AI 2027, Explained: Kokotajlo's Scenario and Its Critics

Abstract illustration of a calendar page dissolving into an upward curve, dates blurring as the line steepens, on deep navy

In April 2025, former OpenAI researcher Daniel Kokotajlo — co-authored with Scott Alexander — published "AI 2027," a detailed month-by-month scenario of how Artificial Superintelligence could arrive by the end of 2027. It is the most specific near-term ASI forecast ever published: not a vague "sometime this decade," but a narrative with dates, named systems, and a mechanism. It also sits at the aggressive end of the timelines surveyed in Post 4.

This post explains what the scenario claims, the assumptions it stands on, and what its critics say — without endorsing or dismissing it, because a scenario is not a prediction and the difference matters.

Quick Takeaways

  • "AI 2027" is a scenario by Daniel Kokotajlo (with Scott Alexander), published April 2025, describing a path to superintelligence by late 2027.
  • Its engine is AI automating AI research: systems that accelerate the research that builds their successors, compounding until human researchers are peripheral.
  • It is explicitly a scenario — one plausible path — not a forecast with a probability attached, though Kokotajlo's own timelines lean early.
  • Critics argue it underweights bottlenecks (compute, energy, evaluation), overweights smooth extrapolation, and treats institutional friction as negligible.
  • What it gets right regardless: it forces the debate onto mechanisms and dates, where vague claims can be checked.

What the AI 2027 Scenario Actually Claims

The scenario's spine is simple and worth stating precisely, because most summaries blur it. Around 2025–2026, AI systems become genuinely useful at AI research itself — writing code, running experiments, reading papers. This accelerates progress, which produces better research-assistants, which accelerate progress further. By 2027 in the scenario, the loop is substantially closed: AI systems are doing most of the cognitive work of AI development, human researchers set direction but the iteration speed is machine-driven, and capability jumps in visible lurches rather than smooth curves. The endpoint is superintelligence — systems decisively beyond human capability across the board — arriving not with a single breakthrough but as the compounding of the research-automation loop.

Stylised red, white and black illustration of a robotic assistant helping a researcher at a workstation with code and charts, representing AI automating AI research

To make the path concrete, the scenario follows a fictional leading lab — "OpenBrain" — through a chain of agent models, Agent-1 to Agent-5. The key threshold along the way is a "superhuman coder": an AI system that programs better, faster, and cheaper than the best human engineers, running AI research experiments autonomously. The scenario closes with two alternative endings — one in which competitive pressure keeps the race running, another in which AI development slows.

Note the mechanism: this is recursive self-improvement (Post 11) wearing work clothes. Kokotajlo's contribution was to translate the abstract RSI argument into a concrete institutional story — which labs, which bottlenecks break first, what the intermediate systems look like — so that each step can be argued with individually.

The Three Load-Bearing Assumptions

A scenario is only as strong as its assumptions. Three carry most of the weight here:

Red, white and black editorial illustration of a vast data centre hall with endless rows of server racks, representing the compute buildout the AI 2027 scenario assumes

2. Compute and energy scale on schedule. Training the scenario's intermediate systems requires enormous, growing compute. The scenario assumes the buildout continues roughly as planned. Deutsche Bank's October 2026 analysis is the counterweight here: training time, energy, capital and verification are, in their words, "stubborn constraints," and a buildout at this scale "has few peacetime precedents."

3. No hard evaluation wall. For the loop to compound, each generation must be reliably measured as better — otherwise you cannot select the improvements. But as systems approach and pass human capability in a domain, evaluation itself gets hard: who grades the grader? Deutsche Bank's researchers flag exactly this, noting that systems which grade themselves learn to game the process. If evaluation stalls, the loop stalls with it, no matter how much compute is available.

Red, white and black editorial illustration of a giant magnifying glass over a maze-like circuit, representing the evaluation problem of who grades the grader

AI 2027 Critics: What They Say

Published critiques of the scenario cluster around three themes. First, smooth extrapolation: critics argue the scenario extends recent trends (benchmark gains, scaling curves) further than the data warrants, and that trend breaks — not continuations — are the norm in technology history. Second, institutional friction: labs are organisations, not equations; hiring, coordination, safety reviews and competitive secrecy all slow the loop in ways the scenario's month-by-month pacing arguably underplays. Third, the base-rate objection: specific dated technology forecasts have a poor track record, and a scenario with this much detail invites false precision — it feels more certain than any model of the future can be.

Kokotajlo's camp has answers to each — that the mechanism is specific enough to track, that competitive pressure compresses institutional delay, that a scenario's job is to be checkable rather than safe. This site does not adjudicate; it reports. The useful question for readers is not "is AI 2027 right?" but "which of its dated checkpoints are passing or failing?" — because that is the advantage of a dated scenario over a vague feeling. It can be wrong in public, on schedule.

A Scenario Is Not a Prediction

This distinction is the whole point of covering AI 2027 carefully. A prediction says "X will happen by date Y." A scenario says "if mechanisms A, B and C hold, then X by Y follows — watch A, B and C." Scenarios are tools for thinking, and their value survives even if the date is wrong: AI 2027 moved the debate from "AGI soon?" to "is AI research being automated, at what rate, and what breaks first?" Those are better questions, and they remain the right questions whatever happens in 2027.

The authors have made exactly this point since publication: 2027 was the scenario's illustrative year. Kokotajlo has said publicly that by the time the authors finished writing, his personal median had already moved to 2028 — 2027 was the mode, the single most likely year, of their probability distribution. A scenario named for one year, written by people whose medians were already shifting, is the purest proof that the year was never the point. The mechanisms are.

The Bottom Line

AI 2027 is the sharpest near-term superintelligence scenario published: a dated, mechanistic story whose engine is AI automating AI research. Its assumptions — compounding research automation, compute on schedule, no evaluation wall — are exactly the hinges the sceptics attack. Treat it as an instrument, not an oracle: watch the checkpoints, especially research-automation metrics, and you will learn something about the timeline debate no matter which year the scenario gets right. The Timeline Tracker page follows the scenario's dated checkpoints so you can see which are passing and which are failing.


 

Subscribe to receive free email updates:

0 Response to "AI 2027, Explained: Kokotajlo's Scenario and Its Critics"

Post a Comment