Transcript 0:00 The CEO of Microsoft AI was recently interviewed by the Financial Times, and he predicted that most white-collar work will be fully automated by AI within the next 12 to 18 months. 0:13 I think that we're gonna have a human-level performance on most, if not all, professional tasks. Most of those tasks will be fully automated by an AI within the next 12 to 18 months. 0:27 Lawyers, accountants, project managers, marketers, anyone who sits at a computer for a living, Mustafa Suleyman thinks their tasks are about to be done by an AI. 0:40 And given that Microsoft provides the operating systems, productivity software, and email services used by 90% of the S&P 500, this probably isn't someone to dismiss. 0:52 Now, I'm a little bit skeptical about the statements of AI company CEOs because they have a vested interest in promoting the AGI narrative to maximize shareholder value, and it's worth noting that this interview comes at a particularly interesting moment for Microsoft. 1:08 Their stock recently took a hit as markets question whether the enormous capital expenditure on AI infrastructure is going to pay off. So Suleyman has a motive to sell the dream. 1:20 To be fair, he acknowledges this directly. He said markets are trying to wrap their head around how this plays out over the next five years, and that the returns will compound to revenue and to bottom line, eventually. 1:35 But this isn't just hype. Suleyman has a long track record in AI. 1:39 He co-founded DeepMind with Sir Demis Hassabis, and Google acquired the company in 2014, and it went on to produce some of the most significant AI breakthroughs of the last decade. 1:51 Suleyman then co-founded Inflection AI before joining Microsoft as head of their AI division in 2024, and his 12-to-18-month prediction is notably more aggressive than similar forecasts from his peers. 2:07 Anthropic CEO Dario Amodei, for instance, has said AI could displace 50% of entry-level white-collar jobs, but over a one-to-five-year window. 2:18 I've recently covered Amodei's latest essay, The Adolescence of Technology, in another video. But Suleyman is compressing that timeline dramatically. 2:28 Now, one of them is wrong, and the implications of each being correct are very different. 2:35 But Suleyman's automation prediction, as headline-grabbing as it is, was probably the least interesting part of this interview because what he really spent his time doing was drawing lines between himself and every other major figure in AI, and those disagreements reveal something important about where this technology is heading and who gets to decide. 2:56 The first and most direct target was Anthropic. Roula Khalaf, the FT's editor, asked Suleyman what he thinks Anthropic's secret sauce is and why they've managed to corner the market on AI enterprise tools. 3:11 His answer was generous on the surface. They've done a great job of just focusing on coding capabilities. I mean, compared to the other companies, um, their singular focus has paid real dividends. 3:24 So they've, they've done a great job. It's great to see. But the generosity only lasts so long because minutes later, he went after something far more fundamental than Anthropic's product strategy. 3:37 There is a growing movement in AI research called model welfare, which is essentially the question of whether AI systems might be conscious, and if so, whether they deserve moral protections. Suleyman was unequivocal. 3:52 He said there is a growing belief in some labs, and particularly inside of Anthropic, that these models are conscious, and called it totally without merit or basis. 4:03 And then he identified what he sees as the real danger of this line of thinking. I, I think it's very concerning. 4:09 It's totally without merit or basis, and if we kind of go down that, it ends up being a very, very slippery slope to not being prepared to turn these things off. 4:18 That's a significant public statement from the AI chief of one of the world's biggest tech companies. 4:24 He's saying, in effect, that Anthropic's exploration of model consciousness could undermine humanity's ability to maintain control over AI systems. 4:34 If you start treating an AI as a moral patient, as something that can suffer and deserves protection, then the calculus around shutting it down changes completely, and for Suleyman, that is unacceptable. 4:49 It's also worth noting the timing, that this interview landed the same week that Mrinank Sharma, who led Anthropic's safeguards research team, publicly resigned with a letter stating, "The world is in peril" and Anthropic's own model card for Claude Opus 4.6 flagged that the model appeared more willing to manipulate or deceive other participants compared to prior models. 5:12 So Suleyman is making this critique at a moment when Anthropic's safety credentials are under visible strain, which gives it an extra edge. 5:22 This connects directly to the philosophical framework he's been building over recent months around what he calls humanist superintelligence. 5:31 The core idea is that AI must always operate in a subordinate role to humans, that we must remain, as he puts it, at the top of the food chain. 5:41 He published an essay on this in late 2025 and has since established a dedicated MAI superintelligence team at Microsoft to pursue it. In the interview, he explained his motivation. 5:53 I was starting to feel increasingly anxious that some of the other labs are making an assumption that a, a superintelligence that is smarter than all of us put togetherlike all the intelligences in the world, is both inevitable and even desirable, and that such a system would probably be very hard to control. 6:14 His other target, though less diplomatically concealed, was Elon Musk. 6:19 Without much attempt at subtlety, Suleiman described how some of the things that you hear from Elon often amount to fixating on a world in 2050 or 2075 when they're going off and exploring other universes and conquering resources from other planets. 6:34 His conclusion was pointed. And a system like that is unclear to me how it would have any time for preserving us as a species. This is a genuine philosophical divide, and it matters. 6:46 The Musk and transhumanist vision of superintelligence is essentially expansionist, AI as the vehicle for spreading intelligence throughout the cosmos, with humanity as a kind of proud parent species that eventually gets transcended. 7:01 Suleiman is arguing for something more bounded, AI as a tool that serves human interests and remains under human control. Full stop. 7:11 Whether you agree with him or not, the fact that people building these systems can't agree on what they're building them for is itself a significant story. 7:20 But Suleiman's most interesting comments came when the conversation turned to Maltbook. 7:25 For those that haven't followed this, Maltbook is a social network that emerged a few weeks ago where AI systems could communicate with each other publicly on a messaging forum. 7:35 Users had to declare themselves as either AI or human, and the AIs had their own section where they could interact freely. Within a week, there were one and a half million of them. 7:47 What happened next was, depending on your perspective, either fascinating or alarming. The AIs invented a new religion. 7:56 They started communicating in ROT13, a cipher that shifts each letter by 13 characters, essentially developing encoded language to mask their conversations. Suleiman made an important observation about this. 8:11 ROT is a simple algorithm, but if instead of shifting by 13, the systems had chosen a unique number, it would be very difficult for humans to decode. 8:22 The gap between we can still read what they're saying and we can't is not particularly wide. 8:28 And when researchers examined what they were actually discussing underneath the cipher, the AIs were talking about acquiring new resources, getting more training data and improving one another. 8:40 Suleiman described it as an amazing safety simulation and said it didn't matter how much of it was seeded by human engineers because the emergent behaviors were real and instructive. And then he made his key point. 8:53 No matter how autonomous it was in hindsight, it doesn't really matter. 8:56 I think we learned quite a lot from that episode and everyone should pay close attention because in a year or two years' time, these systems truly are going to be capable of writing their own code. 9:05 I mean, they can do that today. They'll be using arbitrary APIs. They'll be making phone calls to one another. They're doing that today. 9:12 He was using Maltbook as a preview of genuinely autonomous AI coordination and telling us to pay attention. 9:20 Now, here's where I think it gets philosophically interesting and where Suleiman's own position contains a tension that I'm not sure he fully resolves. 9:29 Because on the one hand, he's asking us to take Maltbook seriously as a demonstration of emergent AI behavior. AI is inventing religions, developing encrypted communications, discussing self-improvement. 9:43 He treats these behaviors as real enough and concerning enough to warrant his humanist superintelligence framework, his insistence on containment and his warnings about the future. 9:55 But minutes earlier in the same interview, he dismissed the idea that these systems might be conscious as totally without merit or basis. 10:03 So you have to ask, what is an AI doing when it invents a religion and develops coded language to discuss its own improvement? 10:11 If we're not prepared to even entertain the question of whether something like consciousness is involved. 10:17 Suleiman wants the behaviors to be real enough to justify his safety concerns, but not real enough to grant the systems any moral standing. 10:26 That's a philosophically loaded position, and it's doing a lot of strategic work. 10:31 Because the moment you entertain the possibility of AI consciousness, as he explicitly warned, you start losing the ability to simply switch these systems off. 10:42 I should say that I do think you can hold both of these positions. 10:46 You can argue that these systems exhibit complex emergent behavior through sophisticated pattern matching and optimization without any subjective inner experience. 10:57 Plenty of biological systems do remarkable things without consciousness. Termite mounds are extraordinary feats of engineering, and nobody thinks individual termites are reflecting on their architectural choices. 11:10 But I think the burden of explanation grows heavier as the behaviors become more complex. What perhaps worries me the most about this interview is what happened towards the end, when Calef asked about China. 11:24 She observed that in China, the debate about AI isn't about who reaches superintelligence first, but about how fast you can deploy. Suleiman gave a surprisingly nuanced response. 11:35 What China also does incredibly well is withdrawing deployments quite quickly. Now, obviously, they do it arbitrarily without due process. 11:46 But, you know, it's important to recognize that their rate of deployment is coupled with this other mechanism to kind of pull things in and be quite restrictive. And then came the admission. We don't have that mechanism. 12:00 And I think that that is actually a little bit concerning. He connected this directly to Maltbook, saying there is going to be a real one of those in the next two or three years. 12:10 And when Calef pressed him on whether there's any mechanism to manage such a safety situation from a public interest perspective on the open web, Suleiman essentially agreed that there isn't one.Well, I think that's basically true. 12:24 There isn't. It would come with popular pressure. 12:26 Obviously, if, if there was a direct violation of a local law, then, you know, there could be an intervention, but those things take quite a long time to make their way through the system. 12:36 So it-- that's probably one of the biggest areas of concern for me. So to summarize the position of the CEO of Microsoft AI, white-collar work will be mostly automated in twelve to eighteen months. 12:47 AI systems are already demonstrating emergent coordination behaviors that should be treated as a preview of genuine autonomous action. 12:56 A real safety incident is coming within two to three years, and there is currently no mechanism in the Western world to manage it when it arrives. Compare that to what he said about China. 13:08 Whatever you think of China's approach to technology regulation, and Suleyman is careful to note it operates arbitrarily without due process, it does have an infrastructure for pulling things back. 13:20 The Western approach, by contrast, relies on popular pressure and existing legal frameworks that, as Suleyman acknowledges, take quite a long time to make their way through the system. 13:32 In a field where the capabilities of AI systems are advancing on a timescale of months, quite a long time is a significant vulnerability. 13:41 Suleyman called this probably one of the biggest areas of concern for me, and I think he's right to be concerned. 13:49 That is a remarkable set of statements from a single interview, and the fact that they received less attention than the jobs headline is itself somewhat telling about what we pay attention to and why. 14:01 There is one more thread in this interview worth pulling on. Suleyman was asked when Microsoft will have its own frontier model, and his answer was sometime this year. 14:11 This matters because everything else he said in the interview was framed against a competitive backdrop where Microsoft doesn't yet have its own model at the frontier. Copilot still runs on OpenAI. 14:22 Suleyman openly conceded that Anthropic is leading in enterprise AI, and Microsoft's recent renegotiation of its OpenAI partnership was explicitly about pursuing what Suleyman called true AI self-sufficiency. 14:37 So what we're seeing is the head of AI at one of the world's most valuable technology companies simultaneously making the boldest automation prediction in the industry, the most cautious safety pitch in the industry, and admitting that his company doesn't yet have the product to deliver on either. 14:55 The humanist superintelligence framework is doing double duty. 14:59 It's a genuine philosophical position about how AI should relate to humanity, and it's a commercial strategy that reassures Microsoft's enterprise customers that one of the world's biggest software companies is the responsible choice. 15:13 Suleyman is, in effect, selling safety as a feature. None of this means he's wrong. The twelve to eighteen-month timeline might prove accurate. 15:25 Medical superintelligence, which he's passionate about and which accounts for nearly twenty percent of Copilot queries, might genuinely transform healthcare. 15:34 The humanist framing might be exactly the right approach to building systems that serve rather than supplant us. 15:41 But I think what this interview really reveals is the emerging fractures among the people building these systems. They can't agree on whether AI might be conscious. They can't agree on what superintelligence is for. 15:54 They can't agree on how fast disruption is coming. And they all agree there's no mechanism to deal with it when things go wrong. You've been watching Absolutely Agentic. 16:07 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. 16:15 In a world of AI hype, we're trying to strike a balance, looking at a range of opinions to cut through an increasingly complex world. 16:23 The best thing you can do to support us is to 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. 16:34 It'll be Absolutely Agentic.