Continual learning is the crux: Dwarkesh's timelines post
Dwarkesh Patel's June 2 essay argues that AGI is not close because models cannot learn on the job, and he backs it with his own failed attempts to use them for podcast production. Reading notes, the two bets he puts on the table, and an addendum on how the same objection turned up in Karpathy's and Sutskever's mouths by the end of the year.
The argument in one paragraph
Patel's claim is that the thing standing between current models and something you could call an employee is continual learning, the ability to get better at a job by doing it. His phrasing is blunt. The lack of continual learning is a huge huge problem, he writes, and while the LLM baseline at many tasks might be above an average human's, there is no way to give a model high level feedback. Everything else in the essay hangs off that sentence.
What makes the post worth reading rather than skimming is that the evidence is his own workflow. He has tried to use models to edit transcripts, find clips for social media and co-write essays. These are short, self-contained tasks with clear inputs, exactly the shape that should suit an LLM. He rates the results at five out of ten. The models are not bad at the tasks. They are stuck at the level they arrived at, and nothing he does across sessions moves them.
The saxophone and the vanishing session
The essay's best passage is an analogy about teaching a child the saxophone. A person learns by trying, hearing what went wrong, and adjusting, over and over. The way we teach a model is to write a prompt, watch it fail, rewrite the prompt, and hand the refined prompt to a different instance that has no memory of the failure. That is not the same process, and Patel's point is that there is no reason to expect it to converge on the same result.
He is careful to note the part that does work. Inside a single conversation the model does improve. When he rewrites one of its paragraphs badly, its next suggestions get closer to what he wanted. The learning is real and it evaporates when the context ends. That observation is doing a lot of work in the essay, because it locates the missing capability precisely. Models can take feedback. They cannot keep it.
What he is willing to bet on
Patel gives two even-odds predictions, which we appreciate because it makes the essay falsifiable. The first is that by 2028 an AI will be able to do a small business's taxes end to end, including chasing receipts, emailing for missing invoices, filling in the forms and submitting to the IRS. The second is that by 2032 a model will be able to learn on the job as organically and quickly as a human, across any white-collar work. Those are seven years apart, and the gap is his whole thesis in numeric form.
He also pushes back on colleagues who are more bullish. Sholto Douglas and Trenton Bricken at Anthropic expect reliable end-to-end computer-use agents by the end of 2026. Patel's doubts are specific: long tasks take long rollouts, there is little pretraining data for multimodal computer use, and ideas that look simple in retrospect, like the reasoning approach behind o1, took years to land. His framing of the overall distribution is lognormal. Either this happens within the decade while compute scaling is still cheap, or the annual probability falls off sharply after 2030 when it stops being cheap.
The replies worth reading
Two responses were appended to the post and they sharpen it. Daniel Kokotajlo says his median for an intelligence explosion is early 2028, and that his disagreement with Patel comes down almost entirely to when continual learning arrives, 2028 in his view against Patel's 2032. That is a useful thing to know. Two people with very different headline timelines agree on which capability is the bottleneck and differ only on its date.
Ryan Greenblatt's reply is more technical. He argues that reinforcement learning fine-tuning on self-verified outcomes could stand in for human-style learning, and that sample efficiency has been improving. We find that the weaker counterargument, because it assumes the environment gives you a verifiable reward, and most of Patel's five-out-of-ten tasks do not. Whether a clip is good for social media is not something a unit test can check.
Addendum, December 2025
We are adding this six months on because the objection has turned into something close to consensus among people who do not otherwise agree. On October 17, Andrej Karpathy sat down with Patel and explained his phrase about a decade of agents as a reaction to overprediction. Asked what the current systems lack, he named continual learning, in almost Patel's words: you cannot just tell them something and they will remember it. He added a mechanism he thinks is missing, a sleep-like phase where what sits in context gets distilled into weights.
On November 25, Ilya Sutskever made the same move from the other direction. He said the models generalise dramatically worse than people and called that a very fundamental thing, and he described superintelligence as a mind that can learn to do every job, deployed the way a new hire is, with a learning trial-and-error period rather than a finished product. His stated timeline for that is five to twenty years. Patel's 2032 sits inside that range. What we take from the three of them together is that the disagreement about AGI has narrowed to one open research problem, and we would rather see a lab publish a small, honest result on weight-level learning from deployment feedback than another timelines essay.
Sources
From the foundation