Based Camp - Is AI Overhyped？ Is AI a Bubble？ (New Models Don't ＂Feel＂ That Much Better, Why？)

00:00:00.000 Hello, Simone. I'm excited to be talking to you today. Today, we are going to be talking about

00:00:03.720 whether AI is hype, whether AI is plateauing, whether AI is over. And by this, what I mean,

00:00:11.260 because I had the head of our development team, Bruno, who comments a lot on the Discord in the

00:00:17.380 comments here, so fans of the show will know him. He sent me an email that we're going to go over

00:00:21.500 as part of this, sort of being like, okay, so here's some evidence that AI doesn't seem to be

00:00:26.640 making the systems level changes in society that you had predicted it would make in the past and

00:00:33.240 that many other people are predicting it will make. And when I, and we're seeing other people say this,

00:00:38.140 when I go out and I interact with AI today, I really struggle to see how having this thing I can chat

00:00:44.720 with is that useful. It may be fun as like a chat bot or something like that, but I don't see its

00:00:50.420 wider utility yet. Now, we'll be going into the arguments around this, because I think that there

00:00:56.780 are some strong arguments, like the AI industry is making almost no money right now. You know,

00:01:01.060 is this industry not, not almost, but almost no in contrast, it was the investment that's going into

00:01:05.780 it. And the amount that we talk about it and other people talk about it as mattering. And then you've

00:01:10.720 got to think about all of this in the context of, yeah, but like 80,000, oh, sorry, around 90,000

00:01:16.580 people just in the U S tech sector had their jobs cut due to AI this year, you know? Yeah. So come

00:01:21.060 on. Do you not matter to them? Um, but, but so what we're going to be seeing here is I think the

00:01:28.940 way that the people are looking at AI and expecting AI to transform the economy is different from the

00:01:35.420 way it actually is. They're looking at how AI is useful to them instead of how AI will replace

00:01:42.860 them. I'd also note here to the question, because Sam Altman, you know, literally like

00:01:48.640 Sam Altman, who one of the, the guy who runs one of the largest AI companies has said AI is a bubble

00:01:54.300 right now. Right. And so people will come to me and they'll be like, well, you know, even he's saying

00:01:58.180 it's a bubble. And I'm like, I would say it's a bubble right now. It is a bubble. It's obviously

00:02:02.820 a bubble right now, but the fact that a thing is a bubble doesn't mean it's not going to transform

00:02:06.820 society. So if you go to the.com boom, right? Like the.com boom was a bubble, right? But the

00:02:14.980 internet still transformed society. The companies at the beginning of the.com boom, you know, like

00:02:21.240 they were formed in the middle, like Amazon and Google and stuff like that. Like if you made the

00:02:26.020 right bets on those companies, if, if anything, what, like, if you wanted to make the best bets

00:02:32.200 possible, wait for the AI bust and then invest in whatever survives. If there is a traditional

00:02:38.320 bust, you know, keep in mind, like what, what I mean by a boom now is a lot of people are

00:02:41.420 investing in AI companies without understanding in the same way, like in the early.com boom,

00:02:45.980 what the internet's actually good for and good at.

00:02:48.440 Well, what kind of sucks is, is also the AI companies that I think are coming out of this.

00:02:54.120 They're not going to be hopefully traded a big shift in AI tech booms. As far as I see is

00:03:00.460 that they're not something you're going to see in a stock market. They're small. They

00:03:04.540 don't have a lot of staff. They're not public. So our ability to participate in the upside

00:03:09.480 is severely limited.

00:03:10.760 Well, the other thing about AI development and you can, you know, back me on this is we

00:03:15.860 can see all these metrics that say that AI is supposedly getting better and smarter. And

00:03:20.180 yet when you consider like the latest model of Grok versus the last model of Grok, you don't,

00:03:25.080 you don't go like, this is like 50% better. Like it doesn't feel that way to you. Same

00:03:29.620 with open AI's model. Same with a lot of these models. You, you interact with the most cutting

00:03:33.220 edge thing. And you're like, this is marginally like three or 4% better, but all the metrics

00:03:38.140 are showing that it's like massively better. So is this a problem in how we develop AI,

00:03:42.460 how we measure AI, everything like that. I'm going to be talking about that in this as well.

00:03:45.700 I'm also going to be talking about the study that Facebook put out saying that you can't

00:03:50.160 get, basically they're saying AI really isn't that smart. No, no, sorry, not Facebook.

00:03:54.560 I want to say Apple maybe put this out, but if it was Apple, it shows why Apple has not

00:03:59.580 developed anything good in the AI space because the people they have working on it are just not

00:04:03.680 that bright. I do have to say though, I'm really excited about their smart house play. I think

00:04:08.400 that if they're, they're going to have any win, them being the ones that make everything that's AI

00:04:14.240 in your smart home connected and work seamlessly and be really pretty, they are going to be the ones

00:04:19.180 that are capable of pulling that off. So we're going to go into an article, start,

00:04:23.440 start by going into an article in future magazine. So you can see that this isn't just Bruno making

00:04:27.480 these claims. And the article was titled scientists are getting seriously worried that they've already

00:04:32.540 hit peak AI. Speaking to the New Yorker, Gary Marcus, a neuroscientist and longtime critic of open AI

00:04:39.100 said what many have been coming to suspect. Despite many years of development at a staggering cost,

00:04:44.760 AI doesn't seem to be getting much better throughout GP though. GPT technically performs

00:04:50.060 better on AI industry benchmarks and already unreliable measure of project. As experts have

00:04:54.700 argued, the critic argues that it's used beyond anything other than virtual chat buddy remains

00:05:00.360 unlikely. Worse yet, the rate at which new models are growing against the dubious benchmarks appears to

00:05:07.200 be slowing down. I don't hear a lot of companies using AI saying 2025 models are a lot more useful to

00:05:13.860 them than the 2024 models, even though the 2025 models perform better on a bunch of benchmarks,

00:05:19.440 Marcus told the magazine. Now here I note here when he's like, I don't see AI being used for anything

00:05:25.220 other than one of the things we're going to get into later in this episode as a chat bot, I'm going

00:05:29.340 to be like, well, then it's just because you're not familiar with any industry other than your own.

00:05:33.820 AI has already invented drugs that are going through the production cycle that are likely to save millions

00:05:39.020 of lives. And not just like one drug, like countless drugs at this point. AI models that

00:05:43.640 are trained on the human genome have already made numerous genetic discoveries. It's the way that

00:05:49.960 you're using AI. And we'll go through these discoveries in a bit. It almost to me feels like

00:05:55.340 with some of these people, like that Monty Python sketch, you know, like what have the Romans ever done

00:06:00.180 for us? What have AI ever done for us? And it's like, well, okay, yes, they invent lots of drugs.

00:06:05.640 And yes, they help with drug discovery. And yes, they help with, but apart from the sanitation,

00:06:09.720 the aqueduct and the road, irrigation, medicine, education. Yeah. Yeah. All right. Fair enough.

00:06:15.800 And the wine. Yeah. All right. But apart from the sanitation, the medicine, education, wine,

00:06:23.120 public order, irrigation, roads, a freshwater system of public health. What have the Romans ever done for us? 0.56

00:06:29.560 Oh, peace. Oh, peace. Actually a fun thing here. Just if you're like, how, how is a way that I'm

00:06:37.800 not thinking about using AI? I'll talk about really quickly the way I use AI in running my company.

00:06:41.980 So, and Brutus is actually the one who implements this is whenever we assign a task to the, one of

00:06:48.440 the programmers, we ask an AI about how long this task should take to complete. And then we benchmark how

00:06:55.300 long it takes them to complete the task against how long the AI thinks it will complete the task.

00:06:59.560 And we can create weighted averages to see sort of how productive a person is. Obviously this isn't

00:07:06.280 going to be perfect and AI is going to overestimate a lot, but it does create a relatively accurate

00:07:11.160 benchmark that we can use to normalize who are the best performers on our team, which is very

00:07:17.980 interesting. And I haven't heard people using AI in this way. And this is from Bruno's email.

00:07:23.600 Ed Zitron has focused heavily on the financial side. He knows OpenAI's current annualized revenues

00:07:28.980 sits around 5.26 billion and Anthropix at around 1.5 billion, but expenses remained outsized. OpenAI

00:07:36.940 alone may be spending roughly 6 billion annually on servers, 1.5 billion on staff and another 3 billion

00:07:44.740 on training. Investors have been asked to sign acknowledgements that profitability may never

00:07:50.660 materialize. Against that backdrop, valuations in the hundreds of billions look speculative at best.

00:07:56.320 So this is really important. 5.2 billion and 1.5 billion for two of the major AI companies are

00:08:05.640 laughably small in terms of what they're making. How are they getting to hundreds of billions in

00:08:14.140 valuation, right? Like why aren't they making more money? And why is this true sort of across the

00:08:19.740 industry? Because we are seeing this across the industry. Before I go further, I'm going to explain

00:08:23.460 this phenomenon because this is actually an important phenomenon. The reason why they're

00:08:29.000 making so little actual money is because their long-term potential is so high.

00:08:35.960 Well, this is how it's pretty much for every tech startup over the past 20 plus years.

00:08:41.520 Yeah. Yeah. So for people who understand how VC works, so VC comes in, it floods the space that it

00:08:46.660 thinks is going to be worth a lot in the future. And then because it's flooding the space with so

00:08:50.780 much money, like because OpenAI and Anthropic are getting so much money to compete against each

00:08:55.060 other, they have to be rock bottom in terms of the prices they're offering or even offer things for

00:09:01.900 essentially free versus the cost to produce them. And people can be like, wait, but then why would

00:09:08.820 they even like do that? Right? Like they're trying to win in a market where you, the customer are not

00:09:15.640 actually the primary customer, where the venture capitalist is actually the primary customer and where

00:09:21.940 it doesn't like part of, for them, you as a user, you are providing them value. You know how like with

00:09:28.940 Google, you provide them value as a user because they get like data from you that they sell to other

00:09:33.720 people and they get like ads from you, but like the data from you is more important. Like some

00:09:38.600 companies like make their money off of the data they collect from you. Okay. You as a user are

00:09:44.040 actually a data point that these companies are able to trade for cash from venture capitalists.

00:09:51.800 That is why you actually like, it's, it's actually a fairer deal. It's not like they're cheating you or

00:09:56.880 you're cheating them. They are trading the fact that you are using them to say, look, I am beating out the

00:10:03.400 other major companies. Hey, you investors know eventually somebody is going to eat this industry.

00:10:09.780 Okay. So that's, that's important to know. Like this is something you'd actually expect if things

00:10:14.920 are going well. So to go back to his email here and none of this is a stupid thing to know. Like 0.98

00:10:20.920 I'm not like saying Bruno was stupid for, for asking this question, right? Like it's easy to add. 0.99

00:10:26.080 Why is it valued so much when it's making so little? Why haven't tons of profits accumulated in this

00:10:30.700 industry yet? To go back to his email for context, compare this with Amazon web services, AWS launched

00:10:36.360 in 2006, reached cost revenue parity in just three years. And in its first decade, accumulated roughly

00:10:42.240 70 billion in costs. By contrast, Amazon itself spent around 105 billion in just the last year on

00:10:49.420 Zitron. So, so AI, the biggest company in the space is making 5.26 billion a year. Okay. And he's

00:10:58.040 pointing out here that in its first decade, Amazon web services made 70 billion. And then he points

00:11:02.860 out that what Amazon has spent on its AI has been more than what Amazon web services has made in the

00:11:08.880 last decade, 105 billion. Right? Zitron underscores that the entire generative AI field, including open

00:11:17.880 AI, Anthropic, Midjourney, and others produces only about 39 billion in revenue. That's less than half the

00:11:25.140 size of the mobile gaming market and slightly above the smartwatch market at 32 billion. These

00:11:30.940 comparisons illustrate the scale mismatch between AI valuations and demonstrated ability to generate

00:11:36.280 revenue. And so, so this is a really apt comparison. It's bringing in about as much as the smartwatch

00:11:42.220 market. That's yeah. Wow. Putting that into perspective, that's pretty sobering. The smartwatches

00:11:48.300 are so pervasive now, maybe not as sobering as you might initially think.

00:11:53.620 Yeah. Well, I mean, it should be more given how afraid people are of it, how much people are talking

00:11:58.940 about it. Right. There is also the issue of market making a product market fit. LLM based tools are not

00:12:08.480 meaningfully differentiated from one another. The average user tries one, thinks this is kind of cool,

00:12:13.920 and then stops using it. This raises a concrete question. How would one sell these products in a

00:12:20.160 way that justifies ongoing subscription fees? How do you... Did they really stop using them? Like

00:12:24.760 here again is where I questioned it. And also like the smartwatch market... No, no, no. This is

00:12:29.320 factually wrong. If we look at usership rates, they are shockingly high. Okay. Okay. Cause I'm just

00:12:35.820 like, you're not questioning this and I'm like, ah, wait a second. Like... No, but I understand how

00:12:41.860 somebody could feel that way if they're just thinking about like, especially if AI hasn't caught

00:12:46.980 you or you haven't found product market fit for AI within your life. Yeah. You're just going to walk

00:12:53.100 away from it. Right. You're going to be like, what's the point? Right. Yeah. It's also really

00:12:56.660 unfair to compare this to the smartwatch market as it is today, because the smartwatch market is in

00:13:01.360 it's now we make money from this era. So a lot of people wear Oura rings, right? When they first came

00:13:07.620 out, you didn't pay a monthly subscription for them. Now you have to, you can't have one without

00:13:13.280 it. I wear a Fitbit. I don't pay a monthly subscription for it, but I'm constantly upsold

00:13:17.620 on it. So now they've switched into the monetization phase, but at the beginning, it was just no,

00:13:22.420 get this on people's bodies and then try to make money from them. And right now AI is in the get this

00:13:27.960 in people's workflows and lifestyles phase. So of course it's not making that much money.

00:13:33.940 Yeah. Well, and, and because that's what VCs are trying to do, because they're trying to capture

00:13:37.780 the market and then make the money by getting the giant company. Yeah. They're, they're not being

00:13:41.560 foolish about this. Like this is actually makes economic sense. Yeah. I mean, it also happened 0.92

00:13:45.840 with Uber and Lyft. They used to pay drivers really well, only have really nice cars and have really,

00:13:50.360 like, they were probably running at a loss. They charged so little. They were running at a huge loss

00:13:54.540 for a really long. Yeah. And so like we had this, this, this generation, which was so nice where you had

00:13:58.940 this like VC subsidized luxury lifestyle where you had like super affordable food delivery and

00:14:05.220 Ubers and smartwatches, but that was because of the growth phase. And right now we're enjoying

00:14:10.240 this short period where we don't have to pay a lot for these AI services.

00:14:13.260 Yeah. I'm actually going to argue that the VCs here might be making a mistake

00:14:17.100 and the mistake that they might be making. And this, this depends on stickiness to particular AI models

00:14:23.800 is they think that they're developing something like the next search engine or the next Uber or Lyft

00:14:28.980 when what they're actually developing is a commodity. By that, what I mean, and I think the way that most

00:14:33.840 people are going to interact with the very best deployments of AI are going to be through skins,

00:14:40.540 basically through windows. Like what we're building was our fab AI, right?

00:14:43.880 Oh, and yeah, people aren't necessarily going to be loyal to Jot, GPT, or Grok. They're going to use 0.76

00:14:50.880 a variety of different services that are going to interchange Grok and ChatGPT based on whichever

00:14:56.180 is best and cheapest at the time. Yeah. Kind of how people switch out, at least in the United States,

00:15:00.420 in many places, you can switch out the utility company that you buy from. So you can buy from

00:15:04.120 this utility company, or you can buy from this one that only does green energy or whatever. And you

00:15:07.800 choose whichever you like based on values and based on price. Yeah. And we actually already see this in

00:15:14.480 the data. People switch between like which AI is the top used one changes a lot, which one is sort of

00:15:19.660 known as the best one changes a lot. And you're, you know, if, for example, like right now, the main

00:15:26.380 AI I use is Grok, right? If OpenAI had a model that I just thought was dramatically better, I would

00:15:32.080 switch to them, which has occasionally happened. Claude used to be my main AI, then OpenAI was my

00:15:36.880 main AI for a while. No, it's the Grok, and I may switch again, right? Like I switch all the time

00:15:42.080 what the primary AI I use is. Now, another argument he made in this email that I thought was really

00:15:49.660 interesting is he's like, the iPod, it's a thousand songs in your pocket. The iPhone,

00:15:56.020 it's the internet in your pocket. Like what is AI in your pocket, right? And I think that this

00:16:03.220 actually is a big mistake. And it's one of the reasons why people are not understanding the value

00:16:07.760 of AI. They're thinking about AI's value to them, not AI's value to say genetic researchers or certain

00:16:16.540 groups of programmers, et cetera, right? Like if AI has replaced a job, we often talk about AI can

00:16:22.520 probably replace, I guess, 25% of law clerks now. Now what it hasn't happened yet, but it definitely

00:16:28.540 has the capacity to do that. And you're like, well, what if it makes mistakes? And I go, then just put it

00:16:33.040 in a chain so that it checks for those mistakes. This is one of the reasons when people are like,

00:16:37.500 well, what about hallucinations? And I'm like, hallucinations, like literally don't matter.

00:16:42.280 First, they don't happen that much in current model AIs. I argue somebody was like, oh, well,

00:16:47.920 I don't trust you guys because you get some of your information from AIs as a fan. And I was like,

00:16:52.960 excuse me, Broseph, do you think that the average thing you read from a reporter is going to be more

00:16:57.360 accurate than the average thing you read from an AI? Yeah. At this point, it's not hallucinations. It's

00:17:01.680 it's sourcing that like from sources that get it wrong. And we check our sources, like we check

00:17:07.740 the sources that AI cites, but we can't always take the time to figure out how reliable those

00:17:13.280 sources are. We're like, well, the New York Times reported about it. Like, right. But the point I'm

00:17:18.180 making here is that inaccurate information is more likely for me, like twisted information is more

00:17:25.040 likely to come from a New York Times article than a Grok 4 output. And I would bet that this is

00:17:30.260 something you could even look at statistically, right? Because these, these, it's not that there

00:17:35.260 isn't like a political bias within AI. It's just less extreme and distinct than the political bias

00:17:41.320 within the reporter class. So, and, and then the, the amount you can reduce hallucinations just by

00:17:46.740 doing one pass. By that, what I mean is you have the AI output and answer, then you take that answer,

00:17:52.040 you put it in to a different AI with the question being like, is anything in this hallucinated or wrong?

00:17:57.460 This is an output from another AI. You do that in the probability I'd argue is like 0.01 that you're

00:18:04.280 going to have a hallucination and whatever that output is, as long as you're using like high-end

00:18:08.180 AI models. But the problem with being like, where is it? You know, if you have like the iPod or the

00:18:14.460 iPhone and you're like, it's this in your pocket, what is AI to you, the end consumer? I think this

00:18:19.600 shows a misunderstanding of AI's role within the market. AI is a tool that at its most productive

00:18:26.880 replaces human beings. AI is a simulated intelligence. It's a simulated human being.

00:18:33.860 That's, that's fundamentally where it's most valuable. It's when it can replace an entire

00:18:38.040 call center. You know, that's like a hundred million jobs. If you replace that, when it can

00:18:42.840 replace an entire coder by making other coders more efficient, when it can replace legal clerks,

00:18:48.440 when it can replace. And one of the things he asked me in this email as well, you know, when,

00:18:52.880 when does the data come down where you change your mind on how much AI is going to change the

00:18:57.260 economy? Like, would you need to see things stop moving as fast in the field? Would you need to see

00:19:02.340 like hurdles begin to come up? And I'm like, even if, and we'll go over the potential hurdles to

00:19:08.240 continue to AI development. Even if one came up, even if AI development completely stopped where it

00:19:13.560 was today, most of my predictions on how much it's going to transform our society would stay there.

00:19:19.120 By that, what I mean is using multi-pass AI, you should be able to replace about 25, 35% of the

00:19:25.580 legal profession right now. And yet that hasn't happened yet. You know, you, you should be able

00:19:30.660 to complete replace 25, 35% of government bureaucrats right now. And yet that hasn't

00:19:35.740 happened. Accountants. And yet that hasn't happened. Right. Copywriters. And, and, and when I point this

00:19:42.580 out, I mean, I pointed out was in my own life, right? Like I have seen, you're like, AI does not

00:19:46.840 replace professions. And I'm like, if you are watching this podcast, you are participating in

00:19:51.080 something where AI has replaced professions because we had an earlier iteration of this podcast. If you

00:19:55.620 go back in our podcast history before base camp, where we paid for an editing team and we paid for

00:20:02.080 a card creation team for title cards. Those tests are now both done by AI. So those are, those are two

00:20:09.300 people's jobs who are now done by AI and dramatically. Well, you still, you still use software and you do

00:20:16.200 the editing and I still use, well, AI image generation and I do the title cards, but yeah, like.

00:20:22.500 AI is an increase in my capacity. I couldn't do it without AI. Yeah. No, same. I couldn't, no way. Yeah.

00:20:29.580 So I think that that's really big. And then keep in mind how much AI transforms the economy. If Elon's

00:20:35.220 move to make like robots that work with AI for like factory labor and stuff like that.

00:20:41.740 And everyone initially was like, oh, this is so silly looking, et cetera. But apparently they're

00:20:45.780 having a lot of success with this from what I've heard, like through the grapevine friend network,

00:20:50.020 stuff like that. If this works out now, it's every factory job, right? Yeah. Well, I've also read that

00:20:57.140 China is investing heavily in AI enabled hardware as well. So things like robots. So if, you know,

00:21:05.020 it's not like only one person is trying this, plus also Boston Dynamics has been at this forever.

00:21:10.000 People are, they're going to be major players and absolutely there's going to be a physical

00:21:13.640 element to this. But Boston Dynamics, I actually feel like this is very much like drones come out,

00:21:19.260 right? And everyone's like, well, that's cool, but that's just a toy. Oh yeah. Now like it's

00:21:24.760 completely changed to warfare. Yeah. And now everyone's like, oh gosh, tanks don't work anymore

00:21:29.860 with this new model of warfare and large ships. And we need to completely change the way we fight

00:21:35.360 wars. Like drones were a toy until they weren't right. Yeah. Yeah. Very much the same with AI and

00:21:42.760 where things are going. Yes, absolutely. And I'd also point out here when you're like, well, okay,

00:21:50.420 but what other industries could AI disrupt other than genetics and science and drug development and

00:21:56.560 copywriting? And well, a big one is my cousin owns the company that created the movie here,

00:22:02.200 which takes Tom Hanks and then puts him at different ages and in different environments.

00:22:05.520 And they're using AI to do this.

00:22:07.320 This time is time and time with your time and his news is captured.

00:22:17.100 And if you look at the, they'll do viral stunts all the time where they'll like create

00:22:20.580 TikTok reels of like the various celebrities and we'll get like millions of views, but it's faked,

00:22:26.220 is faked with their faces. If you can simulate an actor, that's a bunch of industries that you've

00:22:32.500 just nuked. Right? Yeah. And I mean, this is already, I mean, yeah, it's, I was just listening

00:22:39.780 to a podcast on how acting has been disrupted already in that now production companies are more

00:22:48.420 making money off of the IP and the concept rather than the actors, which is why we see so many more

00:22:53.020 actors have side hustles and create companies and start investing and do all these commercials and

00:22:59.740 have a clothing line or a phone company because yeah, this, this whole industry has changed.

00:23:05.300 So I think we also aren't, aren't recognizing how much many industries have already fundamentally

00:23:10.180 changed with only the beginnings of tech enabled industry shifts away from like key man risk and

00:23:17.860 key man risk being defined as any company that kind of depends on unreliable humans for its financial

00:23:24.100 wellbeing. Um, people have been trying to use tech to render key man risk obsolete for a very long

00:23:29.800 time. And AI really handles that well. Yeah. So I'll note here, I don't know. We're going to get,

00:23:35.580 well, a key man risk was in movies and stuff like that. With, with, with tons of other things too,

00:23:40.180 though. I mean, we don't know if it can do it with competence yet, but I mean, keep in mind what we'll be

00:23:45.200 going over the AI that came in second place in that coding competition and stuff like that. Like

00:23:49.000 AI can clearly handle very advanced tasks. Yeah. But one of the things that's often hidden if you're

00:23:54.560 colloquially using AI is how rapid recently the adoption of AI within corporations has been.

00:24:00.860 So if you look at, and I'm going to put a graph on screen here, AI usage at work continues a remarkable

00:24:07.600 growth trajectory in the past 12 months alone. So this is for 1-1-2025. So like

00:24:14.900 recently, right. Usage has increased 4.6 X. That's in 12 months, 460% increase in usage.

00:24:26.260 And over the past 24 months, AI usage has grown as an astounding 61 X, 61 X in growth in usage.

00:24:37.640 This represents one of the fastest adoption rates for any workplace technology, substantially

00:24:43.580 outpacing even SAS adoption, which took years to achieve the similar penetration levels.

00:24:50.160 Now we're going to go over another graph here. This is showing AI development. This isn't actual

00:24:53.960 usage of AI, but this is AI medical devices approved by the FDA. So you see it is shooting up. Now,

00:25:01.500 unfortunately it only goes to 2023, but I doubt this trajectory has flown down some.

00:25:06.360 Yeah. But I want to look at now a few metrics where we're looking at like adoption within companies.

00:25:14.720 So if you look at organizational AI adoption, and this is from the Stanford AI index in 2023,

00:25:20.940 55% of industries had done it. And by 2025, it had jumped to 71%. Now note here, we're reaching saturation

00:25:29.440 on adoption by many points of AI, which is a potential problem, but we'll talk about how much AI is.

00:25:35.940 I think what you're seeing is people are adopting it, but they still don't really know how to use it

00:25:39.700 yet. Right? Like if I say AI won a coding competition, people are like, wait, how could I ever get AI to

00:25:46.040 code like that for me? Right? And I'm actually sending you the model that they use. They used a sort of

00:25:52.900 chained model. And the way the chained model worked is it had multiple models in an engine

00:25:59.280 where you would have one model that asked it to plan what it did next. Then a model that asked it

00:26:03.760 to code based on it planning. Then a model that would evaluate the code that it had just created.

00:26:09.360 Then a model that attempted to improve the code it just created. Then a model that revealed all that

00:26:14.340 planned again, moved to the next stage. And this is something that we are building. Cause if you're like,

00:26:20.940 well, how would I do this? Our fab.ai is going to allow you to build chained models like this

00:26:26.340 in a very easy way. So just wait. And you'll be able to do this yourself using multiple AI models,

00:26:32.660 very near in the future. Now the generative AI adoption. So for 2023, you have 33% of companies

00:26:39.240 using this 75% in, in this year. And this is coherent solutions trends, 2025 McKinsey. If you look at

00:26:46.860 the AI user base, we go from a 100 million active user base in 2023 to 378 million users globally.

00:26:55.460 This is Forbes 2025. If you look at job impacts, there were no reported job impacts in 2023.

00:27:01.920 And in 2024, it looks like 16 million people likely had their jobs automated by AI. And in this year,

00:27:09.620 it looks like 85 million jobs will be replaced. And this is from demand sage 2025 AI jobs barometer.

00:27:18.140 Now I'd also note here that people are like, well, AI has reached certain, and we're going to go over

00:27:23.740 where AI has sort of plateaued in its growth. And this is actually kind of an illusion by the way that

00:27:28.000 we're measuring AI growth. But one of the things that we've actually been seeing is significant

00:27:32.580 advancements to the actual underlying model, which leads to jumps in growth within some area.

00:27:38.800 While I will not say I was wrong about AI, because I don't think I was, where I will

00:27:43.460 admit I was wrong was about DeepSeek not mattering. DeepSeek has been very diligent in publishing

00:27:50.340 how they do stuff. Like, like, despite being a Chinese company, they've been very sort of open

00:27:55.100 source in how their new model works. So we understand how they basically reinvented the

00:28:00.240 transformer model in a way that has a lot of advantages. I mean, this is something that's just

00:28:04.520 been having significant bumps even over this last year. So to go over this, they invented

00:28:11.240 something called multi-head latent attention, MLA. MLA is a modified attention mechanism designed

00:28:17.820 to compress the KV cache without sacrificing model performance. In a traditional multi-head

00:28:23.060 attention for the original transformer architecture, the KV cache grows linearly with sequence links

00:28:29.480 and model size, limiting scalability for long context tasks, e.g. processing 100k tokens. MLA

00:28:36.000 introduces low rank compression across all attention heads to reduce this overhead, making inference

00:28:41.480 more efficient while maintaining or even improving training dynamics. It basically makes training way

00:28:47.600 cheaper and is how they achieved what they achieved. Now I'll show here another graph on screen for

00:28:52.860 people who don't think that we're making advancements. This goes only from 2022 to 2024.

00:28:59.380 Okay. So keep in mind, this is not like a, I'm going distantly into the past to show like

00:29:03.380 massive improvements, right? This is the smallest AI models scoring above 60% on the MMLU 2024 to 25.

00:29:13.160 And you can see here now we're at 5.3 mini, but what's really cool here is when you see the big jump,

00:29:18.680 this happened in late 2023. This was with Mistral 7b. With RFR AI, I've actually found Mistral 7b is

00:29:24.860 astoundingly good given how inexpensive it is to use. We might be able to sort of chain the Mistral

00:29:30.300 B model. I'm thinking to get responses that are near the quality of Grok 4, even though it costs

00:29:36.920 150th to run. Wow. So yeah, very fun to see how we might be able to attempt that. Now let's look at

00:29:45.640 how many employees, because I want to keep this all very recent so people can see like, this is,

00:29:50.700 this is, this is actually happening today. So I'm putting a graph on screen here, which is how many

00:29:56.640 employees use AI tools, contrasting 2024 with 2025. In financial services, just in the last year,

00:30:02.880 it went from 4.7% to 26.2%. Ooh. In healthcare, 2.3% to 11.8%. In manufacturing, 0.6% to 12%.

00:30:14.520 Where you see big ones retail, 1.1% to 26.4%. Oh, wow. And you can look at the others here,

00:30:22.380 but it's, it's huge, right? So now I'm going to put up a graph on screen here of different types of AI

00:30:29.200 tasks and how they have jumped in them. This is from the Stanford index, select AI index of technical

00:30:35.740 performance benchmarks versus human performance. And what you will notice here is where human

00:30:40.860 performance is a hundred percent mark, AI has been shooting up in their proficiency across the board,

00:30:45.740 but you also notice here that it appears that AI gets really dumb after it passes the human benchmark.

00:30:50.740 Like it stops going up as quickly. And then here we have AI benchmarks have rapidly saturated over

00:30:57.080 time. So here we have a number of different AI benchmarks and you can see they all sort of

00:31:01.200 taper off after a human. And this creates an illusion for a lot of people that once AI gets

00:31:07.460 smarter than a human, it stops getting smarter after that. And what's actually happening is the

00:31:12.740 benchmarks that we creating are saturated because we didn't have to deal with entities this smart.

00:31:19.340 And humans are unfortunately very bad at telling when an entity is significantly smarter than them.

00:31:24.340 Where you can see this really loudly is our open AI. Oh, by the way, any thoughts? I've been

00:31:30.880 just rattling here, Simone. No, I'm really enjoying this, but also I've had trouble comprehending why

00:31:39.560 people think AI is plateauing. So. Well, I mean, I do think the perception, like the current model of

00:31:47.880 opening AI I'm using doesn't feel that much better than the previous models. And in some cases,

00:31:52.520 even worse, right? I can understand that. I can understand somebody being like, what do you mean

00:31:58.540 this is like 50 or 60% better? It feels three, 4% better. Based on how they use it. Sure. Yeah.

00:32:04.820 Based. I understand your snarky remark there. You're accurate, Simone. But I think if you want to see

00:32:09.700 where you can see this really loudly, you can see this on the difference between the special version

00:32:15.220 of ChatJPT4 and ChatJPT5 and all of the romance people. If you watch our episode on like the AI dating

00:32:21.240 people. They're so mad. They're so mad because it no longer talked to them like a dumb romance author.

00:32:27.460 It didn't put a bunch of emojis in things. None of this florid poetic language. 0.96

00:32:33.580 Yeah. You see this on the meme where people are making fun of it, where it doesn't like give a bunch

00:32:38.040 of emojis and flowery stuff when somebody gives a baby announcement. It's just like, congratulations,

00:32:42.760 have fun. Where the other one used to do like, you're going to have a welcome to the bipedal

00:32:47.940 moment. Like you're going to have a little one running around. Oh my gosh. But really,

00:32:52.520 I'm so excited for you. But basically it was acting like an idiot. But unfortunately your average 1.00

00:32:58.620 person's intelligent level, your average, it capped out at GPT 4.5. And so when AI became smarter and

00:33:06.560 more sophisticated and more, I mean, sophisticated, that's the word. When it became more intellectually

00:33:12.280 sophisticated and understood that this is not an appropriate way to communicate with your normal

00:33:17.260 person, you don't send them long, lavish love poetry, right? Unless you're prompted to intentionally

00:33:23.160 be cheesy, it stopped doing that. And people freak the F out. So in many ways, one of the

00:33:30.520 phenomenons we're seeing here is people stop being able to judge how smart an AI is when the AI is

00:33:36.020 significantly smarter than that. Now to note here, how much we have saturated our benchmarks at this

00:33:41.520 point. Here, I am reading from a Substack post by Ash K. Curry or something called no AI progress is

00:33:49.660 not plateauing. And he notes here talking about one of the metrics that they were judging on.

00:33:55.220 And to their credit, they created a really difficult benchmark. When they released this benchmark,

00:33:59.960 even the smartest AI models were only able to solve 2% of the problems. This was two months ago.

00:34:05.200 In November 2024. So this post came out a little bit ago. So in two months ago, I love it. I have

00:34:10.680 to say a little bit of a go of this months ago at this point, right? So November 2024, it could solve

00:34:15.480 2% of the problems. And here's a graph of how many it could solve. Great. Except so far, with only a

00:34:21.540 two-month time difference, OpenAI announces O3. So keep in mind, this was not the O4 model yet.

00:34:26.860 Their smartest model at coding math later in December 2024. How did it do? It got 25% right.

00:34:33.680 Now, I note here that 99.9% of the population cannot solve even 1% of the problems on the frontier

00:34:44.840 mass test. Yeah, these are really difficult tests. And here we have an AI that solved 25% of it,

00:34:50.680 though. Five years ago, the state-of-the-art AI was ChatGBT2, which could sometimes write a fully

00:34:57.000 coherent paragraph in English. And if we look here, we can see another test being saturated here.

00:35:03.220 This is ARC AGI semi-private V1 scores over time. And you can see we went from like basically getting

00:35:09.920 none of it right with GPT4 in 2023 to 2025. But when I say none of it, I mean, it's getting like

00:35:15.840 two to 3%. It's getting near 100%. So they have to shut it down and create a new test.

00:35:21.920 Yeah. And the benchmark used to just be, hey, could I not tell the difference between you and a human

00:35:28.180 in conversation? Just keep moving to goalposts.

00:35:31.920 Yeah, yeah. So we're now going to go to this AI competition for coding, right? We talked about

00:35:39.440 this multimodal model. It did really well. This happened recently. So what was this contest that

00:35:44.480 I'm talking about? What happened? So the contest focused on creating good enough heuristics to

00:35:48.500 complex computationally unsolvable problems, like optimizing a robot's pass across a 30-30 grid

00:35:53.880 with the fewest moves possible. Under strict rules, no external libraries or documentation,

00:35:58.980 identical hardware for all, and a mandatory five-minute cool-down between code submissions.

00:36:04.120 A Polish programmer named Pedsmarie Dobrynski, known online as Psycho, who was a former OpenAI

00:36:09.820 employee, so no, really, really smart people were competing in this competition, took first place

00:36:16.460 after a grueling 10-hour marathon session. The OpenAI model debuted, OpenAI HC, finished a close

00:36:24.660 second, with Deepak edging it out by 9.5%. Final scores, 1.81 trillion points for the human versus

00:36:34.960 1.65 trillion for the AI. The AI beat 11 humans in total, so that was the rest of the field right

00:36:42.020 there. The event featured the world's top 12 human coders as qualifiers, with the AI added as an extra

00:36:48.680 competitor. Psycho was the only human to outperform the AI, while the other 11 humans placed third or

00:36:54.120 lower. As for how they ran the AI to make it competitive, it wasn't a standard publicly

00:36:58.300 available model like GPT-4 or even O1 that just spits out code in one go. This was a secret

00:37:04.820 internal OpenAI creation described as a simulated reasoning model similar to the O3 series, an

00:37:10.640 advanced successor to O1. It ran on the same code, at coder provided hardware as humans to ensure

00:37:16.480 fairness, but its strength came from its iterative multi-step process. And I mentioned how that went,

00:37:20.540 like plan, code, blah, blah, blah, blah, blah. Okay, right. So now we're going to talk about a paper that

00:37:26.540 Bruno cited for me in the thing he reached out to. This is an Apple research paper titled The Illusion

00:37:32.440 of Reasoning Makes the Case That Language Models... This is Bruno writing here. Oh, and other people have

00:37:36.880 asked us to comment on this too. This is great. Yeah. Cannot reason as marketed. The critique dovetails

00:37:42.280 with other signals of caution. Sam Altman has called this field a bubble. As I mentioned, it technically

00:37:46.460 is. And Elon has raised concerns about looming energy constraints, which might happen. Basically,

00:37:51.940 Elon's big bugaboo is energy is a bigger constraint than chips. He's not saying the industry is

00:37:56.760 overrated. These warnings are not isolated. So you point to structural issues, both technical and

00:38:02.580 economic, or they point to structural issues. Okay, let's go over this paper because this paper is

00:38:07.060 ridiculous. It's actually ridiculous. So what they did is they gave AI a number of puzzles to do. 0.97

00:38:17.360 And the AI outperformed humans by orders of magnitude at these puzzles. But they didn't

00:38:25.500 like the way it outperformed humans orders of magnitude. And so it could have been more efficient.

00:38:31.680 And I'm just like looking at them like with a guffaw on my face. Like, how can you be this unfair

00:38:38.860 to AI? Like, here, AI, do this puzzle. It does it at like 10 times the speed of a human or at 10 times

00:38:45.940 an advanced level of what an average human can do. And they're like, it's like, I notice you forgot to

00:38:51.480 dot your eyes. I guess I'm gonna have to mark you. It's not sentient. Imagine if teachers did that.

00:38:57.700 Like, it's like a super prejudiced teacher. But let's go in. Let's go into this. Okay. So we've got the

00:39:02.680 Tower of Hanoi. Okay. So the average human limit on the Tower of Hanoi can solve with up to three to

00:39:09.200 four disks. Seven to 15 moves with trial and error. Doubles minutes mentally with physical disks

00:39:15.680 disks at five to seven disks if you move to physical disks. Okay. So AI, when did AI do this?

00:39:22.720 So models like O3 Mini, so not a particularly advanced model here, right? It was able to do it

00:39:28.900 up to 15 moves. And I'll note here, but it did, it did, it did break down. Okay. So, so let's,

00:39:39.420 let's look at Claude 3.7 Sonnet. So we're saying, okay, but we're not looking for how high can it do it?

00:39:44.660 We're looking at how high can it do it flawlessly. Okay. Okay. So your average, you know, 95 IQ human

00:39:52.740 or whatever, right? They're, they're at three to five disks. Claude 3.7 Clonet can do it flawlessly

00:39:58.840 up to five disks. Okay. All right. So why did they get mad at the AI? Yeah. Why? Please explain this to me.

00:40:10.780 Saber argues this is an illusion because even at medium end traces show incoherent exploration

00:40:17.520 and effort peaks, then drops, e.g. fewer tokens spent despite budgets indicating no true recursive

00:40:24.600 understanding, just pattern extension until it breaks. Okay. My, my, my brother in Christ,

00:40:30.860 did you have an EEG hat on these humans? You don't know how their reasoning was working during this.

00:40:37.280 You don't know that this wasn't happening in the humans. Exactly. But also you didn't even use

00:40:43.100 humans as a norm in this. When they did this, they didn't use humans. I'm using other studies to look

00:40:48.040 at how humans perform on this. You just assume that the human brain doesn't work that way.

00:40:52.720 That's what always gets so, when people are like, AI is just a token predictor. And I'm like,

00:40:58.100 a lot of the evidence suggests human brains are a token predictor. So your episode on this,

00:41:01.280 and more episode evidence has come out since our episode on that, that I've got over in other episodes

00:41:06.280 because it annoys me so much. There's just like voluminous evidence, a huge chunk of the human

00:41:10.740 brain is probably a token predictor. But I just hate so much when they're like, humans don't make 0.98

00:41:17.080 these types of mistakes. And I'm like, well, first of all, even if you're considering them a mistake,

00:41:22.580 note that the AI did better than the humans at its task. So if the way it did, it was a mistake,

00:41:29.040 then clearly it understood its resource allocation and limitations and performed with it in a way that

00:41:35.280 out-competed its competitor, right? Who are you to say that you know better than it about how it can

00:41:40.520 do this? And if it could do it better, why didn't you add that to the token layer? You could have done

00:41:45.860 that. All right. So next, we're going to go to river crossing. The average human limit, the classic

00:41:51.040 here is three, is solvable through hints. Though the average might need trial and error to avoid

00:41:57.440 constraints. And at number four, 20 plus moves, complexity explodes. Most would fail due to tracking

00:42:02.620 multiple states mentally. All right. So humans, average human intelligence, three. Clots on it,

00:42:09.180 3.7, fails beyond three as well. And errors step up at four or higher. But it says that in humans,

00:42:16.360 four becomes near impossible. Okay. And so it collapses it around where humans do. Okay.

00:42:21.440 So here, AI is performing similar to humans. So why do they say that this proves it is dumber? 0.99

00:42:27.980 Well, they highlight, to highlight the illusion, they say, despite self-reflection, AI can't 0.88

00:42:32.800 consistently apply constraints, leading to invalid moves early, proving no deep understanding of

00:42:38.440 safety rules, just probabilistic guessing which falters. But you could change the way the AI model

00:42:44.080 works to do this. If the human brain is a token predictor that evolved, it almost certainly has

00:42:49.680 pathways to check for these types of mistakes if these are common mistakes within token predictors.

00:42:55.220 But you have locked the AIs that you are using out of doing that.

00:43:00.620 Oh my God.

00:43:01.260 Also, like, these are tests, you know, how far you get.

00:43:06.360 It's not like I've taken the SAT or some other standardized test and then been told,

00:43:13.640 oh, but you know, you took way too long on like these three problems or like you went back and

00:43:18.800 changed your answer.

00:43:19.340 Or like lowering your entire score.

00:43:21.080 Yeah. Like you're right or you're wrong. You get this many questions answered or you don't.

00:43:26.520 The fact that people keep not only moving the goalposts, but then going back into these tests

00:43:31.380 and evaluations and nitpicking the methodology used just seems like massive amounts of denial.

00:43:38.780 Well, this is how Apple is explaining why they can't make an AI because AIs aren't real.

00:43:42.800 But okay, so here we go to their blocks world test. Average human intelligence, you can get up to three

00:43:49.000 to five blocks. The AI breakpoint got up to 40 blocks. But LRMs collapse at high ends. AI vastly

00:44:01.180 outperforms humans on this one. But the paper points to an illusion via trace analysis. At medium

00:44:06.480 in corrections happen late. At high exploration is random and incoherent, not strategic.

00:44:12.120 Showing reliance on brute force patterns, but it's working, not adaptable planning. But if it's

00:44:17.960 working, it's a good strategy. You are demanding that it solve it the way that you solve it.

00:44:23.480 So throughout the paper, the way that you want to solve it. Yeah. Basically what they show is that

00:44:29.200 AI exhibits flaws like overthinking, exploring the wrong paths unnecessarily, which you could put in

00:44:36.380 the token layer for it not to do, or have another model that checks it to stop it from doing this.

00:44:42.220 Inconsistent self-correction and a hard cap on effort, collapsing incoherent at high complexity,

00:44:47.200 much higher than human, without adapting. Unlike humans who might intuitively grasp rules or persist

00:44:52.700 creatively, even slowly, AI doesn't build reusable strategies. It just delays failure and medium regimes.

00:44:59.260 So like, when I look at this paper, I'm honestly, I read a lot of romance mangas that take place in

00:45:07.680 fantasy worlds. And you'll have the evil, you know, stepmother or whatever, or concubine who will like

00:45:14.080 arrange all the tests. So her clearly incompetent son can beat the clearly much more competent person.

00:45:20.760 And then the, the, the, the bribed vizier will come out and say, well, do you not see that he took too

00:45:26.620 long on question number five, which proves auspiciously unlucky. And it's like, come on, my friend, what are

00:45:34.820 you doing? Like, clearly you're just begging the question here, right? Like the AI is outperforming

00:45:41.960 people and you are using its outperformance. This reminds me of the hilarious test that some people

00:45:46.680 have been like, they released this paper saying, oh, well, yeah, there was this paper done on Claude

00:45:52.280 that showed that it didn't know the logic, the internal logic it had used to get to certain

00:45:56.940 outcomes, right? Like when you could look at this internal logic. Humans don't know the internal

00:46:00.700 logic they use to get to certain outcomes. I point, a lot of people think that humans know,

00:46:04.740 but if you look, there's been a lot of experiments on this. Look at our LLM models where we go over all

00:46:09.040 the studies on this. It's just so stories. They're post, they're adding post-hoc reasoning and

00:46:13.120 they add post-hoc reasoning. Basically you make up how you came to a decision if that decision is

00:46:18.640 changed in front of you. So a famous example of this is people will think they chose like one woman

00:46:23.460 as the most hot from a crowd and they'll do sleight of hand and then show you another woman and say, 1.00

00:46:27.080 why'd you choose this woman? And people will provide detailed explanations. And they've done

00:46:30.380 this with political opinions. They've done this with like, this is a well-studied thing in psychology.

00:46:35.020 You have no idea why you make the decisions you make, but they assume because our intuition

00:46:40.400 is that we think we know. It's not even that it's our intuition. It's that our minds are token

00:46:46.660 predictors, like both on a technical, but also like more philosophically. And when someone asks

00:46:53.220 us a question, we want to be able to answer it. We see this with our kids all the time. Like last

00:46:57.440 night, Toasty, our son was telling us how Tommyknockers, which are these like monsters we

00:47:02.600 made up for them. Yeah. He was like, Tommyknockers cannot exist in this house. And we're like,

00:47:09.240 well, how do you know that? And he's like, well, it, my granddad said it to me at his house when

00:47:14.660 I was a baby. He's not been at his grandfather's house. This doesn't make sense. But humans like

00:47:24.080 to give answers for things. And I get that. That's totally respectable, but like, he hallucinated.

00:47:28.500 He literally hallucinated. Yeah. Like we do that too. So stop people. Stop. You're embarrassing

00:47:35.500 yourselves. Now I'm not going to go too deep into some of the ways AI is being used for medical

00:47:39.800 research, because I don't know if people fully care, but I will at least go into some of the

00:47:44.080 drugs and, and some of the methods where it's been used. It's been used for genome sequencing and

00:47:48.540 analysis. It's been used for variant detection, disease prediction. It's been used for clinical

00:47:52.720 genetics and diagnostics. It's been used for drug design and target identification. It's been used for

00:47:57.540 predicting interactions and toxicity. It's been used for streamlining the development in clinical trials.

00:48:02.180 Now, if we're going to go into some of the specific ones that have been developed,

00:48:05.760 one called Rintocertib. I'm, I'm, I'm, I-N-S-O-18-O-5-5. You know what? I'm not going to

00:48:15.840 list these designations for the future ones, but this was developed by Encelco Medicine using their

00:48:20.780 generative AI platform, pharma.ai. This small molecule inhibitor targets TNIK for pulmonary fibrosis,

00:48:29.200 a rare lung disease, which my family has, and has killed multiple family members of mine.

00:48:34.280 So we might've actually funded this research because my family does fund a lot of stuff in

00:48:38.140 that industry. Then another one co-discovered by Existentia and Sumatoa Pharma using AI driven

00:48:44.400 design. This serotonin 5HT1A receptor agonist treats obsessive compulsive disorder. Now note here for

00:48:51.880 this first AI drug development, right? This could literally save my life one day. My, my, I think

00:48:59.520 my aunt died of this. I know my grandfather died of this. My dad has this. So I could easily get,

00:49:05.600 like, I'm very, like, this is like the number one killer in my family. And AI might've developed a

00:49:10.800 solution to it. Like, you can't understand when you're like, AI has done nothing meaningful. It's

00:49:15.360 other than this drug that saves people in my family's life. Yeah. Like maybe, you know,

00:49:20.160 let's say you have like a serious risk of Alzheimer's in your family. You're going to

00:49:23.980 feel very different about AI once AI cures Alzheimer's. Actually, by the way, another

00:49:28.620 Existentia Sumo collaboration was dual 5HT1A agonist 5HT2 agonist, which targets Alzheimer's

00:49:35.400 disease. Amazing. They're really going for those pithy names. The same company, Existentia,

00:49:42.240 developed a cancer treatment, a tumor fighting immune response thing.

00:49:47.000 Oh, point for Simone, because of course the cancer is coming for me.

00:49:50.720 Yeah. And then in terms of DNA stuff, like what's it finding in genetics,

00:49:54.720 the novel autism linked mutations in non-coding DNA using deep learning on whole genome sequences

00:50:00.620 from thousands of families, researchers identified previously undetected mutations

00:50:03.720 in non-coding regions associated with a disorder, autism. Rare DNA sequences for gene activation.

00:50:09.040 AI analyzed vast genomic data to discover custom tailored downstream pro-motor regions,

00:50:14.680 DPR sequences active in humans, but not fruit flies and vice versa.

00:50:18.500 I also think all this like better genetic sequencing with autism might actually

00:50:21.920 fix the autism diagnosis problem of like too many different conditions being grouped into autism.

00:50:28.660 Like, you know, we're participating as a family in autism genetic research.

00:50:31.940 Yeah. But like our kids don't have any of the like genes for autism and that's because they have

00:50:40.520 Asperger's.

00:50:41.740 Even though they've all been diagnosed, you've been diagnosed. Do you have any of the genes?

00:50:44.960 No. And that's the thing. It's like, I think that when, when AI helps us better understand autism

00:50:51.740 and like the genetic components of it, they're going to be like, all right, so these are actually

00:50:56.240 super different things. And on this technical level, we can demonstrate and show it how probably

00:51:01.860 low functioning autism and different forms of autism are going to be seen as very different

00:51:06.120 from what used to be called Asperger's. And it's now just, but my point here being is people are

00:51:12.140 already being fired over this, right? Yeah. If you're looking at AI and it not just that it's

00:51:17.400 already developing life-saving drugs. It's already developing game-changing scientific

00:51:22.200 development. Yeah. People like, I guess if it doesn't immediately affect them, if they are not

00:51:26.580 married to their AI boyfriend and husband now, if they aren't having a personally scary disease.

00:51:33.060 Which is happening, by the way. So a study which surveyed a thousand teens in April and May

00:51:37.120 showed a dramatic rise in AI social interaction with more than 70% of teens having used AI chat

00:51:43.420 companions and 50% using them regularly. 50% of teens are using them regularly.

00:51:47.880 I can't even wrap my head around that.

00:51:49.900 No, no. You want to hear a crazier statistic?

00:51:51.480 Sure. Despite the widespread use, 67% of teens say that talking to people is still more

00:51:59.360 satisfying overall. Wait, wait, wait, wait, wait. So 33% think talking to AI is more satisfying?

00:52:06.100 That's a huge chunk!

00:52:09.700 But no, we've hit a plateau. It's all a bubble.

00:52:14.580 Okay, guys, have fun being left in the dust. Enjoy it.

00:52:18.660 No. But I see. Because when people are thinking about a product, they think about it from a consumer

00:52:25.340 level. They think about it like an iPod or a, you know, why isn't this in my pocket, right?

00:52:30.460 It can also be hard for people to wrap their heads around it. You know, like when cars started being adopted, it's like, oh, this is just a rich person thing. Like, they break down all the time. They, you know, it's just better to have a horse. Just keep your horse. People couldn't imagine not having horses on the roads. You know, it's similar to the hallucination arguments. Like cars break down. AI's hallucinate. Why? How is that going to transform society?

00:52:52.200 Yeah. Buckle up, guys. You don't have to. But yeah, if you go into this not wearing your seatbelt, this is on you.

00:53:01.760 Yeah. And I could go into like technical things where it looks like parts of AI development have slowed down recently. But in other areas, it looks like it's sped up recently. Like that's the problem with a lot of this is you can say, well, it's slowed down here. It's slowed down here. And then, well, it's sped up here, here and here. Right? And then you'll get some new model, like DeepSync's new model. And they'll be like, oh, and now we have some giant jump. Right?

00:53:22.180 And then we've just been seeing this over and over again. I hope it plateaus. It's going to be scary if it doesn't plateau. But we're not seeing this yet. We're seeing what is kind of a best case scenario, which is steady growth. Steady, fast growth. Not fooming. Okay. But steady, fast growth.

00:53:40.740 Yeah. So yeah, multiple people actually requested this discussion in the comments of the video we ran today, which was on how geopolitics will be changed after the rise of AI and more accelerated demographic collapse. So I'm glad that you addressed all this.

00:53:59.200 There's a lot more. I mean, we're only just getting started. And a lot of people also chimed in in the comments. They're like, well, give me specific dates. I need to know like, you know, what by when we can't do that. Like we can give you dates. We're going to be wrong. Like it's, it's really hard to predict how fast things are going to be. And there are so many factors affecting adoption, including regulatory factors, social factors, that it just makes it really hard for us to say exactly when things are going to happen. Are heuristic with these things? If you're just trying to be like, well, yeah, but like, how do I know when to start planning?

00:54:29.200 This is your reality now. Just like accept it as reality and live as though it's true. That's how we live our lives. We live our lives under the assumption that this is the new world order. And we don't invest in things that are part of the old world order in terms of our time or dependence. And we do lean toward things that are part of the new world order, if that makes sense.

00:54:48.440 Yeah, no, I, I, I absolutely think it makes sense. And I'm just, I totally understand where people are coming from with this, but my God, are they, they, it's, it's like hearing computers transform society and only thinking about the computers that you use for recreation instead of the computers that are used in a manufacturing plant and to keep planes connected and to, you know, like the,

00:55:18.440 and even if the development stopped today, the amount that the existing technology would transform societies in ways that haven't yet happened is almost incalculable.

00:55:30.440 Like that, that's the thing that gets me. I don't need to see AI doing something more than it's already done today. Like I, I, I, I don't need to see something more advanced than Grok 4. Okay.

00:55:43.340 Than opening AI 5. I, I, with these models, I could replace 30% of people in the legal profession. That, that's a big economic thing. Okay.

00:55:53.440 Yep. And I mean, again, we, we can't say how fast this is going to be impactful or not because there are already states in the United States, for example, that are making it illegal, for example, for your psychotherapist.

00:56:08.440 AI therapy. Yeah.

00:56:09.920 Even though AI outperforms normal therapists on most benchmarks.

00:56:12.620 Well, it just to use AI to help themselves, like, and they're going to cheat anyway, but like, so people are going to try to artificially slow things down in an attempt to protect jobs.

00:56:23.220 Or protect industries because they don't trust it. So again, things will be artificially slowed down. Sometimes things will be artificially sped up by countries saying, okay, we're all about this. We need to make it happen.

00:56:33.680 Like China. 0.94

00:56:34.300 Yeah.

00:56:35.080 Or Trump. I cannot believe the Democrats have become such the Luddites.

00:56:39.200 Oh, whatever. Anyway. Yeah. Thanks for addressing this. And I'm excited. We're in the fun timeline.

00:56:46.740 Oh, absolutely. I, I often, I watch AMVs from, you know, the zombie show about the office worker, the, the Japanese office worker. 0.97

00:56:55.780 Oh, where he's like, yay.

00:56:57.200 He's having a blast. We are undergoing like multiple apocalypses right now. And I'm like, I am here for every one of them.

00:57:03.320 Yeah.

00:57:05.500 This is, this is a fun time to be alive during the, the AI slash fertility rate apocalypse, because I get to do the things that I want to win.

00:57:14.440 Have lots of kids and work with AI to make it better.

00:57:17.740 Yeah. All right.

00:57:20.140 All right. Love you.

00:57:20.960 Sending you in the next link. Get ready for it.

00:57:24.900 What do you mean?

00:57:26.360 Make sure it's backed up.

00:57:28.500 Ah, okay.

00:57:29.920 Okay. Oh, do I look okay?

00:57:34.360 You, you.

00:57:37.980 Yeah. Ish.

00:57:39.320 I need to cut your hair, but I will.

00:57:43.100 Your hair doesn't.

00:57:44.040 You are a lovely wife. And I love that you cut my hair now. It feels so much more contained.

00:57:49.100 The more things we bring into the house, whether it's you making food or cutting my hair, it dramatically improves my quality of life because I don't have to go outside or interact with other people.

00:57:58.880 And I really hadn't expected that. And it's, it's pretty awesome.

00:58:02.840 Yeah. I get now why for many people, it's a luxury to have everyone come to your house to deliver services, but it's even better if you don't have to talk with someone else and coordinate with someone else and pay someone else and thank someone else.

00:58:16.820 And not like I'm not appreciative of what other people do and the services they provide, but it's just additional stress. Like this is a generation of people that can't answer the phone, like me included. It's just, and so like the anxiety that you have to undergo to like have a transaction with a human is so high.

00:58:35.720 Even if they, even if they're doing a great job and they're happy and you're happy, you still have to like go through the whole, thank you so much. And, oh, can I have this? And, well, this isn't quite right. Can I have this adjusted? And like, no, I would rather use my mental processing power to just keep our kids somewhat in order.

00:58:52.700 Somewhat in order. That's a tall, what do people think of the episode today?

00:58:58.460 What did they think? They, I think they liked it. I'm trying to think of like, if there was any theme in the comments, a lot of people had small quibbles here or there about birth rates in certain areas. And I think that's because the data is so all over the place. And a lot of people have anchored to old data.

00:59:16.220 And then they're really shocked to see how much the birth rates have changed. Some, I haven't gone deep into it, but some people have questioned why you think like growth in certain areas in populations won't matter due to them sort of being technologically just not online yet and not developed.

00:59:38.220 Yeah. This to me, I just find like a comical thing. Like they, they, they think that they're going to get a Wakanda, right? Like this is not going to happen. Right. You, you, you can't just, when we've seen populations like jump in technology and industry levels, it happens because of some new form of contact or some new form of technology being imported to the region. Like we saw in East Asia.

01:00:01.580 It's very unlikely that you're going to see something like Somalia, which has good fertility rates, just like suddenly develop. It doesn't, it's, it's, and we've tried to force it, right? Like this is fundamentally what the U S tried to do with Iraq, right? Like we tried to force them to become a modern democracy and a modern economy in the same way we did with like South Korea and Japan and Germany. And it just didn't work. 0.98

01:00:26.920 What do you think about the city-states that like Patrick's working on in Africa? Couldn't you theoretically create Wakandas?

01:00:35.160 You could, you could. I think one of his city-states would be most likely to do that, but that's not going to have an impact on a wide spread of the region, right?

01:00:43.040 Yeah. Basically just those who can get in.

01:00:45.280 Yeah.

01:00:45.640 Will thrive.

01:00:47.720 Yeah.

01:00:48.240 So anyway, I am going to jump in here.

01:00:53.600 Tate, can you tell me about your dream last night?

01:00:56.920 Yeah.

01:00:57.780 Yeah.

01:00:58.120 What happened?

01:01:01.940 Nothing happened.

01:01:03.780 Was it just black? 0.84

01:01:07.100 Wow. That's a very exciting dream.

01:01:10.040 So you didn't dream about spiders or Tommy knockers or anything else? Just black?

01:01:20.640 So just black, no spiders or Tommy knockers?

01:01:24.460 No spiders or...

01:01:26.340 Tommy knockers.

01:01:28.820 What are Tommy knockers? 0.97

01:01:30.080 Tommy knockers are but...

01:01:31.700 Hybrids. 0.57

01:01:33.120 Tommy knockers.

01:01:35.060 Are they dangerous?

01:01:36.320 Yeah.

01:01:37.300 What happens if you get near your tummy knockers? 0.92

01:01:39.060 Yeah.

01:01:41.000 The tummy knockers are gone forever. 1.00

01:01:45.740 Well, where do they live?

01:01:47.460 In the tunnel.

01:01:48.700 No.

01:01:49.200 In the tunnel.

01:01:50.240 In the cave.

01:01:50.540 So we stay away from the cave in the tunnel, right?

01:01:52.060 In the cave.

01:01:52.240 We stay away from the cave, right?

01:01:53.120 Yeah.

01:01:53.680 Yeah.

01:01:53.980 So if you Tommy knockers want to get... 0.97

01:01:56.980 What happens if you get too close to the cave?

01:02:00.600 I don't have to get too close to the cave.

01:02:05.460 Because otherwise they'll get you.

01:02:06.740 What happens if they get you?

01:02:08.660 Yeah.

01:02:09.940 What happens if the Tommy knockers get you? 1.00

01:02:12.560 The Tommy knockers didn't get me. 1.00

01:02:15.560 What if they do get you, Titan?

01:02:18.440 They do dance.

01:02:21.620 Will they drag you into the tunnels?

01:02:23.680 No.

01:02:25.100 What happened?

01:02:27.220 Oh my God.

01:02:30.260 You don't know?

01:02:32.000 Do the Tommy knockers have tentacles?

01:02:35.160 No.

01:02:36.440 Do octopuses have tentacles?

01:02:39.580 Sorry, Mommy.

01:02:41.920 Did you spill a little?

01:02:43.500 Yeah.

01:02:44.540 Oops.

01:02:45.560 I'm sorry, Mommy.

01:02:50.980 You're a concentration face. 0.79

01:02:56.280 You're cheating.

01:02:58.600 Oh, thanks, Indy.

01:03:00.900 I make a little smell.

01:03:02.780 Thank you.

01:03:03.680 Oh.

01:03:04.760 I know.

01:03:05.560 You're very good at using water.

01:03:07.840 Yeah.

01:03:08.440 Yeah.

01:03:08.720 I didn't spill water.

01:03:10.940 Thank you, Indy.

01:03:13.340 No.

01:03:14.200 No.

01:03:14.540 Do you want a kiss?

01:03:15.840 Yeah, I do want a kiss.

01:03:18.300 I love you too, Dad.

01:03:23.320 Yeah.

01:03:23.360 I love you too, Dad.

01:03:23.560 Yeah.

01:03:23.700 Thank you.

Based Camp - August 21, 2025

Is AI Overhyped？ Is AI a Bubble？ (New Models Don't ＂Feel＂ That Much Better, Why？)

Episode Stats

Summary

Transcript