DeepSeek: infra, pretraining Kimi: data, persona zAI: post-training Xiaomi: integrations StepFun: products MiniMax: marketing That would be the lab to end all labs… AI-2027 stuff
@teortaxesTex If only we have DS handling the infra and architecture; zai handle the post training and data, we might really get there
Some users dismissed claims about Chinese AI labs forming a Minimax dream team for a 2027 model as pure marketing hype.
No Digg Deeper questions have been answered for this story yet.
Ofc DeepSeek can do it all on their own, and so can most others. Their data is arguably better than Kimi or GLM's. But I'd still allocate more of Kimi's attention to data (and basic research). They have the best taste.
DeepSeek: infra, pretraining Kimi: data, persona zAI: post-training Xiaomi: integrations StepFun: products MiniMax: marketing That would be the lab to end all labs… AI-2027 stuff
@teortaxesTex minimax: marketing 💀💀💀
@teortaxesTex feel like a paradigm shift is coming that replaces resource scaling to generate least wrong answer, with maximize info extraction (ie. surprises) all else equal. in that regard, data will no longer be viewed as an input but as an output, to be measured in terms of expressitivity.