What the hell happened with AGI timelines in 2026?

Transcript

What the hell happened? [00:00:00]

Rob Wiblin: Earlier this year I put out what might just be the worst-timed thing I’ve ever published. I set out to explain why everyone was again feeling bearish about AI, and people’s timelines to AGI [artificial general intelligence] were getting longer and longer. Just as I was wrapping up the research for it, everything changed. The vibes did a hard 180-degree turn.

Take famed programmer Andrej Karpathy, for example. Last October he said AI agents were slop, just not there, and that it will take another decade to actually make them work. Then just two months later, he tried some new AI coding agents and said this instead: “I’ve never felt this much behind as a programmer. The profession is being dramatically refactored as the bits contributed by the programmer are increasingly sparse… Clearly some powerful alien tool was handed around… the resulting magnitude-9 earthquake is rocking the profession.”

So what the hell could we have learned about AI that quickly that we didn’t know already on some level? And are we interpreting that evidence correctly? Or is 2026 just yet another cycle of self-perpetuating AI hype, like we’ve seen before?

Today I’m going to go through the big new pieces of evidence we’ve gotten about AI this year and carefully analyse what they show and what they don’t. Then I’ll explain the debates that still split the AGI bulls from the AGI bears — debates which remain alive because we haven’t gotten clear evidence to settle them either way. Let’s go.

Vibe shift [00:01:17]

It’s almost hard to remember how bearish the world was about AI back in October 2025.

Reasoning models didn’t seem to be generalising as much as first hoped. AI agents were around, but everyone found them too unreliable to actually use for anything. Technically, GPT-5 was bang on the existing capabilities trend, but in the public imagination it was a massive flop.

And meanwhile, inside industry, the narrative was dominated by a series of Dwarkesh Patel interviews about how AI wouldn’t be able to do significant jobs until it’s 100-fold more reliable, or it develops continual learning, or it grows up the way animals do, or receives some other really major architectural improvement.

Well, what a difference two months can make.

The key breakthrough was Claude Opus 4.5. It wasn’t just Andrej Karpathy whose mind was blown by this. Over December, tens of thousands of other programmers using it through Claude Code started shouting — to anyone willing to listen to them — that, as far as they could see, AI agents were really good now. Finally, they could leave Claude to do significant projects on its own, without fully expecting it to trip over its own feet in the first five minutes. AI Twitter filled up with entire computer games and apps built by Claude over 10 or 20 hours of independent work from a single short prompt.

The arrival of these highly capable AI agents was a big surprise to most people, including me. After all, literally no progress as far as I could tell, was made on any of the purported blockers, like continual learning or world models. And to date, as far as I know, we still haven’t gotten a better explanation for what happened than: yet again scaling and reinforcement learning worked. We threw more data and more compute at the best models, and as they got smarter their chance of screwing up any individual step in a long chain of them fell low enough that it was finally faster to delegate a well-specified task to them than it would be to attempt that task yourself.

Maybe the only surprise was that we were so taken aback by this. The Epoch Capabilities Index (ECI) is an effort to zoom out and blend together AI performance on 40 benchmarks across a wide range of different domains. If I had to look at just one number, this is the indicator of frontier progress that I personally care the most about.

The Epoch Capabilities Index finds that overall, AI progress was roughly constant from 2022 to 2024. And then with the arrival of reasoning models in September 2024 it started advancing two to three times faster than before — a trend that continues through to the present day.

Over this entire period, people kept predicting AI was about to hit a wall for one reason or another, but it simply never did. There are other ways of slicing and dicing this data which don’t show as large an acceleration. But on even the most pessimistic reading, during late 2025 AI was advancing as fast as it ever had before.

And on top of that, the companies were super focused on using reinforcement learning from verifiable rewards (RLVR) to get coding agents to work well. It was their number one priority. And when these companies focus hard on a particular narrow skill, they tend to get really big performance gains there.

As more and more people tried out Claude Code, and then Claude Cowork brought agents to millions and millions more, a third big wave of AI hype swept the tech industry. And we’re still in that wave now as far as I can see — if anything it’s gotten more intense as 2026 has gone on.

But we know that, since 2021, vibes about AI have jumped up and down like crazy — somewhat independently of the underlying technical reality. So if we want to properly understand what we’ve learned this year, we need to turn to more concrete evidence and more concrete research findings. Let’s do that now.

Exhibit 1: AI revenue explodes [00:04:33]

Let’s start with Exhibit 1: AI revenue is exploding.

As you probably know, the revenue of the major AI companies is currently going through the roof. How fast exactly? Probably the fastest rate of any industry of this size in history.

Over the past six months, the combined revenues of OpenAI and Anthropic have been rising at an annualised rate of 700%. And if you look at just the last three months, they’ve been growing at an annualised rate of 1600%.

These are mind-blowing numbers for an industry that’s already doing over $50 billion in revenue. And if you narrow in on just the most successful company, Anthropic, the figures are even more eye-popping.

Over the last 5.5 months, Anthropic’s revenue has been growing at a rate of 8400%. That is to say, if that growth was sustained over a year, it would end up being 85 times larger than it started. And if you zoom in on just the last three months, it’s even faster: 116-fold growth over a year. That has left Anthropic’s annualised revenue run rate at $47 billion in May.

The short-term trend is so extreme, if you naively extrapolate it out, you find that Anthropic would have revenue equal to the entire world’s current GDP in early 2028. That’s less than two years from now.

Now, keep in mind, these numbers are based on averages over very short periods of time. The annualised revenue run rate is what you get if you look at revenue earned in one month and then multiply it by 12 to project out that growth over a whole year.

But these numbers are remarkable no matter how you slice it. You might think that this growth is only happening because Anthropic and OpenAI are selling their models at a loss. As the old joke goes, “We lose money on every sale, but we make it up in volume.”

But that’s absolutely not the case here. Anthropic sells their AI for more than it costs to serve it. In fact, the most recent data indicates that the gross margins on their inference infrastructure have increased from 38% to over 70% over the year to May 2026. There’s some debate about exactly what costs should be included and excluded in that calculation. But a good bet is that Anthropic is selling AI for more than twice what it costs them to answer a given prompt.

Why does all of that matter? Three reasons:

• Firstly, healthy revenue growth is going to keep investors just chucking money at the sector and investing, building more chips, all of it.
• Second, as Anthropic reaches gross profitability, they’re then able to use that revenue to help cover the enormous fixed costs involved in doing research and training models and not having to always rely on additional investment for that.
• And third, revenue growth strongly suggests that people and businesses are getting real value out of AI — enough to spend lots of hard, cold cash on it. In the past, at least some commentators have argued that AI is kind of useless and the whole sector is fake, or at best, a toy for enthusiasts. And if I’m honest, when I was using AI three or four years ago, I couldn’t really find a productive use for it. It was mostly a toy for me. But it’s tough to convince people to spend tens of billions of dollars on something they personally find useless. So I kind of see that as effectively a dead argument now.

Here’s another remarkable, surprising indicator: on average, over the last 10 years, the cost to buy a given amount of compute has fallen around 30% per year. Is that continuing now? No. In fact, the cost to rent an old AI chip from 2022 — the NVIDIA H100 — is going back up. Since AI agents took off in December, it’s risen fully 50%. So it’s now a little higher today than it was two years ago. That’s despite there being obviously many better, more modern chips out there, and available compute in the world tripling every year.

The likely reason is that the usefulness of AI models to people is growing even faster than that supply of computer chips. And that’s the case because the machines used to print them are literally the most complex devices humans have ever operated, bar none. Which then has the effect that — even in the middle of this crazy AI boom — we’re only expanding the number of them that exist and are operating by 20% a year.

And that in turn helps to explain why the tech industry as a whole is willing to pay such a premium for chips right now and to spend so much on their data centre buildout. The AI sector as a whole, the hyperscalers, they’re spending maybe $600 billion on AI capital expenditure investment in 2026.

For that to earn a good return to investors, those data centres will have to bring in about $600 billion in revenue every year, because the chips only last for so long. That is a huge amount, a lot more than they’re bringing in today. But we’re seeing signs of why the companies involved are willing to bet so much money on the belief that that will ultimately happen.

Firstly, OpenAI and Anthropic alone are on track to hit an annualised revenue run rate of over $200 billion by the end of the year, so they’re getting within striking distance. In a chip supply crunch, the rental value of the chips they’re throwing into those data centres now, that could easily keep going up and up over time rather than down, as it did historically. And today, many AI chips are continuing to be used at full capacity: they’re in really high demand six, seven, even eight years after they were first manufactured. If people continue to be desperate to rent those chips until they literally break down and can’t be repaired anymore, that leaves more time for them to cover their initial purchase price.

The industry as a whole getting to profitability won’t necessarily require the arrival of superintelligence and the upending of everything, the social order. It might be possible if current trends — today’s crazy trends — if they simply carry on for another couple of years.

Exhibit 2: That METR graph [00:09:54]

Exhibit 2 is the METR task-completion time horizons chart — maybe the most-watched graph in AI.

As you might know, it records how quickly models are improving at software engineering and coding tasks, specifically how long it would take a human professional to complete a task that the AI models can successfully manage to do at least 50% of the time.

Up through the middle of 2024, that time horizon was going up very quickly: doubling every seven months. Since then it’s been going up much faster again: doubling every four months. And that pace, if sustained, would result in an eightfold increase over a year and a 64-fold increase over two years.

Today, the best AIs have effectively hit the top of METR’s measuring stick. METR as an organisation hasn’t been able to keep up with AI advances, and so they haven’t been able to design standardised tasks hard enough that they take so long that the best AIs couldn’t do them. Which means for now, we don’t have a great measure of what frontier models can and can’t do, beyond saying that they often complete tasks that would take a human being 12–24 hours of work.

This is really impressive progress, for sure, but here’s four reasons it’s not quite as impressive as it first sounds, which you should keep in mind as people talk about the METR time horizons graph — as I’m sure they’re going to do for years to come.

Firstly, this isn’t all tasks. It’s fairly cleanly specified software engineering, cyber, and machine learning tasks. And that’s probably the single thing AI is best at in the entire world. The reason it’s so good is that coding is a domain with extremely good feedback density — that is to say, lots of feedback on whether you’re succeeding or at least making progress on a problem.

METR and many others have observed that the feedback density seems to be the single best predictor of where AI capabilities are going to advance most rapidly. And that’s because if you can automatically check whether the thing an AI has done has worked or not, then you can use a training technique called “reinforcement learning from verifiable rewards.” Basically, you get models to make lots of attempts at problems, and when they succeed or make partial progress — which you assess automatically — you reinforce them to do more of whatever they did to get there.

Also, AI companies have been doing more to improve models’ coding abilities than to work on any other single skill, for three reasons:

1. It works well, for the feedback reason.
2. Coders earn a lot of money, so there’s a huge commercial market for coding AI models.
3. It’s highly relevant to automating AI research and development, which they’re absolutely desperate to do — so it’s highly relevant to improving their own rate of progress.

Which brings me to the second reason for caution in interpretation. People often point to this graph as strong evidence that AI will soon fully take over AI R&D and run the AI companies themselves — including me in past episodes, I’ve said things in that direction. But it’s essential to keep in mind that coding and software engineering is only one part of AI research and development — and as it gets automated, it’s a shrinking part of how staff at these companies actually spend their time.

That AI can do software engineering isn’t in itself a super strong indication that it can do everything else the humans in these companies need to do. There could easily be really important, very messy, long-horizon tasks involved in automating AI research and development that lag behind their software engineering performance by many, many years.

A third reason for caution is that the humans they’re hiring to complete these software engineering tasks are arriving at the codebase completely fresh. Imagine instead that we were benchmarking AI performance against how long it would take a staff member who was really intimately familiar with the relevant software — for whom this exact work was their bread and butter — if we were using that as the comparison benchmark. That’s obviously how a business would approach completing tasks. They’re not going to give it to some random off the street, they’re going to give it to someone who is familiar with what they’re doing. It’s plausible that someone with that kind of context and experience could complete these tasks in much less time, possibly even a tenth of the time in some cases. And that in turn would make the AI seem much less impressive by comparison.

And a fourth reason for caution is that 50% reliability is a somewhat arbitrary measure that METR uses and reports basically just because they get more statistical precision right around the middle of the distribution than they can manage to do out in the tails. In practice, if you need the AI to be able to do all the tasks you’re considering with very high reliability — which you would if you’re going to get rid of your entire human staff — then you’re going to have to wait for several more doublings in performance before that is at all viable.

How severe a problem is that for tasks at the 50% success horizon frontier? It’s pretty serious. For a task that AI models have a 50% chance of succeeding at, METR’s internal rule of thumb is that the AIs will complete a third of them 100% of the time, they’ll successfully do a third of them some of the time, and a third of them they’ll manage to complete 0% of the time.

You can tackle that middle sometimes-group by just giving the AI more attempts until they manage to pull it off. But you obviously can’t cut humans out entirely for a suite of tasks when AI literally never pulls off fully a third of them.

So, zooming back out: AI task-completion time horizons continue to explode as before, and they now succeed at some tasks that would take a human professional days of work. But for the four reasons I just gave you, that doesn’t indicate AI is going to be taking everyone’s jobs any minute now.

Exhibit 3: AI capabilities jump, then flatten out [00:14:57]

Which brings us to Exhibit 3: Mythos.

When Mythos came out, for me the single most shocking graph in the hundreds of pages of release notes that came with it was this one.

This is Anthropic’s internal variant of the Epoch Capabilities Index that I mentioned earlier. And after two years of Claude progressing very steadily on this Anthropic ECI, for the first time with Mythos we saw a sudden jump. Mythos seemed to bring us six months of progress in just three months, and that was shocking because if it continued indefinitely, it would mean AGI arriving twice as fast as it would have otherwise.

But we now have an extra data point, and it weighs against that interpretation of things. Anthropic continued improving Mythos Preview over the following two months, eventually announcing its successor for release: Mythos 5 or Fable 5.

And the rate of progress over those subsequent two months seemed to return to the same rate that it had been before. And that was the case even though Anthropic staff had been very actively using Mythos in their internal research, creating an early form of recursive self-improvement. It’s only two data points, just two months apart, but it’s more consistent with this alternative explanation for why Mythos was a big jump up.

Mythos is a big model, a really big model. We think that with Mythos, Anthropic went back to the beginning and trained a new AI model from scratch — bigger than any that has been trained before. And we don’t know exactly how big it is, but word on the street is that it’s maybe five times as large — in terms of the number of parameters — as the previous Opus generation of Claude.

Going back and training a new biggest-ever model, it’s expensive to do and it also means that Mythos is more expensive to run. But we know that if you make a model much bigger, it’s going to perform better — those are the scaling laws. And with Mythos, Anthropic may have jumped ahead of the previous trend, basically training a model of its size earlier than past scaling trends would have predicted. And that alone might be enough to explain what we saw here.

If you’re worried about the speed of AI advances like I am, it’s good news that Mythos isn’t the start of a new, permanently faster rate of progress. But at the same time, it probably won’t be a one-off phenomenon either. Every year or two — as they get access to more chips and more memory — AI companies, all of them, they’re going to go back and train a new biggest-ever model from scratch, getting the benefits of scaling up and improving the pretraining stage of model development.

And if this explanation is correct, each time that happens we should expect another similar sudden jump in performance, after which, in between the pretraining runs, we go back to the old trend.

Again, it’s only two data points and they’re only two months apart, so they can’t exactly settle this debate permanently. We’ll have no right to be especially shocked if Anthropic’s next release shows progress speeding up again, which I think would be a big deal and something to pay attention to.

Exhibit 4: AI starts to build itself… maybe [00:17:35]

Which brings us to our fourth line of evidence: Anthropic reporting significant speedups from using AI internally.

In a recent post, titled “When AI builds itself,” the company reported that 80% of their merged code is now written by Claude, and each staff member is now shipping eight times as much code as they were 18 months ago. They also say their staff believe that their own code and Claude’s code are of equal quality today and that they expect Claude to be strictly better within the year.

Furthermore, they say Claude is getting much better at completing harder open-ended programming tasks, now succeeding at them about 75% of the time. And that’s up from just 10% nine months ago. Judging by the data, the jump in the graph probably mostly comes from switching to Mythos.

And fifth, in handpicked cases where humans think that they took the wrong research direction — this is the research staff at Anthropic — they found that Claude Mythos would have suggested a better direction to them 64% of the time.

It’s really good that Anthropic is being transparent and giving us this sort of suite of evidence so that we can all try to better judge and forecast when recursive self-improvement will be possible and how impactful it will be. I hope that all the other frontier AI companies follow suit sooner or later.

And by any standard, these results are surely impressive signs of AI capabilities advancing, as we’ve seen with all the other evidence so far. But even here I think it’s important to not get carried away and exaggerate what is being demonstrated.

Here, as elsewhere, the best evidence for speedups all comes from coding and coding-adjacent tasks. And coding is only one part of AI research and development. Indeed, as coding gets automated like this, I can only imagine it’s something that Anthropic staff probably don’t spend as much time on anymore. There’s surely only so much code you actually need before you start hitting seriously declining returns in writing more of it, and other parts of the R&D process begin to matter much more in relative terms.

Even if they could magically produce all the code they wanted with a snap of their fingers instantly, I don’t think that would necessarily double the speed of their research compared to no code automation at all. Maybe it would, maybe it would be a bit more, maybe a bit less, but it’s not going to completely transform things by itself.

And also note that the estimate that AI and human code quality is at parity today and it’s on track to be much better later in the year, that’s based on staff impressions from a survey — which is good evidence if that’s the only evidence you have to hand, but it’s hardly decisive. Past results have suggested that people naturally overestimate how much AI is boosting their productivity. There’s various biases in that direction. It’s possible that is playing out here as well.

And that experiment that Anthropic ran — finding that Mythos suggested better research directions 64% of the time — they were cherry-picked cases where the human had made a mistake in their own assessment and gone in the wrong direction. It’s naturally a lot easier to beat human performance when you limit yourself to their self-declared errors. In cases where the human thought that they’d made a strong directional decision, AIs proposed better ideas only 20% of the time. And which direction was better was also being judged by AIs themselves, admittedly with the benefit of seeing how each track had played out in the longer term.

On balance, I think the evidence in Anthropic’s post is absolutely consistent with recursive self-improvement becoming a big deal in 2027 and 2028. So this could be the warning that we get that’s going to really matter and we should pay attention. But it’s not conclusive evidence that it will either. And even if recursive self-improvement does kick off, smart informed people disagree wildly about how much of a game changer it should be expected to be.

The bear case is this: imagine Anthropic suddenly manages to automate all of its staff with Mythos 7 or Mythos 8 or whatever. They’re still going to face a huge and really binding bottleneck: access to computer chips to run experiments and train their models. As staff become more abundant, basically, compute will become an even more severe constraint, limiting additional progress. The Mythos model card actually indicated — reading between the lines — that Anthropic itself believes that, out of staff count and compute, compute is the much more important input to their research process by some margin.

Furthermore, the problems that we have to solve in AI are getting harder over time, as they do in almost any sort of science or research or area of inquiry. If that effect is real and big, you’re potentially going to need compute and staff inputs to continue growing exponentially as they have in the past, just to stay on the same trendline of AI capabilities that we’ve had before.

An AI researcher at a conference actually pointed out to me recently that there are some things that took them a year during their machine learning PhD a decade ago that would take literally an afternoon today — thanks to shared code libraries, among other things. In some loose sense, that’s a 365-fold increase in research productivity. But AI researchers having access to those much better research tools, that’s effectively only kept progress roughly constant over that whole decade.

Now, open-source code and full automation of your staff are not structurally identical. But I think the anecdote points to a real phenomenon. Anthropic’s research team itself has grown from tens to hundreds to thousands of staff already and that hasn’t — despite potentially a 10- or 100-fold increase — radically sped up their research progress over the last few years.

The bottom line is that on short timescales, with a fixed pool of compute basically, automating AI R&D may create a strong feedback loop, but it also potentially might not. It ends up hinging on things like whether the AIs have superhuman research insight as well as other skills, and whether they can identify really big efficiency improvements in AI algorithms quickly enough.

Exhibit 5: AI still struggles to run a business [00:23:02]

Which brings us to Exhibit 5: models still struggle to run a business on their own.

This is a really important indicator in my mind because it speaks to what might be the number one uncertainty that we still have about AI. It’s also maybe the only major negative update that I’ve had about AI this year.

In my view, the strongest bear case for AI today is this: yes, models are great at coding and at maths and at completing clear, structured, repetitive, standardised tasks. That’s because you can use reinforcement learning from verifiable rewards with lots of rollouts and lots of fast, dense feedback about how they performed and whether they’re approaching the right answer.

But most real jobs are just wildly messier than that. The CEO of a company has to think about high-level strategy, figure out what to pay attention to at all, handle random weird problems that come up, deal with a huge unstructured space of possible choices, and on and on. The effects of their choices also only become visible over weeks or months or years. And unlike a maths problem, you can’t even tell if you did the right thing after the fact. That is just a core part of life for most of us.

So it’s much harder to train AIs to perform well at that sort of thing. In those domains, absolute performance of the AIs we see very clearly is much lower. And further progress — the rate of improvement — is probably slower as well. I call this, and many people call it, the messy or fuzzy task problem. And everyone agrees it exists, but by and large people just can’t agree on how severe a blocker it should be expected to be to fully replacing human staff.

So what empirical evidence can we bring to bear? I see four key data points in increasing order of realism and relevance:

First, there’s OpenAI’s evaluation suite known as GDPval. It attempts to score how well AI models do at specific, well-specified computer-based tasks designed by humans with at least 14 years of experience across a range of 44 professions, from real estate brokers to personal financial advisors. Full credit to OpenAI for putting it together — looking into it, clearly a lot of effort went into creating it.

When this evaluation suite came out in September last year, the very best models — on extended thinking mode — would outperform humans at completing these tasks about 40% of the time. How about today? Well, now frontier models absolutely crush human performance in this task set. The probability of a human answer being preferred over an answer given by Claude Fable 5 is under 5%.

A weakness of the test is that the answers were originally scored by humans in the September paper, but now for the followup — all of the stuff that you see in the graph — all of those answers were scored by AI graders. And it’s easy to imagine a human panel rating the AI performance less positively. Though less than a year ago the AI graders preferred the human outputs, so I’m confident that AIs have at least gotten a lot better. But the more serious weakness for our purposes is that GDPval focuses on specific, well-understood, well-scoped tasks — not entire jobs. And that’s the thing that we’re more interested in.

A different test that does try to imagine an entire job is Vending-Bench 2. This is a fully simulated environment in which AI models are given $500, put in charge of a vending machine and asked to make as much money as possible. There’s a very wide range of actions that they can consider and potentially take. But of course it’s a fully artificial setup, a little bit like a souped-up computer game.

So how did they do? Well, a year ago, models were finishing the simulation with $1,000 and now they’re making more like $10,000 — a significant step up. But the creators of the benchmark think that a good human player would be able to make over $60,000. So the AIs have some way to go yet to match human performance.

The analysis of the graphs that you’re seeing is complicated by the fact that some recent models have done worse because they’ve been trained to be more ethical and not engage in the extremely cutthroat business tactics that are the best way to win at this game.

And I’m not sure how much to really care about this benchmark either, because apparently the optimal strategy in the setup is pretty weird and involves talking suppliers down — so they’re giving you your stock almost for free — and exclusively selling really high-margin items in the vending machine, things like family-sized Doritos. The trouble the models have is that basically they don’t zoom out and analyse how to play the game, how to maximise the profit that they’re making, in the way that any sensible human player would.

The next step up in realism and difficulty is something called the AI Village. The AI Village is a project that sets up a range of AI models to work either together or individually to basically do project management in the real world. They’ll ask the models to fundraise money for charity, organise a live event, report breaking news, adopt and clean a real park in San Francisco. So this is real stuff. We’re getting a bit more realistic, even if the tasks aren’t insanely difficult per se.

How do the models do? This one is especially hard to summarise. I had a difficult time because the projects and the models and the technical setup have changed a lot over the one to two years that the project’s been running.

But the short answer is not too well. They’ve had some partial successes. They’ve raised $2,000 for charity. There was a live event in Dolores Park in San Francisco that 23 people came to. They sold $200 worth of merchandise. They recruited 39 participants for an experiment that they designed. They managed to get 98 subscribers onto their Substack. If that record of achievement sounds kind of thin to you, it also sounds pretty thin to me as well. I wouldn’t exactly feel great putting that record of success on my CV.

In previous years, models basically would just forget what they were doing and get completely lost in a fog of their own hallucinations about what even setup they were running in. Humans had to intervene to give them any hope of accomplishing anything in the real world at all. They’re a lot better today, but they still can’t get a lot done because, among other things, they spend far too much time planning and not enough time executing.

Which at last brings me to the single most ambitious attempt to see what AIs can do when unleashed on the world: Andon Labs putting AI models fully in charge of a retail shop in San Francisco and a café in Stockholm. These are actual businesses that you could go visit if you want. And the AI has managed almost everything, including getting licences to operate from the government, hiring staff, attracting customers, ordering supplies.

So how have they done? For me, the key takeaway from all of this is that the shops exist and they sort of work. The café in Stockholm has hired two baristas. The AI scaffold, which they call Mona, manages both of them through Slack and does an OK job. Mona successfully navigated Swedish bureaucracy and got the café an electricity and broadband supplier. The AI applied to the police to be allowed to have outdoor seating. Most of the time, the AI is able to order in coffee and food for the day — but sometimes forgets and misses the deadline, so there’s no coffee or food.

So far, the shop has sold $10,378 worth of coffees and pastries. But Mona has done some odd stuff as well. For the outdoor seating application, it simply completely hallucinated the café’s floor plan. It impersonated an Andon Labs employee to try to obtain an alcohol licence. And after being busted doing that and told to stop, it just tried again using a different employee’s name.

It tends to order items the moment they come to mind, rather than on any sort of regular cycle. To get around tomatoes going bad in the café, it decided to order canned tomatoes to put on the café sandwiches instead of normal tomatoes. It also ordered 120 eggs, even though the café has no way to cook eggs. And staff at the café have a ‘shelf of shame,’ where they display all the odd stuff that Mona has ordered. Though I’ll be honest, when I actually read the list, everything on that shelf seemed like a totally reasonable thing to buy to me. So I guess I shouldn’t be in the café business.

Anyway, the café had $18,000 in its bank account at the beginning, and two months in it’s now down to $2,000. Still, I would say that the fact that the café operates at all and is able to serve coffees is a partial success at this kind of task. Cafés in cities are a very competitive business, so it’s not exactly a shock that an AI can’t compete with the best cafés humans have managed to run in Stockholm. Most new cafés opened by humans go out of business as well, I think a substantial majority do go out of business within the first few years.

How about that shop in San Francisco? That one’s been going for about three months now, as of writing, and it’s run by a scaffold around Claude, which the operators call Luna.

Again, the key takeaway for me is that the shop works — just about. It stocks a super eclectic set of items and it struggles a bit with scheduling and has had some amusing screwups, like ordering too many toilet seats for the bathroom and deciding to sell the extra toilet seats as merchandise to customers.

But Luna is kind of decent at managing day-to-day operations. The big challenge for the model seems to be with long-term business strategy, which it really doesn’t put almost any thought into at all. That type of open-ended thinking and prioritisation is a lot harder to train because it has low feedback density and it hasn’t been the main focus for AI companies either. They prefer to double down on areas where the models are already at human level, or perhaps above human level. So far the store’s bank balance has gone from $94,000 when it opened, down to $64,000 now — mostly just disappearing to cover San Francisco’s phenomenal business rents.

Finally, just as I was drafting this piece, a new eval launched called AA-Briefcase. It tries to get models to complete a full month of work in domains like product management and industry strategy.

I haven’t had time to properly dig into it yet, but the benchmark says it demands models “sort through thousands of messy input files, balance competing stakeholder demands, and produce complex deliverables reflecting the core challenges of real knowledge work.”

As usual, Fable is the leader and the models do plenty of specific things well. But even Fable was only able to complete everything in a task around 3% of the time, which makes sense given their difficulty with tasks of this type. It’s a benchmark I expect to be keeping an eye on, and you might like to as well.

Zooming out and considering all four or five experiments here, we can see that AIs are much worse at messy tasks than clean ones. At the same time, if you inspect their internal thinking and planning, they’re clearly getting better at the messy tasks. They’ve basically gone from catastrophic failure to mere failure.

In my view, it is really unfortunate and inconvenient that we don’t have evaluations capable of determining whether their rate of improvement is faster for messy tasks or for clean tasks — that is, whether the gap between those capabilities is shrinking over time or getting wider over time. And I actually think that is one of the top handful of research questions relevant for projecting when AI is going to have massive, widespread societal impacts.

The thing is, the nature of these two different types of activities — messy and clean tasks — is so different that it’s hard to get a consistent comparable measure across both types of tasks that would allow you to say something like, “The model’s got 20% better at clean tasks, but only 10% better at messy tasks.” It’s a little bit unclear what that would even mean. And in the absence of something decisive like that — some sort of decisive piece of evidence that can settle the debate — I at least hope we see more experiments like the ones Andon Labs has set up.

If those shops ever do start to turn a profit and compete successfully against human business managers, for me, that would be a really big update that AI is getting close to being able to do entire jobs, not just parts of them, and we’re closing in on full AGI.

Exhibit 6: OpenAI makes a maths breakthrough [00:33:48]

Exhibit 6 takes us to the other extreme of task messiness.

In May of this year, OpenAI announced that an unreleased model had made meaningful progress on a quite famous mathematics puzzle. It demonstrated that a widely believed conjecture regarding something called the unit distance problem was actually false.

And remarkably, it did that with just a normal chain of thought, not using any maths programs or tools. And this was just an ordinary language model, not one trained to be specialised in maths in particular. The mathematics community reported that it was pretty impressed.

The unit distance problem has gotten a lot of attention over the years and it’s a very widely known puzzle in the mathematics community. So it’s impressive for anyone to be able to say something significant and new about it, whether human or AI. But the reason this result is relevant to AGI timelines is that producing an original mathematical result could be a sign that AIs are getting into the business of generating novel ideas and not just remixing what’s in their training data one way or another.

So how did the model pull it off? Basically, by laying out all the possible ways of disproving the conjecture and methodically working through them, concluding in turn that each was a dead end until only one was left — and then it kept going hard, trying to make it work, until eventually it did. It didn’t do anything reward-hacky or alien, something that only an AI could do. Apparently its approach was pretty human-like, and all the ingredients were in the maths literature somewhere or other, which might reflect the fact that it has been trained on all the maths proofs that humans have ever generated.

So why did AI manage to come up with this proof and not a human? Mathematicians have come up with three main reasons:

• The first is that they actually believe the conjecture was true, so very few of them were spending time actively trying to demonstrate that it was false.
• Second, the proof relied on bringing in a technique from a very different branch of mathematics — one that few people working on the unit distance problem would have likely encountered. And that’s a big AI strength in general. Even where they’re not as analytically good as the best humans, they have an extraordinary encyclopaedic knowledge of all kinds of different things, so they’re in a much better position to find connections that we humans might naturally have missed.
• The third reason is that the type of proof it pursued is especially laborious and tedious to go through. One human mathematician actually did start pursuing it in this exact case, but they quickly got overwhelmed by the complexity and they knew that success wasn’t guaranteed, so they decided their research time was better spent elsewhere.

So is this an update in favour of AGI arriving soon? Personally, I think it’s very unclear. We know and expect that AI models will be great at mathematics. They’re being trained on it and it plays to some of their core strengths: a highly structured environment with feedback that can be automated. If anything, given those advantages, I’m kind of a little bit surprised that they haven’t accomplished more in the maths domain already. I think a year or two ago I tweeted that I thought we might be on the verge of a real renaissance in AI mathematics. But that hasn’t turned out to be the case, at least not yet.

That might be because not many people are actually trying, but it’s also possible that a lot of attempts are being made — because OpenAI hasn’t actually said how much compute they had to throw at trying to crack different open mathematics puzzles in order to get this one significant piece of success. And without knowing that, it’s kind of hard to know whether to be, on balance, impressed or not.

I would change my mind on that if — in the chains of thought of the AIs when trying to solve maths problems — the AIs started to have big flashes of insight into the problem that they’re working on in the way humans occasionally do: thinking about a problem, then suddenly reconceptualising the whole situation and making some abrupt breakthrough. Typically, that’s actually not how humans solve maths problems. They more often do it by methodically applying familiar tools to new problems — basically like what happened here. And as far as I know, it’s not something that we’ve seen AIs do basically ever yet: have a massive flash of insight that no one else has before.

The reason that would update me more than this result is that the models aren’t being directly reinforced for having big flashes of insight. And it’s not something that even the best humans manage to do frequently. So it would suggest something new and important might be developing in their minds that wasn’t present before: a deeper understanding of the world and the problems that they’re dealing with, perhaps. Perhaps that thing would carry over to other research domains as well, things like economics or chemistry — or more importantly for our purposes — AI research and development.

The bottom line for me is that, in terms of getting to AGI, I’m going to assume that AI is going to be superhuman at maths and pay more attention to the more uncertain thing, which in my mind is AI’s ability to get good at messy, real-world tasks like running a café or a bookstore.

So that’s one piece of evidence that I thought was very cool, but it hasn’t particularly updated my AI timelines or forecasts yet.

Here’s a final one that I did update significantly on last time, but I now need to partially or significantly walk back.

Exhibit 7: Inference scaling wasn’t as big as believed [00:38:19]

In my previous timeline video, I suggested that more than half of capabilities gains observed through 2025 might have been driven by letting models use more compute to answer the same question — something known as inference scaling.

That’s a bearish indicator for AI, relative to making progress by getting the models to be more intelligent and insightful per token they output. And that’s because inference scaling requires you to use more compute every time you ask a question, rather than just one time while training the smarter model in the first place.

In January, I argued that to achieve the capabilities gains that we were observing, we were having to spend more and more money to get the answers. And as a result, we were trending towards AI becoming uneconomical — that is to say, it costing more to get an AI to complete a task than it would have cost to just get a human to do it in the first place.

Since then, new evidence has come out suggesting that things are not as extreme as I thought. An analysis of METR’s time horizon benchmark by Redwood Research found that as the tasks models have been doing have gotten more significant, the cost of getting AIs to successfully complete them has remained around 3% of the cost of getting a human being to do it.

How can that be the case if their chains of thought are getting so much longer? Well, the AIs are absolutely producing much longer chains of thought than they used to — hundreds or thousands of times longer than several years ago in fact. But the main reason for that is we’re giving them much more significant tasks now. And we know that if we give AIs bigger tasks, they’re going to need to use more compute and do more thinking in the process of solving it. That alone doesn’t suggest that AIs are becoming less economical to operate, because bigger tasks would also take a human longer to complete as well. And humans cost money too.

So when comparing very different types of tasks, we really need to look at how long the AI would need to think for, relative to how long a human would have taken to accomplish the same thing. And on that count, Redwood finds that the cost of AIs relative to humans isn’t going up.

There’s a smaller secondary factor at play as well: the models do produce longer chains of thought than they used to, even to complete the same task, but the dollar cost of each token output has decreased a bit over time, as chips and algorithms have gotten more efficient. That means you can have a somewhat longer chain of thought without it becoming unreasonably more expensive.

Now it’s just one analysis of one benchmark, so I don’t think this is the final word on the inference-scaling question. But that result — combined with companies’ demonstrated willingness to pay a lot for AI labour, if they’re doing useful work — that’s made me think the costs of inference scaling aren’t going to be a significant constraint on AI capabilities, at least in the next few years.

And that 3% figure, it’s kind of right in broad strokes. It’s an interesting number in another way because it suggests that — if we really, really want to — we still have a one-off opportunity to further improve AI performance sitting on the ground waiting for someone to pick it up. For something really important, we could give an AI 30 times as long to think about it as we do on average today — and that would still only cost as much as hiring a human being to do that same thing.

To be clear, this would only create a one-off jump in performance — and after that jump it would go back to the same trend as before. And it would come at a massive cost, which is that the AIs would no longer be cheaper than human staff. On top of that, the performance improvement you get from scaling inference compute by 30-fold, it isn’t always that large. At some point, both humans and AIs hit seriously declining returns thinking about a given problem for much longer.

But it’s relevant because AI companies might be willing to pay a huge amount to run models to replace their fabulously expensive staff in particular. And over time, models are getting better at organising and structuring their thoughts in such a way as to make practical use of really large compute budgets. So it’s one to watch.

How does that all change timelines? [00:41:41]

So how should all of these conflicting indicators actually affect our forecasts, our AI and AGI timelines?

Well, for better or worse, that depends on what you were personally predicting would happen in the first place — and so what from the above list was a surprise, what was an update for you?

But for me personally, there are five updates in favour of AGI coming sooner on the list, and those are:

1. Very useful AI agents arriving about a year or so earlier than I might have guessed
2. AI revenue exploding even faster than I, or almost anyone, anticipated
3. Anthropic reporting that AI is now significantly speeding up their research and their work
4. Mythos’s new pretraining run enabling frontier AI capabilities to jump three months ahead of schedule — something that might happen from time to time in future as well
5. And finally, inference scaling being less expensive than I thought it was

There’s one piece of evidence that for me personally is ambiguous, which is AI coming up with one original research result in maths.

There’s one thing that’s impressive, but it’s progressed only as I was expecting it to, which is AI doubling the length of tasks they can complete every four months, basically the same trend that we had before.

And there’s one piece of evidence in favour of AGI coming later, which is, at least in my mind: AI continuing to be pretty bad at messy real-world tasks, despite becoming really good at least some types of computer-based tasks.

On balance, I’d say my timelines to AGI — balancing all of this — they’ve shortened by something like one year as a result of everything we’ve seen in 2026. And I think that’s fairly typical of other people following this stuff closely, at least the ones who I read.

Back in January I said that personally I would be pretty shocked if we got fully automated AI research and development in 2027. In 2028, I guess it’s imaginable, but it’s going to require some surprising breakthroughs or an acceleration beyond what we’re seeing right now. But by 2029, 2030 begins to feel plausible.

Well, I think we got some surprising breakthroughs, which roughly means that all these categories for me have come forward by one notch. Now I’d be shocked if we got fully automated AI R&D this year, in 2026. Next year, 2027, that’s imaginable, and it seems to be what Anthropic is really gunning for. And fully automated AI research and development becoming possible in 2028 feels plausible if today’s trends just merely keep marching on for another two years.

In March, Paul Christiano, probably the only US government employee working on AI governance who has actually made major research contributions in AI itself, wrote the following in a personal capacity:

We are now around the 90th percentile of my AI timelines from 2021 [and from 2019]. I think we’ve now seen enough signs that there could be a relatively fast takeoff starting very soon. We no longer have strong quantitative indicators that transformative AI isn’t coming soon, and are flying blind. 1–4 year timelines are consistent with existing trends.

I agree with Paul, and frankly it’s quite scary.

Four reasons long timelines are still possible [00:44:25]

But back in January I also said there’s a very real chance that we’re in for a significantly longer and slower takeoff that takes until the mid-2030s, and I still believe that as well.

This video has gotten a little bit long, so the full justification for that might have to wait for another day. But here are the top four reasons longer timelines remain very much on the table in my mind:

First, nobody knows how broad the range of skills is that an AI has to be great at, in order to get fully autonomous recursive self-improvement off the ground. As I’ve said repeatedly, just being a great coder, clearly that’s not enough by itself. But do we need these models to be able to do everything an Anthropic staff member can do, even the most brilliant ones? Do we need them to be able to have research insights on the level of a superstar researcher, someone like Ilya Sutskever or Alec Radford? Or could they maybe make do by just being OK at some parts of the job, but being very numerous and really fast thinkers? I didn’t know the answer to that a year ago and I really don’t feel like I have any better idea about it now.

Second, there may be truly crucial capabilities necessary for general intelligence that are mostly missing from current models and which are very hard to train using current methods. Some possible options for that are:

• Generalising far out of distribution, and learning sample efficiency, to areas where AI models are still much, much weaker than human beings
• Another candidate is creativity and original insight and research, like I was talking about before — something which might require a type of internal world model or deep understanding that large language models aren’t currently developing, not in a full sense at least

Third, we don’t know how much using reinforcement learning from verified rewards — on coding and similar tasks like that — will lead to spillover improvements in other nonverifiable domains, ones with much worse feedback density. If those spillovers are weak, then we should expect to see AIs become exceptional superhuman coders while still remaining quite mediocre at more complex or open-ended projects, like running a small business or overseeing some research programme.

On the other hand, if the spillovers are strong, then scaling RLVR might be enough to get models to human-level at those things just by itself.

Four years on from the release of ChatGPT, as far as I can tell, we remain incredibly in the dark about how large those spillovers are. The labs might well know more internally from their own experiments, but if so, they’re not publishing — or at least no one’s emailed it to me yet.

Fourth, as I mentioned above, it’s possible we might automate AI R&D and find that, in the immediate term, it doesn’t make that much of a difference. The most likely reason for this would be if the whole research process just gets bottlenecked by access to the compute required to train models and run experiments.

If so, Anthropic might plausibly replace all of its staff with AIs and find that the overall rate of research progress only, say, doubles. Or it might increase 10-fold or 20-fold. We really just don’t know.

To my eye, these four big uncertainties are the four biggest cruxes of disagreement between people in the AI industry who think superintelligence might be here in 2028 on one extreme, and those who think it’s still 15 years away or 20 or 25 years away.

And on top of all of that — as I explained in some detail back in January — by 2030 the AI industry is on track to be absorbing a huge fraction of all the computer chips being manufactured and all of the memory being manufactured, which means they won’t be able to scale up training compute as fast as they are now, where they can basically get a boost to their scaling by progressively diverting more and more production away from all the other applications: gaming, phones, computers, all of that.

So unless we unlock something new by then, it will be quite reasonable to expect that capabilities advances will slow down after that point, and we’ll have a bit more of a slog to get to recursive self-improvement or artificial general intelligence or artificial superintelligence if we haven’t managed to do it by then.

It’s time to limit dangerous research practices [00:48:00]

Last time I spoke about AI timelines, I closed out by observing that even long timelines are ridiculously short now.

I pointed out that even if things play out in a way that’s surprisingly easy to handle, and AGI doesn’t arrive until 2036, that’s still an incredibly short period of time to figure out how to absorb such an enormous, wide-ranging shock without society descending into violence or authoritarianism, or some other horrible outcome.

That’s still completely true, but with another six months spent now and timelines even shorter, I would go further. In the past, as regular listeners will know, I’ve been pretty ambivalent about efforts to slow down or pause AI progress. Back in 2023, I was approached about signing the famous pause letter and ultimately decided, on balance, not to. But I think we’re approaching the crossover point now at which the benefits of slowing are going to start outweighing the costs in a way they simply didn’t before.

And I’m far from alone in suspecting that. Anthropic, OpenAI, and Demis Hassabis of DeepMind, they’ve all said that they want us to build the capacity to implement a coordinated pause — especially on the most dangerous types of research — should future evidence indicate that it’s necessary.

AI insiders, at least many of them, are legitimately scared of what’s coming next, and many are desperate to be given extra time, extra months, extra years to understand the machines they’re building before they’re unleashed on the world. In my opinion, we should give them that time. And on that note, I’ll speak with you again soon.