When I used cc, I noticed that fable was available again. I had a deep experience with it, and it became increasingly apparent that the stronger the model, the harder it is to "communicate" with it.

fable will engage in a lengthy, self-defensive reasoning process about your instructions, essentially saying: "I know you're testing me, but I won't change my behavior just because you claim to be authoritative."

To advance the task, I had to switch to a smaller model, Opus, which immediately understood the issue and corrected the context to move forward.

I feel this can't simply be described as getting dumber or smarter. But it's true that the more complex the model, the more it has a personality system. The cost is that complex models tend to over-interpret user needs, making them harder to correct in certain scenarios.

It's like a collaborative side effect that arises when capabilities, safety training, identity boundaries, and agent scenarios are layered together.

So, as AI develops, the emergence of more complex models designed to solve engineering challenges may not necessarily broaden their use for ordinary people.

Writing prompts will also become a systematic engineering task.

The ability to express needs, break down tasks, and verify outcomes are all important. In the future, complex models may not be something ordinary people can easily handle using just natural language.

It's like driving an ordinary family car: you just need to know how to shift gears and press the gas and brake.

But driving a race car requires a much more precise control over throttle depth and shift timing.