What this covers
Dwarkesh Patel examines why artificial general intelligence is not as imminent as many recent forecasts suggest, taking a skeptical stance toward timelines that cluster in the mid-to-late 2020s. Rather than arguing that AI lacks raw intelligence or sufficient data, he identifies continual learning — the ability to accumulate knowledge and improve through feedback over time — as the primary bottleneck. The 17-minute video walks through three main areas where he sees progress stalling: why current language models cannot function as reliable workers despite automating individual tasks, why computer use agents are much further from deployment than some labs claim, and how solving continual learning could trigger rapid economic transformation once achieved.
Patel's argument rests on concrete capability gaps. Large language models can assist within a single conversation, but they shed all context and learned preferences once a session ends; tweaking prompts cannot replicate the organic improvement a human employee builds up over weeks. He contrasts this with the way computer use development faces missing training data, long rollouts that slow reinforcement learning, and image-processing overhead that compounds sample inefficiency — conditions he likens to training GPT-4 on only 1980s text. He singles out recent forecasts from Anthropic researchers claiming reliable tax-filing agents by the end of 2026 as implausibly optimistic. On timelines, Patel forecasts continual learning arriving around 2032, with AGI following shortly after—a horizon he frames as reasonable given how much changed in the seven years from GPT-1 to present. He also acknowledges that the probability distribution is wide enough that preparing for earlier, misaligned outcomes remains prudent.
Dwarkesh Patel argues that AGI is not imminent because today's LLMs lack continual learning—the ability to build up context and improve on the job like humans—and that solving this bottleneck (not raw intelligence or more data) is the true gate to transformative AI, which he forecasts for the early 2030s.
- LLMs cannot learn organically from feedback the way human employees do, making them unable to function as real employees even if individual subtasks are automatable
- Computer use agents face a missing pretraining corpus, longer rollouts, and sparse rewards, so reliable end-to-end agentic tasks are further off than labs claim
- Once continual learning is solved, AI copies amalgamating learnings across all jobs could trigger a broadly deployed intelligence explosion
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
When continual learning is solved, there will be a huge discontinuity in the value of these models: even without a software-only singularity, AIs deployed broadly across the economy learning on the job—and amalgamating their learnings across all copies so one AI learns every job—could rapidly become a superintelligence even with no further algorithmic progress, producing a broadly deployed intelligence explosion.
“When we do solve continual learning, we'll see a huge discontinuity in the value of these models.”
After 2030 AI progress must mostly come from algorithmic progress, but the low-hanging fruit will be plucked at least under the deep learning paradigm, so the yearly probability of AGI collapses—meaning on the longer side of his bets we might see a relatively normal world up to the 2030s or even 2040s, while in all other worlds we should expect truly crazy outcomes.
“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.”
AGI timelines are very lognormal—it's roughly 'this decade or bust', better described as lower marginal probability per year—because AI progress has been driven by scaling training compute (over 4x a year) which cannot continue beyond this decade given limits on chips, power, and the fraction of GDP spent on training.
“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).”
The lack of continual learning is a huge bottleneck: there's no way to give a model high-level feedback, so you're stuck with the abilities you get out of the box; tweaking the system prompt does not produce anything close to the learning and improvement human employees experience.
“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.”
Compacting session memory into a text summary (like Claude Code's /compact every 30 minutes) will be brittle outside software engineering, because coding is text-based and already has an external memory scaffold in the codebase; titrating rich tacit experience into a text summary loses hard-earned optimizations whose rationale doesn't make it into the summary.
“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.”
Patel forecasts 2032 for an AI that can learn on the job as organically and quickly as humans for any white collar work—e.g. an AI video editor that after six months has as deep an understanding of his preferences and audience as a human would—reasoning that 7 years is a long time (GPT-1 was only 7 years ago) so finding a way to make models learn on the job is not implausible.
“if I hired an AI video editor, after six months it would have as much actionable, deep understanding of my preferences, our channel, what works for the audience, as well as a human would. I'd say this would come in 2032.”
Today's LLMs would NOT be more economically transformative than the internet even if all AI progress stopped, because it is genuinely hard to get normal humanlike labor out of them—not because management is too stodgy—due to fundamental capabilities these models lack.
“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.”
If AI progress totally stopped today, less than 25% of white collar employment would go away, because while many subtasks get automated, the inability to improve over time and learn preferences means firms still hire humans; without progress in continual learning we'd be in a substantially similar position for most white collar work.
“If AI progress totally stops today, I think less than 25% of white collar employment goes away.”
LLMs do get smarter and more useful within a single session—after you rewrite its bad paragraphs and tell it what you actually wrote, it starts giving good suggestions—but this subtle understanding of the user's preferences and style is lost by the end of the session.
“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.”
A smarter model could in principle build a dedicated RL loop for itself—generating verifiable practice problems and rehearsal environments from high-level feedback—but this sounds really hard and it's unclear how well such techniques would generalize across different kinds of tasks and feedback; it's hard to see it happening within a few years given no obvious way to slot continuous learning into current LLMs.
“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.”
Patel forecasts a 50/50 bet that an AI capable of end-to-end small-business taxes (chasing receipts across websites, emailing for invoices, filling and submitting the form) as well as a competent general manager could in a week arrives by 2028; he frames computer use as currently in the 'GPT-2 era' with no pretraining corpus, sparse rewards, long horizons, and unfamiliar action primitives, analogizing the leap to the 4 years it took GPT-2 to GPT-4.
“An AI that can do taxes end-to-end for my small business as well as a competent general manager could in a week ... This I'd say 2028. I think we're in the GPT 2 era for computer use.”
Teaching LLMs is like trying to teach a kid saxophone by sending each student away after one mistake and refining written instructions for the next student—no matter how honed the prompt, no kid learns saxophone from instructions; this text-only, non-adaptive modality (even with RL fine-tuning) is fundamentally different from deliberate human learning.
“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.”
There is no large pretraining corpus of multimodal computer-use data, and just as the internet text corpus cracked NLP but wasn't enough for reliable agents (analogous to trying to train GPT-4 on only 1980's text data), it's unclear that text-only training gives a good prior over UIs or that RL fine-tuning is sample-efficient enough; there's no public evidence these models got less data hungry in unpracticed domains.
“We don't have a large pretraining corpus of multimodal computer use data.”
The reason humans are so useful is not mainly their raw intellect but their ability to build up context, interrogate their own failures, and pick up small improvements and efficiencies as they practice a task.
“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.”
Even algorithmic innovations that seem simple in retrospect take a long time to iron out—DeepSeek's R1 RL procedure looks simple at a high level, yet it took 2 years from GPT-4 to o1—which suggests we're underestimating the difficulty of the much gnarlier computer-use problem operating in a different modality with much less data.
“The RL procedure which DeepSeek explained in their 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.”
Computer-use progress will be slow because as horizon lengths increase rollouts must become longer (the AI must do hours of agentic tasks before you can verify success), and computer use requires processing images and video which is already more compute intensive even before factoring in longer rollouts.
“As horizon lengths increase, rollouts have to become longer. The AI needs to do two hours worth of agentic computer use tasks before we can even see if it did it right.”
Because his probability distributions are genuinely wide, work to prepare for a misaligned 2028 ASI still makes a ton of sense—a misaligned superintelligence by 2028 is a totally plausible outcome.
“work to prepare for a misaligned 2028 ASI still makes a ton of sense. I think this is a totally plausible outcome.”
LLMs are only about 5/10 at simple, self-contained, short-horizon, language-in-language-out tasks (rewriting transcripts, identifying clips, co-writing essays), which should be dead center in their repertoire, because they don't get better over time the way a human would.
“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.”
Anthropic researchers Sholto Douglas and Trenton Bricken forecast reliable computer-use agents by the end of next year—able to be told 'go do my taxes', go through email, Amazon orders, and Slack, email people for invoices, compile receipts, decide business expenses, ask approval on edge cases, and submit Form 1040—a forecast Patel is skeptical of.
“they said that they expect reliable computer use agents by the end of next year.”
Modern reasoning models (o3, Gemini 2.5) are actually reasoning—breaking down problems, thinking through what the user wants, reacting to their own internal monologue, and self-correcting when pursuing unproductive directions—and the most proximal, concise, accurate explanation of Claude Code zero-shotting a working app is that it's powered by a baby general intelligence.
“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”
We will not see continual learning announced as fully solved in a single livestream; because labs are incentivized to release innovations quickly, we'll see a broken early version of continual learning (test time training) before something that truly learns like a human, giving lots of heads up before the bottleneck is fully 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.”
Human editors became extremely good not through bespoke RL environments but by noticing small things themselves, thinking about what resonates with the audience and what content the host likes, and improving their day-to-day workflows—a kind of organic learning LLMs lack.
“My editors have gotten extremely good. And they wouldn't have gotten that way if we had to build bespoke RL environments for different subtasks involved in their work.”
Some people are too pessimistic because they haven't played around with the smartest models in the domains where those models are most competent.
“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.”
This was originally a blog post inspired by his podcast discussion with Sholto and Trenton, and he produces similar written artifacts after episodes (e.g. thoughts on Stephen Kotkin's writing about Stalin) available via his newsletter at dwarkesh.com.
“This was originally a blog post that I published on my website at dwarkesh.com.”