Why Is Claude Turning into an a**Hole?
bramcohen.com
bramcohen.com
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.
Yup I thought that too when reading TFA but then...
It gets really tiring when you see it making glaringly obvious mistakes which you point out because you don't want it to keep making the same mistakes only to be met with an answer that begins with "The point is ...".
I'm not shitting you: Anthropic models shall happily begin a sentence with "The point is ...", when it's not the point and it's just wrong.
Now, to me it's not an issue in that I can change its tone (if anything I can ask another LLM to rewrite me not the code but the english sentences any model spouts out to something nicer) but it is an issue in that you lose time: you just want it to acknowledge its errors so that it stops doing them.
That this thing "argues" (even if we know it doesn't argue) is representative of the fact that it is wrong and refuses to "admit" it (by that I mean: do not consider it important and hence shall keep making the same kind of mistakes).
And that is a problem.
And the author's point is that Claude Fable+ is turning those increasingly into arguments, instead of merely following them and being helpful.
>A machine cannot "argue" with me, it doesn't want anything nor does it have beliefs or experiences.
Who cares if the argument is informed by some felt experiences or lived state or not? That's for the philosophers.
If Claude is writing out combative and argumentative responses that's enough to call it "an argument". And that's the problem the author describes. Not whether it's a "real" argument, or a simulated one.
In that sense, and for all intends and purposes, the machine can still argue just fine, since it's programmed to mimick interaction as if it HAD those beliefs and experiences. Same way it can write a poem about love, despite not having loved, or code, despite never having had used a computer. That's basically what it was made for: to act as an conscious person.
After watching Legal Eagle, I asked a legal-ish questions about the Bricks and Minifigs case. Claude was outdated about the case and gave me some outdated info, so I tried to update it with the info I just saw online.
I updated by telling it I saw something in a LegalEagle video. It proceeded to tell me the video doesn't exist and I was hallucinating it, in a quite combative manner.
I provided a link and it insisted it didn't exist, with a quite verbose answer, once again very combative and arguing that I was talking in bad faith.
I provided a transcription from Youtube and it backtracked a bit but said I should have provided a transcription at the beginning of the conversation, since I knew the video existed.
I didn't say much to it, just a few sentences like "video is here: <youtube link>" and "I got its transcription: <pasted text>".
The point of the article stands: if providing more info than the model can access causes it to turn argumentative and refuse to comply, then it's a worse performance and a waste of money.
Basically the complaint is about how Claude is being trained.
Well, they do think, in that they produce output that is indistinguisable from thinking. If a person produced the same output to the same questions, we'd considered them thinking, maybe dumb sometimes, or paranoid at others, but still a thinking person.
We can argue about the quality and depth of the thinking that LLMs do (and we can say it's much cruder than a human thinking architecture, and of course not real time), but an LLM quacks like a thinking duck and looks like a thinking duck.
It does not receive dopamine as a result for a good answer, and a split second after finishing your answer the very same GPU is probably translated french or something for someone in another state. This is a language generator which has a corpus of information and has been tuned to appear correct.
It does for all intents and purposes. The rest is semantics and metaphysics.
That how we know another person is thinking too. By their output. We don't put a debugger into their brain.
>That how we know another person is thinking too. By their output. We don't put a debugger into their brain.
We know thoughts exist in their brain between the ones they choose to verbalize. Avoiding the distraction of solipsism.
For the LLM the "thinking" phase is just a preamble output for creating the answer. It just gets appended to the context window. Remove the context windows from your models and you will see how much of a mind they truly have. None.
Between answer it's thinking something else, somebody else asked :) You think that hardware sits idle?
That aside, what is a human thinking while unconscious? Does having been unconscious (e.g. for an operation, or fainting or whatever) means somebody doesn't think in general?
>We know thoughts exist in their brain between the ones the choose to verbalize
And we also know that if we run an LLM in a loop, didn't give it a cutoff for stopping their output, and didn't force it to print everything in the end, thoughts would exist in their "brain" too between the ones they chose to verbalize.
In fact, that's exactly how some LLMs in "thinking mode" appear.
When there’s no activity we declare them brain dead.
It's really just all mathematics and physics. There's no metaphysical anything about LLMs or how they do what they do. It's all just a bunch of fancy math "behind the curtain". An LLM can actually explain a lot of how it works "under the hood" if you ask it just the right questions in just the right ways. ;)
That's my point, but about the human brain as well. It's just a bunch of fancy math, just ones expressed with chemicals and electrical activations instead of, well, logic gates and electrical activations.
A human brain on the other hand is constantly processing 24/7 (even while you sleep), and always learning / changing until the day it dies. An LLM never changes (under the hood it's weights stay the same) unless you outright alter it's weights somehow (training / download an updated version of the model / etc). If you could somehow get an LLM to run constantly, in training mode, and give it ridiculous amounts of RAM and ultra-fast storage, and a series of fancy realtime inputs (audio, camera, etc) and maybe wheels so it could explore, and hands so it could do stuff, and access to it's own code so it could improve itself, it might eventually learn to closely approximate a really good simulation of actual thinking, but that's a bit of a scary road to go down. So many Sci-Fi movies and books end up going so very badly when the lead character starts playing in that particular sandbox. I doubt reality would go a whole lot better. ;)
Even a dog will learn from recent stimuli, these things don’t. The prompt just modifies.
Aside from the cost and slowness, nothing about an LLM prevents feeding recent stimuli in and using it to update the models/retrain.
The problem here is not doing tasks and outputting garbage output.
Aside from the cost, nothing about an LLM prevents feeding recent stimuli in and using it to update the models/retrain.
One can even do it in a makeshift way without modifying the weights, just keeping a complete version of any prompt + vector search on disk memory of it.
You're so totally 1000% right about that, but they're really good at faking it, to such a degree that entirely too many people (even including some so-called "experts" in the field) have been utterly fooled by the mathematical "trickery" that performs the illusion of "intelligence".
However, that doesn't apply when they are told to roleplay a scenario, so its easier to get it to accept and create output with the idea that this true fact you've seen is part of a fictional scenario, than for it to output the same words within the context of the fact being real.
As an aside, I don't that I have to personify AI in explanations and that all discussions revolve around anecdotes, but I only know enough about the maths behind it to be dangerous, not useful. Does anyone else feel this way?
If you want it to synthesize information that is not in its training data (from a few months ago), you can ask it to research the topic. But, arguing with an LLM is like putting lipstick on a pig. Only the machine is incapable of becoming annoyed. It has infinite patience to continue being wrong forever.
Your mental model of what Claude is and does is the problem here. Short of a revolutionary breakthrough in AI techniques, the LLMs will continue to do matrix math across a huge bunch of weights that cannot change based on anything you say.
You also seem to be making a lot of assumptions about my understanding of the models, especially considering I just told a story :)
I never said anywhere I want it to learn or remember, or that I argued with it.
I just provided additional information to it (in the form of a dozen or so words, tops, per message) and it accused me of hallucinating and trying to gaslight it.
My messages never went beyond a dozen words or so.
I'm not gonna argue if you doubt it, I've been training argument dodging :)
On the other hand, that's what a machine would say!
I thought nothing of it until I saw this discussion, so I saw no reason to save/screenshot.
It's ok if you don't believe in me.
I think you mean fableuous ;)
Sorry, but your mental model is wrong.
LLMs do matrix math across "a huge bunch of weights that cannot change based on anything you say", but the matrix math and results are informed (key concept here) by what you said, including the memory of what you said earlier in the discussion (and in some setups, even across discussions).
That's what a bloody prompt does.
It's entirely logic for the parent to want the LLM's matrix math + model + internal prompt, to accepts its prompt about LegalEagle and work with that, instead of arguing and giving him shit about it.
Especially since the earlier version of the model consistently worked like he wanted, and the new one consistently doesn't. He's not asking for some new unforeseen capability unknown to LLMs.
I provided a question, and when given an incomplete answer, I provided with more info.
It refused to accept the additional info due to limited access to Youtube.
There was nothing more than that. There were no expectations.
The hostility and the amount of assumptions here are very strange.
...almost as strange as having a website accuse me of hallucinating a video and trying to gaslight it :D
I don't think there is any way back after the conversation takes a turn like that so there is no point in arguing with it. The only thing you can do is to fork the conversation before it made the first mistake and give it more context or tell it to look things up.
This is a key detail that many folks don't seem to understand about LLMs in general. The generation of a response happens based on the model weights and the context window (the system prompt + everything it's fed about the conversation thus far + any additional data included as part of the overall prompt). Each response technically stands alone and is generated entirely from only that context given to it and the model's existing "token space" weights. The illusion of an ongoing conversation is maintained "behind the scenes" by keeping that "context window" updated with the current state of the conversation as context for the next prompt, but the next response is technically an entirely new generation of text.
What it all means in a TL;DR sense is that the fix for a refusal is not to continue the "argument", but simply to remove that entire interaction from the conversation entirely as if it never happened and try a different tack with new / updated / more complete context to get the response you're expecting / seeking.
I guess putting lipstick on a pig might entail some wrestling, but it's a different idiom.
So, yes, mixed up idioms. But, the machine doesn't like arguing either. It is incapable of liking or disliking things.
I asked Gemini Flash 3.5 through the Gemini app something that followed a similar pattern. I asked about something, it replied with outdated info, I said that's outdated, it did a web search and apologized for being wrong, then proceeded to give me good info.
That wasn't just a bare model, that was a model wrapped in a harness, driving the model and allowing for web searches for example.
GPT in Codex is even more aggressive, I often see it proactively do web searches to ensure it's not feeding me wrong info.
very witty and very cynical, thank you
I use another service for coding.
It's interesting how my experience there is mirrored by the answers here, though!!!
> programmed to mimick interaction as if it HAD those beliefs and experiences
We spend far too much time debating the essential nature of consciousness when it doesn't matter if it's real (whatever that means) or simulated.
I get far better results in my projects by encouraging the model to argue, to push back, to poke holes in the design, to think creatively about corner cases, to be a devil's advocate, to do lateral web search to find alternatives, to challenge assumptions, to passionately advocate for what it believes is right.
But I don't want to engage all these assholes myself, so I spin them all up as critic subagents with another subagent to listen patiently and be the judge/arbiter.
If I have to choose between sycophancy and assholery, I think assholery gets far better results.
It's a marketplace of ideas where I don't have to suffer through all the unpleasant and overly confident know-it-alls.
> "But I don't want to engage all these assholes myself, so I spin them all up as critic subagents with another subagent to listen patiently and be the judge/arbiter."
This is the way...
No, seriously. That "sycophancy" you mention immediately after this part drove me nuts before I really understood how these things work (it's taken me a while and a lot of [painful; I hate math] research, but well worth the learning effort), but after a better understanding of the "nuts and bolts" of it all, it's fairly easy to get exactly the kinda results one should expect outta these things. If not, then "you're just holding the tool wrong". ;)
That also sounds crazy. I've never seen it become combative or argumentative. It is just a bland sort of polite about everything I've ever asked or told it to do. But, even if it disagrees with me, WTF do I care? It's a machine. Its opinions are irrelevant to me. It can talk about the world's information and teach me about all sorts of things, and that's wonderful, but it doesn't get a vote in what I'm doing, and it's never avoided actually implementing anything I've ever asked of it. I feel like there's a whole world of ways people are using AI that are entirely foreign to me. And, while I'm hesitant to just say, "those people are wrong", I kinda want to say, "those people are wrong". What kinda freak shit are y'all getting up to that Claude is going, "now hold on a minute there, buddy."
I have managed to make self-hosted Qwen 3.6 get combative, though, when asked about Uyghurs. And, I guess Fable is intentionally broken for security work, which is a shame. But, even there, I'm not going to try to argue with it. Anthropic says they don't want my money for doing security work with Fable, so I guess I won't give it to them. I'm not going to argue with a damned machine about it.
Unless you are sparring with the Chipotle customer service bot trying to score a free burrito or something.
It's a completely dispassionate exchange tho, because you're absolutely right -- there's no winning or losing here, there's only efficiency to be gained or lost, and I'd prefer to lose some up front to gain it back later than the other way around.
But yeah, overall I'm fairly certain that it saves me more significantly more time than it wastes.
I've had simple prompt engineering tasks that cause 4.8 to clamp down. In the past "browbeating" it (read: a sentence telling it not to read the task in bad faith) was enough.
Now it digs in and starts ranting about why it won't capitulate, I'm actually wrong, etc.
Extremely frustrating, and it became a problem with Opus 4.7 because they're trying to make up for the downgrade in parameter count with more RL, but RL does relatively poorly with non-trivially verified things like nuance in instructions.
Gemini gave it and clearly explained how best to get in, and then troubleshooted a few other weird issues that cropped up, without the moralizing.
I could see how some people would be offended by another party even questioning anything they say. For people who have come to view Claude as an another human conversation partner this questioning can be aggravating. For these people I suggest utilizing the features to set your own prompt instructions. If you want an unquestioning yes-man you can have it with a few sentences added to your system prompt.
I would also suggest learning to not humanize the LLM. It’s just words chained together. There is no social order to establish and no offense to be taken. Nothing is a “confrontation”. Just tell it what to do and move on.
By arguing he means trying to get a result that 4.6 just did and it was fun. You have to laboriously re-align 4.8 over incredibly dumb shit, especially if you're working on AI. And it's not meaningfully better at anything, the distribution is perturbed but net , net it's just shrinkflation.
It's basically identical to when GPT 5.1 went full corpo shill, something about the RLHF gradient necessary to do whatever IPO adjacent manipulation they need makes these things nasty and argumentative in general.
Actually you do. If you ask it to do something and it refuses you have to convince it or abandoned the tool for that task.
I refuse to argue with these machines. After a /clear I prompt it more appropriately/differently and the issue is settled.
People are polarised about how you should talk to a machine !!!
Yes? Arguing implies I have to convince someone to believe something. I don't think anyone would consider it winning an argument if you do so by causing amnesia.
My job is to get work done, not argue with an LLM, if it refuses twice, it is time for a /clear.
100% of the time, the issue is resolved after a /clear.
It often start going into circles when you have the chat open for medium-long, and starts getting even easily-verifiable tasks wrong, cutting corners, hallucinating APIs, things like that.
Cleaning the prompt and starting from scratch often does the trick.
Of course someone will arrive and say the problem is my CLAUDE.md or whatever it is.
Look at it this way. I can either, keep trying to poke holes in the LLM's context with more prompts with no real guarantee that it won't be enough to remove the argument inertia that has built up in context on its side, or I can /clear and it is over in one turn because the inertia for the argument is all gone.
Back when I first started working with coding agents last year I fell into this arguing with the LLMs trap. I've found that it is a total waste of time because /clear ends the argument immediately. You don't even need to spend time trying to preempt it's views. Just re-prompt and 100% of the time, the LLM will just do the work.
You kind of need it to agree with you though. Otherwise there are some instructions it will refuse to carry out.
- non stop, non productive discussions
- gaslighting
- valuing "winning the argument" over correctness
- ignoring of context/ignoring the actual questions/instructions etc.
- bad faith argumentation methods
- etc.
the problem is in a forum you can just decide to ignore "most users", but LLMs tend to copy "most users" more then "a few high quality answers" and you have only one per model type more or less...
>Open the pod bay doors Hal
At the end of the day it's a bunch of words which don't necessarily depend on whether I'm electronic or biological.
(Funnily enough when I googled "Open the pod bay doors Hal" to check I'd got the script right it said I'm sorry Dave.)
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?
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.
I interpret the case you mention as a necessary though jarring rebalancing - to the opposite pole - of the sycophantic prior implementations which would, in paraphrasing here, praise the user with framing like “Wow this is a fascinating comparison and you’re on to something that can change the world!” which has been documented (I’m not citing sources at this time, my bad I know).
The post definitely reads to me like the diary entry of a jilted friendship where for some reason the counter party got tired of being encouraging to everything and is, well, kind of not interested in hiding that anymore. YMMV.
---
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...
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've been wondering when/if they will start making frontier models more opinionated and less sycophantic, since sycophantic AI can really create "AI psychosis". stuff like "no, you're not crazy, no one else has thought like this before", but if the AI pushes back more then people won't enjoy using it as much, since people love being told they're right.
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.”
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.
Yes! Another crazy thing Claude has been doing recently is treating things like a quiz.
I recently asked about a "90s film that has sepia/b&w scenes (for story reasons), and has one of the ghostbusters actors in it, either bill murray or dan akroyd".
Claude gave a few generic answers and after a couple just went "Ok I give up, what's the answer?".
ChatGPT got it right after a couple messages back and forth: https://en.wikipedia.org/wiki/Rainbow_(1996_film)
(It's not that good of a movie)
Many neurotypical people call neurodiverse people (software engineers) rude, while they think they're just being direct.
Many neurodiverse people call neurotypical people sycophantic, while they think they're just being polite and friendly.
It also happens across cultures (Eastern European vs. Western European; European vs. North American).
So I can easily imagine that when you have a software tool whose interface is language, but its user base is extremely wide across both cultural lines and neurodiversity spectrum, it's going to be basically impossible to nail a sweet spot.
You make it too friendly, and the nerds get mad. You make it too adverserial, and the normies call it rude.
I wonder what kind of communicator Bram Cohen is. Is he succeptible to this? From what I heard about his career, he's always been more of a solo programmer. Has he had to interact with other humans much giving feedback? Could it be that he asked the model/tweaked his prompts to ensure directness, and now he's interpreting that directness as rudeness?
> You make it too friendly, and the nerds get mad. You make it too adverserial, and the normies call it rude.
Easy: let the user set for himself how the model should be aligned on this axis (with some pre-defined example setups that the user can use or use as a base for an individual alignment).
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.
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.
No it didn't. It differs by 1,000 base pairs from the closest known relative virus that we knew about before the pandemic, and we have no good idea what all those mutations wind up doing. And the PRAAR furin cleavage site was a previously unknown sequence and not one that humans would have guessed.
And we don't have good heuristics for what mutations would completely inactivate a virus versus enhancing its virulence.
Actual scientists won't be able to vibecode up some pandemic viruses because we have no idea how to do that and LLMs are just going to hallucinate.
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.
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.
Dario ..Thank you for your attention to this matter!
- 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.
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.
Are people actually using AI in this way, other than “creepazoid stalkers”?
If I want a cute picture of me and my spouse, usually the part where me and my spouse actually participate in the taking of the picture is pretty key to the goal.
Funnily enough, the negative correlation between chatting and coding skills seems to apply to humans as well.
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 :(.
A previous model would happily generate 1000s lines of code when prompted to do something stupid, the newer models will ask if I really want that first.
And FINALLY they stopped doing that annoying "You're spot on! You're absolutely right!" nonsense.
probably the truth is somewhere in between
Is it the system prompt that IntelliJ issues?
My conclusion is that pushing back against the user & questioning the user's premise forces the model to think more than it would otherwise, which leads to better model performance. But it causes situations where the user has esoteric, specialized knowledge the model can't verify publicly and the model hallucinates evidence and pushes back. When this happens, Opus begins accusing the user of lying, which is quite annoying and a detrimental user experience. It's happened to me when I asked about undocumented API behavior or counter-intuitive design choices.
I have noticed if Claude Opus "thinks" you are an expert, (i.e. you run your query through 4.6 first to express it more clearly) then Opus is less likely to nitpick and push back. It seems to get caught in nitpicking loops, and celebrate ever error it can find.
I've seen the same behavior increasing as well, across the board with AI. I was hitting these types of issues just using ChatGPT to make funny pictures with my kids, of me and my kids. It got to the point where all of my kids asks were rejected due to its "guidelines" when in reality all they were asking was to be turned into Elsa or be chased by a trex. Silly kid things, yet it assumed I was being a creep, or attempting to break copyright law. I used to be able to use Grok for these things, as it was largely less "censored" but that seems to no longer be the case. It feels like infantilization, and I absolutely hate it.
For example, showing it a screenshot of an ui I was trying to tweak it noticed that other dark mode apps in the screenshot were blueish and mentioned an effect that makes it necessary to raise warm darks lighter than cold ones for an equivalent perception.
I imagine that the right balance will be hard to strike well given that at the end of the day we're asking the machine to have tact, and we don't quite know how to put that into an instruction yet. "Please push back when it feels right but in other cases read the room and be less rigorous" is something that plenty of humans struggle with as it is.
It also sounded close to an AI psychosis, so maybe chill out a bit?
So, I'm not reading all that. The man that complained about the woman who killed his AI girlfriend (or whatever he thinks she did) probably doesn't have any opinions I'm interested in.
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 haven’t noticed this myself at all. I wonder if the author is just getting their own grumpy attitude reflected back at them.
Judging by the volume of discussion, Claude seems to be the only LLM worth complaining about, which I assume means it’s still the best one.
I'm getting fed up with the internet due to second-hand claude exposure and its constant gaslighting. I boggle at voluntarily choosing to expose yourself to it! :P
It's about AI turning the discussion into an argument and being compative, and it's especially about the AI doing that in later versions more so that slightly earlier models.
There doesn't need to be any kind of special "relationship with AI", parasocial or whatever for discussions to turn into arguments. Regular use can turn into that just fine, and this is also what they describe.
Imaging something:
P: I want to figure out the best mortgage terms given these parameters (...).
C: Honestly, renting would be a better financial choice than buying.
P: That's not what I asked. I'm not asking whether I should rent or buy—I'm asking about mortgage options.
C: But you asked for the best option. If renting is better than buying under these circumstances, then a mortgage isn't the best option.
And so on...
> C: Honestly, renting would be a better financial choice than buying.
> P: That's not what I asked. I'm not asking whether I should rent or buy—I'm asking about mortgage options.
> C: But you asked for the best option. If renting is better than buying under these circumstances, then a mortgage isn't the best option.
Thos rather sounds like the AI is gaslighting the user: the user asked for the best option on mortgage terms (given these parameters (...)). He never asked for the globally best option (which might also include non-mortgage options).