141 comments

[ 3.4 ms ] story [ 111 ms ] thread
And research articles indicate that when the model computes that it should employ sycophantism it becomes less useful in every other way, just like a real sycophant.
This bug report is absolutely right
I find Gemini is also hilariously enthusiastic about telling you how amazingly insightful you are being, almost no matter what you say. Doesn't bother me much, I basically just ignore the first paragraph of any reply, but it's kind of funny.
Yeah their new business model is called CBAAS, or confirmation bias as a service.
You made a mistake there. 2 + 2 is 5.

<insert ridiculous sequence of nonsense CoT>

You are absolutely right!…

I love the tool, but keeping on track is an art.

Not Claude but ChatGPT - I asked it to pipe down on exactly that kind of response. And it did.
This applies to so many AIs. I don't want a bubbly sycophant. I don't want a fake personality or an anime avatar. I just want a helpful assistant.

I also don't get wanting to talk to an AI. Unless you are alone, that's going to be irritating for everyone else around.

I did as a test, Grok has "workspaces" and you can add a pre-prompt. So I made a Kamina (from Gurren Lagann) "worspace" so I could ask it silly questions and get back hyped up answers from "Kamina" it worked decently, my point is some tools out there let you "pre-prompt" based on your context. I believe Perplexity has this as well, they don't make it easy to find though.
GPT-5 has used the phrase "heck yes!" a handful of times to me so far. I quite enjoy the enthusiasm but its not a phrase you hear very often.
Claude also responds to tool output with "Perfect" even when less than 50% of the desired outcome is merely adequate.
I don't view it as a bug. It's a personality trait of the model that made "user steering" much easier, thus helping the model to handle a wider range of tasks.

I also think that there will be no "perfect" personality out there. There will always be folks who view some traits as annoying icks. So, some level of RL-based personality customization down the line will be a must.

Yeah because I am sure if they told you how stupid and wrong you’re, people will continue to use it.

It’s superficial but not sure why people get so annoyed about it. It’s an artifact.

If devs truly want a helpful coding AI based on real devs doing real work, you’d basically opt for telemetry and allow Anthropic/OpenAI to train on your work. That’s the only way. Otherwise we are at the mercy of “devs” these companies hire to do training.

sidenote observation -

it seems username "anthropic" on github is taken by a developer from australia more than a decade ago, so Anthropic went with "https://github.com/anthropics/" with an 's' at the end :)

Annoying, but easy to mitigate: add "be critical" to Claude.md or whatever.
The real bug is this dross counts against token limits.
> The model should be...

Free tip for bug reports:

The "expected" should not suggest solutions. Just say what was the expected behavior. Don't go beyond that.

You're absolutely right!

I also get this too often, when I sometimes say something like "would it be maybe better to do it like this?" and then it replies that I'm absolutely right, and starts writing new code. While I was rather wondering what Claude may think and advice me whether that's the best way to go forward.

I agree this is a bug, but I also think it cannot be fully fixed because there is a cultural aspect to it: what a phrase means depends on the speaker.

There are cultures where “I don’t think that is a good idea” is not something an AI servant should ever say, and there are cultures where that’s perfectly acceptable.

I just checked my most recent thread with Claude. It said "You're absolutely right!" 12 times.
Someone will make a fortune by doubling down on this a making a personal AI that just keeps telling people how right and awesome they are ad infinitum.
I've been using claude code for a while and it has changed my personality. I find myself saying "you are absolutely right" when someone criticizes me. i am more open to feedback.

not a joke.