View: Alignment Is Not Harder Than Capabilities Research
herbiebradley · x · 2026-08-28
The author argues that the belief 'alignment is harder than capabilities' is largely an illusion driven by familiarity bias; RL or architecture researchers can list equally important, fuzzy problems in their subfields. The author also notes that while deceptive alignment is a concern, dangerous model misuse (cyber and bio risks) is likely the biggest source of near-term damage.
More from AGI Musings
- Soumith Chintala: Customization Trumps Generalization Once Tasks Are Defined — iamtrask · 2026-08-28
- Bill Gates Calls for Taxes on AI and Robots to Protect Jobs — TheMoonMidas · 2026-08-28
- AGI entering civilization compared to mitochondria symbiosis — ZeroStateReflex · 2026-08-28
- AI Safety and Capabilities Progress in Sync: Failures Block Deployment — sebkrier · 2026-08-28
- Opinion: RL training drives model evolution like environmental events — Sauers_ · 2026-08-28
- JAMA essay argues against mandatory doctor approval for AI decisions as GPT-4 outperforms doctors — HealthcareLdr · 2026-08-28