"That's hilarious!" was my first reaction as well, when I heard about it the first time. When I came to HN and saw this story on top I was hoping this was the top comment. I was not disappointed.
US AI folk were leading for two years by just throwing more and more compute at the same thing that Google threw them like a bone years ago (namely transformers). They made next to no innovation in any area other than how to connect more compute together. The idea of additional inference time compute, looping the network back on its own outputs, which is the only significant conceptual advancement of last years was something I, as a layman, came up with after few days of thinking why AI sucks and what can be done to make it able to tackle problems that require iterative reasoning. They announced it few weeks after I came up with the idea, so it was in the works for some time, but it shows you how basic idea it was. There was nothing else.
Suddenly when there comes a small company that introduced few actual algorithmic advancements which resulted in 100x optimization which is something expected with algorithmic optimizations, the big AI suddenly went into full "dog ate my homework" mode. Blaming everyone and everything around.
Let's not mention the fact that if full outputs of their models could enable them to train a better model at 1% cost then it puts them in even worse light that they didn't do it.
It’s not often you get 100x optimization with some small improvements so I’m kind of skeptical.
We have and apples and oranges thing here which deepseek is intentionally leaning into. They get very cheap electricity and are bragging about their cheap cost, and OpenAI etc typically brag about how expensive their training is. But it’s all pr and lies.
> They get very cheap electricity and are bragging about their cheap cost
The cost of $5.5 million was quoted at $2/GPU-hour which is a reasonable price for on-demand H100s that anyone in the US could access, and likely on the high side given bulk pricing and that they are using nerfed versions. OpenAI might be all pr and lies but everything I've seen so far says that deepseek's claims about cost are legit.
US AI folk were leading for two years by just throwing more and more compute at the same thing that Google threw them like a bone years ago (namely transformers). They made next to no innovation in any area other than how to connect more compute together. The idea of additional inference time compute, looping the network back on its own outputs, which is the only significant conceptual advancement of last years was something I, as a layman, came up with after few days of thinking why AI sucks and what can be done to make it able to tackle problems that require iterative reasoning. They announced it few weeks after I came up with the idea, so it was in the works for some time, but it shows you how basic idea it was. There was nothing else.
Suddenly when there comes a small company that introduced few actual algorithmic advancements which resulted in 100x optimization which is something expected with algorithmic optimizations, the big AI suddenly went into full "dog ate my homework" mode. Blaming everyone and everything around.
Let's not mention the fact that if full outputs of their models could enable them to train a better model at 1% cost then it puts them in even worse light that they didn't do it.