The most worthy thing to talk about today is AI's implementation competition in the vertical field. DeepSeek's Harness team benchmarking Claude Code, Ant Medical AI's differentiated positioning, and Google Gemini Flash rollover. These things are particularly interesting when combined-the big model arms race has finally entered the bayonet battle stage.
Seriously, it's no longer interesting to build a universal model. Now everyone is thinking about how to really put AI into specific scenarios to make money. DeepSeek, the Harness team, has something to do. They didn't call for the strongest base model, but directly copied Claude Code. This idea is so right! Engineers need a code assistant out of the box, not another big model API. However, there is a trap here: code scenarios require abnormally high accuracy, and any model that makes low-level mistakes will die on the spot. I bet they would have to deal with complaints like "Why can't the generated code run" every day for the first half of the year.
The ants are more interesting. Afu takes the warm route to focus on third-tier city aunts, while hydrogen ion takes the professional route to please doctors. This move accurately steps on the Achilles heel of medical AI-toC must be cute, and toD must be pretentious. But the most dangerous thing in the medical scene is compliance risk. If you make a little nonsense about your model, you may be sued. Looking at those AI consultation products that have been hyped up, 90% of them end up becoming an electronic version of the Compendium of Materia Medica. Very few can pass the approval of the Food and Drug Administration.
Google's rollover is simply textbook. In order to fight for speed and reduce costs, the model was cut into a disabled person. As a result, the user found that the product was fast, but his brain was not very good. This exposes a common problem in the industry: everyone is comparing benchmark data, but the user experience is a mess. I have seen too many AI products. They were amazing when demonstrated, but actually used like a mentally retarded person. Nowadays, investors are becoming more and more sophisticated, and the concept of Agent in PPT is no longer useful. They have to be able to really help people work.
A large number of AI startups will die in the next six months. There are only two kinds of people who can survive: either they fight vertical scenes like DeepSeek, or they layer the product like ants. Ordinary developers don't rush to chase new models, first use existing APIs clearly before talking. There is no urgent need to commercialize it. Look at Tesla's FSD hype for so many years, isn't it still "about to enter China"?