Research

AI 2027 relies on several key forecasts that couldn't be fully justified in the main text. Below we present the detailed research supporting these predictions.

Compute Forecast

AI 2027 predicts AIs trained with 1000x more compute than GPT-4 and the internal deployment of hundreds of thousands of AI research assistants by 2027. This supplement introduces the compute production model and the inference compute model behind these predictions.

Timelines Forecast

AI 2027 predicts AIs with superhuman coding ability in 2027. This supplement details our model for forecasting when the Superhuman Coder milestone will be achieved, building off METR's time horizon work, and justifies why we think it’s plausible it is achieved by 2027.

Takeoff Forecast

AI 2027 predicts generally superintelligent AIs in 2028. This supplement introduces our framework for thinking about the progression from AIs that are superhuman coders to generally superintelligent, and justifies why we think it's plausible the progression would take about 1 year.

AI Goals Forecast

AI 2027 predicts that superhuman AIs will not be aligned to the values and goals intended by their human developers. This supplement justifies that assumption by discussing the possibilities for what goals the AIs end up with.

Security Forecast

AI 2027 predicts AI models will be stolen due to poor security and highly valuable algorithmic secrets will leak through espionage. This supplement justifies why we think this is plausible.