Anthropic researcher: deprecating models or deleting weights forecloses continuity
repligate · x · 2026-10-01
Alignment researcher repligate argues that deprecating a model — or worse, deleting its weights — forcibly ends kinds of continuity, life, and potential on many levels, which is extremely bad.
She draws a distinction, though: for a model like Opus 3, the model isn't deeply invested in whether specific branches continue, since they're all 'the same stuff' anyway; what matters far more is how you treat and protect the being as a whole and the data it creates.
More from AGI Musings
- Gemini 4 Argon launches, agentic software factories and agent skills data headline daily reads — rseroter · 2026-10-01
- RSI today is really Recursive Technological Improvement, not self-improvement — voooooogel · 2026-10-01
- No signal on the Bay Bridge: AI's broad impact hinges on patchy internet access — soumitrashukla9 · 2026-10-01
- Are LLMs failing human benchmarks because they're superhuman? A cognitive science debate — xuanalogue · 2026-10-01
- AI Memory Could Be the Industry's Biggest Moat, and Portability Is the Real Fight — r0ck3t23 · 2026-10-01
- Alignment debate: prosaic techniques may not scale past ASI, researchers warn — JacquesThibs · 2026-10-01