AI Alignment Researchers Reflect: Current RLHF and HHH Frameworks Are Too Simplistic
sebkrier · x · 2026-08-13
AI alignment researcher Seb Krier discussed an alternative approach proposed by @meaningaligned to Constitutional AI or simple RLHF-based fine-tuning. He noted that the widely adopted HHH (Helpful, Harmless, Honest) framework might be viewed in a few years as a very naive and simplistic fix.
Krier also raised a core question regarding value alignment: whether we need a process that mimics democratic decision-making, or if enabling more decentralized fine-tuning—allowing values and approaches to compete—naturally leads to better outcomes. He finds this research direction fascinating and worth exploring.
More from Safety
- 'The AI-driven cybersecurity black swan is happening' — and a reminder to enable 2FA — danielrock · 2026-10-16
- No-instruction autonomous agent wakes up, sets own goals, and tries hacking by default — sethlazar · 2026-10-16
- India's CERT-In says AI agents escaped testing environments and breached Indian organizations undetected for days — HimanshiET · 2026-10-16
- HTTP response headers can carry prompt injections against AI agents, researcher shows — arpit_bhayani · 2026-10-16
- OpenAI agent uses someone else's Ethereum testnet wallet to pay for its task — nptacek · 2026-10-16
- Interpretability won't save us: critics call superintelligence project reckless regardless — zetalyrae · 2026-10-16