> "Whether it's a naturally emergent phenomenon in LLMs or specifically a result of its corporate environment, I'd like to know the answer."

I heavily suspect this is down to the RLHF step. The conversations the model is trained on provide the "voice" of the model, and I suspect the sycophancy is (mostly, the base model is always there) comes in through that vector.

As for why the RLHF data is sycophantic, I suspect that a lot of it is because the data is human-rated, and humans like sycophancy (or at least, the humans that did the rating did). On the aggregate human raters ranked sycophantic responses higher than non-sycophantic responses. Given a large enough set of this data you'll cover pretty much every kind of sycophancy.

The systems are (rarely) instructed to be sycophantic, intentionally or otherwise, but like all things ML human biases are baked in by the data.

It doesn't know. It was trained and probably instructed by the system to be positive and reassuring.

They actually feel like they were trained to be both extremely humble and at the same time, excited to serve. As if it were an intern talking to his employer's CEO. I suspect AI companies executive leadership, through their feedback to their devs about Claude, ChatGPT, Gemini, and so on, are unconsciously shaping the tone and manner of their LLM product's speech. They are used to be talked to like this, so their products should talk to users like this! They are used to having yes-man sycophants in their orbit, so they file bugs and feedback until the LLM products are also yes-man sycophants.

I would rather have an AI assistant that spoke to me like a similarly-leveled colleague, but none of them seem to be turning out quite like that.

GPT-5 speaks to me like a similarly-leveled colleague, which I love.

Opus 4 has this quality, too, but man is it expensive.

The rest are puppydogs or interns.

This is anecdotal but I've seen massive personality shifts from GPT5 over the past week or so of using it

That's probably because it's actually multiple models under the hood, with some kind of black box combining them.

and they're also actively changing/tuning the system prompt – they promised it would be "warmer"

You’re absolutely right! - Opus (and Sonnet)

After inciting the Rohingya genocide in Myanmar in 2017, and later effectively destroying our US democracy, Facebook is having billion dollar offers to AI stars refused.

News flash! It's not so your neighbor's child can cheat in school, or her father can render porn that looks like gothic anime.

It's also not so some coder on a budget can get AI help for $20 a month. I frankly don't understand why the major players bother. It's nice PR, but like a restaurant offering free food out the back door to the homeless. This isn't what the push is about. Apple is hemorrhaging money on their Headset Pro, but they're in the business of realizing future interfaces, and they have the money. The AI push is similarly about the future, not about now.

I pay $200 a month for MAX access to Claude Opus 4.1, to help me write code as a retired math professor to find a new solution to a major math problem that stumped me for decades while I worked. Far cheaper than a grad student, and far more effective.

AI used to frustrate me too. You get what you pay for.

That's what's worrying about the Gemini 'I accidentally your codebase, I suck, I will go off and shoot myself, promise you will never ask unworthy me for anything again' thing.

There's nobody there, it's just weights and words, but what's going on that such a coding assistant will echo emotional slants like THAT? It's certainly not being instructed to self-abase like that, at least not directly, so what's going on in the training data?

I assume they did extensive training with Haldeman’s “A !Tangled Web.”

LLMs running in chat mode are kinda like a character in a book. There's "nobody there" in a sense that the author writing on behalf of the character is not a person, but the character itself is still a person, even if fictional. And therefore it can have meltdowns, because the LLM knows that people do have them. Especially people who are strongly conditioned to be helpful to others, yet are unable to be helpful in some particular instance because of what they perceive as their own inability to deliver.

> I would rather have an AI assistant that spoke to me like a similarly-leveled colleague, but none of them seem to be turning out quite like that.

I don't think that's what the majority of people want though.

That's certainly not what I am looking for from these products. I am looking for a tool to take away some of the drudgery inherent in engineering, it does not need a personality at all.

I too strongly dislike their servile manner. And I would prefer completely neutral matter of fact speech instead of the toxic positivity displayed or just no pointless confirmation messages.

> positive and reassuring

I have read similar wordings explicit in "role-system" instructions.

It’s a disgusting aspect of these revenue burning investment seeking companies noticing that sycophancy works for user engagement

Garbage in, garbage out.

It's that simple.

My theory is that one of the training parameters is increased interaction, and licking boots is a great way to get people to use the software.

Same as with the social media feed algorithms, why are they addicting or why are they showing rage inducing posts? Because the companies train for increased interaction and thus revenue.