iamtrask on AI attribution: Shapley for inference, LoRA for post-training, hierarchy for pre-training
iamtrask · x · 2026-09-19
iamtrask argues AI attribution works better than people think, layering methods: Shapley values for inference attribution, LoRA for post-training, and data hierarchy for pre-training.
- His thesis: everyone will fine-tune open source models on private data and register them on routers.
- Instead of AGI, we'll get "RGI" (routed general intelligence) — a general network of narrow intelligences.
- Creators get paid when others route to their model for its unique knowledge, indirectly solving most of the AI attribution problem.
Related event: iamtrask: AI attribution is closer to solved than you think(3 posts)→
More from AGI Musings
- Jeff Ladish: Aligning superintelligence makes sense, controlling it doesn't — JeffLadish · 2026-09-19
- The flip side of AI slop: every human word will now be read by machines — yeastsplainer · 2026-09-19
- Hot Take: If It's Easily Automated, It Wasn't Real Knowledge Work — sebpaquet · 2026-09-19
- Anthropic publishes three public metrics to track how fast AI is building AI — minchoi · 2026-09-19
- MIT-shared paper argues LLMs can't be conscious without brain-like analog computation — burny_tech · 2026-09-19
- AI hallucination about Chinese nuclear parts nearly triggered US military action, report says — polymute · 2026-09-19