Tech in 2026: Inside the AI bubble

0:01George, when did you get the feeling that there might be an AI bubble? It's hard to pull out the signal from the noise in a place like Silicon Valley where everything is so overwrought with positivity. But there have been a few moments where enthusiasm has maybe tipped over into over exuberance. And I think in particular of a funding round for Miramorati's company, which at this stage was very secretive. No one had any idea what they were doing. Still limited information on what they're doing. This was back in the summer. and we spoke to one investor who said, yeah, we've met the company, but we weren't even allowed into the offices.

0:35We just sat in a vestibule at the entrance to their offices. And this was in a funding round that I think they raised around$2 billion,$10 billion valuation. So very serious funding round where you think you might be given, as a serious investor, some information and some insight on what was happening. According to this investor, they were given nothing of the sort and just sat in this little coffin office outside and told, look, we're a promising company and if you want to invest, then you should. but we can't tell you much more than that. Never crossing the threshold of the office or given a single tangible piece of information about your potential investment, how many millions would you have pledged?

1:11But that startup, Thinking Machines Lab, founded by Mira Murati, the former chief technology officer of OpenAI, hasn't had any issues enticing investors. And it's just one of many companies that has benefited from the incredible amounts of capital flowing into artificial intelligence. This episode is all about the evolution of that AI industry and whether 2026 is going to be the year the AI bubble bursts. I'm Murad Ahmed, the FT's Technology News Editor, directing the work of our tech correspondents in San Francisco, London, Beijing and beyond. And in this two-part series of Tectonic, we'll be looking into some of the biggest questions facing the tech industry this year.

1:58I'm joined today by George Hammond, the FT's venture capital correspondent. He set that Silicon Valley scene for us. Hi, George. Hello. As well as our AI correspondent, Melissa Heikeler. Hi, Melissa. Hey. And Sarah O 'Connor, columnist and writer of the AI Shift newsletter. Hi. So we'll get going on the AI bubble. George, this is your specialist subject. The optimistic take on all of this is these businesses are transformative. look guys you've never seen anything like this this is going to completely transform society and industries and therefore the money will catch up we're all going to get rich at the end of it that's the optimistic view the pessimistic view is that the fundamental business realities will catch up to these frothy valuations and there'll be a spectacular collapse so if we if 2026 is a spectacular collapse how do we get there what are the steps that we might see that will cause that Yeah, I think that's exactly the right framing.

2:56And there's a few things that could cause that or at least cause some jitters on the way to that. And one is that this underlying picture, which is very rapid revenue growth, stops or slows. And we've seen tremors of that in the past year. And if we have that again, going forward, then we could see stock market valuations correct. it could be that one of the massive AI labs in the private markets fails to do a funding round I mean OpenAI is the most notable example here they've had no problem raising money they've raised 42 billion in the last year alone but they have a capital requirement which means they have to go out and do that again and again and again and there's only so many investors in the private market who can supply that kind of money so then they have to do a public listing and as we've seen with previous highly fated private companies such as WeWork, when they try and go from the private markets to the public markets, the degree of scrutiny can be quite challenging.

3:50We might also have something like a deep seek moment from early last year when a Chinese open source model came out that looked like it could contend with some of the best US models. And that caused major tremors in the market, albeit briefly. Melissa, you're nodding. You spend a lot of time talking to the AI companies, particularly the startups, mainly the kind of the scientists and the technologists behind them, but do they talk about whether or not there's an AI bubble or are they happy with all the froth? I mean, it benefits them, or the froth, I mean. I think one thing to bear in mind is that all of this is based on one particular type of AI technology, which is LLM's large language models.

4:28AI is that massive umbrella term covering things from facial recognition to autonomous driving to scientific models like AlphaFold, which is a deep mind model used to determine how proteins fold. And so when we talk about AI, what kind of AI are we hoping will work? And LLMs are the hottest thing. And I think people are sort of tired a bit, or like we've been promised this massively transformative technology, but four years on, it's not clear how that'll transform things necessarily. We're seeing a lot more excitement in other alternatives like world models, Jan Le Kunz, met his former chief AI scientist, has a new startup building world models which understand the physical world, not just language.

5:16Fei-Fei Li, another one of these super important AI researchers, also has a new startup pushing similar ideas. So yeah, I think we are seeing this sort of diffusion. Sarah you've covered the economy and the world of work writ large for a long time now given that you've seen a few booms bust crashes along the way do you see red flags in this big AI explosion I think what's interesting when you talked about the kind of the disconnect between the sort of the expectations and this idea that it will sort of transform everything and now people starting to say, well, hang on, how much has really changed?

5:59You know, the people in Silicon Valley who you talk to all the time are testing their models against these kind of benchmarks, and they are blasting through them. So all of these kind of benchmarks of like human capability on whether it's coding tasks or whether it's passing the bar exam to be a lawyer in America, you know, it seems as if you look at these sort of lab type conditions that AI ought to be transforming everything and displacing everyone. And while that might be problematic in terms of rising unemployment, it would at least mean that there would be huge kind of profits to be made for the companies that are implementing this.

6:35But actually, I think that what the technologists don't understand is like the real world of work is very complicated. And, you know, what the models are technically capable of is not the only thing that matters when it turns into like how the economy changes. So like on our newsletter that I write with my colleague John Bear Murdoch, but we recently wrote about radiologists as an example. So back in 2016, Geoffrey Hinton, who's sometimes described as like the godfather of AI, he was at a conference and he said, look, we should just stop training radiologists now. We've already got enough because we know that these models can read radiology images more accurately than humans.

7:12So let's stop training them. So John and I went and looked at like, okay, we're now more than five years on from that point. What's happened? What's actually happened is we have more radiologists than ever. and not only that people are screaming that we have a shortage we need more of them and that is not because the AI has been useless you know I think there's this narrative that either AI is everything or it's just a bullshit bubble but actually it's clearly neither of those things but what you have to get into is like in every single profession there are reasons why it may not be as transformative as the kind of pure technical capabilities appear with radiologists yes it can look at certain images and do a very good job of spotting things like nodules very tiny nodules and lungs that might indicate cancer but actually radiologists do so much more than just looking at images and even with those images there are certain things where there's enough training data for them to be accurate and there are other things where actually they just simply do not have enough data to be as good as humans and so what we have is that AI is kind of supplementing what radiologists do and it has become a very common part of a radiologist job but it is nowhere near reduce the need for radiologists.

8:19And I think we're starting to see that in a lot of different professions. And, you know, tech companies love to talk about these benchmarks, how this model is way better, 10x better, whatever. But these are kind of made up measures. You can measure whatever you want. And that doesn't necessarily translate to real world use cases or benefits. You know, we've seen lots of trials like companies rolling out, having massive FOMO, wanting to roll out AI technologies. And then the state technology is just not working. Yeah, there's also just like the kind of the cultural practices of organisations, right?

8:48So another profession we looked at is lawyers. So again, you know, if you look at what some of these models are capable of, they can apparently do a lot of things that junior lawyers can do. And a few years ago, everyone was like, Oh, my God, there's not gonna be any junior lawyers anymore. And then people were thinking, well, how are we going to train the senior ones? And that seemed like a really interesting problem. And I remember writing about it at the time. But actually, what you see is that there is a pay battle going on for junior lawyers. Junior lawyers' salaries are going up, not down.

9:16And as any economist will tell you, that does not indicate that there is a lack of demand for junior lawyers right now. And part of that is, again, that being able to pass a contract is only a small part of what a junior lawyer does. But also, this is a sector where accuracy is crucially important. And LLMs are still not great at being reliable, being consistent, avoiding hallucinations. Well, I would love to ask you that. Because it seems to me that this, at the moment, there are lots of professions and medicine and law would be two of them. Journalism is another one where I think if LLMs can't be relied upon, that puts a hard ceiling on how much they can transform those professions.

9:54But is that a problem that's going to get fixed or is that something that is inherent to the technology? I think it is inherent. It's like a statistical mathematical reality that they will always have some level of error. and so the best you can do is sort of add ways to verify things, source things, have citations, but I don't think it's something we can ever get rid of. You can't use one model to fact-check another model. Does that work? I mean, that is one way to do it, but then, you know. The risk still exists. Exactly, yeah. It's the underlying training data, right? And I think it's memorably described as bullshit in, bullshit out.

10:28And I think unless you can eliminate all of the bullshit, which means it's the corpus of human data, You know, that's quite a hard thing to do. So, yeah, I don't know. I mean, you're the expert in this, but it feels very difficult with the structure as it is of an LLM to completely eliminate that. Is this discussion being had in the Valley, George? Because my experience of talking to venture capitalists based there and heads of Silicon Valley companies is that when you ask this question, are you worried about the AI bubble? They kind of look at you blankly like, what language are you speaking?

11:03This is transformative. it's going to change the world of course we're piling in i mean i think you start with a set of base assumptions in silicon valley which is that tech is great tech is going to transform everything and it will happen much much quicker than anyone else all these idiots outside of the valley think but i think the interesting thing talking to investors in the valley is that they readily admit that there is a bubble they say yeah of course there's a bubble and bubbles are wonderful because they bring in all of this talent and all this capital and this is a new language bubbles That was good.

11:32That was good. That was great, actually. You know, we should want a bubble. So I think that is the framing that you encounter a lot in Silicon Valley. You sit down with three venture capitalists in a day, and you've had the experience of coming to visit, and all of them will say, yeah, there is a bubble. And some of these chumps down the road are investing in companies which will blow up, and more for them. We'll probably do all right out of this, and the one company we back in 10 that is explosively good will justify all of the losses on the other nine. So that is the kind of ethos and attitude of Silicon Valley.

11:59And that's the optimistic take. I would say to temper the scepticism, because I've been burned by this, by covering years and years ago, Google's rise and Facebook's rise and seeing this and writing very sceptical stories about their valuations before they IPO'd and then become the businesses that they are. The advice of Richard Waters, our former bureau chief, still a technology writer at large based on the West Coast, is that the FT historically has failed to understand how transformative these businesses are. And Richard has spent so much time out there. We'll say, well, thing is, these aren't businesses.

12:38These are industries. Google created the search industry. Facebook now Meta created the social industry and they owned it monopolistically. and if open ai turn up and they own the quote-unquote ai industry or the llm industry then they'll own it all and they will be a like a transformative business even if the other ones won't so there's individual cases like you're saying george where we might see them rise and rise and rise and they'll be with us for a very long time and then there'll be things that fall away or we we're just about to see an almighty bust and we have all these giant warehouses being built around the world and they're not going to be used particularly well do any of you get a sense whether this year is the year that it falls apart or it seems like all the businesses that you're talking to all feel like that we're going to experiment more we're going to play with this technology more and therefore this might have more room to run in the coming year personally I don't I don't think so I think while there's clearly a sort of financial valuation bubble.

13:44I think the technology isn't itself a fantasy. It clearly is very useful. And I think a lot of companies are still very committed to trying to figure out how to make it work. And I think that there are various sort of moves underfoot within companies to sort of transform their own structures and the way they use data in order to gradually make better use of the tools. So personally, I think it's never going to be the thing that transforms everything. It's never going to eliminate the need for radiologists or lawyers. I think that's too far, but nor do I think that there's going to be a sudden moment where everyone says, oh, do you know what?

14:18Yeah, we were just living in the fantasy world. That's not my expectation for what will happen. I think this year, I mean, there'll be immense pressure to find a sort of killer app, one that really works. I think a problem here is kind of a timeline problem where the AI industry has sold us this magic tool that is supposed to just work and it just doesn't. And it's relatively new technology. I mean, when ChatGPT came in 22, people thought LLMs were experimental. They were kind of a joke, not a joke, but they weren't this productivity super tool that they are today, or they're being sold as today.

14:52And so I think realistically, the technology will be useful, but probably not in the timelines VCs would expect it to be. And so this year, AI labs are really pushing LLMs in scientific research as a sort of research assistant, but it kind of hinges on whether that can be something you monetize, whether you can add ads into it, whether we see AI Labs injecting this ad-based surveillance capitalism model into LLMs and chatbots, and whether that will lead to something productive. I mean, I think I agree with both of you that it's not going to run out of road this year. The capital is certainly there.

15:26The demand is there. I think there are a few potential hazard moments, and most notably if capital is not sufficient in the private markets for OpenAI and Anthropik to raise what they need and XAI as well, then we will have this moment where one or more of them tries to go public. Could happen this year, might be a stretch. But if it does, then you have this moment of enormous scrutiny, everyone going through your books. You have to demonstrate that you are a financially sensible company in a way that I think for OpenAI's many merits, they have slightly struggled with to date. So all of a sudden, this kind of finger in the air, optimistic accounting that works for VCs who are interested in your growth become something very different, which is that you have to demonstrate to the public markets that you can be a sustainable business over a long period.

16:09I think also there are increasingly ethical concerns. There are major lawsuits around child safety. We've seen issues around the sexualization of images on Grok, which is the AI chatbot. And copyright. Copyright. Exactly. I mean, there's major lawsuits pending. There are increasingly ethical concerns. So any of those could be a tripwire for the big AI labs. Okay, so we see that there are potential hurdles ahead. And yes, we might be in a bubble. But I think this general agreement, the sober FT view from all of us seems to be that the technology is most likely useful, and therefore will last a long time.

16:51When we come back, let's talk about challenges for the labour market and how the technology will develop.

17:08Sarah, you write the excellent AI shift newsletter with our colleague John Byrne Murdoch, where you've been looking at how AI is already affecting the world of work. And you've already mentioned some of the industries that have been affected. What are the industries that will also get more and more affected in 2026? Yeah. So, I mean, the big story from the labour market so far is there really hasn't been that much to see, if we're honest. You know, like we have done a lot of analyses of the economic data on employment, on unemployment. Even when you go through with quite a fine tooth comb and you kind of try and single out professions that might have been particularly displaced by AI.

17:53And you kind of narrow it down to certain age ranges like younger ones. then you start to see some impact in some sectors like coding, for example. But when you zoom out, you know, unemployment is not surging. People are not losing their jobs left and right because of AI. And although there have been some companies that have made big layoffs and have sort of linked it to AI, it's always been, I don't know if you'd agree with this, it's always been slightly sort of vague terms. Like we need to get AI ready. Again, I'm a professional sceptic, right? But part of me thinks that that is, you know what, the business isn't doing that well.

18:29We need to lay off some people. But if we say it's to do with AI, it makes us look really forward looking and dynamic rather than we're just doing layoffs because that's what companies do sometimes. And so, yeah, the big picture is really we haven't seen a huge amount of labour market disruption. But my personal view is that we probably will start to see a bit more labour disruption. And to go back to what I was saying before about how there are some types of work where I think the unreliability of large language models kind of puts a bit of a ceiling on how transformative they can be. There are other types of work where that's less of a constraint.

19:02And so my prediction is that probably the creative industries are going to be the ones that start to see more disruption. And that is because, you know, you can use Sora or another of these generative AI models now to create short bits of video or images, writing. None of it so far is as good as like what the best human professionals can do. But there are lots of places where actually good enough is kind of good enough in certain sectors, whether that's advertising or marketing or copywriting. And so I think those are the places where we will start to see, particularly for companies that are aggressive about cutting costs.

19:42And my sense is that we will start to see that happen in some of those sorts of sectors. But you see that as a job displacement or the fundamentals of the work changing? I'll give you one slightly terrible anecdote. Well, I was in a private roundtable with banking chief executives, actually. And they were talking about how, like you were saying, Sarah, AI is a great excuse to cut jobs. Thousands of jobs, we're going to get rid of them, even though we know that the technology isn't precise, we're still experimenting. We're not quite sure if it's going to be useful. But bloody hell, it's a great moment to cut jobs because we can say, well, AI is going to take over these jobs.

20:23But what are the industries that you think are, this is a displacement story? And what are the industries, maybe like the creative industries, where actually what we're going to see is maybe it can be a super tool and an aid and change the very nature of work itself? I mean, I mentioned coding before. That's one where I think there have been quite big changes to the way work is done. Over time, I think we'll probably see that to an extent in lots of jobs. It's very rare that an AI model can do everything that a human would do in a particular role. It might be able to do certain tasks. And then the question becomes like, OK, so if AI is going to do some bits of your job, is it going to be the bits of your job that you like or is it going to be the bits of your job that you don't like?

21:09And for me, this is quite a critical question because while economists try and judge labour market disruption by numbers, you know like number of jobs lost or number of jobs gained what also matters is like do we enjoy our jobs do they fulfill us do we get paid adequately to do them and a lot of that will come down to whether the way in which our jobs change is a way that is sort of what we would like or what we wouldn't like I think I mean you see it on the on that point I think one of the areas that you've seen AI come into the workplace is not just displacement but it's allowed more people to do more mediocre stuff so i mean slop is the trendy term for it but i think you'll just see a massive escalation in the amount of content that gets put out and a lot of it is rubbish and this is already flooding social media it's very good at creating lots of powerpoint slides you can mediocre yeah really naff and i don't think that's a hugely positive thing i mean the only other area that comes up in conversations that i have with investors is looking outside of the us and the UK is for instance you know if you're in a call center in India that job could completely get wiped out by an AI customer service agent and customer service is the other the other big area outside of coding that I think you're starting to see some traction but yeah it's uh I don't know how much that's showing up in the statistics at the moment not loads I mean it's definitely true that call centers have been implementing I mean they're always implementing new technology to try and kind of drive down costs it's one of those sectors that is like it's very labor intensive And so the more you can drive costs out, the better.

22:40And so they have been using text chatbots for years. And a lot of them are kind of experimenting with voice chatbots and stuff like that. But there is a fundamental issue that when you call a call centre and the chatbot hasn't been able to help you, at some point you're just jamming the zero key and you're like, I want to speak to a human. And most call centre operators are aware of that. And they know that they still need humans to deal with the really difficult kind of knotty problems that require some sort of institutional knowledge of the way the company works or the systems work. And also sometimes I think customers just want to feel heard that they've actually had somebody to rant at.

23:17And I don't think they're going to be satisfied purely with a machine to rant at. And so that's an area where there will be job disruption, I think. But the idea that there won't be any humans doing these jobs in the future, I think, is for the birds. Another company, Chief Exec, was telling me that one of the problems they're finding is if AI comes and sweeps up all the easy customer service queries, the things that make your day really like fly by, it's relatively admin heavy, but you just kind of do it. And AI comes and wipes all of those queries up. And so all you're left with is the edge, really difficult cases.

23:53And they're really cross customers. That's the person that's just going to rant at you for half an hour. That makes that job even worse. Like it's already not necessarily the most fulfilling role, but it's actually incredibly hard and not well paid. So there are risks attached to the workforce the more and more you adopt the AI. Yeah, and this is, I mean, this is something that we see throughout the history of automation and work, you know, that quite often what technology does is it might kind of make some people redundant, but also intensifies the experience of work for other people because it cuts out some of those kind of more monotonous things.

24:27And yeah, that can be a good thing in some ways, and it can be a bad thing in others. Yeah, there's also, I mean, I think we talk about deploying the technology into these workplaces as if it's kind of, you know, here it is, and you just pop it in, and then it's free and off you go. I mean, the reality is, for the companies who are building these customer service agents is that they haven't figured out how to price them in a way that is ever going to be profitable. So you can price for outcome. But that means, to your point, if you're the angry customer who hasn't had the outcome you want, you just keep them on the phone.

24:57and at some point it costs so much in the use of tokens for the company that it is a loss on each query and I think that's the issue that a lot of them have had so it's not as much as we say AI could do customer service it could but the economics of that have not yet been worked out either yeah I think that's a really good point and one that I've wondered about a lot when chief executive is a sort of blue sky thinking they're like oh you know well in future I won't need to pay people $100 ,000 a year because I can get a bunch of agents and each one, I can rent each one for like$15 ,000 a year.

25:28It's like, okay, and what's going to happen when the company that rents it to you decides to raise the rent? That surely becomes a business risk in itself. If you start to rely on one system or a set of systems, particularly if it's not your own proprietary model, but something that you're buying in, then the price of labor, even though it's now kind of AI labor, could shoot up and be out of your control no and going back to those dastardly banking execs that's one question they were putting to me which was can we trust open ai like we have to we have to use their models to build out all these useful tools but what if they turn around jack up the price as they're probably going to need to yeah and so then they're stuck in using these things particularly if you've sacked all your human workers but it's not even the price right i mean they could update the models or roll out a new one and that might be trained differently and suddenly everything breaks.

26:24Everything breaks. And you have to start again. That's already happening, right? Right, exactly, yeah. When OpenAI launched its latest model, these power users were complaining that, you know, they'd formed this attachment to this AI model and that OpenAI had killed the personality and it was too sycophantic and the company had to roll it back to make it less so. Melissa, just on the potential optimistic view, these companies really do say there are productivity gains to be had i know that i felt them i use generative ai tools every day and they definitely have saved me time what are the potential routes to saving time in work do you think the more and more people adopt this technology i'm really excited about um applying this these technologies in science because you know science is a really slow, laborious process.

27:14And if we can use these tools to help researchers come up with new hypotheses or analyze source material, that could lead to amazing breakthroughs. So that's just a net positive, I think. And we've also seen some really impressive AI breakthroughs in the science field, like AlphaFold, which won the Nobel Prize. So that's positive and exciting. What about all these AI assistants that are going to come and, you know, book our travel and deal with the emails that we don't want to deal with and all of that when when is that happening i mean i would love for someone to deal with the emails i want to do my expenses i know please please but i don't think they're there yet they're not reliable enough you know what one thing that saved me massive time gmail has rolled out this new feature where like there's a lot of generic emails we have to send like many thanks blah blah blah blah and i think it's learned how i email people and you know my use of exclamation marks and vocabulary and it's like rolling out these suggested replies to emails which is amazing i'm like this is great send i thought that was a personal touch from you yeah i mean or they require tiny tweaks but generally i've been really impressed yeah so muir this is a question for you right you say that you're using generative ai and it's saving you a lot of time presumably in your job so what are you doing with the extra time because this is the critical question for economists is like are you just having a long tea break because that's not going to transform the economy or are you like adding more value to the financial times with your saved 30 minutes you want me to transform the economy no i so i naturally in this job i've had to think about it as well right so there is the powerpoint problem which a lot of people have said like instead of creating 15 powerpoints you use the same amount of time to create 30 and you have an even longer meeting as a result i mean that is a terrible use of of time i genuinely have saved a lot more time and it's allowed me to do things which i tend to think improve the article in some way so a lot of my job is handling the actual copy that you guys write melissa too too many exclamation marks will get cut but it means that i can do things like have an interesting chart i can go over and talk to the great colleagues that we have and develop something that will look better I can think a little bit more about the headline in a way that readers might want you know these are little things but I actually rarely have time genuinely and this will sound a bit naff but I have now had more time to think and because I've been in this job and it's the news industry and there's something happening all the time and you very rarely get a chance to sit and think strategically or just like pick up the phone and have a conversation which has no agenda at the start and then by the end of it a smart idea has come about all these things that we've read about about how creativity works so i think more enlightened company bosses out there will hopefully see any productivity gains as a way of giving back time to the worker.

30:24My fear is that that isn't generally how industrial progress works. Actually, we'll just work people more and we'll produce more and more. And as we're all worried about, the quality of that work might actually go down. But Melissa, how do you envisage the technology itself evolving in the coming year? I think we'll see way more innovation. Until now, until the past few years, we've really focused on pre-training, which is learning from the data set, this massive data set that's inscribed to the internet. I think in the coming year, we'll see more on more innovation in post-training, meaning how you tweak the model after that initial training has happened and what happens at inference time, which is when the model generates stuff.

31:06And that's really exciting. Like that's where Gemini, Google's newest model, showed really impressive improvements to OpenAI's models. That's where a lot of applications can really thrive. We'll also see developments in reasoning models, these models that can solve problems step by step. I'm also really excited about potential new architectures, you know, whether these new AI heavyweights going into world models, what that'll lead to, whether we'll see new ways of doing AI. So just describe world models to us. So language models, LLMs, they work by predicting the next likely word in a sentence.

31:42World models predict the next likely thing to happen in the world, I guess, in the physical world. They can generate images or scenes that are anchored to our physical reality. You can walk in a 3D environment and it'll know that, right, George is here and there's a door and I can open the door and move through it. And that movement would seem seamless. You just did a very expensive lunch with EFT with Yan Lecun, who's met as outgoing AI scientist. And it sounded like he's basically fallen out with his company. He calls them LLM-pilled, I think. So they're so focused on trying to develop LLMs.

32:22They can't see that actually the next breakthrough is a more important thing for him anyway. This is the concept of world models. Do you buy that argument? Do you think that we're not going to be able to move towards this concept of artificial general intelligence or super intelligence, this thing, this all kind of all-knowing AI, unless they can build things like world models? I think he does make a very convincing point, right? If you think about intelligence in humans or even animals, right, it's not just language-based. We do understand if you lift a glass and I pour the water, it'll make a mess, right?

32:55And a language model, if it hasn't seen it, how can it know? So I think the problem is that we don't really fully understand intelligence to begin with, but that's a different philosophical debate. We need to be thinking about these things in a more holistic way. I mean, even DeepMind, they're looking at a whole bunch of different ways on how to improve intelligence in machines. So I don't think LMs might be one way that could help us get there, but I don't think it's the way. George, as our representative of Silicon Valley, you not only do you speak to a lot of venture capitalists you speak to a lot of startups are there interesting ideas that people are working on which aren't just something in and around creating a chatbot i think world models is the is the one that has caught my imagination as well but i think you talk to founders and they have slightly different views of what that might look like and how it might be achieved and what it might be used for you know ilia suskeva who was the research leader at OpenAI and this very kind of gilded AI researcher who's been involved in many of the breakthroughs to date, went off to set up his own company called Safe Superintelligence, and he is working on a straight shot to superintelligence, which is this all-powerful AI you were talking about.

34:06Do we know anything about this at all? So his argument is basically LLMs are not going to get you there, so it's something different. And of all of the very secretive San Francisco and Silicon Valley startups, it's probably top of the pile. They've kept their Costco to their chest, but their investors have kind of told me I mean, I think this is as far as they know, or at least as far as they think they know, that it's an AI model trying to extrapolate from first principles. It can look at one table and know these are the features and characteristics of a table, and I will be able to understand that in any domain, in any setting.

34:36So in the way that a baby might learn, you know, you don't have to say, here is every table in the IKEA catalogue, and you need to know that they're all tables. You show them one, and they can probably figure out the features through a small number. So is that a version of a world model? I'm not sure but it's it's along the same track. The thing about these world models presumably is that if they can become proficient then that opens the door to humanoid robots and robots that can move around in the world in a way that at the minute we don't really have I mean I know that there are videos online of robots doing really cool like dance routines and stuff.

35:10Great dancing Chinese robots. Yeah but when you see them like actually trying to interact with the real world they're often like comically bad still so presumably that's one of the potential applications of these world models? Well, I was just going to say, absolutely. I mean, one really exciting avenue is embodied AI, which is just a fancy way of saying AI in physical things like robots. We're seeing a massive sort of sea change in the field of robotics where robotics is very expensive, traditionally being very slow to train, and these breakdancing robots, they can do impressive stuff, but in very limited environments.

35:45And, you know, it requires, traditionally has required lots of manual tweaking. But with AI, these models can generalize, which allows them to be more flexible in the real world. It also have really big job market implications. I mean, if there comes a point where a lot of the jobs we've been talking about in this episode have been like white collar professional jobs, because that's where LLMs make the difference. But if you end up with genuinely safe and useful embodied AI, then that opens up a huge number of sort of blue collar professions to potential disruption or enhancement, depending if your glass is half full or half empty.

36:19One area we're definitely going to see that is robotaxis. They're already on the streets of some US cities. We're going to see them increasingly in London and across Europe and probably across the Western world as we go on. And that is going to be a clash of the US and China. We haven't even mentioned, other than DeepSeek, some of the Chinese advances that are coming. But Baidu are up there making some of the best robotaxis. They're going to be on the streets of London at some point soon. Waymo are also doing that as well. and people will start to see these kind of embodied AI things. Yes, it's still a car that is moving around, but we're going to interact more and more with that in the coming year and get used to it.

36:59Things like that will be when people start to see the kind of the power of this technology and how every day it will become. Like my experience of being in the driver's car is the first 30 seconds is revolutionary. And after that, I'm very bored and playing on my phone. and that's how AI will kind of bleed into the background. So I'm going to wrap up by asking each of you something that AI is very good at but humans are very bad at, which is predictions. So Melissa, what's your big take that you think we're going to get out of AI in 2026? Okay, I think we'll realise that AI is useful but it's a lot more boring than we need to hold.

37:38My dream is for AI to be boring because boring works. Boring means it's reliable and that's great. That's what we want from a technology. So I'm hoping AI gets boring. I hope you're wrong because it makes a bad copy. George? In the interest of good copy, I hope that one of the big labs has a crisis and I think that's highly possible with everything that we've been talking through, the competitive dynamics, the risk of commoditization, issues around ethics, the use of AI in pornography, copyright, that one of these cases goes to fruition and trips up one of the labs, I think is highly possible this year, if not likely.

38:21And in the interests of my reporting, that might be quite a fun thing to hope for in 2026. So yeah, that can be mine. Yes, there's a slightly grim way that reporters and editors think about that. I would think that's terrific too. And Sarah? I hope that over the next year, AI becomes agentic in a way that means it can fill in my expenses for me and reply to your AI written emails for me. And my fear is that we will start to see quite definitive signs of disruption in some parts of the labor market. And I think that's, as I say, most likely to be in the creative sectors. and I don't think it'll be because AI is fantastically better than humans but simply because some companies will decide that it's cheaper.

39:09Terrific. Well, I think we've covered a lot of ground. It feels incumbent on me to say we're all going to have to gather here as and when that AI bubble bursts. Thank you, Sarah, George and Melissa. Thank you. Thank you. Next time on Tectonic, the power and the politics of Silicon Valley in 2026. How will the likes of Mark Zuckerberg and Elon Musk seek to change their companies and the world this year? That's all in part two of this series about what to expect from tech in 2026. Tectonic is presented by me, Murad Ahmed. Our producer was Josh Gabbard-Doyon. Flo Phillips is executive producer. Mixing and engineering by Sam Giovinco and Breen Turner.

39:54Music by Metaphor Music. Cheryl Brumley is the FT's global head of audio. Thank you.