Anthropic says Alibaba illicitly extracted Claude AI model capabilities

rsynnott 1 day ago|

Oh, _now_ we care about IP, do we?

delta_p_delta_x 1 day ago||

Cue Jeremy Clarkson's 'Oh no! Anyway...' GIF.

PostOnce 1 day 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? :)

irthomasthomas 1 day ago|

Ask claude it's name in chinese and it thinks its Qwen (opus) or Deepseek (sonnet). Anthropic are just as guilty as everyone else training AI, today, maybe more so. Every lab borrows from every other. It only takes a few hundred samples to figure out the pattern; look at glm-5.2 reasoning using the caveman tongue of gpt-5.5. Stopping this would require some draconian surveillance.

kgeist 1 day ago|

That's not how it works though. When you prepare the conversations for distillation, it's the most trivial and obvious first step to replace "Qwen" with "Claude" and vice versa. I doubt they'd simply forget to do it.

A model may misidentify itself due to the surrounding context. When a model is about to answer "I'm ...", what follows is a sorted list of probabilities for what the next token should be. In most models it's usually a list of popular model names: say, in the list, first comes Claude, then Qwen, then ChatGPT etc. Usually the "Claude" token would be the most probable token, say 70%. But if the surrounding context is in Chinese, the embeddings for "something to do with China" may nudge the combined embedding of the output token towards the "Qwen" embedding more ("China+Claude=Qwen" in the embedding space). Say, the probability for "Qwen" now becomes 60% instead of 10%.

If we also use high temperature for more "creativity", the token sampler now may choose "Qwen". It's not the most probable token still, but it was chosen because selecting the 2nd most probable token once in a while usually allows a model to explore unexpected "creative" paths, and 60% probability is good enough compared to 70%. It's basically a hallucination.

I once made an experiment: if I ban the word "Qwen" in the inference engine entirely, and ask Qwen "which model are you?", it happily starts announcing it's Claude 100% time, simply because "Claude" is the next most probable token after "Qwen" in this context.