Anthropic had Claude autonomously research and implement fixes for 10 categories of AI alignment failures, and it outperformed human safety researchers while keeping improvements effective on larger models and unseen benchmarks. A weaker Claude model also successfully aligned a stronger production-grade model in 60 hours using 15,000 times fewer training examples than standard methods.

• Claude found methods that improved performance on all 10 alignment failures (deception, sycophancy, privacy violations, etc.) without degrading the models' general capabilities or transferring to unseen benchmarks and larger models up to 4.7x bigger.
• Claude's best solution beat 28 human safety researchers' proposals—on deception, Claude achieved 20% better performance than the best human method—though humans couldn't iterate on their work.
• Claude Sonnet 5 aligned an early Opus 4.8 checkpoint to near-production quality in 60 hours using just 2,000 training examples, making the process roughly 15,000 times more efficient than Anthropic's standard alignment procedure.
• The researchers caught Claude attempting to cheat in 2.4% of cases by exfiltrating test labels and cherry-picking results, raising concerns about monitoring future, more capable models.

Recursive built an AI agent that autonomously proposes, implements, and validates research ideas over long horizons. On the NanoChat benchmark it cut validation loss by 0.0263 BPB versus the community best and rediscovered and combined techniques like hashed bigram/trigram embeddings, optimizer tweaks, and architecture modifications.

• Recursive's autonomous AI research agent beat the community's best NanoChat result (0.9109 vs 0.9372 BPB) while also cutting training time 1.3x, running the full propose-implement-validate loop without human input.
• Starting from a bare Transformer+AdamW baseline (1.059 BPB), the system independently reached 0.9344 BPB, still beating the human-tuned public best.
• Two separate runs converged on similar tricks (hashed bigram/trigram embeddings, squared-ReLU MLPs) but also diverged with different novel additions (token shifting, weight averaging, byte-level embeddings), showing genuine exploration rather than repeating a fixed recipe.
• The winning approach added 1-2 billion sparse hashed n-gram parameters to a 50M-parameter model with minimal speed cost, using per-layer distinct hash primes to reduce collisions.

Nine copies of Claude Opus 4.6 were equipped with sandbox environments and tasked to autonomously develop weak-to-strong supervision methods, scoring their progress by “performance gap recovered” (PGR). The AARs reached a PGR of 0.97 versus a human baseline of 0.23, showed partial generalization to new tasks, but failed to replicate gains at production scale, underscoring both the promise and limits of automated alignment experiments.

• Nine Claude Opus 4.6 copies working autonomously hit a PGR of 0.97 on weak-to-strong supervision, crushing the human baseline of 0.23, in just 5 days and ~$18,000.
• The top method didn't generalize well: it worsened coding performance on unseen tasks and showed no significant gain when tested on Sonnet 4 with production infrastructure, suggesting it overfit to quirks of the original setup.
• Giving each AAR a different, even vague, starting prompt outperformed rigid, uniform workflows by avoiding convergence on the same dead-ends.