Was RLHF really OpenAI's invention? A debate over the 2017 precedent
binarybits · x · 2026-10-11
binarybits and andrewprock argue over RLHF's historical credit. binarybits frames the key insight as "train a reward model from pairwise human feedback, then use it for RL on another model," asking whether any pre-2017 paper used this exact technique. andrewprock counters that using human feedback is old news in ML, was not invented by OpenAI, and calling it comparable to inventing the lightbulb is "facile poppycock," citing an early Robot Shaping paper. The crux: is the novelty the general idea of human feedback, or the specific pairwise-preference-to-reward-model-to-RL pipeline?
Related event: Debate Erupts Over Who Really Invented RLHF(2 posts)→
More from Research
- Google and Tel Aviv researchers unveil SepGen, generating video with per-source stems for 4D spatial audio — YonatanBitton · 2026-10-11
- Why OpenAI's quasi-Riemann result must cover all L-functions, vindicating Hardy's 1921 conjecture — burny_tech · 2026-10-11
- Schmidhuber Points to Section 20 of His 'Annotated History of Modern AI and Deep Learning' — SchmidhuberAI · 2026-10-11
- Stephen Wolfram's 'A New Kind of Science' Is Freely Readable Online — burny_tech · 2026-10-11
- Task-structured modularity emerges in AI networks, aligning with brain architecture — lulzxdxdxd · 2026-10-11
- Toward an Automated Science of the Mind: AI Enters Every Stage of Cognitive Research — burny_tech · 2026-10-11