Zephyr notes ByteDance's DeepMind hires skew toward computer vision.
> It’s infact trivial for Chinese to produce a model like that it's not, but hilariously enough not because of compute limitations. There are clusters in China that could, theoretically, train Mythos/Fable in a few months. Is there capital for that? What about years of data work?
Bytedance could probably do it But they hired too many people from GDM, so they are great at vision but suck at agentic stuff and coding
> It’s infact trivial for Chinese to produce a model like that it's not, but hilariously enough not because of compute limitations. There are clusters in China that could, theoretically, train Mythos/Fable in a few months. Is there capital for that? What about years of data work?
@zephyr_z9 spiritual pollution lol should have kept them away from coding I guess
Bytedance could probably do it But they hired too many people from GDM, so they are great at vision but suck at agentic stuff and coding
@zephyr_z9 Regarding code, we can patiently wait for deepseek (4.1 is coming out this month, focusing primarily on multimodal + code). ByteDance is responsible for the vision model; it is a different division of labor.
Zephyr notes ByteDance's DeepMind hires skew toward computer vision.
> It’s infact trivial for Chinese to produce a model like that it's not, but hilariously enough not because of compute limitations. There are clusters in China that could, theoretically, train Mythos/Fable in a few months. Is there capital for that? What about years of data work?