Regardless of whether this specific claim is true, enterprises are becoming much more cautious about developer tools that can read large portions of proprietary codebases.
Wasn't one of the big promises the AI labs made "uncopyrighting"? Ie. the ability to reconstruct large works, including source code, without actual access to the source code? Everything from movies to operating systems.
Becoming? We've moved entirely in the opposite direction.
When these tools first appeared the overwhelming conversation was about the risk of letting a remote tool siphon your code and intellectual property (where eventually they're going to add that to their training). Now everyone is using them, and that fear seems to have dissolved. Every corporation is sprinkled with Claude Code, Antigravity, Copilot, Codex, and so on. Even the long fear-mongered Chinese providers are being heavily used in many spaces.
In this case this is a PR battle between two firms, and it isn't much more. And Alibaba isn't worried about the "proprietary code" (the truth is that there is incredibly little interest in most orgs code), but that the tool is a backdoor, or at least that is the claim.
1. LOL I've just downloaded literally whole internet and copyrighted books and put them through a neural network. Now I have this whole knowledge in my LLM.
2. Hey? Are you using my NN for training your NN? you're a thief!
i can see why they want to stop it but
1. you have to pay for the "attack"
2. these AI companies trained on copyrighted content without permission or attribution to anyone who's data was used to train.
when i was in hongkong, chatgpt and gemini were disabled. Maybe this has changed though. When I was in China, the corporate vpn (zscaler) routed traffic through hk
Can't say they are wrong, after the latest backdoor, or let's say, undocumented functionality that leaks some data that was pushed in Claude Code few days ago
Considering their massive distillation, if US companies stop publishing new models to the public, would China still be able to develop new open weight models?
I don't think China would strugle to scrape the internet for fresh data.
And they constantly publish state of the art LLM research (see DS4 context compaction and cache tech).
They have very capable tech giants. So while not being able to distill western models would probably have some impact, it's probably becoming lesser as time passes.
We might even see Western LLMs distilling Chinese models soon. If they aren't already to some extent.
More than a year ago, when Anthropic and OpenAI started to gide the reasoning bits from the output, a lot of people here on HN predicted that Chinese models days were numbered.
Fast forward to today, and models such as DeepSeek and MiMo are nothing short of excellent. I haven't used GLM or Qwen but heard very good things about them as well.
This "massive distillation" sounds a lot like anxiety about how companies from outside the US can develop very good models themselves.
China has most probably already achieved "escape velocity" on the software side. Now if they achieve parity, to some degree at least, on the hardware side with Nvidia it is very possible they'll overtake the US.
Regardless of whether this specific claim is true, enterprises are becoming much more cautious about developer tools that can read large portions of proprietary codebases.
Wasn't one of the big promises the AI labs made "uncopyrighting"? Ie. the ability to reconstruct large works, including source code, without actual access to the source code? Everything from movies to operating systems.
Becoming? We've moved entirely in the opposite direction.
When these tools first appeared the overwhelming conversation was about the risk of letting a remote tool siphon your code and intellectual property (where eventually they're going to add that to their training). Now everyone is using them, and that fear seems to have dissolved. Every corporation is sprinkled with Claude Code, Antigravity, Copilot, Codex, and so on. Even the long fear-mongered Chinese providers are being heavily used in many spaces.
In this case this is a PR battle between two firms, and it isn't much more. And Alibaba isn't worried about the "proprietary code" (the truth is that there is incredibly little interest in most orgs code), but that the tool is a backdoor, or at least that is the claim.
What Claude Code did is absolutely mindboggling tho, if Chinese harness did that probably POTUS would lose sleep.
It didn't seem like much info compared to what's collected by modern websites and apps, though?
Maybe if they didn’t farm all the data from Claude to train their own trash models. Anthropic wouldn’t feel the need to do it.
Anthropic stole the entire internet. Excuse my language, but they can fuck right off.
And I'm the king of France
Wait what do you mean "if"?
Exaggerate much? If you think POTUS would lose sleep about a date format timezone marker, I don't know what to tell you.
Translation: Alibaba will continue distillation attacks using accounts that aren't directly attributable to it's own corporate infrastructure.
(Mis)anthropic already performed "distillation attacks" on the internet.
Did Anthropic perform "distillation attacks" when they hoovered up the entire internet?
i can see why they want to stop it but 1. you have to pay for the "attack" 2. these AI companies trained on copyrighted content without permission or attribution to anyone who's data was used to train.
As long as they're paying for the tokens, there's no attack . Otherwise you have to call training on copyrighted material theft.
They are not paying for most tokens. The actual users in China do. All they need is the logs.
How exactly the word attack fits in that phrase?
i gonna ask: how can they still use claude? i thought all users in china are banned
Alibaba has engineers in Hongkong, Singapore, North America. It’s a global corporation
when i was in hongkong, chatgpt and gemini were disabled. Maybe this has changed though. When I was in China, the corporate vpn (zscaler) routed traffic through hk
https://www.chinatalk.media/p/how-to-buy-cheap-claude-tokens...
Same way every ban is evaded, smurfing
There is a reason Singapore tops the rank on Claude usage
the government also actively promotes AI usage in work environments
Does Alibaba only have developers in the China?
Did china invent VPNs yet?
Cc can be used with non Anthropic models.
The same way they buy "banned" and "sanctioned" NVIDIA GPUs.
> how can they still use claude?
Workarounds aside, it says Claude Code not Claude.
i.e. they are using the CLI running any model. You can for instance run GLM with it.
Can't say they are wrong, after the latest backdoor, or let's say, undocumented functionality that leaks some data that was pushed in Claude Code few days ago
https://news.ycombinator.com/item?id=48759754
That’s not what a backdoor is…
Rear entrance then
When a company can remotely push code without explicit user approval, and code that was hostile / almost malicious, it is a backdoor
Another reason to use open source coding agents and local language models.
Claude Code is neither and it is literally info stealing malware.
Considering their massive distillation, if US companies stop publishing new models to the public, would China still be able to develop new open weight models?
I don't think China would strugle to scrape the internet for fresh data.
And they constantly publish state of the art LLM research (see DS4 context compaction and cache tech).
They have very capable tech giants. So while not being able to distill western models would probably have some impact, it's probably becoming lesser as time passes.
We might even see Western LLMs distilling Chinese models soon. If they aren't already to some extent.
Probably yes.
More than a year ago, when Anthropic and OpenAI started to gide the reasoning bits from the output, a lot of people here on HN predicted that Chinese models days were numbered.
Fast forward to today, and models such as DeepSeek and MiMo are nothing short of excellent. I haven't used GLM or Qwen but heard very good things about them as well.
This "massive distillation" sounds a lot like anxiety about how companies from outside the US can develop very good models themselves.
China has most probably already achieved "escape velocity" on the software side. Now if they achieve parity, to some degree at least, on the hardware side with Nvidia it is very possible they'll overtake the US.
Yes, 100%. GLM 5.2 is capable of RSI. It's too late to stop.