Anthropic vs OpenAI: did Anthropic's bet on alignment qua alignment pay off?
AaronBergman18 · x · 2026-11-20
Commenter hamandcheese offers a medium-credence take: Anthropic invested early in alignment qua alignment, treating Claude as analogous to a biological mind, while OpenAI underrated it, treating GPT as a tool shaped by narrow RL policies. Both reportedly realized the inadequacy of their approaches amid the RLVR scale-up, but Anthropic has been more adept at holistically handling their models' "psychology," whereas OpenAI pushes forward with extremely capable math solvers that still can't be trusted to babysit your repo.
The post is part of a discussion around METR's RSI report, questioning whether OpenAI's lack of a trend-break on OECI is a methodological artifact—or a sign that AI is further into recursive self-improvement despite similar ECIs, which would slightly lengthen timelines.
More from AGI Musings
- Alex Stamos: humans can't write safe code in the age of superhuman bug-finding AI — GaryMarcus · 2026-11-20
- Why doing the reading on AI safety is the fastest way to stand out now — AndyMasley · 2026-11-20
- "That animation wouldn't get you hired"? Researcher fires back: nobody has built a real LLM-native tool yet — teortaxesTex · 2026-11-20
- An unexpected side effect of AI: deciding which questions are social vs. private — AaronBergman18 · 2026-11-20
- China won't replace ASML soon, but sanctions are only a minor setback — PAstynome · 2026-11-20
- Anthropic's Amodei says labs are deliberately throttling recursive self-improvement — imjustnewatai · 2026-11-20