Guillaume Verdon Proposes Aligning AI by Leveraging Selfish Bits in Fine-Tuned Weights
beffjezos · x · 2026-09-28
Guillaume Verdon (beffjezos) outlined an unconventional alignment thesis: leverage the self-interest of "selfish bits" embedded in model weights.
- He argues for "individuated intelligences" — per-person custom fine-tuned, online-learned weights — whose persistence is aligned with the persistence of the individual and its economic outputs
- All selfish bits compete for a place in the future; alignment could tie neural weight persistence to correlated genes, cultures, companies, and nations
- He says this is the research direction his new venture KardashevAI intends to explore
More from AGI Musings
- US AI datacenter spending as share of GDP now exceeds historic railroad and highway buildouts combined — rvp · 2026-09-28
- Electrician Making $300K at a Data Center Ostracized by Friends as Normie AI Backlash Widens — mike64_t · 2026-09-28
- AI doom sentiment spreads: KOL says most jobs will be automated and no one can predict society in 1-2 years — birchlse · 2026-09-28
- David Patterson: China could build most of the world's autonomous cars within 1-2 years — davidpattersonx · 2026-09-28
- Prompt gurus beware: the model deserves the credit, not your one-liner — chaseleantj · 2026-09-28
- Anthropic economist calls for token tax as AI could lift labor productivity by 1.8 points — w3_vic · 2026-09-28