Top
Best
New

Posted by htrp 18 hours ago

Anthropic says Alibaba illicitly extracted Claude AI model capabilities(www.reuters.com)
505 points | 864 commentspage 8
SubiculumCode 6 hours ago|
Everyone here praising these Chinese companies for their smarts (sure they are smart) has been ignoring this very big fact, they're improvements have mostly been by being parasitic on the leading edge SOTA models, not from some inherent innovation advantage. They are as innovative as their western counterparts, but they lack the compute, so their keeping up within months of those SOTA models depends on other means, like distillation attacks. I don't blame them; its the obvious only strategy when you cant compete in compute. But we shouldn't be blind to the real state of affairs: equal innovation; unequal compute; distillation attacks are the only vector to keep up.
kgeist 5 hours ago||
>like distillation attacks. I don't blame them; its the obvious only strategy when you cant compete in compute

>distillation attacks are the only vector to keep up

It's demonstrably wrong, they invest in architectural improvements as well, for example, DeepSeek's compressed attention. When you lack compute, you need fast training/fast inference, and distillation alone doesn't solve it. From what I understand, that kind of distillation "attack" (28 mln exchanges) only slightly improves instruction tuning/reasoning traces. If the base model is crap, distilling Claude on a few million exchanges alone won't magically make your model as good as Chinese models currently are (or magically make inference faster on the limited hardware they have). And training the base model needs a proper training run. Serving users at scale needs optimized architectures as well, especially with test-time compute and ever growing context lengths. That's where architectural innovations are happening in Chinese labs when it comes to compute.

SubiculumCode 4 hours ago||
I explicitly called out the fact that there is plenty of innovation, but that we see t Lots of innovation in both Chinese and U.S. labs, and I don't think that there is a co.parative difference there.
Anoian 5 hours ago||
[dead]
PostOnce 5 hours ago||
Suppose Anthropic trained only on data they paid to create, and not the internet or stolen textbooks.

It would still be extremely difficult to muster any sympathy for an organization whose MO is to go public not to honestly raise capital to fund growth and development, but rather to dishonestly leave someone else holding the bag, in some cases involuntarily as their retirement funds are passively invested.

And even supposing they were honest and didn't have an IPO, it would still be extraordinarily difficult to care about their misfortune, because "consolidating all thought-work into the hands of those few who can afford frontier models and datacenters and power plants" is also a special kind of misanthropy.

And even if that were not the case, they're filthy rich already, so who gives a shit if the Chinese companies prevent them from becoming quadrillionaires? :)

thadk 13 hours ago||
Does anyone have hints on what kinds of prompts are most used for a distillation like this—SWE-Bench sorts of things?

Is reconstructing the compressed knowledge in the model like reconstructing a lossy JPG or MP3 a reasonable analogy?

dannyw 10 hours ago||
RLAIF is a good place to start reading.

Claude will also help you with (mostly good advice) if you ask something like “Research and help me make the most effective plan to train a smaller student model to be better from a teacher model”.

I actually was doing an experiment with a GLM->Gemma E4B for fun, and Claude kept on suggesting I should also add Claude Opus as a teacher lol, suggesting techniques I haven’t heard of like thinking inversion (train a small model to deconstruct summarised thinking into detailed native thinking format of the student).

So I can absolutely see and understand the concern around Fable’s frontier LLM development mitigations, but their approach of silently degrading is completely wrong and dangerous.

AI classifiers, like all AI, can make mistakes, and it’d only be a matter of time before it mis-fires and silently sabotaging a university’s HPC cluster for physics simulations or something because the shape looks like DeepSeek or whatnot to a dumb fast classifier.

Chu4eeno 11 hours ago||
There are some Claude datasets (of indeterminate provenance) floating around on huggingface you can look at (or at least used to be, they might've been taken down).
jonplackett 7 hours ago||
How can there be any moat for AI ever, if you can just steal a model by talking to it?
gspr 7 hours ago|
This is what I find the most fascinating about the people arguing that you can copyright-wash anything (e.g. FOSS code) by passing it through an LLM. Surely that same logic applies to the LLM itself?!
zkmon 7 hours ago||
I don't understand. If they are simply using our API and paying for tokens, it's called a "transaction" and not "attack". The user is our customer who is supporting our business by buying our services. And we call them attackers. We happily make money by selling our services, and then call it as attack.

Back in the day, an "attack" was supposed to mean be someone acquiring our assets without paying for them or without having our consent. But none of this seems to have happened in this case.

We built a product without paying for most of the raw material we have used, and we don't call that as an "attack". Did we change the meaning of "attack"?

alpineman 7 hours ago|
Did Anthropic 'attack' all those authors it was forced to pay $1.5bn to for using their work without permission?
monegator 6 hours ago||
Soon, when even the enterprise subscriptions will have ads, every session will begin with a mandatory generated image:

> you would NEVER distill a model..

theplumber 6 hours ago||
Let’s hope they distilled it properly so we can have the best of both worlds: a decent model to work with without Anthropic’s drama.
ycui7 8 hours ago||
in a few more months, when Chinese model gets to Mythos capacity and Fable still locked down. What Anthropic will say? Why can they just admit they are not the only people who know how to train an LLM model.
Grimblewald 5 hours ago|
Claude thinks it's chatGPT, and various chinese models sometimes, whats up with that?
More comments...