Anthropic data reveals Chinese AI firms distilled its models, with Zhipu showing characteristics of unreleased Claude 4.5

"punches above its weight" here is a vague metric that seeks to capture performance divided by (available human resources, capital, compute and focus on products). Minimax is the least impressive in this bunch, but Alibaba has vastly more resources. If they need so much Claude…

@teortaxesTex funny you say this because you can smell Opus 4.5/4.6 the most in Zhipu models

@teortaxesTex I would like to read the letter. Senators Warren and Scott got copies as well as the White House, no one released it. I thought there would be an Anthropic blog post that same afternoon, but nothing. Kind of strange.

@scaling01 Then what does Anthropic even publish the numbers for?

@teortaxesTex That's not at all my argument.

I'm saying the numbers we see are not correct and it's much more for all of them, and that all of them are buying Claude data.

Can Dario share how much Zhipu has distilled? Why the trickle of information? They're visibly the closest right now, so by Lisan's logic should be the biggest, right? Does it take time to retroactively find the last strongest Chinese lab's traces?

@teortaxesTex If it was so easy, we wouldn't have meta and xai flailing about hopelessly.

And lets' not for a minute pretend that the big labs haven't distilled their competitors at some point.

@teortaxesTex bro doesnt understand how distillation works

@teortaxesTex Anthropic just want to push for restrictions, the actual numbers who distills how much doesn't matter to them

it's just very hard to prove which tokens are used for "distillation"

@teortaxesTex Alibaba probably getting blamed because they own datacenters where people use VPNs to run Claude. Similar to Singapore's AI adoption really being Chinese VPN hosting.

@teortaxesTex Lisan reminds me of that account with cat pfp, ultimate bearish with Chinese things.

@AndrewCurran_ @teortaxesTex anyway this thread could be translated into plain English for us non-techical types?

@teortaxesTex what do they mean by "illicit access"

for all we know this could be alibaba devs using claude code?

@teortaxesTex Dario seems completely unable to understand that more interactions mean a stronger feedback loop and better model improvement. By setting those prices, he has driven customers away from his own product and instead of admitting that the problem is customer loss caused by pricing..

@teortaxesTex ..а DeepSeek 150 тысяч раз отказывал Claude в доступе..

@teortaxesTex true🤣

@teortaxesTex Also if we're talking about unfair advantages government protection via compute restrictions >>>> doing SFT on API outputs.

@teortaxesTex 阿里确实在大模型这块显得有点尴尬

@scaling01 @teortaxesTex I'd disagree

The inverse correlation is the important finding here.

If distillation volume were the whole explanation, the companies copying most would rank highest. They don't. That means something else is doing the work — post-training, data curation, method — and distillation is a tool, not the source of the advantage.

"They're only good because of distillation" and "less distillation correlates with better relative performance" can't both be the full story. Teortaxes' data is the more precise read.