Transcript 0:00 [gentle music] The CEO of Anthropic, the creators of the much-loved Claude chatbot, just published an essay with a big claim. 0:07 Dario Amodei believes that powerful AI systems smarter than Nobel Prize winners across most domains could arrive as little as two years from now. He's not alone. 0:20 Sam Altman claimed last year that OpenAI were confident that they know how to build AGI as they traditionally understood it. Demis Hassabis at Google DeepMind puts it at a 50% chance by 2030. 0:35 Elon Musk thinks we'll have smarter AI than any individual human by the end of this year. Now, I've made a couple of videos about the AGI debate of late, and to some degree, I think we're already there. 0:47 But there's also a strong case against. It struggles with a definition problem, for one. 0:53 But Meta's former chief AI scientist, Yann LeCun, thinks the current approach is fundamentally limited and that we're decades away and need major breakthroughs we haven't even discovered yet. 1:05 In this video, I really wanted to zone in on Amodei's essays and examine what he thinks unlocking powerful AI really means. 1:13 It's probably the biggest question of our time, [bell rings] and he's not just got a ringside seat, he's in the ring itself, fighting it out with the other AI companies. 1:21 [gentle music] Now, Amodei's first big essay, Machines of Loving Grace, was published back in October 2024. 1:31 The title is taken from a 1967 poem called All Watched Over by Machines of Loving Grace by Richard Brautigan, and it contains a utopian vision of the future. Amodei's essay is about the upsides of AI if things go well. 1:47 But his latest one isn't quite so positive. He published The Adolescence of Technology in January, and this focuses more on the risks and how we might address them. 1:58 Let me start with the timeline, because it's a little bit concerning. Amodei describes what he calls powerful AI, and he's careful to avoid the term AGI, calling it a marketing term. His definition is more vivid. 2:13 In Machines of Loving Grace, he imagines a country of geniuses in a data center, millions of AI instances, each smarter than a Nobel laureate, thinking 10 to 100 times faster than humans, working around the clock. 2:27 He predicts that we could see AI replacing most software developers within a year and AI conducting Nobel-level research within two years. 2:36 So he's unambiguously claiming that we're on the cusp of a historically transformative AI superintelligence. At Davos in January, Amodei doubled down. 2:46 He said, "I'm more confident than I've ever been that we're close to powerful capabilities in the next two to three years." 2:53 Now, I often think with statements like this, isn't this guy just hyping his own company's products? It's a fair question, but Amodei has historically been one of the more cautious voices in AI. 3:04 He left OpenAI specifically because he was worried about safety, and his essays spend as much time on the risks as they do on the benefits. 3:14 In Machines of Loving Grace, he outlines five areas where AI could transform human life, biology and health, neuroscience and mental health, economic development, governance, and human meaning. His optimism is genuine. 3:29 He talks about compressing a century of medical progress into a decade, eliminating most infectious diseases, and dramatically improving mental health treatments. But he's not naive about it. 3:40 He acknowledges what he calls return on intelligence, the idea that not all problems benefit equally from more thinking power. 3:48 Some things are bottlenecked by physics, by real-world data, social consensus, and the messy reality of human systems. You can't cure Alzheimer's in a day just by thinking about it. 4:01 This brings us to the darker essay, The Adolescence of Technology. Here, Amodei lays out five categories of risk that keep him up at night. 4:11 The first is autonomy risk, the scenario where AI systems start pursuing goals that conflict with human interests. This is the classic sci-fi fear, but Amodei frames it more subtly. 4:24 The danger isn't necessarily that AI systems decide to harm us. It's that as these systems become more capable, we might lose the ability to verify what they're doing and what we think they're doing. 4:37 His proposed solution is something Anthropic calls constitutional AI. 4:41 The basic idea is that instead of trying to hard code every rule an AI should follow, you give it a set of principles, a constitution, and train it to reason from those principles. Think of it like raising a child. 4:55 You can't anticipate every situation they'll encounter, so you teach them values and judgment rather than an endless list of rules. 5:02 Anthropic also publicly released their updated constitution in January, an 84-page document that's expanded dramatically from the original 2,700 words. 5:13 It establishes a priority hierarchy, safety first, then ethics, then compliance with guidelines, then helpfulness. 5:22 The idea is that Claude should understand why it should behave in certain ways, not just memorize what's allowed. 5:29 What's fascinating is that this constitution includes provisions for Claude to refuse requests that would help concentrate power illegitimately, even if those requests come from Anthropic itself. 5:41 It's an attempt to build in safeguards against the very company that created it. 5:46 Alongside this, they use output classifiers, essentially AI systems watching AI systems, flagging potentially dangerous outputs before they reach users. It's layered defense. 5:58 The constitution shapes the model's reasoning. The classifiers catch what slips through.Will it work? Well, we don't know yet, but it's at least a concrete attempt at the problem. 6:10 Now, the second big risk in the adolescence of technology is bioterrorism. 6:15 AI could lower the barriers to creating dangerous pathogens, providing synthesis instructions, and helping with delivery mechanisms, things that currently require specialist knowledge. 6:27 Amodei takes this seriously enough that Anthropic has hard-coded prohibitions in their systems. Certain categories of information are absolute no-go zones, regardless of context. 6:39 The Constitution distinguishes between soft-coded defaults that can be adjusted and hard-coded behaviors that cannot be overridden under any circumstances. Bioweapons assistance is in the hard-coded category. 6:53 Third is what he calls authoritarian capture, the possibility that powerful AI ends up controlled by governments or entities that don't share democratic values. The geopolitical stakes here are enormous. 7:06 Amodei's proposed solution is maintaining Western leadership through compute advantage, specifically continuing chip export controls that limit adversaries' access to advanced hardware needed to train frontier models. 7:20 It's controversial, essentially arguing that the best defense is making sure the technology is developed by people who share certain values. 7:28 Critics would say this is just Western technological hegemony dressed up in ethical language. Fourth is economic disruption. If AI can do most cognitive work, what happens to the people who currently do that work? 7:42 Amodei doesn't sugarcoat this. The transition could be brutal if we don't manage it. His suggestions here are more conventional: progressive taxation, economic monitoring, safety nets. 7:54 But he acknowledges these are harder problems because they require political consensus, not just technical solutions. 8:01 We've made another video on the question of jobs in the AI future, so make sure you check that one out. 8:07 Now, fifth in Amodei's list is indirect effects, the ways AI might destabilize society even without any single catastrophic event. 8:17 Imagine AI-powered disinformation at scale, or the erosion of shared reality when anyone can generate convincing fake evidence of anything. 8:26 What I found interesting is that Amodei actually proposes specific solutions for each category. He's not just hand-wringing about risks. He's trying to think through what we might actually do about them. 8:37 But here's the problem I keep coming back to. If these systems are really arriving in two to three years, do we have time to implement any of this? The Constitution was released a couple of weeks back. 8:50 The economic and political structures Amodei describes would take years to build. Are we building the plane while it's in the air? 8:58 And not everyone shares Amodei's timeline, and I think it's important to take the skeptics seriously. 9:04 The most prominent voice on the other side is Yann LeCun, Meta's former chief AI scientist and one of the founding figures of deep learning, a Turing Award winner, which is basically the Nobel Prize of computing. 9:17 LeCun thinks current AI systems are fundamentally limited. His argument is that large language models learn from text, but text is an incredibly impoverished representation of the world. 9:28 A four-year-old child processes more sensory information about how physics works and objects behave, and how cause and effect operate, than all the text on the internet contains. 9:40 "We can't even build cat-level intelligence yet," LeCun has said, and he's not being dismissive. He's pointing out that cats understand the physical world in ways our AI systems simply don't. 9:54 LeCun's prediction is several years, if not a decade, with a long tail. He thinks we need fundamental breakthroughs in how AI systems build world models, new architectures that haven't been invented yet. 10:07 While Amodei and Altman have talked about years, LeCun has talked about the need for scientific revolutions we can't yet predict. 10:15 Francois Chollet, the creator of the Arc AGI benchmark that's become a key test for reasoning ability, is even more skeptical. 10:23 He puts human-level AI at 2038 to 2048, and has argued that the current focus on large language models may have actually set back AGI research by drawing talent and resources away from more promising approaches. 10:38 So who's right? Honestly, I don't know, and I'm increasingly convinced that anyone who claims certainty on this question is either fooling themselves or trying to fool you. But here's what we actually know. 10:50 The optimists point to a genuine shift in AI capabilities over the past year. OpenAI's o3 model scored 87.5% on the Arc AGI benchmark, a test specifically designed to be difficult for large language models. 11:04 For comparison, humans score around 85%. Previous AI systems couldn't crack 30%. 11:11 Reasoning via reinforcement learning, teaching AI to think longer about problems before answering, has produced dramatic improvements in verifiable domains like maths and coding. 11:22 This wasn't predicted by the skeptics, and it's given the optimists confidence that continued scaling will keep working. 11:28 It's also led investment fund Sequoia Capital to recently state that AGI is here, now, and the deeper reasoning on this point is covered off in my previous video, so make sure you give that one a watch. 11:40 But the skeptics point out that progress in verifiable domains doesn't automatically transfer to messier real-world problems. 11:48 An AI that can solve maths problems isn't automatically an AI that can navigate social situations, make wiser judgments under uncertainty, or understand the physical world well enough to operate in it. 12:00 Demis Hassabis at Google DeepMind occupies an interesting middle ground.He agrees progress is rapid in coding and mathematics, but emphasizes that scientific discovery and creative reasoning remain more difficult. 12:13 His test for real AGI is can it discover new science and come up with something like general relativity from scratch? Market predictions are also interesting. 12:24 Forecasting platforms like Metaculus and Manifold Markets show peak probability for AGI around 2027, with a median of 2031. Now, that's not proof of anything. 12:35 Prediction markets can be wrong, but it suggests the informed consensus is somewhere between the extreme optimists and the skeptics. So where does this leave us? 12:46 I think the honest answer is, in a period of profound uncertainty, the question of powerful AI demands serious attention, regardless of which timeline proves correct. 12:57 If Amoday is right and powerful AI arrives in two to three years, we have almost no time to build the governance structures, safety measures, and economic cushions we'll need. 13:07 The decisions are being made right now, in a handful of AI labs, by a small number of people, and they could shape the trajectory of human civilization. 13:18 If Lacun is right and we're decades away, we have more breathing room, but also the risk of complacency, of assuming the problem will solve itself. But even the skeptics think transformative AI is coming eventually. 13:31 The disagreement is about when, not whether. That means even on the conservative timeline, we're talking about the most significant technology in human history arriving within our lifetimes. 13:42 Amoday ends Machines of Loving Grace with a reference to Ian Banks' Culture novels, science fiction stories about a post-scarcity civilization where humans and AIs coexist. 13:53 He argues that the culture's values are a winning strategy. 13:57 But as one critic pointed out, a consistent theme in those books is the problematic nature of imposing your values on others, even when you're convinced those values are correct. 14:08 The question isn't just whether we can build powerful AI, it's who gets to decide what it does and how we ensure the benefits are shared rather than concentrated. What seems clear is that we're not going back. 14:20 The question isn't whether AI will become dramatically more capable, it's how quickly, how safely, and to whose benefit. 14:28 The title of Amoday's recent essay is The Adolescence of Technology, and I think that metaphor is apt. 14:35 Adolescence is a period of rapid change, of capabilities outrunning wisdom, and potential for both extraordinary growth and serious mistakes. 14:45 How we navigate this adolescence will be the most important challenge of our era. If we get it wrong, there are expert predictions of AI takeover and even human extinction, something I've covered in an earlier video. 15:00 But if we get it right, perhaps we will all get to experience utopia. You've been watching Absolutely Agantic. 15:07 We're a media consultancy and startup focused on AI, and we create weekly YouTube videos and a newsletter we send a few times a week. 15:15 In a world of AI hype, we're trying to strike a balance, looking at a range of opinions cut through an increasingly complex world. 15:22 The best thing you can do to support us is sign up to the newsletter in the description or subscribe to the channel. But also, don't forget to like and leave us a friendly comment. See you next time. 15:31 It'll be absolutely Agantic.