It’s been almost a full year since DeepSeek sent the AI industry and tech stocks into a tizzy with the release of its low-cost, high-performance R1 model. Now the Chinese tech company says it has a new trick up its sleeve.
What happened: DeepSeek published a research paper describing a new way to train high-performance AI models more cost effectively than traditional methods, a technique one AI researcher described as a “striking breakthrough.”
We won’t pretend to grasp the technical details of the new training method, dubbed Manifold-Constrained Hyper-Connections (mHC), but the gist is that it allows AI developers to create better AI systems without spending much more money.
Why it matters: Another “DeepSeek moment” that shocks markets may be brewing.
The mHC method was likely used to train DeepSeek’s R2 model, which is expected to be released in the coming months.
This paper suggests that DeepSeek may have found a way to create a model that matches or exceeds the performance of state-of-the-art systems at a much lower cost. If so, the company’s next model could trigger another panic among tech investors.
Yes, but: Tech stocks recovered (and then some) after the last DeepSeek moment as investors bought into the argument that cheaper AI would expand the demand for AI rather than lowering its cost. If that narrative holds, markets may not be rattled again by DeepSeek’s innovations.