DeepSeek's Underrated Role: Scaling and De-risking Ideas Academia Published Years Earlier
teortaxesTex · x · 2026-11-03
teortaxesTex argues DeepSeek's underrated contribution is making 'Things Work' common knowledge: techniques published years earlier by academia or megacorps, cynically assumed not to scale, were validated and de-risked by DeepSeek and became industry backbones. Examples: V2's fine-grained MoE with shared experts (Apr/May 2024) from Microsoft's DeepSpeed-MoE (Jul 2022); V3's MTP (Dec 2024) after Meta FAIR's first version (Apr 2024); R1 building on an internal RLVR project that produced GRPO by Feb 2024 from MATH-Shepherd work started in late 2023.
More from AGI Musings
- Giraffe evolution meme: AI's low-hanging fruit keeps hanging higher every year — itsOmSarraf_ · 2026-11-03
- Bain's $6T vs Goldman's $0.5T: How Big Can AI Revenue Really Get — luisdans · 2026-11-03
- Databricks: Agents now run AI research loops, training jobs grew 8x in a year — matei_zaharia · 2026-11-03
- AI elites might hoard intelligence like mercantilists hoarded gold, argues Venkatesh Rao — The Decoder · 2026-11-03
- Same AI, cents vs tens of millions: the value gap is in the question, not the answer — r0ck3t23 · 2026-11-03
- AI researchers: one success in an environment loop is enough signal to run RL on it — cephaloform · 2026-11-03