Why I don’t think AGI is right around the corner

Summary

• Dwarkesh’s core call: continual learning, not raw intelligence, is the binding constraint on AGI — LLMs are “stuck with the abilities you get out of the box,” with no way to give high-level feedback that compounds like a human employee’s on-the-job learning. His hundred-plus hours building LLM tools for his own podcast pipeline “extended my timelines.”
• He directly disputes Anthropic researchers Sholto Douglas and Trenton Bricken: if AI progress stopped today, less than 25% of white-collar employment goes away — versus their claim that current models are economically valuable enough to automate white-collar job tasks within five years.
• Three reasons to bet against their end-of-2026 “reliable computer-use agents” forecast: longer rollouts slow progress, no large multimodal pretraining corpus exists (“imagine trying to train GPT-4 on all the text data available in 1980”), and even “simple” ideas take years — GPT-4 to o1 took two years.
• His 50/50 timelines: AI does small-business taxes end-to-end as well as a competent general manager could in a week by 2028; AI that learns on the job as organically as a human — a video editor with six months of accumulated taste — by 2032. “We’re in the GPT-2 era for computer use,” and GPT-2 to GPT-4 took four years.
• The bullish flip side: solving continual learning triggers a discontinuity — one AI amalgamating learnings across all copies “is basically learning how to do every single job in the economy,” and might still get something that looks like a broadly deployed intelligence explosion even without a software-only singularity.
• Structural asymmetry for investors: “AGI timelines are very lognormal. It’s either this decade or bust” — 4x/year training-compute scaling cannot continue beyond this decade on chips, power, or the raw fraction of GDP used on training, so yearly AGI probability collapses afterward. If he ends up on the longer side of his 50/50 bets, the world may look “relatively normal” into the 2030s or even the 2040s.

Deep dive

1. The bottleneck isn’t intelligence — models don’t keep learning your job

• Dwarkesh’s ground truth from over 100 hours building LLM tools for his post-production setup: transcript rewriting, clip identification, passage-by-passage essay co-writing — “simple, self contained, short horizon, language in, language out tasks” dead center in the repertoire — and they’re 5/10 at them. Humans are useful not mainly for raw intellect but for building context, interrogating failures, and picking up improvements through practice. The Fortune 500 aren’t slow adopters; “it’s genuinely hard to get normal humanlike labor out of these LLMs.”
• The signature analogy, kept verbatim in spirit: teaching saxophone by sending each student away after one mistake and handing the next student refined written instructions. “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.”
• His editors got great by noticing small things and internalizing his taste — not via “bespoke RL environments for different subtasks.” RL fine-tuning isn’t deliberate or adaptive like human learning. He can imagine a smarter model building an organic RL loop for itself from high-level feedback, but says it sounds hard and may not generalize across tasks and feedback. Models do get smart mid-session (after he rewrites four bad paragraphs — “your shit sucked” — suggestions improve), but the tacit understanding evaporates at session end. Rolling compaction à la Claude Code is “brittle” outside text-native software engineering; Claude Code itself will often reverse a hard-earned optimization that they engineered together before /compact.

2. Disagreeing with Anthropic: automation without learning caps out low

• Against Trenton Bricken’s on-podcast claim that white-collar tasks get automated within five years even if progress stalls, Dwarkesh’s counter: under 25% of white-collar employment disappears. Claude 4 Opus can technically rewrite his transcripts, but without improvement over time “I still hire a human for this” — subtask competence doesn’t make an employee.

3. Why the “do my taxes by end of 2026” forecast is doubtful

• Sholto and Trenton foresee agents that email for invoices, sort business expenses, and submit Form 1040 by end of next year. Dwarkesh’s three objections: longer horizons mean two-hour rollouts before we can even see if it did it right, plus compute-heavy image/video processing; no large multimodal computer-use pretraining corpus — quoting Mechanize: internet text “was enough to crack natural language processing, but not for getting models to become reliable, competent agents”; and history — the RL procedure DeepSeek explained in its R1 paper seems simple at a high level, yet GPT-4 to o1 took two years. “That’s precisely my point!”
• His hedges preserved: maybe text training gives a good UI prior, maybe RL is sample-efficient, maybe models generate millions of toy UIs to practice on — “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.”

4. Enough cold water — “it’s actually working”

• Read o3 or Gemini 2.5 reasoning traces: “It’s actually reasoning!” Watching Claude Code zero-shot a working app from a vague spec, the most accurate explanation “is simply that it’s powered by a baby general intelligence.” Pessimists “haven’t played around with the smartest models in domains where they’re the most competent.”
• And his distributions are wide: preparing for a misaligned 2028 ASI “still makes a ton of sense” — a totally plausible outcome.

5. The dated bets — and why it’s this decade or bust

• 50/50 lines: taxes for a small business end-to-end as well as a competent general manager could in a week — 2028 (“GPT-2 era for computer use” — no corpus, sparser rewards, unfamiliar action primitives, but more compute and researchers “might even out”); human-grade on-the-job learning 2032 — “7 years is a really long time! GPT-1 had just come out this time 7 years ago.” Cool demos in 2026–27, yes — GPT-3 was cool but not practically useful.
• When continual learning cracks: copies amalgamate learnings, “one AI is basically learning how to do every single job in the economy” — might rapidly become a superintelligence even with no further algorithmic progress. Expect a “broken early version” first, not an OpenAI livestream announcement — “lots of heads up.”
• The closing structure: >4x/year training-compute growth “cannot continue beyond this decade” on chips, power, or the raw fraction of GDP 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, so yearly AGI probability collapses — lognormal, “either this decade or bust” (his own caveat: really “lower marginal probability per year — but that’s less catchy”). If he ends up on the longer side of his 50/50 bets, the world may look “relatively normal” into the 2030s or 2040s; “in all the other worlds… we have to expect some truly crazy outcomes.”