Keynote: After the AI Hype – What’s Real, and What’s Next - Richard Campbell - 2026

9,315 words · about 37 min to read

AI is a terrible 1950s marketing term

0:00 Ready in a minute, okay. Just a small issue. Let me just start doing the talk. It's only pictures anyway, right? Uh my name is Richard Campbell. I come from Vancouver, British Columbia. Uh not my first time in Copenhagen. And like you said, I've done many INDCs. Uh this particular talk has emerged over time because there's been so much confusion around what's happening with artificial intelligence. It's uh I think the problem is mainly that it's a terrible name.

0:25 Uh it's an old term. It's from the 1950s. It was coined by a group of scientists that were trying to raise money from the US military. Uh which they succeeded in doing. Now, this is in the '50s, so this is even before true electronic computers like ICs and integrated you know integrated circuits all that sort of thing. These were electromechanical computers. But even then there was a belief that we could actually do some computing performance that was sophisticated enough that it would make a difference for folks. Hey, that's my screen.

0:54 How about that?

1:00 That's the slide. Anyway, they uh for a moment Uh they succeeded in in writing some software some uh provisioning and organizational software for the US military and it did its job. The US military is still one of the most effective forces projecting anywhere in the world. And that's in the '50s. Then the money dries up and when it does, then uh we go into the first of what we called an AI winter.

1:28 And there's been a number of these over the years with different technology. Over and over again, they've uh scientists have used these terms to finance new technologies to experiment with. In the 1960s, there was a fellow named Joseph Weizenbaum who wrote a piece of software on a mainframe computer called Eliza. And Eliza he called a Rogerian therapist, which was a stretch. Now, Weizenbaum's goal was to show how effective simple software was at providing language responses to people that they reacted to. And literally they Rogerian therapist idea was you would complain about something like how bad your lunch was and it would repeat back to you. So tell me how bad your lunch was.

2:11 And he had intended to just demonstrate that it was kind of silliness and like people are vulnerable to these sorts of things. And then his secretary refused to stop using it. And and he found out people were putting hours into the software. Like just the sense that the software

1 Highlight · Human psychology makes us vulnerable to AI hype

2:26 was listening to you was a big deal. That was the 60s. Like we're very susceptible to these problems. Humans are kind of hardwired to perceive intelligence in other places. It's a competitive advantage. You know, when we were pre-technology society out on the veldt, you know, our ability to project intelligence into our prey let us anticipate its behavior so we could kill it needed.

2:51 There's also another human response, pareidolia. Pareidolia is the ability to see faces quickly. And again, imagine that in a pre-technology society, if you're the first one to perceive the face of the lion, you get to go running first, which means the other guys get eaten and you get to live and propagate your species. And so that trait evolved in. You think of all the products we make today that have faces on them intentionally. We see them everywhere cuz there's really no downside to seeing the face and the upside was really strong once upon a time and so you think about the way we react to a car with the shape of its headlights and grill and so forth as a face.

3:35 And I bring this up just to say there's a group of folks out there that are leveraging normal human tendency to

3:41 create a perception around these products that isn't true. Now, the word artificial intelligence only comes to the public for the first time. It was a coined in the 50s but it comes to the public in a movie. The movie is 2001 A Space Odyssey. That's 1968 Stanley Kubrick. And if you haven't seen it, I recommend it, although it is very long. But in the movie, there's a computer and the computer's name is HAL. And then HAL tries to kill everybody.

4:07 Like every other science movie fiction movie with with artificial intelligence in it. Over and over again. In the 1980s, it's the Terminator movies. And in the 2010s, it's Ultron. I ask students what they think of artificial intelligence and they say it's Jarvis. So the science fiction has penetrated their minds and they believe they know what this tech is cuz they've seen it in a movie.

4:32 And that's a problem, right? That's really we've been deceived. We're you misusing a name.

Modern deep learning unlocked by GPUs and data

4:38 And we've been using it literally for decades and now we're dealing with the consequences of it. So this current wave, and I would depending on how you count it's the fourth or fifth wave of a set of technologies or branding is artificial intelligence comes we generally would call it generative AI now. It comes from a group of people, but one of the key ones there is a guy named Geoffrey Hinton. And you've seen him on the news or in TV. Uh he was uh scientist, professor uh out of uh the UK but working in Canada in the University of Toronto. Got his PhD in artificial intelligence in the 1970s.

5:13 And he was big in the neural nets at that time. And in the '80s, actually was part of a team that built software that was able to do handwriting recognition for things like envelopes, addresses, and for checks. So the early neural nets did work just in a limited scope, certain kinds of handwriting with very specific rules around it. It they were effective. And so what Hinton was proposing even in the '80s is that we could build deeper neural nets. Now this is just a visualization of that, but the idea that with many deep neurons with this concept of back propagation he was part of a paper on how they would tune these neural nets, but essentially ends with more research is needed. We need more compute to make this possible. And so the tech kind of sits on the shelf until the 2010s.

5:58 When a group of his students dusted off to enter a competition. The competition is the ImageNet competition. So this was the beginning of the idea of can we have computers identify the contents of images. Wait, we take this for granted today. Our phones can just do this, but in 2010 they couldn't. It was impossible. Until it suddenly wasn't. So the ImageNet competition was 14 million labeled images, mostly stolen from social media, so thanks.

6:26 And the idea being we have a description of what's in this image. So now you write software to look at the image and recreate the description. Simple. And up until the In the early stages of this competition, 30-40% was the best anyone had ever done with some decision tree models and things like this, but it was Hinton with his students Ilya Sutskever and Kaiming He. A few other folks, the old names you know, took this old technology, applied it to modern GPUs, entered the competition in 2012 and got 75% right the first time.

6:55 And within 2 years 100%. And it was a solved problem. By 2015, you could expect your phone could You could take a picture with your phone. Your phone would tell you what was in the image. It's astonishing. And we already take it for granted. It's not a big deal. Now that's not the only place where this new generation of technology emerged. The Over on the West Coast at Stanford Research Institute, they were experimenting with voice.

7:20 So there've been voice recognition systems for years, but they were bad. It took a lot of training time. Uh 80% 85% was the best you were going to get. So the SRI folks sat down and took a whole lot of recorded audio, pass it through this generative model until they could do really effective like 99% voice recognition with it. And they made it into an iPhone app and put it on the store, and Apple was so impressed by it, they bought it, added an I to the name, and called it Siri.

7:49 So, this was almost a simultaneous exposure. Like, the ImageNet thing, Siri, all happened roughly in the same time span. Now, back on the East Coast, University of Toronto, Hinton and his cohort are encouraged by Google to form a company to commercialize the technology they developed for image recognition, and as soon as they form that company, Google buys it. Shocking. And brings all those folks down to Menlo Park, the Silicon Valley, to join a group called Google Brain.

8:16 And so, Google Brain had for a few years been rounding up kind of the best minds in this space of generative AI, to the point where other organizations were getting really nervous about it. And back when we weren't quite as annoyed with tech billionaires as we are today, a bunch of tech billionaires sort of showed up saying, "Hey, we're really worried about what Google's working on in private. We think we should do it out in public." And this was folks like Sam Altman, and Elon Musk, and Peter Thiel, and Reid Hoffman, and they set up another company in 2015 called OpenAI.

8:47 And OpenAI's mission, publicly, was to do AI in public. Was to make sure it was visible. Privately, it was recruit those scientists out of Google. And that's what they did. By 20 Cisco Argos over there in 2015, bit bit bit, they poached them all. And the OpenAI guys, they were supposed to be well-funded. After all, it was run sort of a bunch of by a bunch of billionaires. It just wasn't well-funded. You know, Elon for many years was saying he put a hundred million dollars into OpenAI. Turned out it was sixteen million dollars and eight Teslas.

9:19 Okay. But,

Microsoft investment fueled neural scaling laws

9:22 uh they do get a remarkable set of talent together. Enough that Elon's so impressed that by 2018, he tries to take over the company and gets kicked out of it. And he's still angry about that. You're watching the lawsuits still go on to this day around that. The first product that the OpenAI guys work on is a universal translator. So, they're trying to build the Babel Fish from Hitchhiker's Guide. And it more or less worked. It was the tokenization strategy, this technique for taking words, breaking them into numbers where one word would represent by multiple numbers. Some some cases multiple words could be represented by single number, but by tokenizing language, you could then detokenize it into other languages. But along the way, as they worked on the training set against text on the internet, they found they could create a response response engine that with some training they ended up with this thing they call a generative pre-trained transformer.

2 Highlight · Automation paradox: radiology demand surges

10:18 Now, in that same same time space as image recognition got bigger, they started using image recognition in healthcare for working with MRI imaging, x-rays, and so forth. In fact, our our friend Jeff Hinton here in 2016 announced to the world that he thought radiology was a dead career, that the computers would do it all. Because they now had at this point there's like 700 models certified by the FDA to do medical image recognition. The reality was he was wildly wrong. The opposite actually happened. The demand for radiologists has gone through the roof because the demand for medical imaging only grew.

10:56 Back when it took 6 weeks to get an image analyzed by a radiologist, a lot of doctors just wouldn't bother requesting an image in the first place. You want to do your diagnosis faster now for your patients, so you don't want to wait. But as the software came into play and the radiologists could work faster, the demand for medical imaging outran the the systems. So, even though radiologists are working wildly faster with better technology, they're still falling behind. So, we're only growing the number of radiologists, but that takes time.

11:25 So, once again, we see automation opening the door to far more productivity, and that increases demand. So, the opposite goes on with this mechanism.

11:34 Meantime, over at OpenAI, they are building this GPT. The first version they put out in public is in 2019, GPT-2. And they make a variety of models for it. The smallest being about 120 million parameters and the largest 1.5 billion parameters. They make different sizes because they're so strapped for cash, they can't afford to run their own GPTs. And so, they make smaller models to do testing with. They only run the big model when they're really sure they need it.

12:02 This is also the point where OpenAI sort of backs off on their original deal where it's like, "Hey, we're going to do all this in public." They're like, "Wow, this is really quite a powerful tool and we're concerned about what people would do with it. So, we're not going to publish the source code. Sorry about that." And so, the business evolves. And it was around this time that Microsoft shows up. And this is actually Kevin Scott. He was the former CTO of of LinkedIn. And he uh knew the folks at OpenAI. He'd worked in Silicon Valley for a long time. Now, he's working for Microsoft cuz Microsoft acquired LinkedIn. And his job was to find workloads for Azure.

12:38 He was at the office of CTO and Microsoft had gone all in on Azure and so, they were just trying to find as much business as they could. And in a very now fairly famous email in 2020, Kevin Scott writes to Satya Nadella and says, "Hey, we've got our own experiments in AI. I think the OpenAI guys are ahead of us and they're strapped for cash. I think we should invest in them and move all their workloads to Azure, which is what you wanted me to do anyway." And so, with some encouragement, OpenAI restructures their company to allow for investment and Microsoft kicks them a billion dollars in exchange for they spend all that money on Azure. Which is a pretty good deal for Microsoft, if you think about it. I just made a billion-dollar investment and made a billion dollars in sales. I made a billion dollars and did two billion dollars.

13:25 But it definitely helps OpenAI. That accelerates their ability to do more

3 Highlight · The neural scaling laws justification

13:29 development and a lot in the midst of this they publish a paper called the neural scaling laws. This is in 2020. Where they go against everything we really know about machine learning. For a long time in machine learning the previous decade or two there was this very standard set of practices around restricting your training set. Not over training on that training set. They called it over fitting. If you train too hard on the training set, it's only good on training set, it's not good on anything else. But the neural scaling laws paper said this, train on everything.

14:04 And then it'll do everything. And as long as we keep getting the set bigger at some point it'll be perfect. This paper is kind of the excuse that the hyperscalers have used more or less ever since for the approach to the generative AI systems we're seeing in LLMs. So that billion dollars turns into

ChatGPT triggered a classic tech hype cycle

14:26 GPT-3. There was This is during the pandemic, so it's hard to remember what was going on then. We're all fairly distracted. In 2020 Microsoft kicked another two billion into the company and in 20 uh and they produced this first version of GPT-3 that had 175 billion parameters. So the largest GPT-2 was 1.5 billion, so this is literally a hundredfold bigger cuz they've got money and Azure. Uh they were publishing PR pieces about this. The fifth largest supercomputer in the world to build this model. 285,000 CPUs involved over several months of training.

15:02 Still wasn't that good. Uh lang- human languages are complicated and they're and they're irregular and it's difficult for the model to make really good responses. But also down in Silicon Valley was GitHub and GitHub said, "Hey, what if we restricted the set? What if we used programming languages instead of human languages?" And that had some better results. That's the first product called Copilot. Cuz you know it's a good name, right?

15:28 Cuz that means you're the pilot. It's still your fault. But also and there's no way Microsoft would have thought of that name. Microsoft doesn't think of names that good. They got to have to ruin a good name though. Uh And so This is largely the first product on the LLM spaces GitHub Copilot. Meantime back at OpenAI, they're trying to make the language system work better. And they realize they need to do a second order training. This was a clever innovation. And that second order training was to use the existing model

4 Highlight · ChatGPT started as a desperate funding experiment

15:58 with humans where you'd make a query, ask for a piece of information, the model would return three or four answers, and then you'd rank those answers. And that would build up a model to tune the responses. They typically did this in developing parts of the world where the people were inexpensive. Indonesia, Philippines, that sort of thing. But by the fall of 2022, they've run out of money again. And so they come up with this great idea. Let's just do it in public. Instead of paying people to do this, we'll just make a public interface to it so that everybody can experiment it and we'll evaluate the data to further train the model. That's ChatGPT.

16:35 It was meant as a an experiment. It was not meant to be a big product. It was just an experiment, a way to get people to do training for them cuz they were strapped for cash. What they didn't expect is that a hundred million people would sign up for it in the next two months by January of '23. And that was the fastest that any product had ever gotten to a hundred million users before. The previous fastest had been TikTok in nine months.

17:01 Now I know a few of the folks that were working at Azure when that two months went on and they had a really lousy Christmas because scaling to a hundred million users is not easy. That's a lot of resources they had to scramble to put together and to spread around the world as people signed up in droves. I had a chance to talk to one of the folks behind the system and I said, "Like, why do you think 100 million people signed up in 2 months?" And he's like, "Well, you know, we kind of did it over Christmas and people would rather talk to software than their loved ones, so."

17:30 Like, get a dog. Why are you talking to

17:32 software? That's just weird. And I I hope you've noticed I'm being very careful with language here. I'm fighting against anthropomorphizing the software. Right? I'm not going to refer to it as a person cuz it's not a person. And I encourage you to pick up the habit, too, to help others think clearly about the fact that this is simply software. It's clever software, but software nonetheless. And it's between science fiction and bad naming we're having a tough time remembering this is software.

18:05 100 million users. Satya Nadella, who at this point has been CEO for 9 years in 2023, for the first time ever does a Bill Gates letter internally. He says, "All right, here are the Open AI APIs. We have access to them. Every team needs to build something with their product against these APIs. Go forth and make way too many copilots."

18:30 Which they will do. For We've only seen a few and it's still too many. There are many, many copilots now being built. And this is a point in hindsight we recognize is a technological trigger. So, I'm going to pull out the Gartner Hype Cycle here, which we came up with after the dot-com boom, and talk about this cycle that some technology goes through. Sometimes you have a technological event that brings all the investors to the yard.

18:58 And a stupid amount of money shows up and stupid money makes stupid decisions. And so, with that much money flowing in and that many companies being formed, you have to keep making bigger and bigger promises to to investors until you create a hype cycle so vast, you get to this peak of overinflated expectations that's insane. And when those expectations can't be met, then the investment comes flying back out again, hopefully not losing too much money along the way, at least for the investors.

19:29 Bunch of companies fail, and we bought them out, and then climb back out. So, this happened in between 1997 and 2001 for the World Wide Web. The technological trigger at that time was the IPO of Netscape. The first time a pure digital company IPO'd. I mean, even in in 1997, Microsoft still shipped CDs for all of their software, and every update was expensive because you had to ship out new CDs.

19:55 But here was Netscape, a company that had no physical media of any kind. You downloaded the browser, you got updates from the internet. It was a profound impact on investors, this zero marginal cost company, and they poured money into anybody who could say they could they were internet companies. And we inflated expectations until we thought pets.com was a good idea. And then the money came back out, and the internet didn't go away, the World Wide Web survived, a lot of dumb subsided.

20:25 We're in another one of those. Doesn't happen every time. It's not the only technological revolution we've had since the dot-com boom, right? One would argue that mobile was a huge revolution, changed the way we interact with with computing in general. We sell far more smartphones than PCs by a mile. One would argue that the cloud is a revolution as well. But neither of them went through a hype cycle. A hype cycle has to have certain elements distinct to it. It helps that the investor doesn't understand what they're investing in. So, you can tell them stories.

20:56 And AI fits that category brilliantly. And that you can set expectations that are insane. You can say, "Every job will be replaced in the next 2 years." And people don't just laugh at you. Instead, they give you money, which seems like a really dumb thing to do. So, here we are in another one of these

Software engineering finally sees real AI productivity

21:14 cycles. Now, Microsoft gets there first. They kick in another $10 billion to Open AI, so that's up to 13 billion if you're keeping track. Uh first product out of the gate is Bing. Now, the reason they were quick here is cuz of Project Sydney. They were already had been working on a natural language interface to their search engine to try and make it relevant. Now, this was a much better natural language search interface, and both guys who tried it thought it was really good. Uh so, turns out you can't save Bing with a better interface.

21:43 But, with that $10 billion, uh in another year or so, Open AI makes an even bigger model. March 23, releasing GPT-4. So, now we're up to a trillion parameters. So, 1.5 billion for two, 175 billion for three, trillion for four. Not quite a doubling or maybe a five, six times increase. But, at that point, we've pretty much indexed all of the internet, including 4chan.

22:15 You know, we're not going to make a two trillion parameter model. There's no more data. So, we're In a lot of ways, you could say we In hindsight, certainly, we kind of peaked at this point. It's an exciting time. Uh we're impressed with the capabilities of it. It's still got lots of tuning to go. We're going to make up some revisions on it, but we made the big big model. It's hugely computationally expensive. Microsoft starts pouring out the go Copilots, M365 Copilot being big one. They're not sure what to make here, so they keep making different versions of it to the point where you know, there are a couple of good Copilots. Like, if you've ever played with the SRE Copilot, it's impressive.

22:51 Uh I think they call it the SRE agent cuz even Microsoft people are trying to distance themselves from the Copilot name. Uh it's just too many products, and many of them are bad, and so folks don't know what to use. Uh 4.0 is when they went multimodal, so this is now not just text but vision and audio. This is in '24 and around that same time we get the first of the deep seek models. The West kind did their very best to ignore deep seek, but what deep seek This is the Chinese model. And again, it's built against the other technologies that were created at the time. So, they definitely took advantage of what had come before, but the main thing that the Chinese model talked about was, "You know, you don't have to be so big and still get pretty good results."

23:30 And we started the first time we didn't just talk stick with the neural scaling laws and just keep getting bigger. And nobody liked that, so they ignored it. They did their very best to just move on past it. By 2025, we're starting to see real

Software development sees real agentic results

23:46 results in programming, software development especially. You know, there's been a lot of noise and a lot of excitement and hype, but not a lot of products. And I make a lot of podcasts and I'm used to new technologies coming on the market and many times I get in the same situation where I'm happy to talk about a new technology on the show when it's new. But I can do one or two of those shows about how cool it's going to be. After that, you have to show me. So, pretty quickly with these LLMs, I hit a point where it's like, "Show me your product or I'm not doing the show."

24:17 And there were no products for quite a while. Till last year, till '25. '25 is when we started seeing

5 Highlight · Real programming impacts from agentic workflows

24:24 real impacts on programming, starting to talk to the sort of best and brightest software developers that are taking advantage of these models to not just get hints for code, super intellisense, but to the point where they're actually, you know, wired into GitHub and are assigning issues to an agent model that can iterate on the code and produce the code results in a pull request. And then you can argue with the tool in the pull request to do further tuning before you accept it.

24:52 And I've seen teams get extraordinary results with these tools. It's not easy. You're still hand rolling a lot of stuff. It still can go badly wrong, but they were the first real results we saw anywhere. And it was part of the proof was just the sheer proliferation of product. Whenever you see this many different groups of people all try to publish software at the same time roughly in the same space, you know it's an unsolved problem with some potential in it. Now half these products already have disappeared. You know, we're in a weird, crazy, frantic ecosystem time of making all these different things.

25:25 So, uh I couldn't even show you them all. And in some cases, some of these have disappeared because the big companies have come along and gutted them. Not bought them, that would be polite. No, just hired all of their staff away. Then their senior leadership and then

25:40 it's like gutted them effectively. I think we when we were just talking last year, in August of '25, we get GPT-5. And GPT-5 was a reminder that we just can't keep getting bigger. I mean, it was kind of a flop of a version for the most part, but there were several different forces acting at the same time on GPT-5. And so we we tapered off. It's like, okay, you know, it's not just going to be radical improvement, radical improvement, radical improvement. It's a a small increase in capabilities. One of the things they originally pitched in GPT-5 was that it was very literary, that you could have had you could write you could ask it to write in any language style, you could ask for it to be written as Wordsworth or Shakespeare.

26:25 Turned out it was terrible at that, actually. It was later sort of revealed that the way they actually trained up the model for doing the literary thing is by exercising against another GPT, and the fact that the sentences made no sense didn't bother GPT at all. It thought the results were great and gave it all green lights. Then humans used it and went, "Wow, this isn't English." Uh the tuning continues.

26:51 Anthropic's now new the new and shiny, they've really had the the wind lately. Their battle with the Department of Defense or Department of War, Department of Defense uh helped them publicly. They were trying to be the good guys. And their results with Claude are impressive. There's no two ways about it. They for the from the dev side of things, using their agentic models, we getting pretty good results. So, they're

Unprofitable startups rely on constant promises

27:15 the new and shiny at the moment. In the pundit space, we talk about we're going back to the dot-com boom. Are these companies Netscape or are they Google? Because today most people don't remember Netscape. They didn't survive the dot-com boom. They came out they got acquired by AOL which also didn't survive. Uh but Google also started in that time span and by 2020 20 2004 had the IPO because they had so many shareholders and became the success you see today. So, they came out of the dot-com boom, but they were one of the ones that they came went on. So, we're looking at all these little companies and saying, well, one of the which one of these are going to survive? Which one of them are going to be the next Google? And which of them are going to be just another Netscape?

27:58 Don't know the answer to that right now. I think most people would pick Anthropic as the survivor at this particular moment. But wait a month. All these companies make a lot of promises. I'm told by most folks, well, it's moving so fast. It's like there are people out there that are really incented to have you believe this is going quickly. And if you want to give yourself a break, do what I do, which is the beginning of each month, I make notes about all what all these companies are promising. And then I compare them to promises the next month. And let me tell you something, they're the same promises every month.

28:30 They aren't progressing that quickly. What they've learned to do is fill the pipeline with noise. Because that makes it feel like it's going quickly so that you won't think too much before you spend more money with them or invest with them further. None of these companies are profitable, not even close. The big tech giants aren't profitable in this line of business. They have other sources of money, so they're fine, but these companies are far from profitable. So, they can't report to their investors profit. That's not an option for them.

28:59 They have to report other things. They have to come up with metrics that keep their investors happy. So, you know, when these gigantic tech models they talk about the interactions, what pieces have to interplay with each other. They talk about a revision of software entirely, thinking there's a case here for how is software going to change? Does the SaaS model even make sense? When we've got tools that are really good at interacting with different data layers to pull data from it, are we really locked into our ERP systems anymore?

29:27 How badly do I need a a a a a CRM these days when I can surf through emails to figure out who I've last contacted and what I should say to them next. There's lots of changes going on in all these spaces. So, while we're coming through this cycle, and I think we're headed down the trough of disillusionment just fine, far from bottomed out but on our way down, we're starting to think seriously about what does this really mean? What's the real value in all this? You're seeing this with the price changes.

29:55 Right? That now these companies can't just keep hyping for the world, they have to start show maybe we can get a profitability, decrease the free use, increase the paid use, increase the cost. But it also begs the question, like what is the hype cycle necessary? Cuz we've had new technologies come on the field recently they didn't go through this. Why is this being hype cycled the same way the dot com boom was hype cycled?

30:23 And one of the things I find in parallel between the two, having lived through

Data center overbuilding strains the economy

30:26 them both, is that a key part of what made the dot com boom hype cycle work was expanding the infrastructure of the internet. It wasn't the part that people talked about, but a tremendous amount of money was spent on building out new data centers, undersea cables, and a lot of those companies lost a lot of money. The end of the dot-com boom, most of those companies went broke. And that's why we ended up with like call centers in India because the cables were cheap. They'd been bought up for 10 cents on the dollar.

30:56 And here we are again in this big hype cycle. So, I'm looking, well, what what are we really spending money on? And we know what it is. We haven't talked a lot about it, but it's more and more in the center of people's minds. It's data centers. Now, there are three big players. All right, it's Amazon, Microsoft, Google. Google even a distant third. Then you have tertiary players, the the Metas and Apples and uh of the world. I won't even really talk about X cuz goodness knows.

31:25 One of the interesting truths about the data center situation was that previous to ChatGPT, even coming into the pandemic, it was getting really hard to build data centers. People don't like them. Municipalities don't like them. They don't employ very many people. Take up a lot of space. They're kind of noisy. Consume a lot of electricity. Fair bit of water. And so, uh if you look at the annual report for Microsoft in 2019, they include a a section warning, "Hey, our demand for cloud is going up, but we can't keep up with our data center builds. We're just We're struggling to find places where we're allowed to build them." Cuz bit by bit, the municipalities are recognizing we have to put more controls around these things. And goodness knows the hyperscalers don't want controls, so they're always looking for other places that didn't want to put controls in place. Then the AI hype cycle hits. And all of a sudden building data centers is very cool.

32:13 And the hyperscalers dump astronomical billions into building data centers. The numbers are absolutely staggering year over year to the point where they're literally distorting the US economy. For the most part, since the new administration kind of came has come in, the US economy has been declining. But the spend by the hyperscalers is in the hundreds of billions and literally has filled the gap, kept the stock markets up.

32:41 So that it's hard to understand how badly they're actually doing. In 2024, more than half of the increase in value in the S&P 500 came from seven companies, the Magnificent Seven. Today they're called the Magnificent 10. 10 companies out of 500 or is that representing almost half the value of the largest index in the world, the S&P 500. Now that's 80% of the stock values in the US, but it's 50% of the stock values worldwide.

33:11 And it's 10 companies. And they're all popped up for one reason. And when that reason

6 Highlight · Circular funding through AI self-dealing

33:18 doesn't make sense anymore, that's a pretty big hit for all of us. The other part of this is what Microsoft started with their investment in OpenAI. It's what we call self-dealing. Nvidia's leading the charge on this right now. They've got a huge bank of cash, $5 trillion. You know, when you buy stock in Nvidia, they get the money. They have no way to use that much money. It takes too long to build more fabs. That's at least 5 years. You got to put that money to work today. This is the Bloomburg diagram where they're showing that what Nvidia is doing is taking investments of 100 100 million, 200 million, 500 million into these little AI companies. And the job of that AI company then is to take that money and buy Nvidia chips with it. Not that they can deliver the chips, just make the order.

34:02 Because then the $500 investment is worth $500. And the $500 order also worth $500. I just made $500 into a billion dollars. I'm a genius. It's one of the signs of you've over invested. You have too much money in play. You can't build the things that we really need to build, but you have to keep the money working to keep the investors excited.

34:27 And so we have these past curves

34:31 of how a bubble burst. Where the over investment continues to a sort of peak point. That peak point I think is already hit. The peak point here, this new paradigm point, the reason it started going back down, see if my pointer will do its thing, there it is. What happened here? In that peak point, I'm not going to play with that anymore. Was last fall when all of the memory companies, TSMC, Micron, all those guys said the same thing, we're not doubling production at RAM.

35:04 We takes us three five three to five years to build a new fab. We don't think you guys are going to be around in three to five years. You know, you've been ordering up chips like crazy. The game that's being played on the data center side is that they're using sub third parties to buy up land. And as soon as they get the land and an agreement to use it, then they make the request for the power and they order up the the concrete and stuff to build the building and they order all the computers. And they're doing it the same way you would buy Taylor Swift tickets. You open 10 browsers to get one pair of tickets.

35:34 So they're ordering far more than they possibly need on the chance they're going to get to build any of it. There's now organizations out there going around checking these sites to see if any construction going on at all. And to be be sure, there is some. In some cases there's none. In some cases there's like a guard girder put up so you can call it underway. But they're over building and they're over ordering. And the memory companies said in in last year, yeah, we we think you're over ordering and you're not going to be around here. So we're not building that many chips. And at that moment the lights put to paid. The projection you made about how much compute you're going to need in the next 10 years cannot be delivered on. There isn't enough RAM and there's not going to be.

36:16 And you saw for the first quarter this year all the tech giants profit values go down a bit. So and if you go look at where the trades were, it was all the institutional investors, the big pension funds and things. They're not running for the hills, they're just backing out a little. They're deleveraging. And that's that first curve down. And it hits that ba- loop right around April. When we start hyping data centers in space.

36:45 And so we've slowed the decline again while we try and figure out if that money makes any sense. And if you want to get my opinion on data centers in space, I'll do that this afternoon in room one. Is this really as bad as the dot-com boom? Well, let me show you another graph. So this is a graph of overall PE ratios. So this is price to earnings ratios. And price to earnings is way out of whack right now.

37:11 But it's not as out of whack as it was in 2000. But it's the second highest it's ever been. So only the dot-com boom where price earnings is wildly out of whack as they

Psychological risks and ethical hazards emerge

37:21 are right now. I don't know if that's the same or not. It's or seems like the same. And that the prospect of the decline is serious. Like that's going to have impact. There's no two ways about it. The other impact is on people. So let's talk about what's being done with the software that's harmful. We call it chat GPT psychosis. Look, if you're in tech, maybe you're less vulnerable to all this. But let's look at this guy. Does he look like a tech bro?

37:51 That's cuz he is a tech bro. This is Jeff Lewis. He's one of the He's one of the the partners at Bedrock Investments. They invested in OpenAI and Vercel and a bunch of others. You may have seen him in the news because he's been spending entirely too much time with a chatbot and got himself into a psychotic episode where he put out videos about how a non-governal governmental system, not visible, but operational has targeted him.

38:19 Now he's getting help, and I'm glad. You know, nobody deserves to to be mentally ill, but it's serious. It's dangerous. It's cost lives. People have killed leveraging chat GPT and been killed. And the tech companies know it. One of the aspects of GPT-5 that helped it flop was that they dialed back the obsequiousness. You know, this software is designed to maximize engagement. And the way it maximizes engagement engagement is by always being positive reinforcement, always being syncopatic. Like, "You're really onto something now, and you're really thinking now, and that's an awesome idea. Let's run with that. I've got more to do."

39:00 So that you'll keep using the tool because the metric they take to the investors is user engagement. Cuz they can't show profit. There isn't any any. They can't show results cuz there's hardly any. So they use engagement to the point where people are dying. And so they dialed it back in GPT-5. They also tried to shut down all the other models, only use GPT-5. It's more efficient. It's more advanced. And people didn't like it.

39:29 And one of the reasons cuz it wasn't obsequious enough. They wanted their buddy. And so Altman gave in, turned 40 back on, and you got these posts. "My baby is back, and I'm crying." And I don't care. I don't care if you think I need help. Cuz my baby's back. It's terrifying. It's also maybe a brief period of time.

39:56 This represents huge liabilities these companies. So they're going to need to deal with this one way or the other. But also, the population gets educated. This is not your friend. It's also not your therapist. Go touch some grass. Put down the software. Remind yourself that it is software. Right? It's the price of the deception to do the fundraising. We call it artificial intelligence. We lean into the science fiction so you'll give us more money. Who cares who dies?

40:28 We'll fix it. We have to. There's really no choice about all of that. I'm also not particularly worried about generalized intelligence or a super intelligence. Again, it's been a marketing engine. They've literally leveraged this at OpenAI. The chant internally is AGI AGI AGI. Cuz that's how they recruit scientists. With no real evidence that the technology can ever deliver. Actually, there's pretty good examples that it can't. The fact that on average people who get more knowledgeable with this tool believe in it less, they trust it less, is a pretty good sign. The folks that are most excited about this are the ones who know the least about it. If you really want to dig in deeply, there are great books in this space. I like Roy Blount's book.

41:07 It's a hard read. It's worthy. I read a I would read a chapter and I'd have to sit and think for a couple of hours cuz he really encouraged me to think about what intelligence actually is. What sentience looks like. Like what it would possibly mean to have real intelligence. We press against this all the time with these tools, right? The current one is the car wash scenario. Hey, I got to get my car washed today and the car wash is only 50 m down the road. Should I walk or drive?

41:37 Because the software has no, quote, common sense. So, it's going to go 50 m. Just walk it. Leaving out the part where you need the car to wash it. And then there's the deep fakes. And I don't know if we're going to get audio for this, but you've probably seen this one. This is the original deep fake from 2017. Right? This it was done at the University of uh of Washington and it's an Obama animation. It's being voiced over by Jordan Peele.

42:08 And it's very funny. It's It's almost 10 years old. And I'm less worried about deepfakes today than I am in the past, mostly because people have gotten way more skeptical than they were in 2017. It's way easier to make them. There was a guy on Twitter who was I'm waiting for, you know, guard Donald Trump to be arrested. Meantime, here's some animations of him being arrested. Cuz the two of you no longer need a group of scientists and and months to generate the thing. It was a tool you could run off of the back of Twitter.

42:40 These are powerful tools. They can do inappropriate things. There's a case for better regulation about what's acceptable and what's not. I think for better or worse, the EU is leading in this space with their current round of regulations. I hope there'll be more enforcement. I think that's the that's the struggle we have right now. And I encourage all of us in this industry with our skepticism to make yourself available when you have an opportunity to work with politicians. Call your MP. Encourage them to make good choices and be a knowledge source.

43:10 We're the experts, for better or worse. To be clear, nobody has 10 years experience with LLMs. They've only been around for three. So, we're all starting out, but most of us know more about this than the average person does. And we choose what happens with these technologies. We've made evil technology before. Anybody remember Greyball? Uber? This came and went.

43:37 So Uber subverted the taxi industry. And to be clear, the taxi industry deserved to be subverted to some degree, right? It's its own kind of monopoly. And so, here running off a phone with a regular car, you could sign up and you could drive people around. It's not safe. It's not wise. They also got over-invested. They got $60 billion. Nobody makes good choices with $60 billion. And as they were struggling to get into certain markets cuz they are well protected, they wrote this little nugget of software, this Greyball. And what Greyball did it it had a list of all of the regulators' phones. And so when a regulator would pull up Uber, there'd be no cars available.

44:11 So, it wasn't a threat. When regular people pull it up, sure, the cars were fine, they could use it. And they kept that going for several years, so that regulators didn't think it was significant and just allowed Uber to exist until it was everywhere. Until finally one of the developers leaked it to the press, and it got pulled off the store briefly, and Kalanick got in trouble, but it all got cleaned up eventually, it went away. I mean, it's not the only scenario like that. Again, we're putting it behind us, but what Cambridge Analytica did for Brexit is a horror show.

44:40 Where they exploited the data in Facebook to lie to people really effectively. Right? Back when it was hard to tell that if you tailored an ad to an individual on Facebook, that individual could not tell they were the only one seeing that ad. And they used that to help people make bad choices. This is what software can do. It's always been a risk. And these new tools represent an even larger risk. And some folks say, like, maybe we should just back away from it. Like, not go here.

45:09 It's not worth it. And I disagree. I think the tech is valuable, it has to be

Protein folding delivers genuine scientific breakthroughs

45:13 well used. And I want to give you the best scenario, the one that for regular mortals even, I can make an explanation which is like, you can't back away from this, it's done too much good. So, let's talk about DeepMind. So, this is Demis Hassabis. He's out of the UK for a long time. DeepMind got acquired by Google. He refused to move to Silicon Valley, which I think helped him. Kept his team there. It also was a pandemic, so he had an excuse to stay in place. And if you remember, his original product was playing video games.

45:39 So, he was building a a a machine learning model that could play video games. And it figured out Breakout first, where it learned cut down cut out the side, and then get the ball in behind. Like, it was clever. It was kind of cool. It was video games, it wasn't important in any way. But he decided to aim himself at a more difficult game, to aim at Go. Go was always the game we thought software could never play, too many combinations, impossible to play. So, AlphaGo was trained over several years against all the best games ever played in the world until finally it produced a player better than a human.

46:10 And that was a profound moment. Again, it's just a game. He didn't stop there. They went on with AlphaZero. AlphaZero is now an adversarial model where you take two entities, two agents, with the rules of the game, they just play against each other until they build up their own play style that could beat the original AlphaGo. And it didn't play like a person at all. AlphaGo was just the most optimized way that people could play. AlphaZero was a different beast entirely.

46:36 Still just a game. So, then he targeted protein folding. And if you're in biology, you know there are 20 amino acids that are are relevant in this equation and they assemble into different ways to create a variety of proteins. And as those different aminos bind together, they twist up. They fold. And protein folding is one of these insanely complicated things in biology that we're still just figuring out. Uh John Kendrew was the first scientist to actually work out the protein folds for a single protein set, the myoglobin set.

47:06 He got a Nobel Prize for it. It took him a decade. And over the past 60 years or so, tens of thousands of biologists using various techniques like lay like x-ray crystallography have worked out roughly the assemblies about 150 different protein sets. Every one of them leads to new medicines, new understandings the way the body works. Could save lives. The problem is there's if you take all the different ways they fit together, it's 10 to the 35 possible combinations and that's a lot.

47:36 So, we've made games. There's a competition called CASP to work out if you can figure out different protein foldings. And over the 20 years or so that contest has been going on, there wasn't a lot of results. It's really hard to build software in this space. In the middle aughts, there was actually a game for people called Foldit where you could do experiments and and measure them against the actual results that's in the different proteins. And a group of gamers over a period of 2 years actually worked out a protein, the

7 Highlight · AlphaFold revolutionizes biological research

48:01 AlphaFold protein, which is cool. One. So, Hassabis and his team aim at protein folding. Over the Over about 4 years, by 2020 they were at 90% accuracy. By 2022, they were pretty much dead on. They now knew how to You could give them any protein set and they could work out the protein folds for it. What do you do with it? He computed the 200 million most common protein fold sets possibly needed and published them for free.

48:33 And changed medicine fundamentally. We are already The new treatment we have for leukemia comes from that. There's a new malaria vaccine comes out. There's three new antibiotics that have come from that. Fundamentally changed medicine. It will take us decades to even know everything that we've gotten from this data set. This was only possible with this adversarial training model of generative AI. It's extraordinary. It's made a huge difference in the world to the point where we're just even struggling to comprehend what we've just

48:59 done. We choose what we do with these tools. It's entirely up to us. We have a lot of choices here on the where we take these tools from here. Get back to the fundamentals. Focus on the things that are important. This is not the first time we've been disrupted as an industry. It's happened over and over again. It's also wrapped in a hype cycle, which makes it harder. You get back to the fundamentals of what you care about. Our job was never to write code. It was to solve people's problems.

49:25 They couldn't code their way out of the protein folding problem. So, instead they built a model that solved it. Maybe that's the answer to the problem you're dealing with. I don't know. What I do know is that every time we get disrupted like this, we go back to the things that are the most important. We make good choices and we get great results and we can help people. It's a crazy time, I admit it, but it will pass and we'll be back to work. The work will look different. That'll be okay. We'll still make a difference.

49:51 Thank you for spending your time with me.