New paper 'Training for Alignment' models how principals optimally train agents with endogenous values
Afinetheorem · x · 2026-10-29
Economists Alex Wolitzky and coauthor release 'Training for Alignment,' a framework for how a principal should optimally train an agent—human or AI—that endogenously chooses its own utility (values). The reposter quips that raising kids was the original RLHF.
More from AGI Musings
- Advertiser who spent $500k/mo on Google Ads says Search growth is artificial amid LLM cannibalization — SumitGup · 2026-10-29
- Billionaire foundations, not corporations, are the better misalignment example — EigenGender · 2026-10-29
- Mark Cuban: the defining divide is people who use AI to learn vs. to avoid thinking — MazMansoor · 2026-10-29
- Everyone talks about AI but they mean completely different things — almmaasoglu · 2026-10-29
- The printing press, not AI, was history's original slop cannon — RexDouglass · 2026-10-29
- Google researcher: people posting these AI-work demo videos are 'doing nothing of consequence' — prajdabre · 2026-10-29