one year on
AI 2027 scenario predicts superhuman AI within two years, sparking debate on Hacker News
A detailed, quantitative scenario by former OpenAI researcher and co-authors envisions AI agents transforming coding and AI research by 2027, with alternative endings for a 'slowdown' or 'race' future.
A group of AI researchers and forecasters today published ‘AI 2027,’ a detailed scenario predicting that superhuman AI will arrive within two years, with impacts exceeding the Industrial Revolution. The document, authored by former OpenAI researcher Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland, and Romeo Dean, describes a concrete narrative from mid-2025 through 2027, including two possible endings: a hopeful ‘slowdown’ and a darker ‘race’ scenario.
The scenario assumes a fictional leading company called OpenBrain and traces the rise of AI agents that begin to automate coding and AI research, leading to a 50% faster algorithmic progress rate. It also paints geopolitical tensions, with China nationalizing AI research and attempting to steal model weights. The authors aim for predictive accuracy, drawing on trend extrapolations, expert feedback, and tabletop exercises.
The release prompted a vigorous debate on Hacker News, with 949 points and 621 comments as of today. Some commenters, like Vegenoid, argue that today’s AI is not radically different from 2023 and that the path to AGI is far more uncertain, with no signs of a runaway singularity. Others, such as Enginerrrd, assert that LLMs still fail at long-horizon tasks and require fundamentally different capabilities. Meanwhile, benlivengood points to emerging research on planning in models, suggesting progress may be underestimated.
The group encourages counter-scenarios and is offering prizes for alternative forecasts, aiming to broaden the conversation about steering AI toward positive futures.
The record
Argues that current AI is not radically different from 2023 and that exponential advance to AGI in two years is extremely unlikely, noting lack of convincing evidence behind closed doors.
Cites METR metrics showing progress on long-term tasks and points to Anthropic research on planning in language models, suggesting current architectures may already model concepts beyond token prediction.
Argues that LLMs lack the ability to handle weeks- or months-long tasks, comparing them to fresh college graduates versus experienced engineers, and says fundamental change is needed for AGI.
One year later — open only if you can handle spoilers
By mid-2026, AI capabilities have advanced but not followed the scenario's exact timeline; the debate over whether LLMs are merely next-token predictors or on a path to AGI remains unresolved, with some progress on long-horizon tasks but no clear signs of superhuman AI yet.
The Weekly Replay · free by email
This week, one year ago — every Sunday.
One email each Sunday: the week's replayed AI news, with the one-year-later annotations included. Written like it's breaking — dated like it isn't.
Free · double opt-in · unsubscribe anytime · privacy