

I mean the paper and code are published. This isn’t a heuristic, so there’s no loss of accuracy. I’m not sure why you’re saying this is too good to be true, the whole tech is very new and there are lots of low hanging fruit for optimizations that people are discovering. Every few months some discovery like this is made right now. Eventually, people will pluck all the easy wins and it’s going to get harder to dramatically improve performance, but for the foreseeable future we’ll be seeing a lot more stuff like this.

















It’s a completely different situation in China. This tech is being treated as open source commodity similar to Linux, and companies aren’t trying to monetize it directly. There’s no crazy investment bonanza happening in China either. Companies like DeepSeek are developing this tech on fairly modest budgets, and they’re already starting to make money https://www.cnbc.com/2025/07/30/cnbcs-the-china-connection-newsletter-chinese-ai-companies-make-money.html