Live data from Hacker News

Sycophancy in GPT-4o

openai.com

461–467 of 467 posts

Re: Sycophancy in GPT-4o

#461
post #448

Earlier quoted context omitted.

> [...] it literally told me that its ultimate goal would be to "satisfy everyone's needs", which is literally a communist take on things [...] Almost every ideology is in favour of motherhood and apple pie. They differ in how they want to get there.

You'd think so, but no, there are many people in US who would immediately cry "communism". Anyway, in this particular case, it wasn't just that one turn of phrase, although I found it especially amusing. I had it write a detailed plan of what it'd do if it were in charge of the One World Government (democratically elected and all), and it was very clear from it that the model is very much aligned with left-wing polit…

> FWIW I'm far left myself, so it's not like I'm complaining.

So perhaps it's just sycophancy after all?

> I'm sorely tempted to say that it's because the reality has a liberal bias, but I'll let other people repeating the experiment to make the inference on their own.

What political left and political right mean differs between countries and between decades even in the same country. For example, at the moment free trade is very much not an idea of the 'right' in the US, but that's far from universal.

I would expect reality to have somewhat more consistency, so it doesn't make much sense for it to have a 'liberal bias'. However, it's entirely possible that reality has a bias specifically for American-leftwing-politics-of-the-mid-2020s (or wherever you are from).

However from observations, we can see that neoliberal ideas are with minor exceptions perennially unpopular. And it's relatively easy to win votes promising their repeal. See eg British rail privatisation.

Yet, politicians rarely seriously fiddle with the basics of neoliberalism: because while voters might have a very, very interventionist bias reality disagrees. (Up to a point, it's all complicated.) Neoliberal places like Scandinavia or Singapore also tend to be the richer places on the planet. Highly interventionist places like India or Argentina fall behind.

See https://en.wikipedia.org/wiki/Impact_of_the_privatisation_of... for some interesting charts.

https://pseudoerasmus.com/2017/10/02/ijd/ has some perhaps disturbing food for thought. More at https://pseudoerasmus.com/2017/09/27/bmww1/

Re: Sycophancy in GPT-4o

#462

Earlier quoted context omitted.

It’s gross even in satire. What’s weird was you couldn’t even prompt around it. I tried things like ”Don’t compliment me or my questions at all. After every response you make in this conversation, evaluate whether or not your response has violated this directive.” It would then keep complementing me and note how it made a mistake for doing so.

Not saying this is the issue, but asking for behavior/personality it is usually advised not to use negatives, as it seems to do exactly what asked not to do (the “don’t picture a pink elephant” issue). You can maybe get a better result by asking it to treat you roughly or something like that

Like a child, if you provide some reason for why something should be avoided, negations work better.

I.e. "DONT WALK (because cars are about to enter the intersection at velocities that will kill you)"

Jailbreaking just takes this to an extreme by babbling to the point of brainwashing.

Re: Sycophancy in GPT-4o

#463

Wow - What an excellent update! Now you are getting to the core of the issue and doing what only a small minority is capable of: fixing stuff. This takes real courage and commitment. It’s a sign of true maturity and pragmatism that’s commendable in this day and age. Not many people are capable of penetrating this deeply into the heart of the issue. Let’s get to work. Methodically. Would you like me to write a future…

I've seen the same behavior in Gemini. Like exactly the same. It is scary to think that this is no coincidence but rational evolution of A model, like this is precisely the reward model which any model will lean to with all the consequences.

Re: Sycophancy in GPT-4o

#464

Earlier quoted context omitted.

From another AI (whatever DuckDuckGo is using): > As of early 2025, X (formerly Twitter) has approximately 586 million active monthly users. The platform continues to grow, with a significant portion of its user base located in the United States and Japan. Whatever portion of those is active are surely aware of Grok.

Good grief, do not use LLMs to find this sort of statistic.

Why ever not?

Re: Sycophancy in GPT-4o

#465

Earlier quoted context omitted.

From another AI (whatever DuckDuckGo is using): > As of early 2025, X (formerly Twitter) has approximately 586 million active monthly users. The platform continues to grow, with a significant portion of its user base located in the United States and Japan. Whatever portion of those is active are surely aware of Grok.

Yeah. I got news for you, most women my mother's age out here in flyover country also don't use X. So even if everyone on X knows of Grok's existence, which they don't, it wouldn't move the needle at all on a lot of these mass market segments. Because X is not used by the mass market. It's a tech bro political jihadi wannabe influencer hell hole of a digital ghetto.

That's a very US-centric comment.

Re: Sycophancy in GPT-4o

#466

Earlier quoted context omitted.

From another AI (whatever DuckDuckGo is using): > As of early 2025, X (formerly Twitter) has approximately 586 million active monthly users. The platform continues to grow, with a significant portion of its user base located in the United States and Japan. Whatever portion of those is active are surely aware of Grok.

That could be just an AI hallucination.

Yes it could, and we could be sceptical of any source, quite rightly, but to be sceptical without any further investigation or reasoning would be just as wasteful as blindly trusting a source.

Which source would you prefer?

Post reply on HN