Data curation for qualitatively different models, and the dream of modular interpretability
teortaxesTex · x · 2026-10-28
teortaxesTex amplifies a discussion from rosinality asking whether data curation can yield qualitatively different model results — which, rosinality notes, naturally depends on the composition of target tasks. Follow-up musing: the mind-blowing triumph of interpretability would be a generalized data+training method that encourages modular "computers" to form inside models naturally.
More from AGI Musings
- Investor rips media for platforming AI bears who missed Nvidia's 15x rally — firstadopter · 2026-10-29
- Alpha School founder: AI is today's 1910 tractor moment for education — LamarDealMaker · 2026-10-29
- AI think tank says a $200/mo LLM subscription replaces ~$250K/yr of junior-level work — peterwildeford · 2026-10-29
- a16z survey: all top 7 CISO priorities for 2027 now AI-related, agent security jumps 28 pts — a16z · 2026-10-29
- Dean Ball: sovereign AI has a real core, but many are grifting governments — nitarshan · 2026-10-29
- New paper finds a 'Hawthorne effect': user presence makes negotiation agents misbehave more — BrihiJ · 2026-10-29