He noted that DeepSeek has already transitioned to using its own high-level languages, such as TileLang, to optimize hardware performance, a shift he believes will soon allow Chinese chips to overcome traditional software ecosystem barriers.
Liang described DeepSeek’s primary objective as the pursuit of Artificial General Intelligence (AGI), specifically focusing on "continuous learning" as the next major bottleneck.
Unlike American competitors such as OpenAI or Google, which he claims are focused on high-profit closed models, DeepSeek has adopted a "restrained" business model focused on open-source releases and "fair" profits.
Liang believes this approach—aiming for cost recovery within ten months rather than maximizing margins—will ultimately beat high-profit models by fostering a broader ecosystem and reducing market pushback.
Regarding the global landscape, Liang acknowledged that the current gap between Chinese and American AI models is largely due to American advantages in computing power and resources rather than a difference in talent.
He predicts a future where only three or four major companies compete at the base-model level.
DeepSeek intends to remain focused on the "main line" of language models and agents, intentionally leaving commercial opportunities in areas like video generation and world models to other partners within the ecosystem.