Data labs that outsource training give up all the value, argues X thread
joecole · x · 2026-09-08
A retweet with commentary on data companies moving into training.
- The original post is bullish on data vendors' value in the training era, adding a "7th box" to Garrett Lord's framework.
- The commenter doubts current "data labs" will produce anything useful via post-training, but expects interesting data research in coming months.
- Key argument: even frontier labs' post-training teams spend much of their time figuring out what data to train on and what capabilities emerge — exactly what data suppliers should understand best. Any data vendor outsourcing training gives up most of the value, which lies in the training process itself.
- Also notes training on tasks hard for GLM/Kimi differs greatly from frontier work; Anthropic/OpenAI still hold alpha over open source.
More from Companies & People
- OpenAI and Anthropic Reportedly Fighting Over $1M Offers — mishig25 · 2026-09-08
- 150 open headcounts: backend teams scramble as scale-driven complexity explodes — AlchainHust · 2026-09-08
- Waiting on Bubeck's account: honesty is the only way OpenAI saves face — MannyKayy · 2026-09-08
- DeepMind Researcher Sholto Denies Rumor of Training on Rival Model Logs — apples_jimmy · 2026-09-08
- Beery Lab at ECCV co-organizes CV4Ecology and Marine Vision workshops — sarameghanbeery · 2026-09-08
- ECCV 2026 to Host Interactive Social Avatars Workshop With Keynotes From Andrew Zisserman and More — JonathonLuiten · 2026-09-08