I seem to recall Anthropic going on record saying that they don't do anything to model performance to stretch their compute capacity. I've anecdotally noticed massive peaks and troughs in performance week to week (albeit with Opus, not Fable).
I wonder what their official explanation for this behavior is.
When something is new, its capabilities feel incredible. Over time, those same capabilities become mundane, and you start to notice the flaws.
(Now, if TFA is actually measuring reasoning tokens, that's quite different! It's not entirely obvious to me how he is measuring.)
I don’t think that’s what’s going on. I notice flaws on day one of model releases. But I also notice improvements if the model is truly more advanced than what I’m used to. Then over time the same questions or tasks return worse results.
What is actually stopping these model companies from running a model at full capacity on release then once its name rings out, start serving users quantized garbage?
>What is actually stopping these model companies from running a model at full capacity on release then once its name rings out, start serving users quantized garbage?
As far as the API goes, it would be really obvious. I run a small service that uses LLMs extensively, and if a model suddenly dropped in performance it would be straightforward for us to prove it. We regularly run comparisons where we generate completions with alternative models to e.g. see if we could get away with using cheap models for easy cases, if the baseline outputs deteriorated it would be all over our metrics.
> What is actually stopping these model companies from running a model at full capacity on release then once its name rings out, start serving users quantized garbage?
...I mean, if they were actually doing this despite saying that they don't—promising one product and delivering something else—I think that would be fraud, no?
And, maybe it's one thing to secretly defraud normies like us (although class action lawsuits do exist), but I don't think major enterprises or the US military would take too kindly to it.
is it? it's still the same model, they can claim the quantization down to q4 still retains 98% of the performance therefore it's fine.
nothing on the fine print tells you what the weights are, you're just getting Fable 5, whatever that is
Are you telling me that companies might defraud people for millions and billions of dollars and pay fines that are 1000% less than their profits?" My goodness, you must live on a hell planet.
Sorry there for the smarminess but fraud is just a standard business practice these days and fines are the cost of doing business.
And I really am all for someone suing these companies forcing discovery so we can see how the sausage is made and how many eyeballs are in it.
The question isn't whether the penalty would be less than their profit, it's whether the penalty would be less than whatever they make by secretly downgrading the models (or whatever it is you suspect), which remember also causes consumers to get less value out of the product and more likely to cancel.
The reputational hit, if this was to be confirmed, would also be massive. And I do think it would leak! Some employee would say something.
It's called hedonic adaptation.
> What is actually stopping these model companies
You can say this about any company in the world, selling anything.
It's trivially measurable, and there are people running the same benchmark on the leading models every day and measuring if they degrade. Spoiler: they don't.
But you can always say "the conspiracy goes higher", and that the companies know about these daily benchmarks and are routing them to "quality" envs.
Their exact phrasing IIRC was that they "never intentionally degrade" their models.
This still leaves an absurd amount of wiggle room for arguments like "oh no, our evals show that this quantization has no detectable effect on performance (in the eval distribution) therefore running the quant doesn't degrade quality"
They did. More than once...
Anthropic Walks Back Policy That Could Have ‘Sabotaged’ AI Researchers Using Claude https://www.wired.com/story/anthropic-responds-to-backlash-o...
But it's still happening: https://github.com/anthropics/claude-code/issues/81759
And here's another great example of how a bunch of people who don't know what's going on throw noise into the system. That post is simply confused: the 1m opus calls are the auto-mode classifier, actual agent calls are still in Fable.
>bunch of people who don't know what's going on
Do you know why nobody outside the companies knows what's going on? Because they sell a black box with magic inside while steadfastly refusing to tell you if they are pushing buttons on said box while it is running.
Can you imagine how much fraud would exist in the gambling industry if the gambling commission didn't exist at all? Everytime an industry is unregulated and has high costs of entry the entities in the industry abuse their customers. The incentives are much too high for them not to.
Look at the usage. Fable wasn't being consumed.
Last time they were called out, it was a regression in Claude code itself.
At least that's their explanation. Either way, it wasn't a good look for "vibecoding" but it got brushed over.
They are deploying optimizations weekly (if not daily) with various AB tests. They don't manipulate model performance, but they do actively perform tests.
You're right to push back, and one honest caveat -- they could just be lying.
Your caveat isn’t just a side note, it’s worse than that, they have incentives that go against your best interests!