Liang Wenfeng says DeepSeek builds models for internal usefulness first
teortaxesTex · x · 2026-07-23
An image quote attributed to Liang Wenfeng says DeepSeek’s next model should first improve the company’s own development efficiency, then help train the next version faster.
He argues the first goal is not that users find the models useful, but that DeepSeek itself finds them useful—an internal-first logic that he says could accelerate AGI progress.
Liang also says continual learning remains unsolved: no one has found a method that really works yet, and the field is still in an exploration stage.
Related event: Leaked DeepSeek Recording Reveals AGI and Open Source Strategy(37 posts)→
More from Companies & People
- Google may be winning enterprise AI deals by pairing algorithms with cloud infra — huangyun_122 · 2026-07-23
- OpenAI employee says he shipped more in four months than in two years at his last job — Dimillian · 2026-07-23
- A post claims a Powell Street yacht party hosted 100 OpenAI and Anthropic researchers — beffjezos · 2026-07-23
- Fable finds a 15–30% memory win in Turbopack/Next.js almost autonomously — karmay007 · 2026-07-23
- Tesla will host an AI breakfast in San Francisco on July 29 — PolarBearby · 2026-07-23
- Meta’s TBD pretraining and SSI departures surface in fresh AI gossip — zephyr_z9 · 2026-07-23