What's this about?
Quotes attributed to DeepSeek founder Liang Wenfeng from a recent investor meeting are circulating in the Chinese AI community: on the road to AGI, products are just a byproduct; coding agents are the top priority now; hallucination is a solvable product problem, not the point; and the crucial missing capability is continual learning — the next-generation model must have it.
What exactly is "continual learning"?
Today's models are frozen after training — they don't get smarter from experience; every conversation starts from the same snapshot. Continual learning means the model improves on the job like an employee. Liang says AI has taste and intuition already but can't yet learn continuously — and nobody in the world has cracked it.
What's the community saying?
988 likes on the Chinese thread
- Admiration for the focus: dismissing video generation and world models as irrelevant to the intelligence ceiling is a direct shot at rivals' roadmaps.
- "We want the next model to help us build the one after" — recursive self-improvement as an explicit, stated strategy.
- Unverified provenance: these are meeting notes relayed by KOLs, not an official DeepSeek publication.