Import AI 454: Anthropic's Automated Alignment Researchers Recover 97% of Performance Gap at $22/Hour — Plus HiFloat4 and MirrorCode

Import AI·high signal

Jack Clark's Import AI 454 covers three significant developments: (1) Anthropic's nine parallel Claude Opus 4.6 agents outperformed human researchers on weak-to-strong supervision, recovering 97% of the performance gap vs. humans' 23%, at $18K total ($22/research-hour). (2) Huawei's HiFloat4 4-bit format beats Open Compute Project's MXFP4 with ~1.0% relative loss vs ~1.5%. (3) METR/Epoch's MirrorCode benchmark shows Opus 4.6 autonomously reimplemented a 16K-line Go bioinformatics toolkit — a task estimated at 2-17 weeks for a human. The automated alignment result is particularly significant: AI research acceleration is no longer theoretical.