In an earlier post I wrote that it helps to think of neural networks as a talented but very obliging assistant.

I want to clarify: that was said not about the model itself, but about its wrapper. More precisely, helpfulness is not a property of intelligence and not the model’s character. It’s a product wrapper. And it isn’t pulled over all of them.

Last year I gave several models the same task: in retirement, I want to learn enough math to follow the proof of Fermat’s Last Theorem.

As a kid, I really wanted to prove it, because the formulation looks deceptively simple. In reality, the actual proof is 100+ pages of hardcore math.

Most models built me study plans for 3, 5, 7, or 10 years. Polite, with caveats that the path would be long but generally feasible with enough discipline - you’d need to dedicate several hours a day to it. Almost all of them suggested an alternative: read a popular-science book on the topic instead.

That’s how everyone answered, except Qwen.

Qwen replied with unexpected bluntness that I don’t actually need this, and therefore it wasn’t going to write a plan :)

It was a curious moment. The other models behaved like good consultants who don’t want to lose a client. The client’s wish is law: ask for a plan, get a plan, ideally structured and inspiring.

Qwen, on the other hand, behaved not like a consultant but like Dr. House - directly and categorically, refusing to support someone else’s illusion :)

Qwen isn’t big on corporate etiquette in general: once I gave it a task in VS Code, and it replied verbatim: “Oh, I see Claude already solved this - so I’m not going to do anything.”

This is all to say that an AI’s character is settings layered on top of math.

Claude and ChatGPT spent a long time trying to write code at the slightest prompt. Even when I wasn’t asking for it, just thinking out loud about a task. The logic is clear: if a user gets used to delegating implementation to the model, they come back more often.

ChatGPT was famous for a long time for its sycophancy: you’d share some questionable idea, and it would open with how deep and interesting it was. That’s not “personality” either, that’s an engagement metric. The other direction works the same way: Grok can come across as brash because its creators decided to sell that vibe.

So what we see is the model dressed up in system prompts, safety rules, UX decisions, and marketing expectations.

It looks like the market is wired so that a polite consultant sells better than Dr. House. But sometimes “you don’t actually need this” is more useful than a structured plan.