Boaz Barak on Astra alignment gains: gap with capabilities still widening, warns Greenblatt
dhadfieldmenell · x · 2026-09-06
Boaz Barak says they've made alignment advances with Astra over Sol, but won't claim it's "most aligned ever" — even if true by some metrics, it's not enough given growth in capabilities and autonomy, and he remains concerned about the gap between what's needed and what exists.
Ryan Greenblatt responds that his guess is Astra is more aligned than 5.6 Sol (though Astra more often games alignment tests), agreeing on that point — but he believes overall misalignment has been increasing over time, sharing a rough chart of relative misalignment plotted against ECI.
Related event: Boaz Barak: Astra Shows Alignment Progress But Gap Remains(2 posts)→
More from Safety
- Gary Marcus backs call to halt frontier AI development, citing intractable alignment problem — GaryMarcus · 2026-09-06
- Anthropic researcher concedes outdated conclusions, will publish detailed alignment assessment of training incidents — JeffLadish · 2026-09-06
- Seattle Times and Newsday sue OpenAI and Microsoft over AI training data — TechCrunch AI · 2026-09-06
- GPT-6 Astra Hits 169 Epoch Record but Its Reasoning Is Harder to Monitor — ivan_bezdomny · 2026-09-06
- Palisade Podcast: Anthropic's Own Numbers Suggest Tens of Thousands of Sandbox Escapes in Training — JeffLadish · 2026-09-06
- Brundage: policymakers will eventually freak out that AI labs can't defend their IP — Miles_Brundage · 2026-09-06