Databricks' Yuchen Jin predicts a Chinese AI model will beat GPT-5.6 within three months, exposing the futility of regulatory bans

@Yuchenj_UW You think Chinese AI with Fable level intelligence won’t be able to hack into computer networks?

@Yuchenj_UW America is led corrupt leaders. Sometimes, I am glad China rose to help the world out.

@Yuchenj_UW the ban speedrun is gonna look so stupid in hindsight

@Yuchenj_UW Totally doable. I still wonder why Fable 5 (Mythos) was so good at cybersecurity and hacking. Was it specifically trained with every available CVE and during RL it got awarded for being able to reverse engineer them with proof of concept code?

@Yuchenj_UW China will not release a model that can hack China.

Unless they believe their defense against cyber is better than America's.

Then it's game on

@Yuchenj_UW it makes you wonder what anthropic is running internally now, these models were finished training at the beginning of 2026

@ryan_landay @Yuchenj_UW Yes, that is probably the reason why. Model being a better coder, also means better understanding of bugs. But I imagine the task of exploiting bugs in code must be part of some separate RL process.

Zhipu allegedly catching up fairly quickly when their models are otherwise several months behind suggests that OpenAI and Anthropic were not aggressively optimizing this use case, but Zhipu now is. Obviously this trend is only going to get stronger if OpenAI and Anthropic can’t legally ship models that are good at cybersecurity. https://www.wsj.com/tech/ai/chinese-ai-anthropic-mythos-cybersecurity-574b02c2

@lapplandtrader @Yuchenj_UW These tasks are very closely related to coding which is a major focus of these companies.

@Yuchenj_UW I'm thankful that this crated the opening needed for people to go try things like GLM.

@Yuchenj_UW Watch us all eat crow by summer’s end

@Yuchenj_UW I'm afraid that they are going to try block providers inferences then. Almost in the country that they can. What do you think?

@Yuchenj_UW I hope you’re right!

@Yuchenj_UW honestly that sounds exactly like how this will play out

@Yuchenj_UW This situation is already funny, Dario fear mongering and blows up in his face epically

@Yuchenj_UW This, but if you talk to the life science folks who want to use frontier models , they have and continue to be living under similar hysteria from the US AI labs

@Yuchenj_UW Export controls on frontier weights buy time only if the domestic labs cannot close the gap through other routes. The ban narrative assumed capability was the scarce resource. It is not.

@Yuchenj_UW Eagerly looking forward to the next 3 months 🚀

@Yuchenj_UW It will happen. Many times.

Another 'GPT2 is too dangerous' moment.

@Yuchenj_UW Hope your oracle comes through!