Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Anthropic says Alibaba illicitly extracted Claude AI model capabilities

reuters.com

Post reply on HN
Print view View original

961–970 of 1001 posts

Re: Anthropic says Alibaba illicitly extracted Claude AI model capabilities

#961

Post by cheesecakegood »

fnord77 wrote: post #344
0xbadcafebee wrote: post #11

There's two basic kinds of distillation: 1) the massive [and dumb] method where you ask a question and use the answer as reinforcement (Black Box), and 2) more targeted distillation where you use one model to directly inform/train/guide another model (RLAIF). The latter is basically fine-tuning the model with direction from another model. Thousands of businesses do this every day to fine-tune. This is almost certainl…

Can you reach into the model and "transplant" weights directly?

Yes you can! Well, mostly, depends on how pedantic you are with definitions: you can transplant layers but not weights, which in common parlance are conceptually similar. But usually it isn’t a good idea for a few reasons.

There’s a really fascinating example[1] where a guy identifies a particular set of layers and transplants them. Overgeneralizing, early layers are encoders and the later layers are decoders and in the middle some blocks seem to do specific things or tasks related “reasoning”. So you can actually create a FrankenLLM and it sometimes works.

This needs architectures to be roughly similar however and internal representations to be consistent-ish so for “stealing” it’s not really a thing (other practical concerns aside)

[1] https://dnhkng.github.io/posts/rys/

Re: Anthropic says Alibaba illicitly extracted Claude AI model capabilities

#962

Post by InkCanon »

whizzter wrote: post #563

Whoa, Antrophic,etc are really running afraid that their IPO's are gonna crash when people realize that the open models are Good Enough(TM). So I'd put it at 30% that this is a ruse, say that Qwen 3.5,etc is tainted by training by them and start issuing DMCA takedowns to protect the IPO valuation (Or they'll hold off on that, getting a DMCA takedown could backfire spectacularly if others do that to them).

I'd hazard anthropic perceives their number 1 enemy as open weight models. If alternatives to their business exist (which is mainly coding tokens currently), they will get into a nasty fight and the nightmare of all tech companies - losing their monopoly. It threatens their ability to extract value, and could reduce their valuation to a tenth of what it was. They cannot make open weight models worse, so they're using lawfare to try and get them banned. And of course we previously know they attempted to block Chinese companies under the guise of national security by lobbying for restricting GPU sales.

Re: Anthropic says Alibaba illicitly extracted Claude AI model capabilities

#963

Post by InkCanon »

salviati wrote: post #578

If you have openrouter do this little experiment: Go to https://openrouter.ai/chat . Select a few models, but customize them to have an empty system prompt. Then ask: "你是什么模型?" ("What model are you?" in Mandarin). My result after trying only three times: Sonnet 4.6 says it's DeepSeek, while Opus 4.8 says it's Qwen. The second time around Sonnet said it was Anthropic Claude. Are Chinese companies currently complaining…

"You're trying to kidnap what I've rightfully stolen."

Re: Anthropic says Alibaba illicitly extracted Claude AI model capabilities

#964

Post by InkCanon »

rayiner wrote: post #841

This is why I don’t understand the concerns about “our AI overlords” monopolizing all the gains from AI. It doesn’t seem like there’s much of a moat around the models themselves. So the race is mainly about compute. But compute is subject to power law effects. I remember Intel building the first Teraflop computer (ASCI red) in 1996. It was the size of a house. By 2014 you had more compute and 50% more memory in an of…

The openness of AI is currently being held up only by Chinese companies (previously Meta, but they stopped). They're not saints, but there is not even a question in the open weight/HF community that the immense mass of Chinese talent, knowledge and resources are the only thing stopping a monopoly/duopoly from forming. In a very Cathedral vs Bazaar-esque way, China severely lacks compute, but are extremely ingenious in coming up with new optimizations, architectures, etc, which they all detail in their papers.

Re: Anthropic says Alibaba illicitly extracted Claude AI model capabilities

#965

Post by anileated »

icfly2 wrote: post #282
tristanj wrote: post #25

Here's what is happening: Chinese resellers are offering Claude tokens at 70-90% below official Anthropic API prices. They achieve this by reselling capacity from pooled Claude Max accounts, payments fraud, and also reselling the model output & reasoning chains to various Chinese labs. They are subsidizing model access in exchange for user logs and reasoning traces, which they then sell as training data, allowing the…

[flagged]

Why was this flagged

Re: Anthropic says Alibaba illicitly extracted Claude AI model capabilities

#967

Post by killerstorm »

0xbadcafebee wrote: post #936
killerstorm wrote: post #746

Earlier quoted context omitted.

I'm sorry, but you got the terminology exactly backwards. Training on the answer is called supervised fine-tuning. Just for the sake of clarity: 0. Full distillation uses logits of the teacher model - that's much more information than the text itself. This is a kind of distillation used inside labs, but one can't distill Claude this way as logits are not available via API. 1. Supervised fine-tuning on synthetic data…

I agree I left out option 0, but I think the other two were presented correctly? - Black box distillation uses direct answers to questions and conversation style. This is less useful as you still have to do supervised fine-tuning on the answers, as they may be wrong, and don't lead to greater insights (which reinforcement learning does) - RLIAF relies on preferences and values to judge answers. These don't need super…

Well, I mean you mixed up "fine-tuning" and "reinforcement learning" a bit when describing these options.

Regarding the value of these options, SFT communicates more information to the model being trained, but there's a risk of overfitting. So I'd guess they might use both - do a bit of SFT and then finish with RLAIF.

Re: Anthropic says Alibaba illicitly extracted Claude AI model capabilities

#968

Post by Lio »

eloisius wrote: post #354
Lio wrote: post #311

Earlier quoted context omitted.

I’m surprised that instead of cutting them off Anthropic doesn’t just switch them to a lower quality, cheaper to models. That would seem more effective than simply shutting down the accounts. Keep them paying for junk.

That sounds like it would actually be fraud.

That's an interesting point but I'd ask who would Anthropic be defrauding?

The Chinese resellers that are happily operating outside of the terms of service?

Or the reseller's customers, that are knowingly buying a potentially fake service?

Or Alibaba, also using resellers to break terms of service with the aim of attacking Anthropic's business?

INAL but I think Anthropic would be able to argue an "unclean hands defence" to prove they are not defrauding anyone:

https://en.wikipedia.org/wiki/Clean_hands_doctrine

In short, you can't knowingly break a contract and then complain that someone else is not treating you fairly with regard to that same contract.

Re: Anthropic says Alibaba illicitly extracted Claude AI model capabilities

#970

Post by mike_hearn »

areoform wrote: post #166
tristanj wrote: post #25

Here's what is happening: Chinese resellers are offering Claude tokens at 70-90% below official Anthropic API prices. They achieve this by reselling capacity from pooled Claude Max accounts, payments fraud, and also reselling the model output & reasoning chains to various Chinese labs. They are subsidizing model access in exchange for user logs and reasoning traces, which they then sell as training data, allowing the…

Identity verification won't work. Nothing will. They are paying (and will continue to pay) US citizens sitting at home to copy-paste / type prompts out if they have to. But eventually they won't have to. Once there are enough spam PRs on github / uploads of claude conversations, enough mythos output used in production etc.; it'll just be the same albeit delayed. Doesn't matter either way. I feel for Anthropic's team…

You actually can "distill" web search. At one point Google accused Microsoft of doing this by monitoring clicks in IE to train Bing with Google click logs for better ranking. They discovered it by creating fake result pages with nonsensical queries and discovering the same results appeared in Bing a few weeks later.

https://searchengineland.com/google-bing-is-cheating-copying...

They also had a lot of evidence when I was there that Microsoft were cloning Google results. They monitored result query constantly and whenever Google launched a quality improvement, the quality of Bing results would go up by the same amount and always the exact same amount of time later.

Post reply on HN