Hacker News Reader

Hacker News

Comment by BryceEller | Hacker News Reader

Parent Full thread

BryceEller

I dunno there are some other nerf trackers that aren't showing a regression: https://github.com/ninjahawk/livenerf

View on HN

The problem is that it seems to be an A/B testing thing:

https://news.ycombinator.com/item?id=50422951

So only if you were "selected" for their experiments, your account was nerfed.

But the claim here is "the models are nerfed before a release", not "they A/B test quants / efforts". If it were A/B testing, it doesn't really make sense to correlate it to release windows at all.

The models are nerfed [for some people] before a release

They might want to see how people react to a new model depending on whether they've been nerfed.

If the nerfing is lower effort (seems to be, presumably why they hide the CoT) then they may be wondering if they can get away permanently with lowering effort on the new release provided it gives a wee improvement.

For sure, myself and colleague saw this nerfing. He said - new model coming. It was incredibly obvious. I also started getting a lot of "How's Claude doing today" messages at the same time, which is presumably to get feedback on the A/B

I always skip those. I know it doesn’t matter in the bigger scheme of things, but I don’t want to be classified as a “good feedback provider”.

I’m working on a task. And the task is never to evaluate the agent performance.

And whenever you press a number, even 0, they keep your full conversation.

Source?

I strongly believe it will be shit like this that will have people abandon them. They're being way too cocky for a company that pretty much didn't exist 2 years ago

Its incredibly easy to roll out a compute instance to X% of users, or route traffic between compute instances

so not everyone will have the same experience. If I was saving on costs that would be the easiest way to do it without deploying, without any nefarious intent. The persons in charge of that likely don't even know about how individual users are impacted

The degradation only happens to some accounts but not to all.

https://www.reddit.com/r/ClaudeCode/s/64DXRERLyU

It would be nice to know how they select. I max out my $200 subscription every week and I was selected. On Saturday, the Opus on my account that happened to be supervising a running live process, became unusable.

It would be trivial for them to opt-out accounts that are doing benchmarking from a planned regression.

I really really really do not trust companies on the verge of IPOing from goosing their numbers. Even if the lawyers are up their ass, the social pressures are more than enough to compel the behavior. For every good person willing to do the right thing even at the cost of a delayed career progression, there are 20 people willing to do anything not even for a dollar but for an "atta boy!" from the boss.

That's a major reason to use local models. Or open weights models on a trusted cloud provider. Even if they are way less capable, they're more predictable

If it's a private benchmark, how would they know it's a benchmark?

In the context of public relations they would not need to ID every benchmark only the most credible or most public benchmarks many of which are openly available on github and whose problem-set could be fingerprinted.

You're absolutely right that it is a game of whack-a-mole and eventually we may have leading private + credible + popular benchmarks that ultimately dissuade the behavior, if it can be dissuaded.

Look up eval awareness. There are all sorts of subtle signals an LLM could use to know if something is shaped like a benchmark or not. But that's not a bad thing, every animal behaves differently in play vs real-life.

Yes the model can do that, but it would not be "trivial" (as GGP had said) for Anthropic to opt out accounts performing private benchmarks from this theoretical model downgrade people think is happening.

The most plausible approach would be for Anthropic to train some separate classifiers to detect benchmarks. I'm not saying it's completely impossible but IMO this is conspiracy theory levels of effort.

At what point does this become false advertising, or even fraud? Imagine paying AWS for a high-performance server, but 5% of customers who order one quietly get a slower, cheaper server and still pay the full price. That's a serious claim and makes me skeptical Anthropic is doing something like that without more evidence.

Actually, as a person who rented a dedicated server or two (not from aws) this exact thing happens quite often.

For example a provider quietly underclocks your cpu or puts it is an energy saving plan hoping you won't notice.

Or your "dedicated" CPUs on your VM disappear for minutes at a time, or just process a tiny number of instructions

I think they could easily reduce several parameters to slightly reduce cost (something like "it's the same model, a few less passes, but quality is still 95%+").

Or switch to a faster, cheaper cache mechanism or something that's almost as good, but much cheaper to run, in the name of improving efficiency and keeping up with demand. This switch to cheaper mechanisms is what it feels like to me.

It's a problem for us. When they nerf the model our productivity really drops (due to the specific use case and process we have). We've started using OpenAI and will gradually use it more and more if this continues.

"Nerf tracking" is adversarial and exploratory. It's not something Anthropic wants anyone to do, and it's not something that the "nerf tracking" designers are necessarily doing in a foolproof way. Don't assume too much from them either way.

But do from anecdotal evidence without benchmarks?

Did I suggest that?

Sometimes you just have insufficient evidence to be sure of something either way, and nobody gets to win the argument or shut somebody else down.

In this case, some have reported analytical work[0] that goes beyond "gut feel" anecodotal evidence, but even that's not going to conclusive and may not be trustworthy at all.

[0] https://www.reddit.com/r/ClaudeCode/comments/1xr9x5y/the_ner...

No, you did not comment on it. That is why I asked. Thanks for the clarification!

For example in that A/B test the user got more effort from the model.

Ok, it probably burns tokens faster, but is it really a nerf?

These nerf trackers do not work. I wish people would stop posting them.

They don't detect a quantized model until the quantization gets down to 4 bits or less.

Humans absolutely do detect quantization well before 4 bits...

Try it yourself if you want. Try a small local model, both quantized and not, and I would be very surprised if it isn't immediately obvious which is which.

But quantized local models also dip in evals.

It doesn't matter if it's local or not, I just suggested local so people could test themselves.

But no, evals and benchmarks do not detect quantization, until around 4 bit quantization.

https://developers.redhat.com/articles/2024/10/17/we-ran-ove...

https://pub.towardsai.net/your-model-passed-safety-evals-the...

That is not my experience. I have issues where quantized models fail pretty early so I have evals to catch it.

And the second article seems to reinforce the idea. Yes, broad perplexity tests are not a good proxy for the tests used during training.

Quantizing anything is effectively changing the weights, even if you don’t quantize the weights themselves.

Nothing stopping Anthropic (or other labs) from pulling a VW Dieselgate and ensuring that benchmarks are routed to a non-nerfed model.

Should be fairly routine for anthropic to game a nerf tracker?

That must be on anthropic's benchmaxxing CI test

Nerfmaxxing.

It probably need a whistleblower to uncover what they are doing.

Reply on news.ycombinator.com