Anthropic cofounder says AI self-improvement has a 60% chance by 2028

Anthropic联创定下deadline：2028年AI实现自我进化，没有人类了

Anthropic cofounder Jack Clark says human-free AI R&D has over a 60% chance by end-2028. He cites SWE-Bench, CORE-Bench, MLE-Bench, and PostTrainBench: Claude Mythos Preview reaches 93.9% on SWE-Bench, and Opus 4.5 reaches 95.5% on CORE-Bench. The key signal is longer task horizons and post-training capability, not the “singularity” framing.

Why it matters: HKR-H/K/R all pass: a named Anthropic cofounder gives a 2028 timeline, backed by benchmark numbers. The headline is overheated, but the concrete claims and practitioner stakes justify P1.