Policy researcher Lennart Heim says government regulations could delay this timeline.
Tick... tock... Government has the tools to delay this. Some rules are already drafted. Won't stop it forever, but you can buy time. And time is desperately needed.
I forecast that China will likely have its own Mythos-like model around February 2027, exactly a year after Mythos itself was ready for internal use in February 2026. With a 90% CI of October 2026-September 2027 🧵
Some users praise the solid analysis predicting China's Mythos-level AI model by 2027, while others doubt US efforts to stop Chinese labs and taunt that the rest of the world will benefit.
No Digg Deeper questions have been answered for this story yet.
Heh, I did well baiting Elon to give this prediction. Anyway, Hamish is well-calibrated on the estimate but 1) I doubt any Chinese player will commit to its first Mythos-scale job outside Mainland. The risk of meddling is high. 2) we don't know if they *can* train a multi-T LLM
“… or legally accessed remotely in Southeast Asia,”
It’s funny people talk so much about chip exports to China, and so instead what happens is they legally train models in Malaysian datacenters, and there’s virtually no talk about that anywhere.
Time *for what*? Anyway, I think Lennart is wrong that there's much time to be won on this front. They aren't going to train Mythos in Malaysia, precisely because they know about Lennart. All they'll get is their intermediate checkpoint copied and the original subtly corrupted.
Tick... tock... Government has the tools to delay this. Some rules are already drafted. Won't stop it forever, but you can buy time. And time is desperately needed.
Meta, OpenAI, xAI and Baidu all are known to have trained a >2T model (Behemoth, GPT 4.5, Grok 3/4, ERNIE 5). All have been flawed and eventually got replaced by smaller AND stronger ones. It's not clear to me anyone in China (or outside GDM/Ant) currently knows how to do this.
in fairness, Tencent or whatever may try, they're behind, nothing ventured nothing gained. But I don't expect them to reach the frontier first regardless of compute.
Time *for what*? Anyway, I think Lennart is wrong that there's much time to be won on this front. They aren't going to train Mythos in Malaysia, precisely because they know about Lennart. All they'll get is their intermediate checkpoint copied and the original subtly corrupted.
@teortaxesTex Are we sure GDM has trained a 2T model well? Leaks seem to imply Gemini 3 pro had the same base as 2.5 pro, which also explains the knowledge cutoff
@teortaxesTex super solid work.clean dev work.
@teortaxesTex 核心技术壁垒没那么好跨越的
@ohlennart Rules = MATCH act?
@ohlennart I very much doubt that US can stop Chinese labs. But you can of course stop yourself from enjoying what the rest of the world will have