https://x.com/i/article/2117553026684653569

【X 长文】China AI Export Rules Explained

In October, Moonshot AI released Kimi K3.1 Pro. Along with making the model stronger, the team did something that goes against the usual race for better performance: they made some of its skills weaker on purpose.

According to the company's internal tests, without this step, the model could pass 90.5% of tests on using software security flaws. After the team removed that data from training, the final score dropped to 81.0%. By doing this, Moonshot hopes to lower the risk of people misusing the model once its weights are open.

Meanwhile, Chinese regulators are also changing the rules for sharing models abroad. In October, Beijing released a draft update to China's technology export list and started requiring companies to register these technologies before sending them abroad.

Companies control what a model can do, while the government manages which technologies leave the country and how. Both face the same question: as open models get more useful, and more likely to be used for attacks, how can they stay open?

So, What Did Kimi Do?

Moonshot's research team said the model got better at finding security flaws for two reasons. First, its coding skills improved. Second, its training data and reinforcement learning tasks included content on cyberattacks and defenses. So before opening the weights, the team removed that content from both parts of training, while keeping what the model needed for normal coding.

The score drop mentioned earlier came from an internal test that gave the model three tries at each task. It shows the cut worked, but it does not mean real-world attack risks also fell by 9.5 percentage points. Even after this cut, K3.1 Pro is still much stronger at these tasks than the last version.

This is the tough problem Moonshot faces: as a model gets better at coding, reasoning, and long debugging jobs, its cyberattack skills can grow right along with them. Deleting some specific data does not make those general skills go away.

In July, the UK's AI Safety Institute (AISI) and the US Center for AI Standards and Innovation (CAISI) teamed up to test Moonshot's last model, Kimi K3. They found that its safety rules failed to stop it from carrying out attack tasks. Even so, it was still clearly behind top US models. The report also noted that for closed US models, researchers turned off system-level safety guards to see the most they could do. So those test scores don't show how the public versions actually work.

So wouldn't it be simpler to just keep the model closed for safety?

Moonshot clearly doesn't want to do that. Cyber skills can also help companies find and fix software flaws, especially in older systems that have gone years without proper security checks. Open weights let more companies run the model on their own servers, without relying completely on a few closed-model companies.

That choice makes sense, but it brings up new questions: How much do the skills used by attackers and defenders overlap? After this cut, how much of its ability to run normal security checks is lost? A lower score on an attack test isn't enough to answer these questions.

Users already have mixed feelings about this move. Some complain that Moonshot is handing out a 'watered-down' model. They worry the company could use safety as an excuse to hold back its open-source version and sell an expensive closed version instead. Others worry a weakened model won't be strong enough to test their own servers.

Beijing Draws a New Line

Kimi K3.1 Pro came out just one week before Beijing stepped in to manage large model exports. On October 16, the Ministry of Commerce released a draft plan to update its list of restricted export technologies, adding large AI models as a possible target.

This timing points to an unspoken understanding between Chinese companies and regulators. Officials likely gave companies an early heads-up before the rules came out, and companies adjusted right away to stay in line.

Under the draft plan, companies have to register first while officials keep discussing which technologies to limit. Public feedback runs from October 16 to November 15, giving both sides a month to adjust. However, temporary registration started the day the notice came out and will stay in place until the final list becomes official. Also, registering now is not the same as getting an export license later.

Even for technologies that may face limits, regulators are leaving a door open. The Ministry of Commerce said open-source projects will get fast-track approval, aiming to approve requests right after companies send them in. The rules also won't apply to technologies that are already public through open-source releases or research papers.

Many see this relaxed 'quick registration' policy as a sign of support for Chinese open-source AI companies, though the exact timing and handling in practice remain unclear.

There is also a difference between...

Continue Reading