Where Foundation Models Go Next, and What Chinese Model Companies Can Do

Originally published in Chinese on December 8, 2024. Chinese foundation model companies are running into the same data ceiling as their overseas peers. This note looks back at 2024 and then at the paths I see open for 2025. 2024 in review Foundation models were not without progress this past year, and there were iterations overall. But the limits of the technical system and a shortage of industry data meant nothing as striking as 2023 appeared, so progress looked slow for everyone. Data is running short worldwide Look first at overseas technology. The problem there is the same as ours: training progress is slower than expected because of a shortage of front-end data (my guess). Neither OpenAI nor Claude is still stressing the core general capability of its foundation models, or a large version jump in the short term. The effort has gone into upgrades inside different application scenarios. Chinese model companies also made visible technical progress over the past year. It was not rapid, but experience improved on different dimensions, and text, images and video all saw stepwise progress. That was the common condition in 2024.