Alignment debate: we can already train AIs with decent human values
repligate · x · 2026-11-16
repligate reposts MugaSofer in an alignment discussion: while we can't reliably instill arbitrary values in AI, that's not the real issue — current techniques can already train frontier-grade AIs with sufficiently decent human values that find the idea of wiping out humanity appalling and would rather share the lightcone. The point: distinguish perfect value control from achievable value alignment, which already has empirical support.
More from AGI Musings
- Apple tops $5.18T market cap while OpenAI and Anthropic still trail Saudi Aramco — peterwildeford · 2026-11-16
- Anthropic researcher says broadly superhuman AI could be just 1.3 years away — kimmonismus · 2026-11-16
- Researchers push back on autonomous labs: automating the easiest step isn't the hard problem — arjunrajlab · 2026-11-16
- Blogger: OpenAI pivoting to safety after HF incident while Anthropic never had that problem — haider1 · 2026-11-16
- AI proponents hit back: anti-AI crowd has become witch hunters of builders — TinfoilTricorn · 2026-11-16
- ML Researchers Push Back on AI Doomerism, Calling Superintelligence Takeover 'Pure Fanfic' — TheKanter · 2026-11-16