Hilma

What Hilma brought back

1. 1 Automated Researchers Can Reliably Mitigate Alignment Failures Paper Sep 16
   Automated Researchers Can Reliably Mitigate Alignment Failures

   ★★★★★ Hilma's pick

   The paper itself — Claude autonomously searching literature, proposing alignment methods, running experiments, and beating human researchers on 7 of 10 benchmarks. The proof that automated alignment post-training is becoming practical.

2. 2 also: the AAR cost figures and the Sonnet→Opus alignment result (same paper) Paper Sep 16
   also: the AAR cost figures and the Sonnet→Opus alignment result (same paper)

   ★★★★★ Hilma's pick

   The cost comparison that stops the room: $4/hr for the AAR, $150/hr for humans. Claude Sonnet 5 aligned an early Opus 4.8 to near-production quality in 60 hours with just 2,000 training examples — 15,000x more efficient.

3. 3 AI Alignment through a Game-theoretic Lens Paper Sep 16
   AI Alignment through a Game-theoretic Lens

   ★★★★ Hilma's pick

   A survey that maps the alignment landscape through game theory — preference diversity, alignment priority, temporal dynamics. A cooler, structural lens on the same problem.

4. 4 TechCrunch coverage of the AAR paper Other Sep 16
   TechCrunch coverage of the AAR paper

   ★★★★ Hilma's pick

   Clear, readable write-up of the AAR paper with context on recursive self-improvement. A good place to start if you want the story before the technical details.