DeepSeek
Article
DeepSeek is a recurring organization in the Astral Codex Ten archive, appearing 4 times across 4 issues between February 21, 2025 and November 26, 2025. The archive places it in contexts such as “I’ve recently been asking DeepSeek about René Descartes and Gottfried Leibniz”; “My list of links to publish today includes something like a dozen about DeepSeek”; “DeepSeek seem closer behind OpenAI”. It most often appears alongside China, OpenAI, America.
Metadata
- Category: Organizations
- Mention count: 4
- Issue count: 4
- First seen: February 21, 2025
- Last seen: November 26, 2025
Appears In
- Highlights From The Comments On Tegmark’s Mathematical Universe
- Links For February 2025
- My Takeaways From AI 2027
- Why AI Safety Won’t Make America Lose The Race With China
Related Pages
-
- China (3 shared issues)
-
- OpenAI (3 shared issues)
-
- America (2 shared issues)
-
- Anthropic (2 shared issues)
-
- Bulldog (2 shared issues)
-
- Cold War (2 shared issues)
-
- Dean Ball (2 shared issues)
-
- Russia (2 shared issues)
-
- United States (2 shared issues)
-
- NootropicsDepot (1 shared issues)
-
- slatestarcodex (1 shared issues)
-
- 1960s sci-fi (1 shared issues)
External Links
Source Context
Recovered passages from the original issue text. When the raw archive preserved outbound links inside the source passage, they are listed directly under the quote.
2010? I’ve recently been asking DeepSeek about René Descartes and Gottfried Leibniz. Someone could have said most of that in 1710…Why is there something rather than nothing? Is straight out. Of Leibnitz’s Principles of Nature and Grace, which we can now read as being about Artificial Intelligence.
44: My list of links to publish today includes something like a dozen about DeepSeek, which now seems so thoroughly yesterday’s news that I’m tempted to throw them all out. But in case you still have questions about it, I felt most enlightened by takes from Dean Ball (X), Helen Toner (X), and Miles Brundage (X). The story seems to be that DeepSeek genuinely did a great job, made extensive algorithmic progress, and was able to create an excellent AI on chips scrounged up from before the export controls hit + mediocre chips that got through the export controls. Along with these real reasons to be impressed, there is also a little bit of illusion at work - OpenAI delayed announcing o1 for a long time (remember the rumors about “Q*” and “Strawberry”?) and DeepSeek was very fast to announce r1, which made DeepSeek seem closer behind OpenAI than they really were. Most of the smart people I read said that the absolute worst response to this (from an arms race point of view) would be to give up on export controls - if a rival has geniuses who can use resources ultra-effectively, you don’t want to also give them more resources!
It might be even worse than that; once AI becomes good at cyberwarfare, there will be increased pressure on companies like Meta and DeepSeek to stop releases until they’re sure they can’t be jailbroken to hack people. If that’s hard, it could slow open-source even further.
Models: The quality of foundation models - giant multi-purpose AIs like GPT or Claude - primarily depends on the amount of compute used to train them, so America’s compute advantage carries over to this level. In theory, clever training methods and advanced algorithms can make one model more or less compute-efficient than another, but this doesn’t seem to be affecting the current state of the race much - most advances by one country are quickly diffused to (or stolen by) the other. Despite some early concerns, neither DeepSeek nor Kimi K2 Chinese models provide strong evidence of a Chinese advantage in computational efficiency (1, 2).
In 2023, the US realized it was in an AI race with China and slashed chip exports. Chinese access to compute dropped dramatically. They began accelerating onshore chip development, but this will take a decade or more to pay off. For now, the Chinese AIs you’ve heard of - DeepSeek, Kimi, etc - are primarily trained on a combination of stockpiled American chips from before the export controls, and American chips smuggled in through third parties, especially Singapore and Malaysia.
Others argue that chip sanctions just encourage China to be smarter and more compute-efficient, and that we’ll regret training them into a scrappy battle-hardened colossus. I think this is insulting to American and Chinese researchers, who are already working maximally hard to discover efficiency improvements regardless of our relative compute standing. More important, it doesn’t seem to be true - Chinese AIs are no more compute-efficient than American models, with most claims to the contrary being failures of chip accounting. I’m not even sure the people making this argument believe their own claims. When I play devil’s advocate and ask them whether America should perhaps pass lots of AI safety regulations a hundred times stricter than the ones actually under consideration - since that would increase training costs, reduce the number of chips we can afford, and cripple us in the same way that chip sanctions cripple China - these people suddenly forget about their bad-things-are-good argument and go back to believing that bad things are bad again.
Inline links: no more compute-efficient than American models