Really seems like Deepseek is one of the only vendors actually focusing on performance per unit compute power and not just throwing infinite compute power at the problem. Calling it now, when the bubble bursts they’ll be one of the few to make it out with a usable product.
For sure, they’ve probably dropped more significant papers in the past year than any other groups. It does seem like the mindset in China is very different overall though. In the states, it’s basically a cult at this point where they’re trying to build a god with AGI. In China, it’s just treated like another tool for automation and companies see it as common infrastructure, akin to Linux, that people will build interesting things on. Hence why pretty much all the models in China re developed on open basis. Everybody there seems to realize that there’s no real path towards monetizing the models themselves.
Simon may want to randomize his Pelican/Bicycle test.
There is a long tradition in tech of firms tweaking their outputs to get higher scores on well-known tests. The ultimate example is VW Dieselgate.
But in AI, it’s easy to game benchmarks, by adding the best answers to the training set for the next version.
Slop is slop
I get the AI hate when it comes to a lot of things, but it is genuinely a useful tool for software development.
Usually it’s people who don’t code or understand the complexities involved that go that way.
