AI 2027: month-by-month model of intelligence explosion — Scott Alexander & Daniel Kokotajlo

Scott Alexander and Daniel Kokotajlo describe AI 2027 as an effort to make rapid progress toward AGI and possible superintelligence concrete. Rather than jump from today's chatbots to a distant milestone, the scenario supplies a month-by-month sequence from the present through 2027 and 2028. Its purpose is not only to make that transition imaginable, but to make the forecast feel earned through intermediate steps.

Alexander emphasizes that forecasting accurately is harder than constructing a plausible story: the median forecast is humiliation when events unfold differently. His optimism comes partly from Kokotajlo's 2021 forecast, What 2026 Looks Like, which Alexander says anticipated much of the intervening progress. Kokotajlo is more measured, describing that earlier post as partly right and partly wrong, but useful enough to motivate a better sequel. He had originally planned to continue into AGI and superintelligence, then stopped when the developments and uncertainty became too difficult to model.

Alexander says he joined because he admired the people involved and trusted Kokotajlo's willingness to sacrifice millions of dollars rather than accept OpenAI's non-disparagement terms. Working with the team also made the scenario's concerns more concrete, including the prospect of an arms race with China. Patel describes his own shifting views on intelligence explosion as he tried to engage with the forecast and formulate objections.

The forecast's early stages focus on agents, longer task horizons, and improved coding. The proposed mechanism is recursive: AI systems begin helping with AI research, which can accelerate further progress. The authors summarize this with an R&D progress multiplier, measuring how many months of ordinary progress one month of AI-assisted research would produce. Their scenario reaches a fivefold multiplier for algorithmic progress during 2027, while expecting 2025 to remain broadly similar to current trends and 2026 to bring incremental gains in agents and coding.

Kokotajlo expects computer-use agents to stop making basic mouse and screen-reading errors by the end of 2025, while remaining unreliable over long tasks. A short office-planning task might become possible, but using an agent to run one's life could still produce viral mistakes. Alexander explains why the forecast prioritizes coding: if coding accelerates AI research, it may unlock much faster progress across other capabilities, even those that are not uniquely human bottlenecks.

The speakers disagree about whether observers have been too optimistic or too pessimistic. Patel points to the difficulty of scaling pretraining and the time required to develop reinforcement learning systems such as O1, arguing that economic effects remain less dramatic than benchmarks might suggest. Alexander counters that expert surveys have often assigned timelines of many years to capabilities that appeared much sooner. Kokotajlo similarly argues that people following the field have underestimated both technical progress and diffusion, while acknowledging that some more bullish forecasters have already been wrong. They cite Metaculus as one imperfect aggregate measure whose estimates have moved from around 2050 in 2020 to around 2030 today.

The discussion begins with an AI researcher's report that current models save four to eight hours a week in familiar domains, but roughly a day a week when he must work outside his expertise. Alexander and Kokotajlo suggest that language models may be especially useful when someone needs to learn an unfamiliar technical area, since the models have absorbed a broad range of material. Patel then asks why systems with so much stored knowledge have not made a clear scientific discovery by connecting ideas, as in the example of magnesium supplements helping some migraines.

Alexander distinguishes recognizing a connection after it is pointed out from discovering it independently. Etymology offers a simple example: words such as 'happy,' 'hapless,' and 'happen' look related, but most people do not notice their shared roots until they consult a dictionary. Patel counters with David Anthony's work connecting words for wheels, carts, and horses across Indo-European and Indian languages, which anticipated later genetic evidence about the Yamnaya. Alexander says such discoveries may depend on accumulated expertise, useful heuristics, and a sudden insight rather than merely storing facts.

Alexander imagines a scaffolded process in which agents systematically compare words, record possible matches, and investigate clusters, while acknowledging that brute-force comparisons would face combinatorial explosion. Kokotajlo adds that people have often mistaken a limitation of current language models for a permanent limit of the paradigm. The speakers discuss whether scaffolding, larger models, or training specifically for discovery could help; they agree that current pretraining does not strongly incentivize making novel connections. Kokotajlo also notes that building a suitable reinforcement-learning environment for scientific discovery would itself be difficult.