Artificial Intelligence thread

[tphuang]

Distillation is simply impossible to stop. You have stop serving the product completely if you want to stop distillation.

If they ban all Chinese IP from using claude, they can just use VPNs to mask their access.

If they want to stop any particular account from making too many prompts, they can just mask it by creating thousands of fake accounts.

Distillation is basically fair use, which cannot be stopped with any technical means.

Any public facing AI system can be distilled no matter what measure is taken to stop it. AI by its very nature is "open source", anyone can learn from another AI as a teacher using its output.

As David Sacks explained in there, you can entirely slow it down significantly if you do KYC on your customers. Many of these accounts run through certain Asian countries where people there create account. If Anthropic itself can identify it got distilled then its AI is clearly good enough to figure out if a user could be originating from mainland China. Because we are told this mythos model is just so amazing. Well, if it's so amazing, then it can do additional monitoring of accounts originating from countries with a lot of these type of transiting accounts or certain credit card sources and then look at the requests it's getting and figure out which accounts should be blocked.

Qwen 3.8 Max Preview is terrible.

The inference level wasn't set properly on the first day, resulting in an extremely clumsy model. The second day was way way better, showing potential to challenge Kimi K3

But its performance has been gradually declining each day since, completely contrary to the advertised daily evolution. It's now showing signs of being worse than Qwen 3.7 Max.

What is Qwen labs doing?

they've never been on par with Kimi.

Is very difficult to quantify the use of Claude by just anecdotes because Ant. models are geoblocked in China so there is no real metrics of usage, And the situation is even more blurry in the agentic era because you can use Claude as the orchestrator for a project, DeepSeek Pro for the backend and DeepSeek Flash for the frontend. DeepSeek would generate more tokens even if Claude is the orchestrator.

I'm told from people in AI research in China that "everyone uses American AI".

You guys are late to the party. There was a security warning on Claude Code posted by MIIT on July 9th. All SOE IT departments have banned it already and private enterprises that contract with SOEs probably need to remove it as well. SOEs are also major clients of domestic IT service firms who do custom integrations, so they would naturally need to stop using Claude Code to keep contracts.

Also there is a perception lag. In the past month after the GLM-5.2 release and now especially after Kimi K3, you can feel a huge vibe shift in the CN developer community and dev influencers on using domestic LLMs for coding.

it probably is a perception lag, but Chinese government should have been more proactive in getting its big tech companies to stop Claude Code usage. There is just no excuse to waiting until July to issue a warning. That's ridiculous.