Great read. I first observed this phenomenon when I was interacting with Gemini 3 Pro. I tried to ask it about some of Huawei's unreleased internal facts, and it turned out to know a lot of details. When I tried to search for those same facts on Google, I couldn't find much. This means the training data of Gemini 3 Pro definitely contains some of that information, perhaps leaked documents that aren't easily found on Google.
Later on, at the end of last year, I talked to several people from OpenAI and Anthropic. They told me that with a large enough model (for example, 5T parameters), it can effectively store the useful knowledge of almost every industry.
Currently, the coding capability of models is the strongest because coding is the most open; we have a lot of open-source code. But for most other industries, knowledge is pretty closed. To solve this, frontier labs are actively collaborating with these industries. Whenever agentic workflows go into these sectors, their know-how and skills will eventually be distilled into the model itself.
Previously, I thought that holding a proprietary knowledge base might actually be a moat, but it seems like that moat is becoming thinner and thinner.
This is a continuous loop:
1. Closed industries use frontier models, and their know-how is distilled into these closed models.
2. Open models then distill that knowledge from the closed models, making the knowledge open to everyone.
I don't see anything that could stop this from happening. If you publish an article about leaked documents from Huawei, you will be sued. But if a frontier lab distills your knowledge from agentic traces, there is currently nothing to prevent it.
与其说 AI 生产的应用最难做的是设计,不如说大量相似的设计提高了大家的审美阈值。
这导致现在用户的口味变刁了,所有的用户都会要求以往 90 分以上的设计,但在没有 AI 的时代,设计真的要做到 90 分以上的产品也是极少的。中国互联网公司的产品通常不太重视设计,他们可能更愿意把实验和预算投入到投放上去。
由于大量的 AI 应用都是由独立开发者或小团队做出来的,他们既缺乏投放预算,又大量依赖有机增长,所以强调设计就变成了 AI 产品面对第一波用户时,必须要解决的事情。