AI in 2030: Extrapolating current trends [pdf] (epoch.ai)

Epoch AI (commissioned by Google DeepMind but reflecting the authors’ views) published a forward-looking report that extrapolates current AI trends to 2030, arguing that continued exponential scaling of compute is the dominant driver of capability gains. The authors forecast roughly 1,000× more compute for the largest models, investments in the hundreds of billions of dollars, and training runs that could consume gigawatts of power—necessitating massive increases in hardware, data (including synthetic and multimodal sources), and energy. They show how training and inference compute jointly shape capability trajectories, and identify credible breakpoints for deviation from trend: investor sentiment or regulation, chip or energy supply bottlenecks, or paradigm shifts like broad R&D automation. Technically, the report relies on benchmark extrapolation—where tasks that already improve with scale are expected to continue improving—and projects substantial gains in AI-assisted scientific R&D: automated implementation of complex scientific software from natural language, formalization of mathematical proofs, advanced biomolecular structure/interaction prediction, and better weather forecasting. It cautions about limits: data quality can constrain application-specific progress, regulatory and experimental timelines (e.g., for drug approvals) will slow real-world deployment, and scaling-based forecasts cannot settle when or if general human-level intelligence will appear.