The Man Who Calls BS On AI: They’re LYING About AI, 2027 Is When It All Breaks! | Ed Zitron

0:00I think generative AI is at its heart

0:02con and seeing these ultra rich ultra

0:05powerful people lie through their teeth

0:07turns my stomach. The word con is a

0:09strong word.

0:10>> Well, what do you call something where

0:11from the very beginning they've sold

0:13[music] it in the terms of magic but

0:14it's just a halfass arcery machine. They

0:16are misleading the entire world.

0:18>> You are the first person that I've

0:19spoken to that has that opinion.

0:20>> Well, the fact that this is happening is

0:22insane and the fact it's not a scandal

0:24is insane. And I've been in the tech

0:26industry for 16 years now and I love

0:28technology and I'm enthusiastic about

0:29it, but I don't like being misled. And

0:32this is the largest non-consensual push

0:35of technology in history.

0:36>> So, we're going to play a game, Ed. I

0:38have the things that you consider to be

0:40myths about the AI industry.

0:42>> Let's play it. The AI industry is

0:44creating enormous economic growth. No,

0:46it's not. All of these companies run at

0:47a horrifying loss. Open AI lost $20.9

0:50billion last year. None of these people

0:52can just say, "Yeah, we're on the path

0:53to making this profitable." because they

0:55can't.

0:55>> Next one.

0:56>> AI will replace all human jobs. That

0:58just isn't happening and there's no

1:00economic data to support it. Next, the

1:02United States need to spend trillions to

1:04beat China in the AI race. What's the

1:06race to do for us to constantly piss our

1:07pants worrying about China? But people

1:09keep saying, "What if these models fall

1:11into the wrong hands? They're already in

1:12the wrong hands." Mark Zuckerberg, Sam

1:14Olman, Dario Amade.

1:16>> Mark Zuckerberg says, "We'll continue to

1:18invest aggressively in infrastructure to

1:20meet the demand." God met as a

1:21monstrosity. Makes me think of Shrek

1:23with L fogquad. Some of you may die, but

1:26that's a risk I'm willing to accept. If

1:27only these people gave a about

1:29poverty or actual problems in the world

1:31versus are we buying enough GPUs. If

1:34this continues, [music] what does the

1:35future look like?

1:41This is super interesting to me. My team

1:43given me this report to show me how many

1:44of you that watch this show subscribe.

1:46And some of you have told us according

1:47to this that you are unsubscribed from

1:49the channel randomly. So, favor to ask

1:51all of you. Please could you check right

1:53now if you've hit the subscribe button

1:54if you are a regular viewer of the show

1:56and you like what we do here. We're

1:57approaching quite a significant landmark

1:59on this show in terms of a subscriber

2:01number. So, if there was one simple free

2:03thing that you could do to help us, my

2:05team, everyone here to keep this show

2:07free, to keep it improving year over

2:09year and week over week, it is just to

2:11hit that subscribe button and to double

2:12check if you've hit it. Only thing I'll

2:13ever ask of you, do we have a deal? If

2:16you do it, I'll tell you what I'll do.

2:17I'll make sure every single week, every

2:20single month, we fight harder and harder

2:21and harder and harder to bring you the

2:22guests and conversations that you want

2:23to hear. I've stayed true to that

2:25promise since the very beginning of the

2:26Dire of Sio and I will not let you down.

2:29Please help us. Really appreciate it.

2:31Let's get on with the show.

2:33[music]

2:36>> Ed Zitron,

2:38there are a number of things that you

2:39believe that a lot of other people don't

2:41believe, right? You have, I think, a

2:44couple of controversial opinions and

2:45opinions that are in contrast to the

2:48other guests that I've sat here with.

2:50What exactly are those opinions, Ed? I

2:53think generative AI is at its heart con.

2:57I don't think it is sold as honest

2:59software. I think that they overstate

3:02both what it can do, what it will do,

3:04and the underlying financials to the

3:06point that they are misleading the

3:07entire world. And they're actively

3:09exploiting the weaknesses in journalism,

3:11in our economies, and indeed within the

3:14responsible parties with sellside

3:15analysts, governments, and all over the

3:17shop.

3:17>> The word con is a strong word.

3:19>> Yeah. I mean, what do you call something

3:21where from the very beginning they've

3:23sold it in the terms of magic as this

3:25thing that will replace all jobs, that

3:27will cure cancer, and all of these

3:29things? And when you look at it, it's

3:31boring cloud software that's extremely

3:33expensive and unprofitable and also

3:35unreliable at its core.

3:37>> People will be asking where are you

3:38drawing from in terms of your

3:40references, your personal experiences?

3:41Where were you educate? What you study?

3:43What you write about? What do you do Ed?

3:44>> So that's the funny thing is people say

3:46he's not got a finance experience. He's

3:48not going to take. I've been in the tech

3:49industry 15 16 years now in PR but still

3:52had practical experience and I love

3:54technology and I'm enthusiastic about

3:56it. And this thing just comes along that

3:58everyone is telling me is the best thing

3:59since sliced bread. And it can't even do

4:01the basics. It can't even do search.

4:03Well, whenever you ask an AI person,

4:05well, what's your setup? They describe

4:07this PeeWee's Playhouse thing of like,

4:09well, you got to harness here and you

4:10got to use the right prompt. Well, you

4:11don't want to use that prompt. You want

4:13to use this prompt here with this model,

4:14but don't use this model for the

4:16beginning, but at the end, you're going

4:17to want to use this model. And this is

4:19meant to be artificial intelligence.

4:22It's meant to be smart. It's meant to be

4:24autonomous. It's meant to be something

4:25that you set and forget.

4:26>> We have the sort of six leading AI

4:28companies on the table here. Anthropic

4:30Amazon, Nvidia, Microsoft, OpenAI,

4:32Google. You're saying that their

4:34fundamental business model is a con.

4:37>> Well, their revenues are not really

4:39coming from AI. Up until fairly

4:41recently, none of their revenues were

4:43coming from AI. Like dribbles a bit.

4:44Right now, 70% of all AI revenues across

4:48those three companies are from OpenAI

4:49and Anthropic to unprofitable,

4:51unsustainable companies that literally

4:53cannot afford to exist without these

4:55very same companies giving them money.

4:57Amazon sent $50 billion to OpenAI this

5:01year. They sent $5 billion to Anthropic.

5:03Google sent $10 billion to Anthropic.

5:06And in the next three and a half years,

5:08OpenAI and Anthropic based on actual

5:10sellside analyst evaluations, their

5:12estimates that inform whether stock is

5:14going to go up or down after earnings,

5:16they are expecting 400 or more billion

5:19dollar of revenue, 30 or something% of

5:22cloud growth just from these two

5:24unprofitable companies that will need to

5:26be given the money from somewhere. And

5:28on top of that, these companies have

5:30such low respect for the average

5:33investor, for the analyst, for everyone

5:35really that they don't even disclose

5:36their AI revenues. The few times they

5:38dain us worthy, they use something

5:40called a run rate, an annualized run

5:42rate, which means well, nothing. They

5:44never define it. It can mean months 12.

5:47It can mean month 13. It can mean last 4

5:49weeks time 13. It's different every

5:51time, and they never define it. And then

5:53they sometimes just don't mention it.

5:54So, you've got this big thing that is

5:56meant to be the biggest, most

5:58influential change to software ever. And

6:00whenever you ask them about it, when you

6:02say, "What? How much you making from

6:03this?" They go, "Oh, I couldn't possibly

6:04say. I'm too shy." These are public

6:06companies, or at least the ones that

6:08aren't anthropic and open AI. When they

6:10have good news, they'll tell you. And

6:11when they don't tell you something,

6:13well, that actually speaks volumes.

6:15>> Have you you used these tools, the AI

6:17tools, Gemini, Anthropic, Chat, GBT,

6:20etc., and you found no value in them?

6:22There's some value, but it's not there's

6:25they have spent over a trillion dollars

6:26in capex. What

6:27>> does capex mean for you?

6:28>> Capital expenditures. So, when you are a

6:30business and you have operating expenses

6:32like electricity, for example, those

6:34come right off immediately. Capital

6:36expenditures are long-term investments

6:37that are theoretically one-off. So, a

6:39data center or indeed the GPUs you put

6:41inside an AI data center.

6:43>> Okay? So, you've got a data center

6:45>> and then you have these GPUs which are

6:47like computer chips. So AI GPUs are much

6:51bigger, much more power intensive. They

6:52take a bunch of high bandwidth memory

6:54and they because of how many of them you

6:57need. You need thousands of them, tens

6:58of thousands, hundreds of thousands in

7:00some case. You need a bunch of power. So

7:03an example, OpenAI and Oracle are

7:05building a data center in Texas in

7:06Abalene, Texas. 1.2 GW called Stargate

7:10Abene. Within that, with each one of the

7:12eight buildings, there'll be 50,000

7:15Nvidia GB200 GPUs. So, city of Bristol

7:19takes about 7800 megawatt of power a

7:22year, right? Well, Stargate Abene is

7:26condensing more power than that, 1.2

7:28gawatt into a space around 1,172

7:31times smaller. City of Bristol is about

7:331.2 billion square ft. Star Evelyn is

7:36about 998,000.

7:38So, you're condensing all of this power,

7:40all of this money, all of this labor

7:41into this one spot. And all of these

7:44data centers cost billions of dollars.

7:46All of these companies other than

7:47Microsoft are now to take out debt. And

7:50the thing is they've spent over a

7:51trillion dollars so far and they want to

7:52spend another trillion dollars next

7:54year. And for what? To make tens of

7:56billions of dollars, most of which comes

7:58from two unprofitable companies,

8:00Anthropic and Open AI. One of the

8:02rebuttals to that would be that the

8:04adoption, the customer adoption of

8:07people using Open AI and Enthropic has

8:09been absolutely insane. These are the

8:11fastest growing products in all of

8:12history, especially as it relates to

8:14sort of technology. If we just focus in

8:16on technology, they are, you know,

8:18hundreds and hundreds of millions of

8:19people, billions of people are using

8:21these tools every single day for things

8:23that they have subjectively decided are

8:26problems they need solving. So, you

8:28know, money is a lagging indicator of

8:30value. So, one would argue that they're

8:33just investing ahead of the monetization

8:36options.

8:36>> The first let's start with this

8:38adoption. Is it honest adoption when you

8:41are forced to use generative AI when you

8:43load Google? When you load Google Docs,

8:45Gemini screams in your ear. When you

8:47load Word, co-pilot's bugging you. When

8:50you use Amazon, whatever rofus AI is

8:52wants has opinions on what socks you're

8:54buying. This is the largest

8:57non-consensual push of technology in

8:59history. Chat GPD for example, every

9:01single media outlet has been screaming

9:03about this for 3 years. They've been

9:05saying, "This will take your job. You

9:07must use this. If you don't use this,

9:09you're going to be falling behind." So

9:11people are using it because they've been

9:13told to use it constantly and they're

9:14using it like search predominantly and

9:16that's partly because Google fell behind

9:18search and also because it's better at

9:20ingesting queries sometimes. Sometimes

9:21if you use a generative search it's like

9:23a trolling vessel. It's not very good at

9:25specifics but if you're like does this

9:27thing exist? Has this person ever said

9:29anything like this? It'll still probably

9:30get it wrong but it'll scour the ocean

9:32for you. Nevertheless, that's not worth

9:34a trillion dollars. None of it is. The

9:37amount of money being sunk into this is

9:40just incomparable to anything. Railways,

9:42it blows everything out of the water

9:44because there is no postbubble story

9:48even for this. AIG GPU is not useful for

9:50other things either. There's it's a

9:53directionless egregor of capitalism.

9:56this headless beast that lumbers around

9:59desperate to seek out growth everywhere

10:01in the hopes that if it harasses people

10:03and scares people and demonizes labor

10:07enough, people will be forced to use it.

10:10>> The the reason I I pause is because I

10:13just I think about my own company.

10:14Obviously, everybody thinks about their

10:15own personal situation. So, you have

10:16people listening now that don't use any

10:17AI tools. Then you'll have people that

10:20are using it for everything from coding

10:22new software tools to everything they

10:24write to, you know, images, whatever.

10:26And when you look at the the stats

10:27around enterprise adoption, it says 88%

10:29of organizations regularly use AI at

10:31least once for one particular business

10:34function. And I'd say in our company,

10:3695% of people use a one of these AI

10:40tools like anthropical chatbt or Gemini

10:42every day,

10:43>> right? And that exists on some kind of

10:44spectrum of like the super users that

10:47are using it probably, you know, every

10:49hour of every day for almost everything

10:51to, you know, someone maybe hiring the

10:54executive team that's using it less

10:55because their job doesn't require of it

10:57as much,

10:57>> right?

10:58>> And when you look out into the world,

11:00you know, at how the world is changing

11:03from a content perspective, if we're

11:04looking at generative AI, it is obvious

11:06that these tools are being widely

11:09adopted. Part of the symptom is the AI

11:10slop you see all over the internet,

11:12>> right?

11:13So, I I don't know this this this idea

11:16that it's not being used. I struggle

11:19with

11:19>> it's being used. Here's the thing with

11:21the slop. Before we had AI slop, we had

11:24SEO slop because Google incentivized

11:27doing the lowest common denominator that

11:28would rank well in search. There's a

11:30whole story about how they pulled back

11:32spam guards thanks to Bravagar Ragavan,

11:33which we can get into,

11:35>> where they made the internet worse by

11:37allowing worse content to rank higher.

11:39It's why we have when you used to

11:41Google, oh, best washing machine,

11:43there's 11 different horrible blogs that

11:45read like somebody got a concussion.

11:47They are built to rank rather than be

11:49read by humans or built to be good made

11:52good. So AI helps weaponize that at

11:54scale. Yeah, you can make a bunch of

11:56generic slop. We've had slop for years.

11:59We've just found a slop machine. But

12:01then also there's the problem of cost.

12:03So when you use AI services, you burn

12:05tokens and it's per million tokens. So

12:07>> what's a token? So it's around 3/4 of a

12:10word. So it's characters.

12:12>> So the AI companies have a currency in

12:14which they charge you. Like a taxi in

12:16New York has a meter.

12:17>> Yeah.

12:18>> And they call it tokens.

12:20>> Yeah.

12:20>> And every word, let's just say for ease

12:22it's a word. You're paying per word.

12:24>> About a word. Yeah. And it's per million

12:26tokens. So you'll be charged per million

12:28input tokens. The stuff you feed into it

12:30like a document or a bunch a code base.

12:32And the output tokens are both the stuff

12:34it spits out at the end but also when it

12:36thinks. So, okay, you've asked me to

12:38give you the best restaurants in this

12:40area of New York. I should find the best

12:41restaurants in New York. All of that's

12:43output tokens as well.

12:44>> However, when you're paying for a

12:46monthly service, you don't see any of

12:48that. Put all that crap to the side.

12:50They just have rate limits. So, you can

12:51use them a certain amount and then when

12:53you run out, but they kind of offiscate

12:54what that was. Now, someone recently

12:57found, semi analysis actually found

12:58this, a big analyst group. They found

13:00that on a $200 a month chat GPD

13:03subscription, you can burn $14,000

13:06worth of tokens and on anthropics you

13:09can burn $8,000 for 200 bucks. That is

13:13how most and even on the 20 buck a month

13:15service you can burn $400.

13:18Now most people don't realize that. Most

13:20people have no idea what AI costs. Most

13:22people just think, "Oh, it's 20 bucks a

13:24month." No. All of these companies run

13:26at a horrifying loss. OpenAI lost $20.9

13:29billion last year because people can

13:31burn as many tokens as they want. And

13:33when they tried to move everybody on the

13:36enterprise side, so companies bigger

13:37than 150 onto actually paying the cost

13:40of AI in around March of 2026, to quote

13:43Sam Orman, they said, uh, people have a

13:45big problem with it. I think it's a huge

13:47issue, which is not really what the air

13:50apparent text history is meant to be

13:52saying, but the point is enterprises

13:54immediately started freaking out. Uber

13:56burned through their entire annual token

13:58budget in three months. So suddenly

14:01after everyone saying AI is the most

14:02productive thing ever. It's amazing.

14:04It's changing everything. The moment

14:05people actually had to pay for it, they

14:06go, [snorts]

14:08I don't know actually. Um maybe it's

14:11obviously we all love it. It's all

14:13great, right? But it's costing too much.

14:15So we need to reduce the cost because

14:17people are just dumping stuff into it

14:19being like what do I do here and getting

14:21whatever the median is out because

14:23that's what these things do. they

14:24provide the median answer.

14:26>> So essentially, someone like me who's a

14:28power user of these tools,

14:30>> I could be costing Anthropic or OpenAI

14:34$1,000, but they're only charging me

14:36$100, let's say. So they are having to

14:38subsidize $900 of my usage because of

14:41the electricity costs and the costs at

14:43their data centers. And so your

14:45assertion here is that that is

14:46unsustainable.

14:47>> Yes. And just to be clear, they're

14:49probably not one for$1. It might be 30

14:51for. We don't we don't know. I think

14:53it's unprofitable. These companies don't

14:55disclose them even in their auditive

14:56financials. They play funny games with

14:57how they categorize things. But

14:59nevertheless, yes. And on top of that,

15:01the way that you stand up inference,

15:03which is the thing that creates the

15:05output within these data centers, you're

15:08not just saying, "Okay, turn the

15:09inference machine on. Let's go." You are

15:12standing up the GPUs necessary to take

15:14in the demand, and if you buy too much,

15:17you've wasted the money. You You have to

15:19pay for the hourly GPU use regardless.

15:22If you buy too few, your customers can't

15:23use it. They get pissed off at you. They

15:25cancel. They go with someone else. But

15:26nevertheless, yeah, they would get

15:27demand selling $20 or $40 for a dollar.

15:31And that's what these services do. And

15:33really, the simplest way to explain it

15:34is they were actually profitable if they

15:36were actually just they believed that

15:38these services were worthwhile and that

15:39they were worthy of the cost, they'd

15:41charge it. Regular people wouldn't be

15:43able to get a monthly subscription.

15:45They'd just be paying what it's worth,

15:47unless, of course, there was an economic

15:50problem. And it's very simple. You pay

15:53when you use an LLM regardless of

15:55whether you get what you want. When

15:56these things hallucinate, say you're

15:58doing something, you're coding something

15:59and they go through a code base and they

16:02up a bunch of stuff, they break a

16:03bunch of stuff, you're paying for that.

16:04You're paying for it whether it works or

16:06not, unless of course you're using one

16:07of these subscriptions. I think the the

16:09really interesting point is are they

16:13spending ahead of the value showing up

16:16which is I imagine what they would argue

16:19or are they spending all of this money

16:22and subsidizing all of their users in a

16:24way that's unsustainable and that will

16:26never be justified like does it you know

16:28because you think back through the

16:29history of technology you often get

16:31people

16:32losing money to grab market share

16:34>> right

16:35>> and they're also focusing on bringing

16:37the costs down and making it more

16:40profitable for them as well. But they

16:42can't afford to underinvest.

16:44>> If they were bringing the cost down,

16:46they would have brought the cost down,

16:47which they have not. It seems to be

16:49getting more expensive. In fact,

16:51everyone inference providers don't seem

16:52to be profitable. Even the companies

16:54renting out GPUs don't seem to be

16:56profitable. I imagine that it wasn't

16:59like they started out and they were

17:00like, "Shit, this is unprofitable at the

17:01beginning. We know it. Screw it. We'll

17:03keep doing it any screw." I don't think

17:05it's some big conspiracy. They probably

17:07thought at some point, yeah, this will

17:09go profitable. The chips will catch up.

17:12Customers will pay for the overwhelming

17:13value because you don't know in 2023

17:15where it's going to be in 2026. You

17:16assume it's going to go up. That's the

17:18nature of venture capital. They should

17:20have stopped in like 2024 when OpenAI

17:22lost over $5 billion. They should have

17:24been like, "Yep, this is not going to

17:26work." But they kept going because it

17:28helped number go up so much. It helped

17:30stock values pump. It helped everyone

17:32pump. It helped Nvidia pump, Microsoft,

17:34everyone. and not from the revenues.

17:37Because here's the funny thing about

17:39Google, Microsoft, and Amazon. People

17:41for years have been saying their AI bets

17:43have paid off. Wow, their AI bets have

17:45paid off. As these companies refused to

17:47say how much they're making from AI, but

17:49because their existing businesses

17:50continued to grow and did so, by the

17:52way, through price increases, changes to

17:55how Google and Meta uh did advertising.

17:58Amazon bumped up prices and changed how

18:00they did actually Amazon started a

18:01remarkable ad business during this whole

18:03time as well. and the selling through

18:05Amazon platform anyway nothing to do

18:06with AI but because number go up because

18:08revenue go up everyone went it's AI

18:11because these companies wouldn't spend a

18:12trillion dollars for for no reason right

18:15except in fiscal year 2026 which just

18:18ended for Microsoft annoying I know they

18:21made total according to Bloomberg about

18:23$34.33 billion $24.1 billion of that was

18:27from OpenAI so that leaves them with

18:29about $10 billion in a year when they

18:31spent 115 billion on capital

18:33expenditures just intend to spend 175

18:36billion next year. The math does not

18:39make sense. I imagine their plan was

18:41okay, this is just going to get

18:42exponentially more valuable and at some

18:44point the costs will be outpaced by the

18:46return. Problem is that large language

18:48models need a bunch of money to train

18:50them. They need constant data flow. They

18:52need customized data. It's just this big

18:54expensive monster. And when you try and

18:58talk to people about it and you try and

19:00say, "Hey, look, this is really bad.

19:02Nvidia has sold it was $215.9 billion in

19:07the last fiscal year worth of GPUs

19:08mostly. And you try and go, yeah, that's

19:11to support like $22 billion of revenue

19:15total in the entire world outside of

19:18these two companies that literally

19:19require money being fed into them

19:21sometimes by Nvidia to keep alive. When

19:24you tell people that, they go, "Well,

19:25companies just lose money, right?

19:26Companies because we have this quote

19:29Edson from Prophy Markets. We have this

19:31cult-like worship of the wealthy where

19:33we think that someone wouldn't spend all

19:34this money for no reason. Right? Because

19:36reconciling with that with this idea

19:39that the ultra wealthy, the ultra

19:42powerful didn't get there through big

19:44brains. They didn't get there through

19:47anything other than luck and opportunism

19:49and getting an MBA perhaps with the

19:51right people. That they just got there

19:53because they're regular people and they

19:55just happen to be in the right place at

19:56the right time. reconciling with that

19:57and realizing that the world is not

19:59controlled by people like a meritocracy

20:01is kind of grim. So it's easy to be like

20:03no they're not making a mistake I must

20:05be missing something and that's what

20:07they want. So you know I think back

20:09through the history of technological

20:11breakthroughs and I think about I mean

20:13you can look at different industries and

20:14one of my favorite books on this subject

20:15is the innovator's dilemma. not read it.

20:18>> And one of the things it talks about is

20:19how the the innovation that ends up

20:21taking out or transforming an industry

20:25often starts worse, doesn't make

20:28economic sense, none of your customers

20:30are asking for it. And this is typically

20:32why we end up ignoring it. So like

20:33you've got horse and carriages in the

20:351800s.

20:36>> Amazing form of transport according to

20:37the 1800s, you know, people of the

20:391800s. And then you have this thing

20:40called cars come along. Now the problem

20:42with cars is they broke down all the

20:43time. It's kind of like AI hallucinates

20:45now. um they were more expensive and the

20:47the economics of it didn't make sense.

20:49You might as well walk than buy a car.

20:50There was a law at the time that meant

20:52you had to walk in front of it with a

20:53red flag and wave and someone had you

20:55had to employ someone to walk in front

20:56of it waving a red flag. Obviously, it's

20:58worse. It's like a worse solution.

21:00However, these things that are

21:02disruptive innovations, they have a

21:04higher ceiling of growth and so they

21:07eventually overtake the horse. And I

21:09when I think about that analogy in the

21:11context of all of this, I go, okay, it's

21:13imperfect at the at the moment. the

21:15economic models aren't perfectly ironed

21:18out. They're still figuring out how to

21:19make it cheaper, the infrastructure,

21:21etc. But as if you think about the rate

21:23of improvement versus other you know

21:26let's say coding how much could I train

21:28a human coder to improve and to increase

21:31their output versus an AI agent one

21:33would go if you just imagine any rate of

21:35improvement in these AI tools at some

21:38point if you just imagine a 5% rate of

21:39improvement per month at some point it's

21:43you know and then you imagine a 5%

21:45reduction in cost which is what we did

21:46with the internet what we did with cars

21:48but Mo's law

21:49>> mos law is a mos law is not with GPUs.

21:52So let me let me actually explain. So

21:54Nvidia Nvidia invented I think it was in

21:56the 2000s they put out something called

21:58CUDA which is the underlying software

22:00library and the way to run software on

22:03GPUs. took them solid decade or more to

22:06make it something where they could do

22:07data analytics, one of the early things,

22:09mapper and such. And then when AI came

22:11along, they'd had lots of experience

22:13with it. But nevertheless, this company

22:15has got more money, more attention, more

22:18geniuses behind them, more people

22:20focused on making their things more

22:22efficient than anyone could ever ask

22:24for.

22:24>> And Nvidia, for anyone that doesn't

22:26know, makes the chips.

22:27>> They So, and that CUDA thing I

22:29mentioned, they were the ones with CUDA

22:30and CUDA allowed generative AI to grow.

22:32Okay, so they're chips.

22:34>> Chips and chips are needed. Those are

22:35the things that go into the data

22:36centers.

22:37>> And there specific chips are the ones

22:38where you can run AI software on it. So

22:40the training runs and also the

22:41inference. Now, here's the thing. The

22:44the car example back then you didn't

22:47have pretty much every mathematician and

22:49scientist going into the car industry.

22:51You didn't have the combined world's

22:52governments never shutting up about

22:54this. And by the way, giving them credit

22:57early since 2023, they've been saying

23:00this is inevitable. Even in what you

23:01said, 5% improvement. I don't even know

23:03how you'd measure that because a junior

23:05software engineer can still experience

23:08things and learn things from context,

23:10from how people deal with problems. And

23:12the way that people deal with problems

23:13is not as simple as looking at the code

23:15or reading some emails. It's context

23:17cues from speaking to a person. It's

23:19being in different environments. And

23:21there may there are uses for LLM's

23:23encoding. I don't dispute that. But even

23:26saying 5% uh what does that mean? Is it

23:28better at Rust? Is it better at C++?

23:30>> I'd say productivity just like yeah

23:32shipped. If we did it in the context of

23:34coding, it would be like shipped code.

23:36>> That's the thing that would be like he's

23:38the best writer in the world cuz his

23:39newsletter's really long. That's an

23:41insane way of evaluing it. With coding,

23:43it would be I mean it's even difficult

23:45to evaluate because it's is the software

23:48out there better is actually a great way

23:50of evaluating it. And I would say

23:51uniformly not. I would say the standard

23:54of software across Google, Microsoft,

23:56Amazon, Meta, especially God, Meta is a

23:59monstrosity, is worse. GitHub, GitHub,

24:02someone posted on Twitter earlier today,

24:04we should get a notification when GitHub

24:05is up rather than when it's down because

24:07that would be more reliable. Microsoft's

24:09one of the largest companies in the

24:10world, and they can barely wipe their

24:11own ass when it comes to GitHub. The

24:13quality of software is going down

24:15weirdly enough as more people use LLMs

24:18and more businesses demand and I really

24:20do mean demand that people use these

24:22services. So on this point of if we go

24:24back to this horse and carriage and car

24:26analogy say that we're at whatever point

24:29today if you imagine any rate of

24:31improvement in the technology which we

24:32have seen since tragedy came out

24:34>> I remember when tragy came out and I was

24:36in Asia and I was there showing it to my

24:37fiance I was like look it can do this

24:39and it was hallucinating once in a while

24:40and getting things wrong. I actually

24:42don't have that experience anymore. I

24:45have moments where I believe it's

24:47reasoning is weak, but I don't have

24:49outright hallucinations anymore. See

24:52that? I I disagree. So,

24:54>> give me an example of what you define as

24:55a hallucination.

24:56>> Okay, great one. So, I have a Bloomberg

24:57terminal. Yeah. The very useful thing

24:59they have on there is ask B. So, when

25:01you do a Bloomberg inquiry to like look

25:03up what we think Nvidia's revenue is

25:05going to be next quarter, it runs

25:07something called BQL, which is its own

25:08programming language. Now, instead of

25:10having to learn that, you can just type

25:13into RSB and it will generate it and run

25:14it for you. And so, you get it pulled up

25:16and you know where the data is coming

25:17from. It deals with hallucinations real

25:19well. The other day, I was like, you

25:20know what, get a little spicy. I'm going

25:22to look up the growth rate of stocks of

25:25Microsoft, Google, Meta, and Amazon over

25:28the course of 5 years, I think it was.

25:30>> And I was about to I was copy pasted it

25:32over to something looked at in Excel. I

25:34was about to was writing the newsletter.

25:35I went, Microsoft stocks never been $575

25:39a stock.

25:41You know what? When it's a cute little

25:42thing like, oh, it's a stock price and I

25:44kind of call it was no harm, no foul.

25:46That's fine. But when you're talking

25:48about, I don't know, like a transcribing

25:50tool for a doctor or a financial model

25:53that a hedge fund is dependent on, at

25:56that point it becomes a little more

25:57dangerous. And the thing is a

26:00hallucination with a software package.

26:03For example, you're refactoring a code

26:04base and it leaves a door open

26:06security-wise or it just breaks

26:08something and you I don't know maybe

26:10you've been vibe coding for 6 months.

26:12You haven't really been coding with your

26:13own hands for a while. Maybe you've

26:14forgotten a few things. You had this

26:16slop to look for. I'm not doing

26:18it. And so the problems become

26:21multiplicative. And I don't really know

26:23how you train them out of that. And

26:25they've certainly not succeeded. So on

26:28one hand they have got better but one of

26:31the main ways they evaluate them getting

26:32better are benchmarks that are adjusted

26:35specifically for large language models

26:37because you can't just have them do

26:39tasks. They've got better at that. They

26:41found some tasks they can have them do

26:42on them like meter me they have this

26:45thing where it's like check out this

26:46chart look how much better it's getting

26:48at running tasks. Wow it can go for an

26:50hour and then you look it's like yeah

26:51and successfully completing them 50% of

26:53the time. They they have a hallucination

26:55leaderboard and it really focuses on

26:57basic tasks and it shows that the

26:59four-year trend according to historical

27:00data from the Victaria hallucination

27:03leaderboard shows that hallucination

27:05rates on simple summarization tasks have

27:07plummeted from around 21% 21.8% 4 years

27:11ago down to 0.7%

27:14roughly on today's top frontier models

27:16like Gemini and Chat GPT. Again the

27:19point of nuance here is that these are

27:21on simple tasks which is kind of what

27:23I've experienced. I've experienced that

27:24on day-to-day things that hallucinates

27:26less again rate of improvement thinking.

27:28So if I just imagine the trajectory to

27:30continue there is going to become a time

27:33where hallucinations become rarer than

27:35they are today increasingly and also

27:38what I would say is when I think about

27:39other technologies there's two more

27:40points other technologies at their

27:42inception when they first came to the

27:43world like the internet also had

27:45technical difficulties. I remember

27:47growing up with dialup modems and I

27:49couldn't go on the phone at the same

27:51time as going on the internet. I'd have

27:52to stop Runescape upstairs to go on the

27:54phone. And you thought this is crap.

27:56This is technology crap. All the

27:57>> I I don't know, mate. I loved it.

27:59>> Yeah, I know. You It felt like magic.

28:01And then in hindsight, you go, "Wow, I

28:03now have Starink and 5G internet from my

28:05phone. It's unbelievable." You couldn't

28:07leave the house with internet before.

28:09And that's what I mean by the rate of

28:10improvement thinking. I'd say the last

28:12point is we often compare AI to

28:16perfection,

28:17>> right?

28:18>> Whereas that's not actually the

28:19alternative in the working world. Like

28:22if I wanted to do let's say a simple

28:24writing task, I should compare AI to my

28:27alternative alternative way of doing

28:29that simple writing task which is both

28:31measured in my time right and my ability

28:34to hallucinate as a person who doesn't

28:35know everything

28:37or if I'm hiring someone an intern who

28:40might also be prone to hallucination or

28:42have gaps in their knowledge.

28:44>> So it's not actually like we're

28:45comparing we should compare AI to

28:46perfection. It's AI to the other

28:48alternatives. And if someone

28:49hallucinates 0.7% of the time, but knows

28:52way more and is faster, maybe on a net

28:56basis, that's a good trade. Maybe I

28:58should use AI. So, let's start with an

29:00example. Someone I love dearly, Matt

29:02Hughes, my editor, lives out of

29:04Liverpool. Wonderful guy. I don't pay

29:06Matt Hughes because he knows everything.

29:09I pay him because he has incredible

29:11context and a ton of knowledge and he's

29:13willing to expand it and work with me

29:15and moral sport and he's a great editor,

29:18but he's also someone who gets into the

29:20guts of it and has the experiences of

29:21it. He's a decorated tech journalist and

29:24on top of that a wonderful loving being

29:26with empathy and joy in his heart for

29:29the stuff he loves and absolute

29:30venom for the people he hates. That's I

29:33can't get that from a large language

29:34model. But on top of that, I don't I

29:36push back on just the assumption there.

29:38>> When you say knows everything, what good

29:40is something that knows everything when

29:42it sometimes doesn't know anything when

29:43it's sometimes? And on the thing is, are

29:45you really paying an intern for

29:47something basic? Are you really going to

29:49them and saying, "Yeah, can you look up

29:51what the date is?" No, you're doing that

29:52on Google. Whatever the task is, you are

29:55trying to also train an intern. The

29:57point of an intern is to train them and

29:59turn them in, take them out of Pinocchio

30:01status,

30:02>> but it's also an intern learns. And in

30:04turn gets context and in turn learns

30:05your habits. Learns

30:06>> AI gets context and learns.

30:08>> No, it doesn't. It

30:09>> doesn't learn.

30:09>> I mean, it doesn't. The way it learns is

30:12you create a giant claw. MD file that it

30:14sometimes doesn't read, sometimes does

30:16read. You create a harness. You put it's

30:18like it's Pee-Wee's breakfast machine

30:20from PeeWee's Playhouse. You have to do

30:22all these controversies to mitigate the

30:24hallucinations. And even then at the

30:26end, how much effort have you put in?

30:28>> But so, okay, this is an extreme

30:29simplified example. If I went on my

30:31Claude now and said, "What's my dog? my

30:32dog's name.

30:33>> Uhhuh.

30:33>> It would know my dog's name.

30:35>> Jesus Christ. This this company raised

30:3795 billion.

30:38>> I'm saying I'm I'm using an extreme

30:40simplified example to show that it can

30:41remember things from the past.

30:43Obviously, it knows much more complex

30:44things as well, but I just use that as

30:46an example. So, we we we accept the fact

30:48that it can it does have memory of the

30:50past.

30:51>> It has files it can access that have

30:52stuff on it, but that's not the same as

30:55memory. And it's also just okay. So, it

30:57remembers your dog's name. It might

30:59remember your habits. It might be able

31:01to read things you've said before.

31:03>> Does it know your moods? Does it know

31:05what's going on in the world around it?

31:06Does it have good days and bad days? Is

31:08it there for you? Because it's just a

31:10text machine. And the thing is

31:12the intern example. An intern is

31:15something that can grow. It's something

31:16that you invest in. That's not something

31:18you do through feeding files and text to

31:20it. The way that we store memories

31:22ourselves, the way in which we acrue

31:24experiences is a a milerum of emotion

31:29and feelings and facts

31:30>> completely different. So I think there's

31:32two things here. There's the process in

31:34which something happens and then there's

31:35the output.

31:37>> So the process you're describing the

31:38process of how a human does memory,

31:40>> right?

31:41>> The way that an AI does memory is

31:43different. But the thing that people

31:45care about is there value in the output.

31:47I.e. You know, if I dump all of my files

31:50into Claude, I don't really care how it

31:52processes it as long as when I ask it,

31:54what's my revenue? It has the number.

31:56And one could say the same thing about

31:58training someone. You could say, you

31:59teach them, you put lots of effort into

32:00them. You give them lots of context. You

32:03you educate them and give them

32:04experiences. And then you might come and

32:06say to them, by the way, what's my

32:07revenue? Now, the processes are entirely

32:09different, but the outcome is what I

32:10care about. Do they know the revenue

32:12number when I ask them? And so, I think

32:14that's the part that we sometimes get

32:15lost. we get, you know, cuz I have I've

32:16heard this debate about like can AI be

32:18creative,

32:19>> right?

32:19>> I think like the way to answer that

32:21question is like it's about the output

32:23when I ask it to do a creative thing

32:25does it give me the answer not is the

32:27process the same as a human process cuz

32:30actually no who cares what the people

32:32care about they pay for the outcome the

32:34product.

32:34>> I actually disagree about the process

32:37because Matt Hughes for example

32:39>> your editor

32:40>> Yeah.

32:40>> Yeah. watching him go down a rabbit hole

32:43and being there with him and actually

32:44vice versa him doing the same thing. We

32:46wrote these well I mean we were working

32:48on the research I ended up sitting there

32:50for like the dayong session of writing

32:5311,000 words and he he had given me a

32:55bunch of notes. It was actually just

32:56even describing that process, I feel so

32:59happy cuz it was like us being like I

33:00can't believe how these Jesus

33:02Christ they can't do like just like the

33:04misanthropy of just the horrible cynical

33:07people of asset managers like Blackstone

33:09just learning about them and being like

33:10it can't be this and having a back and

33:12forth with him that is fundamentally

33:14different because we were both learning

33:16together and the learning process was as

33:18much about creating the output as the

33:20output itself. When you learn something,

33:22you're not creating the average, which

33:23really is what these things do, of the

33:26documents it could find. You're not

33:28getting particularly novel outputs. If I

33:31needed a generic slop output, sure, but

33:34I've I've used some of the higherend LLM

33:37harness machines that the hedge funds

33:39use, and they all give the same shite.

33:41It's all the same the same generic

33:43reports, the same, oh, we noticed this

33:45analysis, things that you can find on

33:46any kind of AI slop out there. what you

33:48described to me there, what I heard

33:50anyway is there's two points of value

33:51you're getting from your time with that.

33:53I mean, I mean, there's many more, but

33:54you said you're you're learning and then

33:57you're getting this book edited blog

33:59blog. You're getting a blog edited,

34:00which is the output, and you're getting

34:02learning and you're also really getting

34:03connection and all these other things.

34:05But when I come to when people sort of

34:06think about the value of AI, of course,

34:08they could use it to learn. But in the

34:10example I gave of like repeat my revenue

34:11number back to me or do this number, I I

34:13just care about the output. I could use

34:15it to learn. I could say what if the

34:16revenue number was wrong once you should

34:18have defined deterministic ways of

34:21knowing those numbers you should not

34:23rely on them even with the terminal

34:25running BQL which I trust I will double

34:27triple treble check everything just to

34:30be sure partly because also the process

34:32of learning for me I don't want just a

34:34report I go like that I want something

34:36that I fully understand and also

34:38understand the context around it I don't

34:41think that LLM do that and I just don't

34:43see them getting

34:45in a way that does that because it's

34:48it's just not what they do. And also

34:50there's the other problem of the more

34:51detailed the report, the more likely

34:53there are things to be wrong with it. If

34:54you are with Matt Hughes, for example, I

34:57can trust he's got it right. I can trust

34:59he understood and I can trust that I can

35:01have a back and forth with him that will

35:02inform me if I've missed something. I

35:04can read the stuff that he's read and

35:07actually trust him because there's a big

35:09trust part as well. What is the basis of

35:11your trust in Matt? Could it be his

35:14historical performance?

35:16>> I mean, yes.

35:17>> Okay.

35:17>> And also the fact we've learned half of

35:19this stuff together,

35:20>> but but tenure tenure doesn't

35:22necessarily There's probably people, you

35:23know, for 15 years who you also don't

35:25trust. Yes.

35:25>> So, I think I was trying to figure out

35:26like what is the what is the thing

35:28that's causing humans to trust another

35:29thing. And I guess it would be continual

35:31delivery of a commitment made of sorts.

35:34And so with Claude for example on simple

35:37tasks as we've seen from this

35:38hallucination leaderboard it continually

35:41delivers for people and that's why we've

35:43seen the fast

35:44>> I mean is that what that board says

35:45>> well it's it's saying like is it getting

35:47it wrong is it hallucinating

35:49>> simple task how are those defined

35:51>> I I don't know

35:52>> that's the thing though because this is

35:53actually a very very illustrative thing

35:55of the AI industry they are the what

35:58aboutist masters they have like well

36:00look we got this we got this benchmark

36:02that says we're good at this and look

36:03the numbers higher What's the number

36:05mean? No. What does that mean? And I'm

36:08not using this as a critic against you.

36:09It's

36:10>> when you can't give a direct answer, you

36:12give a side answer. When you as the LLM

36:14industry want to prove your worth, you

36:17can't just be like just use the product.

36:18When the first iPhone came out, go was

36:20Penn State at the time. Oh, I felt like

36:23the uh apes at the beginning of 2001.

36:25official voicemail. It was

36:27immediate. And I showed it to tech

36:29friends. I showed it to the most normal

36:31people in the world. And everyone was

36:32like, "Holy this is They were on

36:34razors. They were on Nokia 3210s. It was

36:37obvious the value." Amazon Web Services,

36:38same deal.

36:39>> It wasn't obvious though.

36:40>> Yes, it was. I mean, I bought it

36:42>> to you. To you, it was.

36:43>> It was. And I also showed it to a bunch

36:45of people because I'm aware that I had

36:46bias when I just love gadgets.

36:48>> But but I remember the famous Steve

36:50Balmer who was the CEO of Microsoft

36:52interview where he was told about the

36:54iPhone and he bursts out laughing.

36:59[laughter]

37:00$500 fully subsidized with a plan. I

37:03said that is the most expensive phone in

37:06the world and it doesn't appeal to

37:07business customers because it doesn't

37:09have a keyboard which makes it not a

37:11very good email machine. You can get a

37:14Motorola Q phone now for $99. It's a

37:17very capable machine. It'll do music.

37:20It'll do internet. It'll do email. It'll

37:22do instant messaging. So, I I kind of

37:25look at that and I say, "Well, I like

37:27our strategy. I like it a lot.

37:30>> He burst out laughing, mocking it

37:32because it was so disruptive. It was way

37:34more expensive

37:35>> and it was way different. No keyboard.

37:37>> Well, phones used to be insanely

37:39expensive and the carriers would cover

37:40them, but you had to sign a long

37:41contract. You were still spending 500

37:43bucks. But the thing I'm getting at is

37:44you didn't have to explain to someone

37:46why perhaps you'd have to get past the

37:48cost part, but you could just be like,

37:49"Look how good this is." And then once

37:51the app was the iPhone 3G with the App

37:52Store, people were like, "Oh this

37:54could actually change things." mobile

37:56web. Even though it was a monstrosity,

37:58it was so bad at first. Even then, you

38:00could get your emails and you could just

38:01look at them. Point is, Blackberries

38:02were also expensive and were still

38:04actually kind of cool, but the way they

38:06worked was not like consumer software.

38:07They didn't have the classic GUI.

38:09iPhones felt like that. It felt like an

38:11a cell phone designed even like a

38:14computer. It was obvious. It was obvious

38:16from the beginning. Everyone I was I was

38:18dating a girl in the center of

38:19Pennsylvania at the time and everyone I

38:20showed it to was like, "Wow, this is

38:21incredible." That to me is the obvious

38:24thing with AI to this day when you're

38:27like, "Okay, why is it so amazing?"

38:28People still dither. People are still

38:30like, "Yeah, you can't run a business

38:32fully with it without this weird system

38:35of pulleys and levers and such."

38:37>> But how come then when you look at the

38:39stats around ChachiBT's growth,

38:42>> 100 million active users in just the

38:45first 60 days after launching? For

38:47comparison, Tik Tok took 9 months.

38:48Instagram took 2.5 years. And the

38:50internet itself for the worldwide web

38:51took roughly 7 years to reach that

38:53scale. Over 60% of the US adults are

38:56integrated into AI tools in their daily

38:58and regular routines within 3 years of

39:00the launch, reaching a 40% of the

39:02population. And that same milestone took

39:05the internet 5 years and personal

39:06computers nearly 12.

39:08>> Okay. So like this is the I think this

39:10is the part that's giving me dissonance

39:11is like when I showed my fiance chachi

39:14okay it was didn't [clears throat]

39:14really work

39:15>> but as a sole entrepreneur who English

39:18isn't her first language

39:20>> who has to write lots of text lots of

39:22copy and generate lots of images and was

39:23paying a graphic designer to help her

39:24make um certain images that she you know

39:27couldn't make herself because she

39:28doesn't have the skills.

39:30>> She would describe it as being

39:32transformative for her business. What

39:35I'm hearing from you is that it's not

39:37transformative and there's no value in

39:38it for people. But she if she was sat

39:40here transformative,

39:42would she pay the per million token

39:44rate? Would she pay the actual rate? Cuz

39:46that's the thing. If this was sold at

39:48its honest cost. Yeah.

39:49>> I would actually if and people were

39:50reacting like that and they were paying

39:5223 $4 every time they did something and

39:54they were genuinely happy. That might be

39:55an argument.

39:56>> What is the what would be the honest

39:57cost if they weren't sub

39:58>> the actual per million token cost? The

40:00actual API cost they should char.

40:02>> Do you know how much that is relative to

40:04God? Depends on it depends on the model.

40:06But there's actually kind of a point I

40:09want to make about the thing you said

40:10with the internet earlier. So when I

40:12first got on the internet 33.4 kilobits

40:14a second modem even back then I was like

40:17if this was faster and that was

40:20like immediate just like if this was

40:21faster cuz it was slow. You go on like

40:23happy puppy or something download take

40:25all bloody day waiting for share word to

40:27download immediately like if I could do

40:29this faster it would be better. And even

40:30back then I'm like, man, you could

40:32probably do video camera stuff with this

40:34stuff that eventually happened. And

40:35actually, there's this guy called Jim

40:36Cavell from Goldman Sachs in a report he

40:39did in 2024 that was geni too much spend

40:41for not enough return. Paraphrasing

40:43there. And he made the point that in the

40:44run-up to the iPhone, there was

40:47thousands of presentations that when GSM

40:49radios get smaller, when Bluetooth

40:50radios get smaller, when Wi-Fi radios

40:52get smaller, it is inevitable that we

40:55will get something like this. And then

40:57he said that there is no such path for

40:59AI. There was no road map to AI becoming

41:03this thing that they promised. And I

41:04must be clear, if these companies had

41:06gone out there and are like, "Yeah, this

41:08is interesting cloud software. It's

41:10generative. It's really expensive. We're

41:12not sure if we can fully not trust it.

41:15Not in the I'm scared way. I mean, just

41:16like we're not sure that this is going

41:18to be a disruptive world changing thing.

41:21It has potential, but we're going to go

41:23slow. It's really expensive. This is an

41:25R&D effort. We're not going to expose

41:26consumers to it." and actually being

41:28like called them like I don't know

41:30language models and no no generative AI

41:32stuff just being not even call it

41:34because it isn't AI it's not autonomous

41:36it's not smart I actually might respect

41:38it but this is not they've gone out

41:40there since 2023 and said it was 2022

41:43this is the best thing since sliced

41:44bread this is changing everything this

41:46is going to do all your work this is

41:48going to take your job you're going to

41:50talk to Bing and it's going to tell you

41:51to leave your wife all of these crazy

41:52things and what's funny is when the

41:55writer uh Kevin Roose I think it was

41:58He was speaking to Kevin Scott, the CTO

41:59of Microsoft, about it. And Kevin Scott

42:01goes, you know, I'm just glad we're

42:02having this conversation. Instead of

42:04being like, "Settle down, Beas. It's a

42:06website. The website told you something.

42:08It's just LLM." They talked it up. And

42:10that's because everyone is talking about

42:12what they wish this was. Rather than

42:14talking about what it can actually do.

42:16This makes it scary to people

42:18deliberately. So, it makes it

42:20environmentally destructive. Look at the

42:21gas turbines poisoning black

42:22neighborhoods. I think it's in

42:24Louisiana. It's one of Musk's data

42:25centers. Look at the incredible energy

42:28draws. It is raising power bills and

42:30also it is creating inflation across all

42:33consumer electronics because of the

42:35massive RAM.

42:36>> You know what's interesting? I almost

42:37feel like so much of what you're saying

42:40is true and also it can be true that

42:45this technology is going to profoundly

42:47change the world. And I think like you

42:49know I think back to the early days of

42:51the internet is maybe the closest

42:52analogy we have of you know in the com

42:55bubble. you know, you wrote this great

42:56essay.

42:57>> Yes. Yes.

42:57>> Which I found really funny um especially

43:00the name the rot economy and you talked

43:02about the rotcom bubble.

43:04>> Yes.

43:05>> Talking about how AI is of less value

43:07than people think.

43:09>> And in that in the sort of com bubble,

43:11what you saw is huge hype, people

43:13overselling the capabilities of their

43:15websites and what they were building.

43:17But in the wake of the dotcom bubble,

43:21yes, 90% of stuff went to zero,

43:23>> but you had generational companies born

43:26that changed the world,

43:27>> right?

43:28>> And so I I do I kind of and that's what

43:30bubbles do, right? Huge hype,

43:32overinvestment, investors get crazy,

43:34delusional. They think it's everything's

43:36going to change. At the same time, you

43:39do have skeptics

43:40>> in these moments. The the dot bubble had

43:42I mean the internet itself had the

43:44biggest skeptics in 1998. Nobel Prize

43:46winning economist Paul Krugman said by

43:502005 or so it will become clear that the

43:52internet's impact on the economy has

43:54been no greater than the fax machine. In

43:561995 astrophysicist Clifford stool

43:59famously I wrote about this in my book

44:01wrote famously in Newsweek. Do our

44:04computer pundits lack all common sense?

44:06The truth is no online database will

44:08replace your daily newspaper. No CDROM

44:11can take the place of a competent

44:12teacher. Commerce and businesses will

44:14shift from offices and malls to networks

44:16and modems. Bologoney. So, how come my

44:19local mall does a roaring business and

44:22the cyber mall gets zero business? And

44:24then I'll give you one more from

44:26Krueger, who was the award-winning

44:28economist. He said, "The growth of the

44:30internet will slow drastically as it

44:32becomes apparent most people have

44:33nothing to say to each other."

44:36That's that that that may actually be

44:38the worst one of those predict like hang

44:41around any bar in middle America.

44:43Honestly, the best conversation,

44:44>> but it's just all the same thing.

44:45>> I actually So, Clifford Stall actually

44:47his piece was interesting cuz that there

44:49were some boner points in it, but he

44:50made points about how like an

44:52overwhelming amount of bad information

44:53out there is bad for society. He's

44:54completely right saying how online

44:56education would not be a great

44:58replacement for regular education. I

45:00think we've seen that. But there is an

45:02economic difference that's vastly it's

45:05just completely different. So.com bubble

45:07was actually two bubbles. There was the

45:08website bubble which was just trash on

45:10trash on trash. It was just like I think

45:13what was it? Excite at home bought a

45:15eury incard company for like a billion

45:17dollars. It was insane crap happening

45:19that was so small. The big thing that

45:22people are thinking about is the dark

45:24fiber.

45:24>> Dark fiber. dark fiber was all of the

45:27wires that put in the ground thinking

45:28we're going to have all this demand for

45:30internet and it turned out that demand

45:32for internet I think the analyst

45:35estimate was it was doubling every 90

45:37days when it was doing that every 6 to

45:3812 months maybe maybe longer and just

45:41thus there was a massive overbuild of

45:43fiber optic cable and indeed the

45:46transmission stations and such just

45:48simplifying to bring that to people's

45:49houses and there was the assumption that

45:52well that would all get lit up and

45:53people would want it immediately didn't

45:54really

45:55Now the post.com bubble thing people say

45:57is well but after that there was demand

45:59from the internet. That's the thing

46:01though that's very different to demand

46:03for generative AI. Right now the demand

46:06we have for generative AI is

46:07predominantly subsidized. Just let's

46:09start there.

46:10>> Yeah

46:10>> predominantly subsidized and most people

46:12experience it are not paying the real

46:14cost.

46:14>> I agree.

46:15>> On top of that we already have all of

46:18the possible marketing in the world. We

46:20have the largest, most disingenuous

46:22marketing campaign in the history of

46:24man, pushing this up the hill. We have

46:27the apex predator of cloud software,

46:30Microsoft. They can only get singledigit

46:32billions from selling AI software. And

46:35Christ almighty, outside of OpenAI and

46:37Anthropic, we barely get $22 billion.

46:40And the thing is, $22 billion is a large

46:42amount to you and me. It's not a large

46:44amount of money when you spent a

46:45trillion plus dollars. When you have

46:47anthropic and open AI with $1.1 trillion

46:50worth of cloud commitments and on top of

46:52that, how does this turn into a post.com

46:54bubble thing? A data center built today

46:57is going to be as expensive to run in

46:592050 as it is today unless there's some

47:01breakthrough in electricity. But again,

47:04that's not happening with AI. AI is not

47:06doing that unless there's some

47:07breakthrough in GPU technology. But we

47:09already have Broadcom, Nvidia, etched.

47:12We have every major chip company ARM

47:15trying to do something about this. And

47:17no one seems to magically be able to

47:18make this profitable or indeed even less

47:21costly. Even Nvidia with Vera Rubin,

47:24their more expensive new GPU system.

47:26Even then, they're like, "Yeah, 10x more

47:28efficient. It's uh more dollars per

47:30megawatt." They're all koi about it.

47:32They don't just say, "Yeah, we worked

47:33with OpenAI and Anthropic and we found

47:35it reduced our cost by 50%." Easiest

47:37thing in the world if it was true. And

47:38that's because it's not happening. And

47:41this isn't a case where

47:42>> So are you saying there's not going to

47:43be the demand for let's say let's you

47:46know there's different types of AI

47:48generative AI we

47:49>> Yeah. And actually that's a good point

47:50to make. The reason they use the term

47:52artificial intelligence is so everyone

47:54would lump everything into it.

47:56>> They [clears throat] would lump uh

47:57protein folding nothing to do with LLMs.

47:59Robotics not LLM.

48:01>> Autonomous weapons even horrible as they

48:02are not LLMs because you couldn't trust

48:04them. But they've mushed everything into

48:06AI so that when you say, "Well, AI

48:09can't," they'll go, "Um, um, sir, you

48:12forgot to give us homework and also AI

48:14it's working on curing cancer." When

48:15it's just like, "No, that's not LLM.

48:17Stop giving them credit."

48:18>> The similarity though is they all need

48:20GPUs, all these.

48:21>> And that's the funny thing. All those

48:23data centers that we're building, all of

48:25them are for just generative AI. They're

48:28not for all of the other stuff. They're

48:30not for the cool AI has been

48:32around for a long time. Google. A lot of

48:35the good stuff that comes out of Google

48:36from the search side is AI but

48:38pre-generative.

48:39>> How would you run the the type of AI

48:42that sits in a robot? Let's say one of

48:44the Optimus robots if you didn't have a

48:46GPU.

48:47>> So Matic Matic has this cleaning robot

48:49for example. That thing is not got a

48:51little GPU in it. What it has and may

48:54indeed have used some GPUs but no year

48:57as many as they need for generative AI

48:59to run the data feed training data into

49:01it so it's able to clean a house. But

49:03when the little buggers going around

49:04cleaning my floor, turdsly I call him,

49:06it goes around mopping my floor, it's

49:08not like burning money the whole time.

49:10But when it comes to these massive

49:12amount of data center, sighteline

49:13climate said in February there's 190

49:15gawatts of data centers under in

49:17planning. Don't know about under

49:19construction that works out if about 12

49:21million megawatt that's what like $1.6

49:23trillion to3 trillion a year in annual

49:26demand you'd need for that. We don't

49:27even have $130 billion worth of annual

49:30demand. And people say, well, it will

49:31grow. how when most of the demand is

49:33coming from Amazon feeding money to open

49:36AAI or anthropic, Microsoft feeding

49:38money to OpenAI and Anthropic, Google

49:40feeding money to Open AI and anthrop

49:42well hasn't fed it to Open AI yet, but

49:44they're a pretty big customer, billions

49:46of dollars. The conside is that we are

49:49building these effiges to capitalism,

49:51these giant GPU data centers, and people

49:53are being told, well, it's for AI, you

49:56know, the thing that's done all this

49:57other stuff that's unrelated. Or the

49:59worst thing I've seen is like, oh, you

50:01don't like you like online banking.

50:02Well, you do like data centers. There's

50:04a big difference between a data center

50:05for regular nonGPU compute for standing

50:08up a server, a content delivery system

50:10like Akami or something that brings the

50:12website to you or how Meta runs

50:14Facebook. That is not the same. It takes

50:16way less power, mostly CPUdriven

50:19compared to these giant GPU data centers

50:21that offer one thing, one thing only.

50:23>> But I was doing the the research and

50:25looking at some of these notes here. It

50:27does say that for tougher types of AI

50:29systems designed to solve concrete

50:30physics, biology, and spatial problems,

50:32they require some of the most intense

50:34data center infrastructure on the

50:36planet.

50:36>> Yeah.

50:37>> AI systems like Deep Mind's AlphaFold,

50:39the protein folding company

50:41>> used for genomic sequencing and climate

50:44forecasting, etc. run on high

50:45performance computing clusters. These

50:47require immense precision and continuous

50:49heavy computing data centers.

50:51>> Yeah. Training the brains for

50:52self-driving cars requires billions of

50:54miles of simulated physics environments.

50:58The AI isn't generating text. It's

51:00learning to navigate 3D spaces and

51:03gravity and relies on data centers,

51:04>> right? And the thing is those data

51:06centers, they might have GPUs in them.

51:08We had GPUs used for this HPC, the high

51:12performance computing before generative

51:14AI. And yeah, that's how AI has been

51:16trained before. That's how Tesla did.

51:18believe they've had their own data

51:19centers when it comes to training the

51:21autopilot system for better or for

51:22worse. That's how we've done it before.

51:24Again, that is not why we're building

51:26these data centers. These data centers

51:28are being built to sell to AI generative

51:31AI companies to either train systems or

51:33run inference. These things are being

51:36built in this brainless way where it's

51:39just well actually maybe this is a good

51:41way of illustrating the con because

51:43everyone saw Google, Microsoft, Amazon

51:47and Meta give Nvidia over call it 800

51:52something billion dollars

51:54because everyone saw that they went well

51:56they wouldn't do that for no reason.

51:57They went we got to build more of these

51:58things. There must be all this demand.

52:00Even though the demand 70% or more of

52:04all that demand comes from these two

52:05companies who were funded by these three

52:07companies and that's the funny thing.

52:10The reason that they don't want to break

52:11out their AI revenues is because it will

52:14become alarmingly obvious that this was

52:16the case. It turns out that the only

52:18real big customers cuz it's not like

52:21they're building a few data centers.

52:22They're building trillion plus revenue

52:25potential. They believe they'll get

52:27speculative. It's entirely speculative.

52:29They're building it because they saw the

52:31biggest companies in the world buy a

52:32bunch of GPUs and they said, "I want in

52:34on that." They must have diverse

52:36customers, right? They wouldn't just

52:37have two unprofitable fail sons that

52:40they're propping up with. Christ,

52:42they've raised $217 billion just in

52:442026.

52:47>> So, we know that some of the biggest

52:49companies in the world are using AI,

52:51generative AI to write a lot of their

52:52code.

52:53>> Mhm.

52:53>> That is a great productivity gain for

52:55those companies, right? I mean, have you

52:58used Google or Facebook or Instagram or

53:00GitHub recently because they are

53:03catastrophically worse? Amazon Web

53:04Services went down multiple times

53:06because of their AI coding tool. How

53:08>> how is how is Google worse?

53:10>> Well, I'll tell the story of a real

53:11guy called Preaggo Ragavan.

53:13Previously, one of the heads of ads at

53:15Google in 2019, Google called something

53:17called a code yellow, which is when they

53:19said, "We've got a problem." And it was

53:21material weakness in query numbers which

53:24means the amount of times that people

53:26were searching on Google search. Guy

53:28called Ben Gomes internal at Google then

53:29the head of Google search says wait a

53:32minute to increase this number of using

53:33Google more.

53:34>> Mhm. We're going to have to I mean you

53:37what you're suggesting would mean we

53:38give worse answers because if someone

53:40got the answer quickly that would reduce

53:41the amount of queries right and people

53:44at Google Shashi Tako was another

53:46engineer was saying yeah can we please

53:47tell Sunda this because this doesn't

53:49seem good. We can't just increase the

53:52amount of queries. That would just mean

53:53that people would have to search more

53:54which would make the product worse.

53:56>> But but it would make them more money.

53:57You saying you'd show them more ads. So

54:00if you're spending more time on Google

54:02because Google's work,

54:03>> but is this linked to AI doing code?

54:04>> Oh, I'll get there. So

54:07>> this is the problem is is that this guy

54:09called Pragar Ragavan who's the head of

54:11ads at the time was pushing pushing and

54:13saying, "No, we need to make more

54:14queries happen. Got to make it happen."

54:16and Nick Fox who was there as well I

54:17believe was actually taking over Google

54:18search got to make them go up this is

54:20our new reality sometime in early 2020

54:23propagar ragavan takes over Google

54:25search from then and this is this is

54:28what I believe can't prove it if you go

54:30and look around the various SEO sites

54:32such journal and the various forums

54:34Google stripped back a lot of the

54:36suppression of spammy sites so that

54:38people would be on Google more and then

54:40over the course of time Google wanted to

54:43create more queries and Google search

54:45became much worse. It's why people

54:47always do like plus Reddit or from

54:49Reddit or what have you. It's because

54:50the actual underlying search results of

54:52Google had got worse. And then

54:53Generative AI came along and Praagar,

54:56wouldn't you know, it gets put to run

54:57part of Gemini. And Google also was

55:00having trouble getting people back on

55:02Google. And what did they think they'd

55:03do? Well, everyone's talking about

55:05this AI thing. We'll just put it right

55:07at the top so people have to stay at

55:09Google. And actually, they'll use it

55:10more because instead of searching

55:12websites and doing that annoying thing

55:13where they click away from Google,

55:15they'll just only use Google. Instead of

55:17generating answers, by which I mean

55:20giving you search results you click

55:21through, now Google is the answer. Is it

55:23right? God know. It might tell you to

55:24eat rocks, might eat poisonous

55:27mushrooms. Maybe it'll give you a little

55:28few links you could click through. But

55:30the ideal situation was that AI was the

55:33ultimate form of Google's evil which was

55:35>> But I'm saying here I'm saying here but

55:36that's not the fact that coders could

55:39code on Google that's made Google worse.

55:40That's human decisions have made it

55:42worse.

55:42>> Yes. And then there's the instability of

55:44Google's platform which is actually I

55:46should have probably led with that a

55:47problem across the whole tech industry.

55:49>> Okay. So you're saying that you're

55:50saying Google is going down more.

55:52>> Yes. Google is less stable. Google Docs

55:55is a bugfest right now and has been for

55:57a while. Google Sheets, same deal. And

55:59the thing is, you're right, I'm being a

56:01little unfair. This is everyone. It's

56:03the same with Microsoft. It's the same

56:04with Amazon. It's the same across.

56:05>> How do we quantify that outside of

56:07anecdotes? Like, is there a way to

56:09>> You're right. I mean, GitHub downtime is

56:11the best example. Amazon Web Services

56:13went down two or three times this year

56:15because of AI tools. And honestly,

56:18you're right. It is kind of hard to

56:20quantify outside of anecdotes. But I

56:22challenge anyone listening to this. Go

56:23and use a website these days and tell me

56:24how well it works. Tell me how buggy it

56:27is. Tell me how many problems even with

56:28my iPhone. The supposed best UX in town.

56:32Even the iPhone is a flipping mess these

56:34days.

56:35>> Okay, so the research says the short

56:39answer is yes. Tech downtime and

56:41software outages have demonstrabably

56:43increased over the last few years and

56:45industry data points directly to the

56:46explosion of AI assisted coding as a

56:48primary culprit. The problem is hitting

56:51the tech industry from two entirely

56:52different directions. The code itself is

56:54getting buggier and the sheer volume of

56:56AI activity is literally crashing the

56:59underlying infrastructure. Interesting.

57:01>> Yeah, that's because GitHub people are

57:03just writing a bunch of code, pushing

57:04it, and thus there's just more code on

57:07there.

57:08>> That's interesting.

57:09>> Yeah, it's it's a real mess as well

57:11because

57:12open source has had this problem as well

57:14because it's well-meaning people.

57:15They're like, I learned a bit of code

57:16with an LLM. I'm going to go out and do

57:18some stuff. I'm going to make this

57:19project better. And these people barely

57:21understand what they're shipping. Or

57:23maybe they understand a bit of code and

57:24they say, "Oh, Dunning Krueger, this

57:26I'm going to I'm just

57:28like, I can understand some of this."

57:29And now the code's all written and just

57:30push it right now. So GitHub is flooded

57:32with AI code.

57:33>> This sounds like it's making humans

57:36complacent.

57:37>> It is

57:37>> because we're going, "Okay, look, I let

57:39it write the the code for the last 100

57:41lines and it was broadly right. So the

57:44next 100 lines, I won't check them as

57:45much."

57:45>> Yeah. Yeah. And that's human nature is

57:48to get sort of to take shortcuts to

57:50spend less energy on an activity if you

57:52can right but the AI's still making the

57:55mistake and we're still making all the

57:56promises of AI that's the thing this

57:59thing is meant to be this autonomous per

58:01you say it can't be perfect I don't know

58:03based on what Samman has been saying for

58:05the last few years clammy Sammy has been

58:07promising the world saying this will

58:09replace software engineers Dario

58:10Ammedday Wario himself has been saying

58:13oh yeah 50% of white collar labor is

58:16going to go away in the next few years.

58:18These people are promising the world.

58:20Again, if they were saying it would be

58:22smaller and they were like, yeah, it

58:23does have issues and we must be none of

58:26this, oh, what if it wakes up and it's

58:28super powerful. Just like, yeah, it's

58:30probabilistic. It's going to make

58:32mistakes and if you don't know what

58:33you're doing, you don't really know what

58:34you're looking at, you're going to miss

58:36those mistakes and it's going to get

58:37multiplicatively worse as you go when

58:40you don't know what you're doing. So

58:41yeah, human nature is part of it, but so

58:44is the marketing. So are the promises.

58:47One of the smartest things a business

58:49can do is build like a bigger company

58:52without actually hiring like one. But

58:54the problem we all face is that most

58:56companies don't have every skill in

58:57house. So when I look at the businesses

58:59seeing real success today, the

59:01consistent pattern with all of them is

59:03how quickly they move. They bring in

59:05specialists with skills in emerging

59:06areas to keep themselves ahead. Even in

59:08our company, we spent the last year

59:10pulling in talent across areas like AI

59:12native strategy, no code builds, and

59:15product workflows. And we find this

59:16talent through our longtime partner,

59:18Fiverr Pro. Their premium service only

59:20shows you vetted talent. So, you've

59:22always got the safeguard that anyone you

59:25pull in to help you with a complex

59:27project has the skills that you're after

59:29and will deliver to the same high

59:31standards as your internal team. And

59:33most importantly, they'll keep up with

59:34the pace. It's a simple strategy, but it

59:36lets us stay agile without compromising

59:38on quality. So, if you need these kind

59:39of skills in your business, head to

59:41pro.fr.com to find pioneering talent to

59:44fill your business's gaps. That's

59:46pro.fr.com.

59:48You know, the little traditional SIM

59:49card that goes inside of our phones.

59:51They haven't changed at all since they

59:53were invented in the '90s. You have this

59:55physical piece of plastic that means

59:57you're locked into one carrier, one

59:59network, and the second you cross a

1:00:01border, that carrier can start charging

1:00:03you whatever they want. But there are

1:00:05alternatives, and today's sponsor, SY,

1:00:07is one of them. It's an eim app that

1:00:10gives you a safe and secure data

1:00:12connection in over 200 destinations. All

1:00:15of their eims have built-in cyber

1:00:17security, which is great if you're

1:00:18traveling for work and looking at

1:00:19confidential material. I've been using

1:00:21SY whenever I travel because the

1:00:23connection is always reliable and it

1:00:25saves me a ton of roaming fees. It also

1:00:27means I don't have to deal with all of

1:00:29the faf that surrounds sorting out a SIM

1:00:31everywhere I go. If you want to give it

1:00:33a try, download the SY app from the app

1:00:35store now and scan the QR code on

1:00:37screen. And if you want 15% off your

1:00:39first purchase, use my code D

1:00:44when you get to check out. That's D O A

1:00:47for 15% off. Keep that to yourself. My

1:00:51my car that drives itself that is AI

1:00:53technology.

1:00:54>> Yes.

1:00:54>> I sat here with Dra from Uber and he was

1:00:57saying that I think in a couple of years

1:00:59time

1:01:01we won't need drivers um for Uber

1:01:04because the cars will drive themselves

1:01:06like they'll be fully autonomous.

1:01:08>> And I think if I'm not mistaken

1:01:11driving is one of the biggest

1:01:12professions on planet earth. So when you

1:01:14hear people when you hear these CEOs

1:01:16saying that there will be job disruption

1:01:19>> you say that they are not telling the

1:01:21truth.

1:01:22>> Yes. Or they're guessing in a way that's

1:01:24very good for them. Think about it from

1:01:26perspective of Microsoft Sachin Nadella.

1:01:28He's not going to be like yeah we don't

1:01:30know if this is going to work mate. Of

1:01:31course he's going to talk his book and

1:01:32he's going to say yeah this is going to

1:01:34replace all workers. It's going to be

1:01:36amazing. He it's going to be so

1:01:38powerful. And then he'll change his tune

1:01:39and say actually it's not going to

1:01:40replace workers. that make him more

1:01:41powerful because the things aren't

1:01:43catching up. Dor from Uber for example,

1:01:45of course he's going to say if this

1:01:47happens then that would be good for Uber

1:01:49because Uber would just become an

1:01:51autonomous taxi service. There's a

1:01:53reason that Whimo's taken I I find Whimo

1:01:56fascinating. I think that it's

1:01:57really cool. I think there are

1:01:57socioeconomic problems that will come

1:01:59from it. I think there are actual real

1:02:00problems that will emerge and also

1:02:02>> what kind of problems?

1:02:03>> Well, I mean socioeconomically there are

1:02:05like you said one of the largest

1:02:07employment centers in the world. I mean

1:02:09just the economics of cabs will fall

1:02:10apart but again we are nowhere nowhere

1:02:12nowhere near that. We're not even close.

1:02:14Whimo has had to do the smallest

1:02:16rollouts and the most control things

1:02:18because the problem with pretty much

1:02:19every AI system but especially driving

1:02:22is not the getting 95% of the way. It's

1:02:25those edge cases. It's raining which is

1:02:27a big problem for them in San Francisco.

1:02:29It's a kid runs across the road but

1:02:30they're wearing a high viz thing. Does

1:02:32it even notice it's a child? Again, this

1:02:34is a really interesting but very very

1:02:36applicable example of uh the right

1:02:38comparison to be made shouldn't be

1:02:40autonomous vehicles versus perfection.

1:02:42It should be autonomous vehicles versus

1:02:44human drivers. I mean, I don't know if I

1:02:46agree because a human driver might make

1:02:48mistakes, sure, but again, not an expert

1:02:51in autonomous cars. Just want to be

1:02:52clear. But if we're pushing autonomous

1:02:54cars out there willy-nilly and we're not

1:02:56doing so in extremely controlled

1:02:58environments, those edge cases will

1:03:00multiply and be dangerous. Yeah, they

1:03:02might be better at human drivers in some

1:03:04ways, but they might also I was in Vegas

1:03:05the other day and I was in a hotel and I

1:03:08watched a bunch of Zuk's cars just get

1:03:10stuck.

1:03:10>> They're autonomous cars.

1:03:12>> Yeah, they these weird boxy things. They

1:03:14just blocked the exit. They just all

1:03:16kind of lined up and just fell asleep. I

1:03:18saw the same thing actually happen

1:03:19outside of a hotel when I got out of a

1:03:21Whimo in San Francisco. Just stopped at

1:03:22the and then a bunch of cars and another

1:03:24Whimo got stuck behind it. And these are

1:03:26kind of

1:03:27>> I've seen some human bad drivers as

1:03:29well. I I agree, but it's just we have

1:03:31control over deploying these bad or good

1:03:34drivers. We have an ability to roll them

1:03:37out slowly, which is exactly what we

1:03:39should do. I'm not saying autonomous

1:03:41cars are bad. I'm saying we need to be

1:03:43so so so careful and treat them as

1:03:46guilty and pro till proven innocent

1:03:48because we can prove and also they have

1:03:50people overlooking them. They actually

1:03:52have people monitoring the roots. It is

1:03:53something they cannot rush out and it

1:03:55doesn't seem like they're rushing it,

1:03:56which is good. and they're not promising

1:03:58the world.

1:03:58>> I do agree. Listen, I I'm a big fan of a

1:04:01big fan of taxi drivers generally in

1:04:02part because I spend a lot of time in

1:04:03taxis and I think I'm not just getting

1:04:05in there because I want to get to from A

1:04:06to B. I'm getting in there for lots of

1:04:08other reasons.

1:04:08>> Yeah.

1:04:09>> However, when I look at the stats

1:04:11>> around what is more dangerous

1:04:14>> driving myself or having an autonomous

1:04:16vehicle drive me, there's an 68% lower

1:04:19overall crash involvement rate when

1:04:21you're in an an autonomous vehicle. Mhm.

1:04:23>> Autonomous vehicles experience roughly

1:04:252.1 police reported crashes per million

1:04:27miles compared to humans that are at

1:04:29roughly 4.68 per million miles. So, a

1:04:3255% reduction when you get in an

1:04:34autonomous vehicle. And autonomous

1:04:35vehicles show an 80 to 81% reduction in

1:04:38crashes resulting in injuries versus

1:04:41human drivers.

1:04:42>> Uhhuh.

1:04:42>> So, you're 85% less likely to be

1:04:45involved in a single vehicle crash like

1:04:47hitting a wall or a tree if you're an

1:04:49autonomous vehicle

1:04:51>> versus being driven by I agree. But

1:04:53>> so it's safer

1:04:55>> in also that data is what's the sample

1:04:58size of human drivers? I mean we've got

1:05:00many many many many many many more years

1:05:02of drivers and many many many more years

1:05:04of accidents and also man does that not

1:05:06have anything to do with generative AI.

1:05:08If we were just talking about that be

1:05:11having a different conversation.

1:05:12>> I guess the question here was really

1:05:13around job disruption. Like you know we

1:05:15we look across industries and we go

1:05:16driving is a massive profession. Is

1:05:18there going to be job disruption because

1:05:19cars can now drive themselves? If we

1:05:21think about white collar, you know,

1:05:22jobs, you know, lawyers and accountants,

1:05:25people sit here and they tell me that

1:05:27lawyers and accountants would the

1:05:29profession, right? I should say some of

1:05:31the skills within the profession will be

1:05:33relegated to AIS to do.

1:05:35>> Here's the thing. Lawyers, for example,

1:05:37great example. Always hearing

1:05:40legal partners talking about AI. Never

1:05:42the associates. The associates are the

1:05:44ones that go out and find the president.

1:05:46They're the ones that go and do the

1:05:47grunt work. They're the ones who are

1:05:48pulling motions half the time. The

1:05:50partner is the one that might be the

1:05:51litigant. It may be the client facing,

1:05:53but the ones that are actually doing the

1:05:54day-to-day work. I'm not hearing from

1:05:56them. I'm not hearing associates being

1:05:57like, "This is awesome." I'm

1:05:59hearing a bunch of well- paid people

1:06:02that have sat on Chat GPT and gone,

1:06:04"Yeah, yeah, I'm the greatest lawyer

1:06:06ever." They're not the ones that I want

1:06:07to hear from the actual workers. White

1:06:09collar labor disruption is not

1:06:11happening. Open AAI had a study that

1:06:13came out I think like a week ago that

1:06:15said there was no corre connection

1:06:17between spending on AI tokens and

1:06:18revenue per employee. Like this is open

1:06:21and that's

1:06:21>> what does that mean? Could you explain

1:06:22that to me?

1:06:23>> As in the more tokens you spend has no

1:06:25no correlation at all with the amount of

1:06:28money you make. It's the second report

1:06:30they've put out. The other one was like

1:06:31hallucinations are mathematically

1:06:33guaranteed kind of almost the one thing

1:06:35I respect about that company that

1:06:36occasion they just put out a study. It's

1:06:38like, yeah, kind of sucks.

1:06:40[clears throat] But the people that are

1:06:42having their lives disrupted work-wise

1:06:44are art directors. It's people, art

1:06:47directors, transcribers, translators,

1:06:49who have bosses that don't care about

1:06:51the output. It's what they consider

1:06:53cheap work. And the problem is is those

1:06:56people would have automated your work

1:06:57away anyway. They would have sold it.

1:06:58They would have taken the cheapest for

1:07:00they would have sold it to the global

1:07:01self. They would have taken the

1:07:02shittiest option they could. That is

1:07:04something that AI is doing. And again,

1:07:05those people are not paying the actual

1:07:07cost of AI. They're using a

1:07:08subscription. The actual white collar

1:07:11labor force might have some things that

1:07:15are slightly changing, but there is no

1:07:17evidence of like productivity gains. In

1:07:20fact, if there were, they would be

1:07:21screaming it from the rooftops. There

1:07:23was an Oxford economics study last year

1:07:25where it's like, oh, young people are

1:07:27finding less jobs because of AI. We

1:07:29actually read the study, which multiple

1:07:31journalists did not. It was a single

1:07:32line that said, "Yeah, we saw some

1:07:34correlation." Didn't give a number.

1:07:37Didn't actually say what the correlation

1:07:38was. We are so conditioned to believe

1:07:41that the rich and powerful know what

1:07:43they're doing that we internalize these

1:07:46narratives about like, well, previous

1:07:48booms lost a lot of money. Well,

1:07:49technology takes time to do stuff. And

1:07:51they are intentionally playing on those

1:07:54mythologies. They are playing on these

1:07:56knowing that journalists, analysts,

1:07:59investors will believe them. And this is

1:08:01partly because our our realities are

1:08:03defined by stock prices. Because the

1:08:05stock prices of these companies went up,

1:08:07we're like, "Oh, look, it must be

1:08:09working, right?"

1:08:10>> Both of those things you said were true,

1:08:11though, right? Like that previous

1:08:12technologies didn't make money at the

1:08:14start and you The other one you said was

1:08:16um they'll get better.

1:08:17>> But that's the thing. Okay. Because

1:08:19another thing got better, this will get

1:08:21better.

1:08:21>> No, but there's there's got to be

1:08:22something that they're saying that is

1:08:24fundamentally not true because those are

1:08:25two true statements that okay,

1:08:27technology often starts

1:08:28>> I know. I get what you mean. What they

1:08:30are fundamentally misleading people

1:08:32about is how possible it is. How many

1:08:34actual signs they have because they

1:08:35don't have the signs. If they had the

1:08:36signs as in the signs of this getting

1:08:38cheaper as in the signs of this being

1:08:40able to autonomously do work without the

1:08:42Rub Goldberg machine and even then in a

1:08:45reliable way that was making the

1:08:47customer more money being productive in

1:08:49a way you can say with your whole chest

1:08:51without a series of asterisks and that's

1:08:54how it is across the board. The people

1:08:56that are most excited about this,

1:08:59psychopaths on Twitter in many cases are

1:09:01people that I believe there really are

1:09:03some I'm sorry, there are some people on

1:09:05Twitter because the other thing about

1:09:07this is this is really unique to the AI

1:09:09industry. I've never seen it any other

1:09:11industry outside of maybe like sports

1:09:13teams. The attachment that some people

1:09:15online have to these companies. If you

1:09:17dare dare to criticize anthropic, it's

1:09:20almost this religious attachment. Good

1:09:23example was this week Bloomberg reported

1:09:25that OpenAI was on track to hit $40

1:09:27billion in annualized revenue. Month

1:09:29times 12, four weeks times 13, we don't

1:09:31know. They don't define it. I saw

1:09:33multiple people and I going actually

1:09:35it's 60 billion. It's actually 60

1:09:37billion. I heard from someone it is like

1:09:40a cult and it's a cult of software

1:09:42driven around growth and this idea that

1:09:45by backing the right horse you will have

1:09:48some grand thing and open AI in

1:09:51particular in particular Mr. Baltman

1:09:54they have been fermenting this that Tibo

1:09:56as well the Tibbo the one of the guys at

1:09:59uh OpenAI they ferment this thing online

1:10:01they build this kind of parasocial

1:10:03relationship with both the large

1:10:05language model themselves and the

1:10:07companies and one's allegiance to the

1:10:09companies is so important it's truly

1:10:12vile if only these people gave a

1:10:14about I don't know Medicare for all or

1:10:17poverty or thing like actual problems in

1:10:19the world versus are we buying enough

1:10:21GPUs Do you know what's interesting is

1:10:23some of what your narrative

1:10:26one would argue actually helps them.

1:10:29How? Because you know the AI doomers

1:10:31that have come here and told you know

1:10:32some of the original founding fathers of

1:10:34AI like Jeffrey Hinton have told me that

1:10:37what they're building is highly highly

1:10:38dangerous and that it will be

1:10:40fundamentally disruptive to society. And

1:10:43it's interesting because some of the

1:10:45CEOs who you've mentioned, their

1:10:46historical narrative was also, by the

1:10:48way, this is really dangerous

1:10:50and there is a significant chance it

1:10:51could f we could up the planet.

1:10:53>> And what we've seen is this slow pivot

1:10:55away from it because now they're getting

1:10:57booed and they're being attacked.

1:10:59There've been this slow pivot away from

1:11:00it. And the pivot almost sounds a little

1:11:04bit like your narrative.

1:11:05>> It now sounds like actually no, it's not

1:11:07going to change anything and you're all

1:11:08going to be fine. And it's now there's

1:11:10just not it's nah it's not dangerous at

1:11:11all.

1:11:12>> But that's the funny thing

1:11:13>> and that's why I'm saying like you're

1:11:14you're not they I actually think there

1:11:16might be a couple PR people at these big

1:11:18AI companies thinking thank god for Ed

1:11:22some [laughter] of it because you're

1:11:23like you're saying actually don't worry

1:11:25everything's going to be fine. It's not

1:11:26going to take your job. It's not going

1:11:27to disrupt the economy. It's just a fad.

1:11:28There's no technology. And I think they

1:11:30don't think that.

1:11:31>> Here's the thing. I think Alman and

1:11:33Amday are some of the most deeply

1:11:34corrupt and cynical people in the world.

1:11:36I don't think of course they were going

1:11:37to say from the it was early 2023 or man

1:11:40said we're a little bit scared about

1:11:41what we're creating. Oh, shut up. I'm

1:11:44just I hear that and I feel so

1:11:46frustrated because I've met so many of

1:11:48these rich liars, these people.

1:11:50And you know why he wants to say that?

1:11:52So you'll invest in his company and buy

1:11:54the software. So you'll be scared that

1:11:56if you don't use AI today, you'll be

1:11:57left behind in the future, which is

1:11:59their continual narrative that if you

1:12:01don't get on the train today,

1:12:03then you'll be left behind. By the way,

1:12:05every single scam and con starts with

1:12:07rushing you. Every single trick in

1:12:10history begins with saying you must do

1:12:12this now. And best piece of advice I

1:12:14ever got was if anyone tries to rush you

1:12:16and it's not literally a mortal thing

1:12:18like you are bleeding or on fire or the

1:12:19house is on fire, slow down. And yet all

1:12:22of these companies saying it's so scary.

1:12:24And now they're talking about slowdowns.

1:12:26But you ever noticed that Amade and

1:12:28Ortman, they say, "Oh, maybe we should

1:12:29slow down progress." And then they

1:12:31don't. Right now, Orman's saying, "Oh,

1:12:33we slow down progress because we're so

1:12:34delayed." No, they're out of compute.

1:12:36Now, they're doing it. I can guarantee

1:12:37you, by the way, their PR people do not

1:12:39like me. I know for I know I don't think

1:12:40OpenAI's PR people are super fond of me.

1:12:43>> But I bet there's elements of what

1:12:44you're saying because you're calming

1:12:46people. You You are theoretically

1:12:47calming down the general public.

1:12:49>> And you know what? I hope I am because

1:12:51>> the fear based tactics is horrible.

1:12:53These companies don't want that. These

1:12:54companies want people scared. I'm 100%

1:12:56sure.

1:12:57>> Uh I don't I just fundamentally

1:12:59disagree. I think it

1:13:00>> can I so the timelines there and I sit

1:13:02here and what I do is I log their quotes

1:13:04over time

1:13:05>> and I read them out from 2015

1:13:08>> to 2026 and the change you see is them

1:13:12going from there could be extinction

1:13:14that's the narrative the early narrative

1:13:16Elon said it himself he says it's the

1:13:17single most dangerous thing in

1:13:18>> Elon and then you track it over time and

1:13:21it evolves to this age of abundance

1:13:23we're all going to have unlimited stuff

1:13:25and then um the the new slogan at

1:13:28trackbt is intelligence for everyone.

1:13:30It's suddenly and all the and and

1:13:32whenever Daario comes out and says, "By

1:13:34the way, it's really dangerous."

1:13:35They attack Daario. Yeah. They hate him.

1:13:38>> That man [laughter] Daario is

1:13:40>> They're like, "Dario, shut the up."

1:13:41>> Honestly, I I've been saying Dario, shut

1:13:44the up for years. But it's But the

1:13:46thing is, I get your point where it's

1:13:47like I don't think they've changed to

1:13:49calm the public down so much as they're

1:13:51desperate to not get regulated, which is

1:13:53laughable. We don't regulate tech. We

1:13:55don't regulate America doesn't

1:13:57regulate We are in the We are

1:14:00still trapped in the hands of Milton

1:14:02Freriedman, Margaret Thatcher, and

1:14:04Ronald Reagan. We're still stuck

1:14:06in the neoliberalistic hellscape, which

1:14:09is growth at all cost, free market

1:14:11capitalism. So, no, no one's regulating

1:14:13the regulation of these companies should

1:14:15have been, I don't know, breaking up.

1:14:17Put these bastards to the side. Break up

1:14:19these for sure. We shouldn't

1:14:20have companies this big. It makes things

1:14:22worse.

1:14:22>> But these technologies are dangerous.

1:14:24>> I mean, they're dangerous, but not in

1:14:26the ways they've been warning about.

1:14:27Let's if we think about cyber hacking,

1:14:30>> right? And just to be clear, those cyber

1:14:32hacking things that happened were not a

1:14:33result of they were like break out of

1:14:35the sandbox and then they set the

1:14:36sandbox up wrong. They set up the server

1:14:39they were on wrong. But I mean, you

1:14:41know, advanced AI models could very

1:14:43easily cuz they can go out onto the open

1:14:45internet as agents. They could very

1:14:47easily go and look at code bases of

1:14:48different websites, find vulnerabilities

1:14:50and exploit those vulnerabilities.

1:14:52>> Yeah. in at scale and arguably um at a

1:14:56higher intelligence and faster and wider

1:14:59than humans a human hacker could

1:15:01theoretically. So that's dangerous.

1:15:02>> Well, here's the funny thing. We don't

1:15:05know how much compute was spent to do

1:15:07the hugging face attack, the open AI

1:15:09one. We also do know that they

1:15:10improperly set up the server to keep it

1:15:12in. They thought they'd turn the

1:15:13internet off and they didn't. That's

1:15:15human error. And that's human error in a

1:15:17sense that yeah, they threw about an

1:15:19indeterminately large amount of compute.

1:15:21This is dangerous, but people keep

1:15:23saying we can't let the the Chinese get

1:15:25a hold of these models. We couldn't

1:15:27possibly because what if these models

1:15:28fall into the wrong hands? They're

1:15:30already in the wrong hands. Mark

1:15:32Zuckerberg, Sam Olman, Dario Amade. The

1:15:35wrong hands are the hands of those who

1:15:37are running these companies. We should

1:15:39not be training these models to do these

1:15:41things. I don't know why the we're

1:15:43doing it other than they've run out of

1:15:45other things they can train on. There's

1:15:46a ton. And the fact that they can do it,

1:15:48it's kind of interesting. But you do

1:15:50would you agree that it's an

1:15:52intelligence and I'll call it that you

1:15:54know you might disagree with that

1:15:55terminology but an intelligence that can

1:15:57go out onto the internet and click

1:15:59around and take actions is inherently

1:16:03there's risks associated with that. Well

1:16:06the second part I agree with the risks

1:16:08we've had people running automated

1:16:10scripts hacking scripts for a while

1:16:11we've had hackers doing that for years

1:16:12and years and years. This is brute

1:16:14forcing it with a bunch of compute and

1:16:16yet it is dangerous. These companies are

1:16:18doing something dangerous. That is not

1:16:21what Jeffrey Hinton at have been warning

1:16:23about. They've been saying, "Oh, these

1:16:25things could destroy society. They could

1:16:26manipulate people." When you actually

1:16:28look at the underlying things, not so

1:16:29much. Jeffrey Hinton as well talking his

1:16:31book still got his Google stock, I

1:16:32think. And weirdly enough, he left

1:16:34Google because he was worried about the

1:16:35AI there, but then immediately made a

1:16:37comment being like, "Yeah, actually

1:16:39though, Google's very responsible."

1:16:40Strange thing. But let's get back to the

1:16:42the cyber security side. I agree this is

1:16:44dangerous. These people should not have

1:16:46access to so much comput. They clearly

1:16:47don't know what to do with it. There's a

1:16:49really easy way of dealing with this.

1:16:51It's not letting them use so much

1:16:52compute. It's regulating that part out

1:16:54of existence. What if the Chinese do it?

1:16:57The Chinese were able to distill the

1:16:58models. And also,

1:17:01I don't know, regulate it and stop I I

1:17:04feel like with this particular thing as

1:17:06well, we got to this point and let the

1:17:09genie out of the bottle to use an

1:17:11annoying Samman term. We let this happen

1:17:14because we let these companies be

1:17:15unregulated and use as much computers we

1:17:17want. We had these enablers

1:17:19allowing them to burn as much computers

1:17:20as they want. And also we for all of

1:17:24these dire warnings about AI dangers, no

1:17:26one seems to have done anything.

1:17:28>> Okay, we're going to play a game, Ed.

1:17:29>> Let's play it.

1:17:30>> On these cards here,

1:17:31>> I have the things that you consider to

1:17:33be myths about the AI industry.

1:17:37>> The challenge is I want you to give me

1:17:39one sentence.

1:17:40on each myth.

1:17:42>> Oh, Christ.

1:17:43>> So, just your first reaction. You're

1:17:44going to pick it up, you're going to

1:17:45read it,

1:17:45>> and then you're going to give me one

1:17:46sentence on your opinion of that

1:17:49>> um belief.

1:17:50>> Okay, let's go.

1:17:51>> So, let's do this.

1:17:56>> What does it say in your says the the AI

1:17:59industry is creating enormous economic

1:18:01growth?

1:18:02>> No, it's not. It's nowhere in the data.

1:18:05>> Okay. [laughter] Like, it's just May I

1:18:07do a second sentence?

1:18:08>> Go ahead. pretty much all of the

1:18:10economics is either Nvidia feeding money

1:18:12to it companies like Corewave or these

1:18:14three companies feeding money to these

1:18:16ones to spend it with the them.

1:18:18>> Okay. And what evidence do you have that

1:18:20there's it's not causing economic

1:18:22growth?

1:18:23>> Just to be clear, other than the spend

1:18:25on semiconductors, so the speculative

1:18:27investment in GPUs and data center

1:18:29infrastructure that's happening, but as

1:18:31far as like spend on AI goes, barely

1:18:33cracking hundred billion. And most of

1:18:35that is just these two running their

1:18:37services and paying these three

1:18:39companies, Oracle, Core, and others.

1:18:41>> But a hundred billion is a lot of money

1:18:43for a relatively new technology.

1:18:45>> Not when you've spent $300 billion in

1:18:47equity funding. And it if we're going

1:18:50with just these three, I think $600

1:18:52billion in capital expenditures.

1:18:53>> Yeah, I get that. That means it's not

1:18:55profitable. But the hundred billion is

1:18:57an expression of consumer demand

1:18:58>> when the compute is mostly driven by

1:19:00subscriptions that subsidized. No, it's

1:19:02not. When you're giving someone $20 or

1:19:04$40 for a dollar, they're going to use

1:19:06it more. If this was all on a per

1:19:08million token basis, we'd be having a

1:19:09different conversation.

1:19:10>> Okay, fair. Fine. Cool. Next one.

1:19:14>> The United States need to spend

1:19:15trillions to beat China in the AI race.

1:19:19Let's see.

1:19:21What AI race?

1:19:23That's actually That's actually my

1:19:25point. It's what AI race is there. Is it

1:19:27to make big scary LLMs? They they did

1:19:30that already without the Nvidia GPUs. By

1:19:32the way, they've got Blackwell GPUs.

1:19:34Kakashi and Jastario, two amazing

1:19:35analysts I love. They've been on this

1:19:37for years. It's like China's already had

1:19:40Nvidia GPUs that they're not meant to

1:19:41have for years. But also to do what?

1:19:43They already got the LMS. What What's

1:19:45the race to do? To make us spend more

1:19:47money than them? For us to constantly

1:19:48piss our pants worrying about China?

1:19:50Because u they won if that's the case.

1:19:53Myth number three, AI will replace all

1:19:56human jobs.

1:19:58that just isn't happening and there's no

1:20:00economic data to support it.

1:20:02>> Will it replace some jobs?

1:20:04>> I mean, it's replaced some contract

1:20:06labor that would otherwise be replaced

1:20:07with cheap labor out in the global

1:20:09south. It's a digital globalization in

1:20:11that sense, but all jobs, most jobs, a

1:20:15lot of jobs. No.

1:20:16>> What about robotics?

1:20:17>> Robotics is not what we're talking

1:20:19about. Robotics is a very different

1:20:20thing. And even then,

1:20:21>> robotics will be powered by AI.

1:20:23>> I mean, yes, but there are tons of

1:20:24different kinds of AI. We're talking

1:20:26explicitly about generative AI. And

1:20:27that's what I this mythbusters piece

1:20:29that was definitely about generative AI.

1:20:31>> Okay. But what about robotics? Like the

1:20:33thing is the Optimus robot that Elon's

1:20:35working on at Tesla.

1:20:36>> The one where even in the demo of the

1:20:39hand he like they had to have a guy

1:20:41controlling it. Wasn't doing it

1:20:42autonomously. Here's the thing. If they

1:20:44can beat all these challenges, yeah,

1:20:46robotics would be really cool. I don't

1:20:48know how long that's that's one I'd

1:20:50actually be willing to believe in a

1:20:52couple decades.

1:20:53>> Have you seen them ch them Chinese

1:20:55robots? I know you've seen them. the

1:20:56uni, what's it called? The one that can

1:20:58dance and that, but they can't really do

1:21:00human things.

1:21:01>> Well, it's just it is pretty

1:21:02mindblowing.

1:21:04>> Robotics are cool. I like I'm

1:21:06not going to pretend. I don't think

1:21:07robots are cool. I wish they were

1:21:09building robots and actually doing cool

1:21:11I wish the tech industry still

1:21:12made fun stuff and interesting stuff.

1:21:14Instead, we get these large

1:21:16language models. But with AI plus

1:21:18robotics is, you know, I was in San

1:21:20Francisco and I went to this massive um

1:21:22incubator there. And when I'd gone there

1:21:24three years earlier, it was all software

1:21:26startups, right? And when I went back

1:21:27three years later, it was all these

1:21:29robot startups. And I remember saying to

1:21:30the founder of the incubator, I was

1:21:32like, "Why is everything robots now?"

1:21:34There was this one robot where it was

1:21:35just the arm and it had a frying pan on

1:21:37it. Yeah.

1:21:38>> And it whole thing is it cooks for you.

1:21:39>> Yeah.

1:21:40>> So it was he was showing me it cooking

1:21:41whatever. And he goes, "Well, you know

1:21:43the arm." He goes, "The the hardware

1:21:45part, the physical parts,

1:21:47>> that's always been fairly cheap." Yeah.

1:21:48>> He goes, "The expensive part was the

1:21:50intelligence. And now that's come down

1:21:52to pennies." So what you're seeing is

1:21:53this explosion in the robotics industry

1:21:55because robotics is a function of

1:21:57intelligence plus hardware. We've always

1:21:58had the

1:21:59>> and a ton of data though as well and the

1:22:00data is very expensive.

1:22:02>> Yeah.

1:22:03>> The thing is cyber cabs rolled out real

1:22:05slow. It's going to take a long time. It

1:22:08could be a threat if they do a robot

1:22:10that could replace a human job. Sure it

1:22:12could. But that human jobs are

1:22:13multifaceted. Human jobs change with

1:22:15environments. And also a lot of human

1:22:17jobs that you might think of like I

1:22:19don't know dishwashing robot for

1:22:21example.

1:22:21>> Yeah.

1:22:22some guy at a restaurant isn't paying 10

1:22:2420 grand for a robot to replace the job

1:22:26that they're already not paying enough

1:22:28for. The point is, yeah, it could if you

1:22:31can replace the jobs. That is not what

1:22:33we're talking about with this.

1:22:34>> Yeah. I I just I just I ask these

1:22:36questions not because I'm trying to be

1:22:38like I actually I'm trying to form my

1:22:39own opinion on these things and

1:22:42>> I I do think, you know, as it's written

1:22:45there, it says AI will replace all human

1:22:48jobs. Obviously not. Obviously, that's

1:22:49Yeah.

1:22:50>> But um I'm trying to figure out if the

1:22:51truth is somewhere in the middle that

1:22:53there's a certain type of job which

1:22:55actually humans probably shouldn't have

1:22:57ever been doing really.

1:22:58>> Um if you think back through history,

1:23:00there was someone's job just to sit in

1:23:01an elevator and press the buttons.

1:23:02>> That's an example of a job that humans

1:23:04probably shouldn't have been doing. And

1:23:05as technology gets more advanced, it

1:23:07takes on a lot of that

1:23:09>> sort of automated monotonous stuff.

1:23:11>> Right? The thing is with this particular

1:23:14thing that I know that this is from,

1:23:15it's a specific blog I wrote. I was

1:23:17explicitly talking about generative AI

1:23:18though. I was explicitly [clears throat]

1:23:20talking about people when they say this

1:23:22they are referring to that.

1:23:23>> So you're not talking about agentic AI

1:23:24which is

1:23:25>> agentic AI is LLMs. Agentic AI is just a

1:23:27fancy way of saying an LLM talking to

1:23:29another LLM with a harness on top. That

1:23:31is still LLM. Agentic AI is one of the

1:23:34big the bigger lies they to tell. It's

1:23:35like when you hear agent you're meant to

1:23:37think autonomous AI can do what you

1:23:38want. It's still LLMs. It's still LM

1:23:40talking to other LMLs

1:23:42>> taking screenshots and putting them in

1:23:44LLM and stuff.

1:23:44>> Oh god. Yeah.

1:23:45>> Okay. But but you know I could I could

1:23:47make the case that

1:23:49I'm just thinking about my personal

1:23:51usage. I definitely use agents to do

1:23:54things that I would have previously

1:23:55asked people to do. It's not to say that

1:23:56I didn't I still don't hire cuz we're

1:23:57hiring like crazy.

1:23:58>> Yeah.

1:23:59>> And I still in that particular function.

1:24:00I'm thinking about like the chief of

1:24:02staff role. So my chief of staff would

1:24:04have triaged all of my inboxes

1:24:06previously and put them somewhere and

1:24:08told me about them or maybe once upon a

1:24:09time shown me a piece of paper back in

1:24:11the day. I guess now my chief of staff

1:24:13is no longer doing that job. You still

1:24:14have a chief of staff though.

1:24:16>> This is what I'm saying. They're doing

1:24:17other things,

1:24:18>> right? But the thing is again what you

1:24:20were describing is

1:24:22fairly basic automation. I don't know

1:24:23what the tasks are triaging.

1:24:25>> Basic spend a trillion dollars on

1:24:27triaging email. Like that's the the

1:24:29promise. If they'd spent $10 billion and

1:24:31this was much smaller and you I go cool

1:24:33software. Yay. A lot of the things that

1:24:35people are impressed with like script

1:24:36stuff as well. It's just LM's doing

1:24:38Python. You should be impressed by

1:24:39Python code. Python's incredible. You

1:24:41can scrape websites. You can download

1:24:43It's awesome. But the point I'm

1:24:45making is none of this would be anywhere

1:24:47near as much of a problem if they didn't

1:24:50ask for all of the attention, all of the

1:24:51money, and promise the world. It's their