Comment by arkadiytehgraet
20 hours ago
> Again, you're just interpreting anything that goes against the "AI bad" grain as shilling.
Once again, putting "AI bad" into my words. No, Anthropic shill, this is not what I am saying. Is your LLM malfunctioning or are you not really getting it? Stop gaslighting, Anthropic shill, and try to stick to the actual words I am saying. I understand that this is hard for you, because then you would have no real argument to make, but please try, Anthropic shill.
> Please show it.
I used an LLM to count actual experiences of people reporting their experience with Opus 4.6 being degraded. There are literally several hundreds of such data points. This is data. People, who are employed and actually use LLMs for coding, unlike you, Anthropic shill, who uses it only to poison online communities. Are you really going to disregard all that to claim it is mass-psychosis or something? I guess you would, Anthropic shill, because that's your job, to peddle bullshit LLM-hype unbased on anything in reality.
> It was an incoherent mess of insults, so I am still not sure what you're trying to say.
Repeat after me, Anthropic shill: being called a shill is not an insult. You are a shill, stop being obtuse and at least take some pride in your work of promoting LLM-hype. So once again you are providing nothing to the conversation except for baseless accusations of insulting, which I did not do, and refuse to answer to the actual arguments I made. I can provide it again, but you would likely ignore it because it just showcases how you are clueless about the topic.
Your words, not mine:
> Third, and most importantly, the actual output measurably changes. Quants have a lower latency, higher TPS, and different token distribution.
I asked if you understood what "different token distribution" meant. I can tell you what it means: models performing worse at coding tasks. So people report models being worse at coding tasks, YOU write that indeed quantization leads to that and then just "forgot" about it? Nice level of "objective" discussion, Anthropic shill.
> So now I'm lying about my employment on an anonymous forum for... what, exactly?
It is not anonymous forum, as much as you would have liked it to be, so that your shilling could not be dismissed as easily, Anthropic shill. For what? So that people would fall for the bullshit you are peddling. Are you really this dense, Anthropic shill?
I must admit I skimmed most of your comment because it is largely an incoherent rant, but I will address some points:
> This is data.
Nope. Because frequency bias is a thing. If you hear on Twitter "model X got nerfed," your brain will look for that pattern and notice it more than usual. This will then confirm your suspicion, which leads to a vicious cycle. Then you tell your friends and the same phenomenon repeats.
None of this requires the model to get worse. It's a well understood psychological phenomenon.
> I can tell you what it means: models performing worse at coding tasks. So people report models being worse at coding tasks
The perception of a model performing worse at some coding task is not what "different token distribution" means. You should ask AI to explain my comment ;)
Latency and TPS can also tell you if you're getting a quant.
Anyways you should really get some help. Praying for you!
> frequency bias
Gaslighting again, Anthropic shill. What does frequency bias have to do with the objective fact that hundreds of people reported their own experiences with LLMs being degraded over a short period of time? The very same tasks that the very same LLMs could do, they no longer can? You seem to ignore this FACT, this DATA, and instead have to gaslight and divert into "frequency bias" nonsense. I do understand, why you are doing it, Anthropic shill, but at least have guts to admit it.
> perception of a model
You once again ignore what your LLM outputted and you typed yourself and divert into "perception", Anthropic shill. You do not need to sample entire output for tokens to notice the distribution moving. If the LLM used to be able to achieve set goals and no longer could, it is already a sign of the distribution shift. And as you said yourself, different token distribution = model being quantized. Which is reported in hundreds of separate instances. Which is more than enough to conclude that the model was, in fact, quantized, and no amount of gaslighting can change that. But you are an Anthropic shill, so you have to peddle your bullshit, trying to twist facts to support your employer's narrative. And you deserve being called out on that, Anthropic shill.
> What does frequency bias have to do with the objective fact that hundreds of people reported
This isn't hard to understand
"Model is nerfed" claim hits social media
Someone else sees it, frequency bias makes them think their model is also nerfed, and they amplify the claim
Now it spreads, like a virus, even if the model never changed
Social dynamics like this are well understood psychologically
> If the LLM used to be able to achieve set goals and no longer could, it is already a sign of the distribution shift.
The more likely explanation is that you're looking at older LLMs with rose tinted glasses, and misremembering what it could achieve
Otherwise you could measure the token shift and see the better tps and latency
Your own evals would trend down
But no one, not one person, has presented empirical evidence of being served a quant. Just vibes.
2 replies →