

9:18

going on inside the organization and to

9:20

figure out where you are in your

9:23

agent journey. That's going to produce an

9:25

agent readiness score that comes with a

9:27

deep set of explanations, strength, weaknesses, key

9:30

findings, and of course, a set of

9:32

very specific recommendations that then we have

9:34

the ability to help you go find

9:37

the right partners to actually fulfill. So

9:39

if you were looking for a way

9:41

to jumpstart your agent strategy, send us

9:44

an email at agent at b super

9:46

dot AI, and let's get you plugged

9:48

into the agentic era. So I think

9:50

there is a time. that's interesting in

9:53

here and I'm going to take it

9:55

in honestly no particular order just kind

9:57

of my set of thoughts as I

10:00

reflect on this piece. The first is

10:02

that I think it's actually a very

10:04

salient point that helps us understand how

10:07

AI is improving things right now to

10:09

recognize that corporations are groups of specialists

10:11

whereas an individual inherently can only be

10:14

a specialist in so many things. And

10:16

so given that an individual is a

10:18

specialist in very few things, and a

10:20

novice in the vast majority of things,

10:23

AI's capability to take them from novice

10:25

to intermediate or novice to adequate in

10:27

terms of design, coding, whatever it is,

10:30

is a greater change in general than

10:32

a person who's just operating in their

10:34

specialist capacity inside the context of the

10:37

corporation they're contributing to. Basically the band

10:39

of use cases that are valuable to

10:41

an individual, operating as an individual, are

10:44

inherently higher than the band of use

10:46

cases useful to a specialist operating as

10:48

a specialist. I think it's a really

10:51

salient point, and I also think that

10:53

it will influence how these things take

10:55

root even inside enterprises and corporations. One

10:57

of the interesting and non-obvious outcomes, although

11:00

it feels like it's getting perhaps more

11:02

obvious now, is a stronger differentiation within

11:04

side companies between core functions and secondary

11:07

functions. And when it comes to those

11:09

secondary functions, more possibility that the capabilities

11:11

of a novice who's a specialist at

11:14

the main thing using AI tools to

11:16

do those things might be good enough.

11:18

In other words companies might decide not

11:21

to hire certain skill sets and expertise

11:23

that they would have before because it's

11:25

not core to the business even if

11:27

it's something that they do yes have

11:30

to do and the AI-fied version is

11:32

good enough. Now obviously if you're sitting

11:34

there screaming agents but agents... I think

11:37

you're correct to recognize that this seems

11:39

especially where we might see the rise

11:41

of people opting to hire digital employees

11:44

instead of human employees. One of the

11:46

ways to embody and embrace that good

11:48

enough is to pick the low-cost agentic

11:51

option that yes, maybe still can't compete

11:53

with best-in-class human versions, but for a

11:55

company for whom that thing is important

11:57

but not mission critical, that's totally sufficient.

12:00

One of the themes that you'll hear

12:02

a lot next week in our interview

12:04

conversations is the idea that a lot

12:07

of where agents are going to start

12:09

is going to be on the margins.

12:11

In sales organizations and marketing groups, not

12:14

that those things aren't important, but they

12:16

are separate from the core function of

12:18

whatever the business the business the business

12:21

is. A second thing, though, that I

12:23

think is important to note is while

12:25

I agree in general that if you

12:27

just take an individual operating in their

12:30

own capacity and an individual operating in

12:32

their corporate capacity as a specialist, AI

12:34

and LLLM specifically have had or have

12:37

the potential to have more broader impact

12:39

on that individual operating in their own

12:41

capacity, I think that Andre's argument perhaps

12:44

might mislead people in terms of understanding

12:46

how big an impact and how fast

12:48

these technologies are already having in the

12:51

enterprise sector. If you look at the

12:53

previous rate of change and rate of

12:55

adoption of technologies in the corporate sector

12:58

and compare it to AI, there is

13:00

absolutely nothing comparable. The speed with which

13:02

enterprises and businesses have radically reoriented themselves

13:04

to at least attempt to adopt these

13:07

technologies is totally unlike anything we've ever

13:09

seen. There is an understanding up and

13:11

down the organization that these are hugely

13:14

disruptive forces. They go beyond just new

13:16

tooling. They are unlocking totally new types

13:18

of efficiencies as well as totally new

13:21

types of opportunities, which we have barely

13:23

scratched the surface of, and relative to

13:25

the pressure of general corporate inertia, the

13:28

adoption is actually happening incredibly quickly. Now

13:30

like I said, that doesn't undermine his

13:32

point that there's something really powerful about

13:34

the fact that people are figuring this

13:37

stuff out on their own faster. In

13:39

fact, a lot of a natural evolution

13:41

and progression we're seeing is people using

13:44

their personal Gmail accounts to figure things

13:46

out and then slowly bringing in what

13:48

they've learned into the office. One of

13:51

the things that really smart organization. are

13:53

doing, that some that are perhaps lagging

13:55

behind aren't, is actively embracing that external

13:58

experimentation to internal new process adoption funnel.

14:00

Organizations and enterprises are constrained on how

14:02

free-wheeling and experimental they can be. However,

14:04

I would say that in general they

14:07

overestimate risk, and they undervalue finding ways

14:09

for people to experiment, even if they

14:11

can't do it with corporate data. The

14:14

last thing that I will say is

14:16

that I think that one other outcome

14:18

of this progression of technology diffusion that

14:21

Andre is noticing. is that we are

14:23

likely to see a radical increase in

14:25

bottoms-up entrepreneurship. The amount of activities that

14:28

seemed inaccessible to people before, or capital

14:30

constraining because they would have had to

14:32

hire someone to do it, and which

14:35

stood in the way of them just

14:37

doing the damn thing that they've wanted

14:39

to do, has decreased radically. And that

14:41

means the number of people who are

14:44

going to just throw off the shackles

14:46

of their normal enterprise and try to

14:48

do the thing that they've always dreamed

14:51

of, is going to increase. I think

14:53

we are going to see an absolute

14:55

Cambrian explosion of small business entrepreneurs, solo

14:58

performers, people building stuff because that becomes

15:00

easier. And that's one of the impacts

15:02

that I'm most excited to see play

15:05

out. Anyways, big thanks to Andre for

15:07

another thought-provoking piece. And thanks to you

15:09

guys as always for listening or watching.

15:11

Until next time, peace.