Yang You's 'Large Partial Derivative Theory': pick research where your marginal contribution is biggest
YangYou1991 · x · 2026-11-21
Professor Yang You responds to a roundup of Chinese 'PhD theories' with his 'Large Partial Derivative Theory', countering the popular 'Large Slope Theory': high-growth fields (like LLM base-model training internships) often reduce individuals to replaceable data-washers. Instead, PhDs should find directions where their own marginal contribution is maximal. He recounts leaving LLM Agent post-training — '99% of academic agent papers are worthless to industry' — for diffusion post-training, which led to DiffusionART (arXiv:2610.12572, aggregation-first reward tilting + trust-region-masked restoration, no EMA tuning or differentiable reward needed), reportedly adopted in video-gen companies' post-training pipelines. His takeaway: in the paper-explosion era, representative work, blogs and open-source projects beat paper counts.
More from AGI Musings
- AI writes 90% of code but plateaus at ~50% of online articles — bendee983 · 2026-11-21
- Jack Dorsey says Block is 'a lot flatter' with 3 core roles after AI-driven restructuring — shivsingh · 2026-11-21
- Superintelligence could solve cancer treatments theoretically before wet labs — Dr_Singularity · 2026-11-21
- Looking back: OpenAI reasoning model autonomously disproved Erdős's unit-distance conjecture — aran_nayebi · 2026-11-21
- Government-funded kids channel looks AI-generated, sparking animator fears — justalexoki · 2026-11-21
- Compute alone won't fix painting or robotics: US labs' data edge debated — teortaxesTex · 2026-11-21