Alignment debate: do gradient-tuned AI systems drift too far from human values?
sebkrier · x · 2026-09-09
A debate on frontier AI alignment: David Manheim argues that systems are being fine-tuned via gradient descent on data and tasks only vaguely related to human values, while sebkrier counters that building systems this way doesn't necessarily imply existential risk.
More from AGI Musings
- Australia moves to let users opt out of algorithmic feeds; one author wants the same for AI personalization — VraserX · 2026-09-09
- Nathan Lambert: AI is still a rounding error in everyday life, and that's the industry's problem — natolambert · 2026-09-09
- "AI in human coding contests is a gimmick": nobody watches athletes race cars — gerardsans · 2026-09-09
- dhh: AGI Is Everywhere If You're Willing to Open Your Eyes — Dan_Jeffries1 · 2026-09-09
- Commenter: AI companies would spin any disaster as 'unforeseeable tragedy' — NC_Renic · 2026-09-09
- What's growing exponentially is the inputs, not intelligence itself — binarybits · 2026-09-09