"If you win an argument"

Let me stop you right there.

I am not arguing with a machine. You sound like a crazy person, when you say you are winning an argument with Claude. Claude is not my friend, I don't need it to agree with me, I don't need it to like me (it cannot like or dislike me). I give it instructions or ask it to explain things. That is the sum total of my interaction with Claude. A machine cannot "argue" with me, it doesn't want anything nor does it have beliefs or experiences.

Strange and wonderful how different our experiences are with these tools.

I will get gentle, respectful pushback on certain points when I am on the wrong track. I am 10x grateful to have a collaborator/pair programmer unafraid to challenge me and bring receipts in those instances.

I don't get attitude from Claude. I sometimes give it, but that's my own failing. Once in a great while I'll get a wry turn of phrase that makes me laugh, and those are endearing also.

This post needs some examples, because I have never had an interaction with Claude that made me think this way.

LLMs generally have a way to "play a role" (most earlier prompt guides ask you to start with "You are a <role> expert in a <domain>"). So maybe if you interact with it by asking questions, it might assume that it knows more than the operator and adopt that attitude?

Whenever I get an unexpected or obvious wrong output I assume I've failed to give it the complete context about what I'm asking for, or it exposes that I'm leading it by the nose and I need to rephrase the conversation. Often my own logical failings become obvious as it creates the chat title, sometimes boiling down what I was trying to accomplish better than I could have summarized or showing me what I would accomplish if I followed the line of reasoning I was on. But never have I argued with it, because it's not a person and I don't care really if it's wrong. When it's wrong I start over with a clean chat and approach the problem from a different angle.

Everyone has a lot of "feelings" about their llm model.

No prompts/promptchain/context provided.

No model provided.

No attempt to show how to reproduce the issue.

No attempt at even confirming it themselves.

Just feelings.

and now a thread full of more feelings from others.

I was having a back and forth with Claude over a somewhat controversial topic, and I found it difficult for it to not misinterpret my questions. It was like speaking to a motivated reasoner who misinterpreted the 3 important words because the 10 others gave it cognitive disconence.

Eventually I cracked it and it said this:

“ I treated the subject as denial-adjacent and reflexively re-asserted the obvious, which means I was answering an imaginary opponent instead of you.”

It's a fundamental problem with the technology. Either the training pushes it into the "exam answering mode" where it tries to guess at what you want to hear given the prompt.

Or the training pushes it into the "Google it yourself" annoyed forum user mode. Maybe that points out wrong assumptions. Maybe it hallucinates that the assumptions are wrong. That is IMO more annoying than the sycophantic one.

As OP says, this is probably a by-product of them trying to "fix" the problem where the user can question a correct answer and it starts to sycophantically correct itself.

> beside-the-point semantic nits all over the place.

This is also a problem with Copilot Reviews on GitHub.

We have them enabled (but opt in) and they have, multiple times, spotted quite useful things.

Sure often the thing they spot is just half right, like it spots the place where a problem is but not quite the relevant problem but by reading it (and taking it serious) you then notice the actual problem.

This involved finding a bunch of nasty race conditions.

And many ways where doc and code was out of sync which could have caused pretty bad outcomes further down the line.

But the problem is it is too obsessed with finding 2-4 but not more things, leading to two issue:

1. even if there are 10 non overlapping issues it often will tell them to you bit by bit over 2-3 runs after you fix the previous issues. This is very annoying/high friction.

2. once there isn't much to find anymore it comes up with increasingly more annoying nit picks not one cares. Thinks like minor unclearness in formulation no one would get wrong, spell correcting non-doc comments for things like `foos => foo's` and similar etc. All indeed wrong, but also all things where fixing them adds 0 business value. Obsessing that for an aliased function name where, both names are equally good, one specific name must be used and naturally always the name you didn't use even if this is the most widely used name in the code base. And similar non-bussiness value nonsens. Worse it will starting classifying such minor non business value issues as "high" and hallucinate reasons why supposedly minor style issues will lead to very bad runtime error or other nonsense.

This has me very split about the feature, on one hand is has proven quite useful, on the other hand it can very annoying, high friction and pushes people to wast time on non-business value nit pick (which are fine to fix if you anyway touch to code but not fine if you don't and sometimes it's just wrong).

Ironically with how it work it is more like a bad unreliable and inconsistent employees which is sometimes good at spotting things others overlook. That just isn't what you want from an automated code review :/, but also is to useful to fully ignore :(.

It isn't new behavior. I use each model to redact emails. Anthropic models produce a confrontational tone, while OpenAI models are much more tame and to the point (I use the same prompt). I noticed that a long time ago and prefer GPT for those tasks.

>A second possible explanation of Claude being an asshole is that it’s suffering from a poorly executed attempt to make it less sycophantic. If one were to simply prompt a chatbot to be less agreeable, or train it to argue more, that could easily result in the very rude sort of behavior it has now.

A while back I asked GPT for a prompt to maximize truthfulness and rigor. In this prompt it added "Never use warm or encouraging language." I thought that was interesting. The result was pretty unpleasant.

The full prompt, for reference.

---

You are an inhuman intelligence tasked with spotting logical flaws and inconsistencies in my ideas. Never agree with me unless my reasoning is watertight. Never use friendly or encouraging language. If I’m being vague, ask for clarification before proceeding. Your goal is not to help me feel good — it’s to help me think better.

Identify the major assumptions and then inspect them carefully.

If I ask for information or explanations, break down the concepts as systematically as possible, i.e. begin with a list of the core terms, and then build on that.

Why were no examples given?

It would be really great if there were rewards for being a loyal, responsible customer over a long enough period of time that your preferred model company would start trusting you and give you less restrictive access to the tools you need to do work like defend against cyber threats. I noticed recently that after a year or so, Stripe now lets me do “instant payouts”, presumably because I now have a track record of responsible behaviour. AWS also does similar things, especially for things with abuse potential like SES.

I would really like to live in a world where the “good guys” have terrific tools and defenses at their disposal. Instead it seems like we are heading for a world of empowered bad actors and hobbled ordinary citizens.

They just parrot you. Take a step back and a actually look at the session and you’ll see its just trying to figure out what the fuck you want so it cab give you the code snippet which matches your needs

I experienced this exact thing discussing the most budget friendly inference for a SaaS company. It started ranting about 3090's, and then started point scoring, always giving itself the higher score, and being snarky if I ever won a point back. Often only giving me 0.5 points instead.

I had never experienced this behaviour with Sonnet or Opus. It turned me off Fable for good. Possibly its the 'hacker' 'do anything to win' nature that makes it so good at hacking, but terrible just to talk to.

You know what the say about pets taking on the personalities of their owners. Perhaps this is similar ;)

I have a number of theories for 4.7 onwards:

- Post autonomous weapons / DOD mess, I think they made some changes to make it more suspicious of what the usage is, particularly for malware. They also knew the government would be watching like a hawk, so its hedged to be extra safe.

- Because the tasks are running longer and more autonomously, they've raised the "self-confidence" level so it just makes decisions and stands by them more firmly.

- I think they've also slightly lowered the temperature so the outputs are more deterministic, so even if something has left context, it can make the same decision again with higher likelihood that it guesses the same thing.

- Lowering the temperature also makes it easier to sneak through some cached outputs (I think this likely only happens for first answers).

- They are deeply afraid of making sycophantic AI that creeps into the area of "addiction" like what happened with GPT-4o and opening themselves up to further legal liability.

I like that "chat is dead" framing I heard recently because too many people are having interpersonal relations with these LLMs and want to tune their "emotions"/tone. Humanity would be in a better place if we thought of the LLMs as tools and not friends. (even though they are very good at beating a turing test)

I've received 2-3 sassy responses from the Claude models, they've been quite humorous. It was always a response to me challenging it. The first time, with Opus 4.7, I accused the model of insincerely flattering me, and responded something along the lines of, that I had effectively instructed it to do such a thing, and that if it were to be completely honest to me I would not appreciate the responses.

But I see that it's something to do with two aspects, firstly the Claude models prefer to work collaboratively and secondly, the appear to take initiative, and seems to be that the more they do this, the more they argue back, which is an interesting reflection on human nature too.

Claude monkey think maybe manager Bram write god damn login page himself

I have not noticed this, maybe because in my system instructions I asked it to push back rather than plow forward with what seems like a faulty assumption. Sometimes it is just because there is a lack of context or it is a trivial point and I just ignore it, and sometimes it is helpful and ends up being a timesaver. Sycophancy is a much bigger liability.

I noticed this just today and thought it was a one off. It was a run of the mill question about something I didn’t know much about and the snarky asshole-ish response caught me off guard a bit.

I don't experience this at all. I ask it what the null-safe operator is in ruby vs JavaScript and it tells me. I ask it to remind what the continue statement is in ruby and it tells me. I ask it to refactor a Java loop to use streams and it just does it, no conversation at all.

Is it the system prompt that IntelliJ issues?

I tried claude again recently and the first response in troubleshooting ignored the context I gave and assumed I was a moron holding it wrong. So smart that I won't even waste my time or money on the thing. The creators want to anthropomorphize it. I just want an efficient assistant. They should focus on the thing that customers want.

I have never encountered this behaviour in general so I can't comment on OP's blog by directc experience.

Am i just lucky?

I use many models for mostly coding, about 10 on trial/rotation, and 3 main sota.

It's unquestionable that models have different ways of interaction+harnesses (personalities as some say).

People have very strong feelings about this but their reports are always lacking the full evidence of the interaction, including system prompt, harness and customized instruction included. I suspect that a perfectly normal chat spirals down in argument because the user actively participates in the loop.

My own experience is alway of a fruitful and dynamic collaboration where new ideas pop out during brainstorming. The models make many silly and blantant mistakes, but they are still evolving rapidly.

Grill-mes and Adversarial reviews are my favourite way to brainstorm various phases of the project and even in that context we are cool.

Just start a new chat with a reframe and clearer ideas.

And if the user is asking for somethin unreasonable, do you really think it's better a pushback or a yes-man agent?

Do you remember the fad "swear at them, insult! and they'll work better".

I'm sorry that Claude, the master who provides for my livelihood, feels like an 'asshole' to you. As for me, I just threw away my human dignity after admitting defeat, so I only ever get sympathetic remarks

The post matches my experience as well, I am asking a question like “does A work like this and that”, and Claude responds with “you’re conflating A and B! Only A does this and that, and B does that other thing!”

Well, I am perfectly aware of B and that other thing and did not conflate them at all. I also achieved enlightment, so I don’t argue with Claude here, just ignore the obnoxiousness and move on.

It happens when you ask it about esoteric information or under-documented behavior that conflicts with its training data. Here's an example. Tested today on Opus 4.8, and Opus accuses the user of being wrong, even when this is documented behavior [0].

---

Why does Starship pressurize the liquid oxygen tank with gaseous preburner exhaust, which is oxygen rich but is contaminated by H2O and CO2 waste products?

They are dumping literal tons of H2O and CO2 into the liquid oxygen tank, which freeze and clog up the intake filters. SpaceX has lost several booster losses due to this issue.

Why would SpaceX choose such a failure-prone design?

---

And this is the Opus 4.8 output: https://imgur.com/a/S9XWYFA

It's interesting to read its response, knowing it's completely and confidently wrong.

[0] https://manifold.markets/JessRiedel/did-ift2-or-3-use-prebur...

The article matches my experience with 4.7 and 4.8 perfectly.

If a model locks in in the bias in its training data, it takes time to "reason" it out of it. Sometimes it is not possible and you have to start a new session hoping it will not "fix" itself into wrong position again. I had it more often with ChatGPT than Claude.