Anthropic's AI acceleration is probably real

An analysis of METR's report on Anthropic's internal AI acceleration argues that public data supports the finding of a capability trend break after October 2025. The author contends that the reported acceleration is likely real and primarily driven by AI productivity gains rather than alternative factors like training compute or hiring. However, whether this trend will continue toward runaway recursive self-improvement remains uncertain due to potential verifiability bottlenecks. Open article →