Bill Gates on why he broke a decades-long habit of optimism to write a memo about AI: "If you're used to my memos being 100% optimistic, I will say this is different."

Sir Demis Hassabis, Chairman of Google DeepMind, on the single idea behind AlphaGo, AlphaFold, and AGI: Go and protein folding look like completely different problems. One is an ancient board game. The other is one of biology's hardest puzzles. @demishassabis sees them as the same kind of problem. "You're effectively navigating there's a needle in a haystack solution, but the haystack is, is unfathomably large, right?" The numbers explain why. In Go, there are more possible board positions than there are atoms in the universe. Proteins are even more complicated: an average-size protein has 10 to the 300 possible shapes. "So you can't find the right shape by looking at every single shape. You're going to have to do something much smarter." His team's approach comes down to three ingredients. 1. Build a model of the domain Use deep learning, build neural networks, and feed them the data you have for that domain. For Go, that means games. For proteins, that means the structures already known. The system has to learn how that world works. 2. Search smartly, guided by the model Once the model exists, you still search, but selectively. "In a Go game, instead of you have 100 possible moves in a position you don't search all of them... The model tells you which ones are the most likely, the most fruitful ones to look at. So you can spend your compute budget on very useful alternatives." That is how an impossibly large problem becomes something that can potentially be solved very quickly. 3. Define a clear objective "The final thing you need is an objective function. You need to have a clear goal you can hill climb against." This is why DeepMind started with games. Before Go, the team used Atari games as the earliest measures of how its systems were doing, because the goals were obvious: win the game or maximise the score. (Demis also admits he simply loves games.) Science has clear objectives too. Minimising the energy in a system, for example, shows up often in biology, chemistry and physics. Put those three pieces together and you get what Demis describes as "a very generic solution to many problems if you can couch it in this way." And in his view, that generality is the point. Demis says this kind of general solution, applicable to many, many problems, is what AGI should be.

Nobel laureate Geoffrey Hinton has a view many AI skeptics won’t like: LLMs think in much the same way humans do.

Mustafa Suleyman on Microsoft AI's Science cover paper, diagnosing very rare conditions from health records, radiology and pathology images: "You know, we can do that better than humans."

Google DeepMind CEO Demis Hassabis on why a reputation for accuracy kept Google from launching the first big chatbot: Before ChatGPT came out, Google already had the technology. Both DeepMind and Google Brain had large language models, and Google Brain even had a very early product called LaMDA. "There'd been a conscious decision not to release that. And it was obviously in retrospect, probably the wrong decision, but you can sort of see." Demis wasn't involved in that call. Back then, before Google DeepMind was formed, DeepMind was running independently, mostly as a research organization. But he understands exactly why Google hesitated, and it comes down to what Google stands for. "Google's known for accurate information. That's what Google is for. Google search is for organizing the world's information. That's its bread and butter." The problem was that those early models were very flawed. Demis points out that even today's chatbots make mistakes and hallucinate, and the early ones did it a lot. For a company whose entire brand rests on getting information right, that was a hard pill to swallow: "So I think it's sort of doubly tricky to think about putting out a bot that 20% of the time gives you the wrong information. I think there's a clear issue there for a company like Google." OpenAI had no such reputation to protect. It didn't have any existing products that relied on reliability, so mistakes carried far less risk: "Whereas obviously a startup, especially someone with someone like Sam leading it, it's like, go for it. You only live once, yolo it, and it's fine." @demishassabis adds that even OpenAI saw ChatGPT as a research experiment. Nobody knew how big it would get, and he says the name gives it away: "ChatGPT is a programmer's name. They would have named it something catchier if they knew how big it was going to get." Here's what Google's accuracy-first mindset missed. Once millions of people started using these early models, they found plenty of uses where accuracy wasn't the point: Summarizing things, helping with homework, brainstorming. "So there were these sort of low hanging fruit that it didn't matter, do some brainstorming, it didn't matter if it made some errors that weren't to do with information seeking. And these were new use cases that hadn't existed before." Demand for those use cases turned out to be huge, and that's what caught everyone by surprise. "I think that's an interesting thing of where you can be sometimes too close to your own technology."

Anthropic co-founder Jack Clark's $44T-by-2030 GDP number will get clipped, but he hand-waves it as "one model predicts" and admits it "sounds very outlandish." The sharper line is quieter: "We are watching it get much more capable, much more quickly than all of our intuitions." The builders are surprising themselves.

Microsoft AI CEO Mustafa Suleyman says a Nobel laureate's prediction about AI and jobs is "way too cautious." Suleyman was asked about an essay by Nobel Prize-winning economist Daron Acemoglu, which suggests that within 10 years, only 5% of what humans do will be replaced by AI. That's a lower figure than many people assume. Suleyman shared the essay himself, and Acemoglu published it in the Humanist Review of AI, a new magazine Suleyman's team just launched. Still, Suleyman disagrees with the conclusion: "I actually don't really agree with him. I think that that's way too cautious. I mean, obviously he's a Nobel laureate in economics, so, like, full respect. But I think that it's going to happen much faster than that." He describes these as "enduring and fundamental changes in the way that we work." He grants that AI will create new skills, capabilities, products and services, and that it's very hard to predict all of them. But he believes some sectors will feel the impact quickly: "I think it is pretty clear that in some sectors there's going to be very fast and quite dramatic reform." His example is call centers: "The models are very personable, they're very patient, they're very factually accurate, they're available 24/7. Like, obviously they're going to be adopted very, very quickly. And many millions of people around the world do those kinds of jobs. So it's very clear that they're going to displace some industries really quickly." For @mustafasuleyman, the priority is responding fast when that displacement happens: "What we need to be ready to do is to intervene quite quickly to support those people through their transition to other types of work and hopefully to other upskilling and stuff." And beyond the next decade, he expects the change to be even bigger: "If you look out far enough, like 15 or 20 years, it is going to look radically different, in my opinion." Suleyman's view is that AI's impact on work won't be gradual or evenly spread. Some industries will shift very quickly, and the people in those jobs will need support well before the economy fully adjusts.

Elon Musk answers Peter Thiel’s famous complaint that the future promised flying cars but gave us 140 characters: “Well, not anymore.”

Nobel laureate Demis Hassabis predicts a new golden era of science: “In the next 10 years we’ll have almost a second renaissance.”

Jensen Huang was asked whether sovereign AI still counts if it runs on American chips, he glides past the question and redefines the terms: "Sovereignty basically means that you have to control your own experience. And in the case of AI, that means control your own data. Your data, your life experience, your company experience, your country's culture, values, the institutional knowledge of your people. The flywheel of creating more data, more experience, more intelligence, that belongs to you. And the computer is just an instrument." Then in the next breath: "nobody could deny you service... you can't ever have a denial of service... you have to control that." If denial-of-service is the test, the chips are the single most denial-of-service-able part of the stack right now. By his own definition, that's exactly the layer you don't control.