LLMs are terrible at designing AI systems

From my perspective, this is reasonable.

The basic working principle of LLM is that the model weights are combined with the current context to predict the most likely Token; however, considering that most of the data in model training does not relate to AI Systems, LLMs are inevitably faced with difficulties when designing AI systems.

But I believe this issue will be gradually alleviated because the increasing number of open-source projects related to AI Systems will increase the amount of data for AI Systems, thus enabling the model to perform better.

However, I think this is a good thing for software engineers, because the emergence of new business needs and new product forms, the time it takes for them to be learned by the model and become a new baseline may be longer than we imagine.