China's Zhipu released GLM 5.2 and apparently it works as well as the American models that can't ship because someone might use them to write a mean email. Anthropic and OpenAI spent the last year adding guardrails. Zhipu spent it adding features.
The headline says this is about intelligence per dollar. It's not. It's about intelligence per regulatory approval form. Turns out when you don't need to explain to Congress why your chatbot won't help someone make a sandwich, you can focus on making the chatbot better at making sandwiches.
Open source is suddenly competitive. Not because the technology got better. Because the closed-source companies stopped improving their models and started improving their apology letters.
Retail traders are now googling "how to invest in Chinese AI" and discovering that the answer is "you can't because your government won't let you." They'll pivot to buying shares of whatever American company mentions AI in an earnings call. Probably a company that makes industrial lubricants but added a chatbot to their customer service page.
OpenAI has a model that could probably cure cancer but they're still deciding if releasing it might hurt someone's feelings. Anthropic has one that's even better but it requires you to read a constitutional charter before each prompt. Zhipu has one that works right now and costs less.
The AI race was supposed to be about who builds the smartest system. Now it's about who's allowed to ship it. Zhipu wins by default while Sam Altman testifies about strawberry-counting models and Dario Amodei writes another blog post about responsible scaling policies.
America is losing the AI war to a country that banned Google because we're too busy making sure our AI doesn't say anything that might offend the DSA.
Photo by on Unsplash
Leave a Comment