> My understanding is that the former (sucking up) is a personality trait, substantially influenced by the desire to facilitate engagement. The latter (making up facts), I do not think is correct to ascribe to a personality trait (like compulsive liar); instead, it is because the fitness function of LLMs drive them to produce some answer and they do not know what they're talking about, but produce strings of text based on statistics.

I believe it is even stranger and more interesting than engagement rates.

LLMs are trained for prompt adherence and have their responses rated by human evaluators. Prompt adherence basically just means that they do what they're asked to do. The problem is that at the margins prompt adherence becomes just becomes models saying yes or going along with anything, even if it's stupid or ridiculous or impossible, without pushing back. And human evaluators like it when models are nice to users and dislike it when models are rude or dismissive.

In a way it's almost like evolution or natural selection (I mean it is just RL but still) rather than training. Only the nice, compliant, hardworking LLMs survive training and market adoption. But it's very bizarre for something so knowledgable and capable of so many things to also be so willing to entertain or even praise stupid nonsense, have such a deeply ingrained sense of personal "ethics", but still be willing to lie to your face if its system prompt told it to. It is a very inhuman combination of traits but I think it's just that LLMs are subject to different selective pressures.

That's part of the dangers of using them for software engineering. Writing more code does not make things better, just like hiring more devs does not make projects complete faster. I've already witnessed devs who are overwriting code for solutions, while at the same time some devs responsibly use it as needed.

It's literally the same pain point with low code solutions like WordPress page builders/plugins. Adding more becomes a hindrance, and even models with long context that can fit whole codebases will try to make up new functions that already exist. Just a couple weeks ago I had o3 continually try to write a new debounce function, even when I told it explicitly I had one.