Dwarkesh Patel
I’ve had a lot of discussions on my podcast where we hash out our timelines to AGI. Some guests think it’s 20 years away. Others think it’s 2 years away. Here’s where my thoughts lie as of July 2025.
Sometimes people say that even if all AI progress totally stopped, the systems of today would still be far more economically transformative than the internet. I disagree. I think the LLMs of today are magical. But the reason that the Fortune 500 aren’t using them to totally transform their workflows isn’t because the management there is too stodgy. Rather, I think it’s genuinely hard to get normal humanlike labor out of these LLMs.
And this has to do with some fundamental capabilities these models lack. I like to think I’m “AI forward” here at the Dwarkesh Podcast. I’ve probably spent over 100 hours trying to build these little LLM tools for my post-production setup. And the experience of trying to get them to be useful has extended my timelines.
I’ll try to get an LLM to rewrite autogenerated transcripts for me, to optimize for readability in the way a human would. Or I’ll try to get them to identify clips from a transcript that I feed in. Sometimes I’ll try to get them to co-write an essay with me, passage by passage. These are simple, self-contained, short-horizon, language-in, language-out tasks—the kinds of assignments that should be dead center in the LLMs’ repertoire. And they’re 5/10 at them.
Don’t get me wrong, that is impressive. But the fundamental problem is that LLMs don’t get better over time the way a human would. This lack of continual learning is a huge, huge bottleneck. The LLM baseline at many tasks might be higher than the average human’s. But there’s no way to give a model high-level feedback. You’re stuck with the abilities you get out of the box.
You can keep messing around with the system prompt, but in practice this just doesn’t produce anything even close to the kind of learning and improvement that human employees experience. The reason humans are so useful is not mainly their raw intellect. It’s their ability to build up context, interrogate their own failures, and pick up small improvements and efficiencies as they practice a task.
How would you teach a kid to play the saxophone? You’d have them try to blow into one, and then they’d see how it sounds, and they’d adjust. Now imagine if this was the way you’d have to teach saxophone instead: A student takes one attempt. The moment they make a mistake, you send them away and write detailed instructions about what went wrong. Then you call the next student in.
The next student reads your notes and tries to play Charlie Parker cold. When they fail, you refine your instructions and invite the next student. This just wouldn’t work. No matter how well-honed your prompt is, no kid is just going to learn how to play saxophone from reading your instructions. But this is the only modality we have to “teach” LLMs anything.
Yes, there’s RL fine-tuning. But it’s not a deliberate, adaptive process in the way that human learning is. My editors have gotten extremely good. And they wouldn’t have gotten that way if we had to build bespoke RL environments for the different subtasks involved in their work. They’ve just noticed a lot of small things themselves and thought hard about what resonates with the audience, what kind of content I like, and how they can improve their day-to-day workflows.
Now, it’s possible to imagine ways in which a smarter model could build a dedicated RL loop for itself that just feels super organic from the outside. I give some high-level feedback, and the model comes up with a bunch of verifiable practice problems to RL on—maybe even a whole environment in which it gets to rehearse the skills that it thinks it’s lacking. But this just sounds really hard. And I don’t know how well these techniques will generalize to different kinds of tasks and feedback.
Eventually, the models will be able to learn on the job in this organic way that humans can. But it’s just hard for me to see how that could happen within the next few years, given there’s no immediately obvious way in which to slot in continuous learning into the kinds of models that these LLMs are.
LLMs actually do get kind of smart and useful in the middle of a session. For example, sometimes I’ll co-write an essay with an LLM. I’ll give it an outline, and I’ll ask it to draft the essay passage by passage. And all its suggestions up till paragraph 4 will just be bad. I’ll rewrite every single paragraph from scratch and tell it, “Look, your shit sucked. This is what I wrote instead.”
At this point, it will actually start giving good suggestions for the next paragraph. But this whole subtle understanding of my preferences and style will just be lost by the end of the session. Maybe there is an easy solution to this that looks like a long rolling context window, like Claude Code already has, which just compacts the session memory into a summary every 30 minutes.
I just think that titrating all this rich tacit experience into a text summary will be brittle in domains outside of software engineering, which is very text-based, in which you already have this external scaffold of memory that is stored in the codebase itself. Again, think about what it would be like to teach a kid to play the saxophone just from text.
Even Claude Code will often reverse a hard-earned optimization that we engineered together before I hit “/compact”—because the explanation for why it was made didn’t make it into the summary. This is why I disagree with something that Anthropic researchers Sholto Douglas and Trenton Bricken said on my podcast. This quote is from Trenton:
“Even if AI progress totally stalls—and you think that the models are really spiky, and they don’t have general intelligence—it’s so economically valuable, and sufficiently easy to collect data on all of these different white-collar job tasks, such that, to Sholto’s point, we should expect to see them automated within the next 5 years.”
If AI progress totally stops today, I think less than 25% of white-collar employment goes away. Sure, many tasks will get automated. Claude 4 Opus can technically rewrite autogenerated transcripts for me. But since it’s not possible for me to have it improve over time and learn my preferences, I still hire a human for this.
So even if we get more data, without progress in continual learning, I think that we will be in a substantially similar position with all other kinds of white-collar work. Yes, technically AIs will be able to do a lot of subtasks somewhat satisfactorily, but their inability to build up context will make it impossible to have them operate as actual employees at your firm.
While this makes me bearish about transformative AI in the next few years, it makes me especially bullish on AI over the next decades. When we do solve continual learning, we’ll see a huge discontinuity in the value of these models. Even if there isn’t a software-only singularity, where these models rapidly build smarter and smarter successor systems, we might still get something that looks like a broadly deployed intelligence explosion.
AIs will be getting broadly deployed through the economy, doing different jobs, and learning while doing them in the way that humans can. However, unlike humans, these models can amalgamate their learnings across all their copies. So one AI is basically learning how to do every single job in the economy.
An AI that is capable of this kind of online learning might rapidly become a superintelligence even if there’s no further algorithmic progress. However, I’m not expecting to watch some OpenAI livestream where they announce that continual learning has been totally solved.
Because labs are incentivized to release any innovations quickly, we’ll see a broken early version of continual learning—or test-time training, or whatever you want to call it—before we see something which truly learns like a human. I expect to get lots of heads-up before this big bottleneck is totally solved.
1. Computer Use
When I interviewed Anthropic researchers Sholto Douglas and Trenton Bricken on my podcast, they said that they expect reliable computer-use agents by the end of next year. We already have computer-use agents right now, but they’re pretty bad. They’re imagining something quite different.
Their forecast is that by the end of next year, you should be able to tell an AI, “Go do my taxes.” It’ll go through all your email, your Amazon orders, and Slack messages, and it will email back and forth with every single person you need to get invoices from. It’ll compile all your receipts, decide what things actually are business expenses, ask for your approval on all the edge cases, and then just submit Form 1040 to the IRS.
I’m skeptical.
I’m not an AI researcher, so far be it from me to contradict them on technical details. But given what little I know, here’s why I’d bet against this forecast:
1. As horizon lengths increase, rollouts have to become longer. The AI needs to do 2 hours’ worth of agentic computer-use tasks before we can even see if it did it right. Not to mention, computer use requires processing images and video, which is already more compute-intensive, even if you don’t factor in the longer rollouts. This seems like it should slow down progress.
2. We don’t have a large pretraining corpus of multimodal computer-use data. I like this quote from Mechanize’s post on automating software engineering:
“For the past decade of scaling, we’ve been spoiled by the enormous amount of internet data that was freely available for us to use. This was enough to crack natural language processing, but not for getting models to become reliable, competent agents. Imagine trying to train GPT-4 on all the text data available in 1980—the data would have been nowhere near enough, even if you had the necessary compute.”
Again, I’m not at the labs. Maybe text-only training already gives you a great prior over how different UIs work and what the relationship is between different components. Maybe RL fine-tuning is so sample-efficient that you don’t need that much data. But I haven’t seen any public evidence that makes me think these models have suddenly gotten less data-hungry, especially in domains where they’re substantially less practiced. Alternatively, maybe these models are such good front-end coders that they can just generate millions of toy UIs for themselves to practice on.
3. Even algorithmic innovations which seem quite simple in retrospect took a long time to iron out. The RL procedure which DeepSeek explained in their DeepSeek-R1 paper seems simple at a high level. And yet it took 2 years from the development and launch of GPT-4 to the release of o1. Now, of course, I know that it’s insanely and hilariously arrogant to say that R1/o1 were easy—a ton of engineering, debugging, and pruning of alternative ideas was required to arrive at this solution. But that’s precisely my point.
Seeing how long it took to implement the idea of “We should train a model to solve verifiable math and coding problems” makes me think that we’re underestimating the difficulty of solving the much gnarlier problem of computer use, where you’re operating in a totally different modality with much less data.
2. Reasoning
Okay, enough cold water. I’m not going to be like one of those spoiled children on Hacker News who could be handed a golden-egg-laying goose and would still spend all their time complaining about how loud its quacks are.
Have you read the reasoning traces from o3 or Gemini 2.5? It’s actually reasoning. It’s breaking down a problem, thinking through what the user wants, reacting to its own internal monologue, and correcting itself when it notices that it’s pursuing an unproductive direction. How are we just like, “Oh yeah, of course a machine is gonna go think a bunch, come up with a bunch of ideas, and come back to me with a smart answer. That’s just what machines do”?
Part of the reason some people are too pessimistic is that they haven’t played around with the smartest models in domains where they’re the most competent. Giving Claude Code a vague spec and just sitting around for 10 minutes while it zero-shots a working application is a wild experience. How did it do that? You can talk about circuits and the training distribution and RL or whatever, but the most proximal, concise, and accurate explanation is simply that it’s powered by a baby general intelligence.
3. So what are my predictions?
At this point, part of you has to be thinking, “It’s actually working. We’re making machines that are intelligent.” My probability distributions are super wide, and I want to emphasize that I do believe in probability distributions. That means that work to prepare for a misaligned 2028 ASI still makes a ton of sense. I think this is a totally plausible outcome.
But here are the timelines at which I’d take a 50/50 bet. An AI that can do taxes end-to-end for my small business as well as a competent general manager could in a week—including chasing down all the receipts on different websites, finding the missing pieces, emailing back and forth with anyone we need to hassle for invoices, filling out the form, and sending it to the IRS—I’d say 2028. I think we’re in the GPT-2 era for computer use. But we have no pretraining corpus, and the models are optimizing for a much sparser reward over a much longer time horizon using action primitives they’re unfamiliar with. That being said, the base model is already decently smart and might have a good prior over computer-use tasks, plus there’s a lot more compute and AI researchers in the world, so it might even out.
Preparing taxes for a small business feels like, for computer use, what GPT-4 was for language. It took 4 years to get from GPT-2 to GPT-4. Just to clarify, I am not saying that we won’t have really cool computer-use demos in 2026 and 2027. GPT-3 was super cool, but it was not that practically useful. I’m saying that these models won’t be capable of end-to-end handling a week-long and quite involved project which involves computer use.
Okay, and as for the forecast of when AI will be able to learn on the job as easily, organically, seamlessly, and quickly as humans, for any white-collar work: for example, if I hired an AI video editor, after 6 months it would have as much actionable, deep understanding of my preferences, our channel, and what works for the audience as a human would. I’d say this would come in 2032.
While I don’t see an obvious way to slot continuous online learning into the kinds of models these LLMs are, 7 years is a really long time. GPT-1 had just come out this time 7 years ago. It doesn’t seem implausible to me that over the next 7 years, we’ll find some way to get these models to actually learn on the job.
At this point you might be reacting, “Wait, you made this huge fuss about continual learning being such a huge handicap. But then your prediction is that we’re 7 years away from what, at a minimum, looks like a broadly deployed intelligence explosion.” And yeah, you’re right. I am forecasting a pretty wild world within a relatively short amount of time. AGI timelines are very lognormal. It’s either this decade or bust—not really; it’s more like lower marginal probability per year, but that’s less catchy.
AI progress over the last decade has been driven by scaling training compute on the frontier systems. It’s been over 4x a year. This cannot continue beyond this decade, whether you look at chips, power, or even the raw fraction of GDP that is used on training. After 2030, AI progress has to mostly come from algorithmic progress. But even there, all the low-hanging fruit will be plucked, at least under the deep-learning paradigm. So the yearly probability of AGI collapses.
This means that if we end up on the longer side of my 50/50 bets, we might be looking at a relatively normal world up till the 2030s or even the 2040s. But in all the other worlds, even if we stay sober about the current limitations of AI, we have to expect some truly crazy outcomes.
4. Dwarkesh.com
This was originally a blog post that I published on my website at dwarkesh.com. It was obviously inspired by the discussion I had with Sholto and Trenton on my podcast, where I ended up disagreeing with them about timelines, but it took me a few weeks of thinking afterward, sorting out exactly where I disagree and why I had longer timelines. And I do this for other episodes as well. I wrote up some thoughts I had about the many thousands of pages that Stephen Kotkin has written about Stalin, which we were not able to exhaustively cover in that one 2-hour interview.