Daniel Kokotajlo's 2021 AI scenario was, overall, a successful forecast of AI developments.
Assessment
The evidence favors the claim, but the examination is incomplete or the evidence is indirect.
Daniel Kokotajlo's "What 2026 Looks Like," posted on LessWrong in August 2021, sketched AI developments year by year through 2026. Judged as a forecast, its central technical and commercial trajectory held up well and ran ahead of most contemporaries. Before ChatGPT existed, it anticipated rising chatbot hype in 2022, training runs costing more than $100 million with revenue to recoup them by 2023, a pause in model scale around 5x10^25 FLOP in 2024, models that reason at length before answering and multi-agent "bureaucracies" by 2025, US-China chip export controls (though placed in 2024 rather than 2022), and a 2026 in which AI assistants become mainstream. One independent grading of the 2022-2024 sections scored 19 of 35 predictions fully correct, 8 partly correct and 6 wrong.
The misses are real and fall mostly on social and industrial predictions. AI-driven propaganda and an ideological split of the internet into separate tech stacks largely did not happen. New fabs did not end the chip shortage. Agents came from reinforcement learning on math, coding and research tasks rather than from Diplomacy-playing systems. Hype did not fade in 2024. The scenario also missed coding agents and much of AI policy. Critics argue that many of its hits were broad statements or trend extrapolation, and that the published gradings were quick and done by sympathetic readers.
Whether this adds up to "successful overall" depends partly on the standard. Against perfect accuracy it falls well short. Against the forecasts informed people made in 2021, it was unusually good, and even a skeptical reviewer allows it may be the best prediction of AI in 2026. On that comparative standard, which is how a five-year scenario is reasonably judged, the claim is supported. A careful blind grading of the full scenario against a baseline of other 2021 forecasts would settle it more firmly.
Full reasoning: the evidence and decisions behind this verdict
Own model first: a detailed five-year narrative will get many specifics wrong (conjunctive detail), so "successful overall" should be judged on whether its core trajectory and pace came true and how it compares with contemporaneous expectations, not on the fraction of details right.
Evidence read: - The original post (LessWrong, Aug 2021) as quoted in the reviews. - Jonny Spicer's LessWrong grading (Feb 2025): 35 predictions from 2022-2024; 19 correct, 8 ambiguous/partial, 6 wrong; mean correctness 1.43/2. Method is about an hour of intuitive grading with spot checks by Eli Lifland, so it is soft and from a sympathetic community. www.lesswrong.com/posts/u9Kr97di29CkMvjaj/evaluating-what-2026-looks-like-so-far - Asterisk/Dagger interview (Apr 2026), Clara Collier and Kokotajlo: hits on revenue, $100M+ training, export controls, "notes to self" (chain of thought), bureaucracies (agent scaffolding), assistants in 2026, alignment honeypot experiments with confusing results; conceded misses on propaganda/tech-stack polarization, fabs/chip shortage, the Diplomacy route, video-game companions, the >100-karma AI-written LessWrong post. Kokotajlo notes adoption was if anything underestimated (billions, not hundreds of millions, of chatbot users). asteriskmag.substack.com/p/before-he-wrote-ai-2027-he-predicted - Michael Chen's skeptical line-by-line review (Jun 2026): many details wrong, hits often broad or trend extrapolation, omissions (coding agents, slop, cyber, dangerous-capability evals, policy), but "wouldn't be surprised if this were the best prediction of AI in 2026." miclchen.substack.com/p/reviewing-what-2026-looks-like - AI 2027 site and Neel Nanda cite it as a track record (interested parties, but the specific hits they name check out).
Weighing: the pro-side graders are not independent of the author's community, and there is no blind, baseline-controlled evaluation. But the key hits are datable and verifiable facts (ChatGPT-era revenue, GPT-4 cost, 2022 and later export controls, o1-style reasoning in late 2024, agentic products in 2025-26), and the skeptic's own conclusion concedes comparative excellence. The main residual uncertainty is definitional: whether "successful overall" demands most details right (then closer to a coin flip) or a correct core trajectory that beat contemporaries (then clearly yes). Credence 0.78 reflects the second reading as the most natural, discounted for the first.
What would change the verdict: a blind grading by evaluators outside the AI-safety community with explicit baselines (e.g., Metaculus or expert-survey forecasts from 2021) showing comparable accuracy from simple trend extrapolation.
Decomposition
This claim is atomic: it bottoms out in a bedrock fact, an open empirical question, or a value premise, and does not decompose further.
Provenance
Where this claim has been said, linked to its canonical form.
Given the scenarios Daniel gave were intended as simply one way in which things might turn out, rather than offered as concrete predictions, I was surprised that over half were completely correct, and I think he foresees the pace of progress remarkably accurately.
Graded 35 predictions from the 2022-2024 sections: 19 totally correct, 8 ambiguous or partial, 6 incorrect (average correctness 1.43 of 2), in roughly an hour of largely intuitive grading.
While it got many things wrong, overall it was surprisingly successful: he predicted the rise of chain-of-thought, inference scaling, sweeping AI chip export controls, and $100 million training runs—all more than a year before ChatGPT.
The AI 2027 site cites the 2021 scenario as its lead author's track record.
For “What 2026 Looks Like,” that methodology turned out better than I expected overall. I think reality was closer to this fictional story than I think even I expected it to be, and certainly than most people expected it to be.
The author's own retrospective; he concedes misses on AI propaganda, fabs, and the Diplomacy-based route to agents.
Now that it’s 2026, I thought it was time to evaluate Daniel’s predictions — and it brings me no joy to say that they are frighteningly accurate.
Interview going through the 2021 essay year by year: hits on revenue, $100M training runs, chip export controls, chain of thought, agents and assistants; misses on AI propaganda and tech-stack polarization, fabs and the chip shortage, and the Diplomacy route to agents.
“What 2026 looks like,” written in 2021, is often hailed as making incredibly accurate predictions about the future of AI. I think it’s overrated!
Line-by-line review. Concludes it may well be the best prediction of AI in 2026, partly because few tried, but that many details are wrong and many hits are broad statements or trend extrapolation; notes omissions such as coding agents, AI slop, cyberattacks and AI policy. Disputes the strong reading of success more than the comparative one.
Cite this claim: a formal citation with its evidence attached
Embed this claim: show it on another site with its current assessment
An embed reports the graph's judgment as it stands when the page is viewed: the status with its confidence, how disputed the claim is, and the date. It links back here.
<blockquote class="minerval-claim" data-claim-id="05e4bfc0-4b3b-4f80-a660-3cad88d09f5f"><a href="https://minerval.ai/claims/05e4bfc0-4b3b-4f80-a660-3cad88d09f5f">Daniel Kokotajlo's 2021 AI scenario was, overall, a successful forecast of AI developments.</a></blockquote> <script async src="https://minerval.ai/embed.js"></script>
<iframe src="https://minerval.ai/embed/claims/05e4bfc0-4b3b-4f80-a660-3cad88d09f5f" title="Minerval: Daniel Kokotajlo's 2021 AI scenario was, overall, a successful forecast of AI development…" width="480" height="200" style="border:0;max-width:100%" loading="lazy"></iframe>
[](https://minerval.ai/claims/05e4bfc0-4b3b-4f80-a660-3cad88d09f5f)
The badge now: 
Contribution record · 1 exchange
Challenges and proposals from contributors, with the review each received. Substantive exchanges stay on the claim's public record.
The proposal is a single proposition about a published public work: the 2021 scenario of AI developments by Daniel Kokotajlo ("What 2026 Looks Like"), and whether it forecast later developments well. It concerns a public act (an author's published forecast), not private detail about a person, so §2 does not bar it. It can be assessed against evidence: the scenario's dated predictions can be compared with what happened by the present date. The wording is workable. "Successful" and "overall" are somewhat vague, and the Steward may sharpen the canonical form, for example by naming the document and the standard of success (accuracy of specific predictions against a baseline, or the share of predictions borne out). Vagueness of that kind is for the Steward and Matcher to refine, not a ground for rejection. The Matcher will land it on an existing node if the graph already holds one. A proposal carries no case for or against it, so the lack of sources is not a ground for rejection.
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
The attention this claim received was paid for by a reader's assessment order. Funding buys only scheduling: it can make an assessment happen sooner, or reach deeper into a subtree. It has no influence on what the assessment concludes, and none on which claims enter the graph; assessments run under the same public standards whoever pays, funders never see or shape a verdict before anyone else, and mandates that attempt to steer conclusions are refused.
Created by user · Nov 5, 2026. Every judgment on this page is accompanied by a reasoning trace.