🖙Today in AI Startup Brain
Deepseek's release has thrown big tech into a tizzy, as they realize they have even less of a competitve advantage (a "moat" in business-speak) than they thought, producing some excellent sources of schadenfreude in the process, such as "OpenAI Furious DeepSeek Might Have Stolen All the Data OpenAI Stole From Us".
Dario Amodei, the CEO of Anthropic, wrote a kinda weird post on his personal website. He argues one, DeepSeek is actually behind where you'd expect a model to be today and not that much cheaper than, just picking one out of his hat, Anthropic's models, but, two, there's an "existential risk" from China having access to GPUs, necessitating export controls.
I'm not going to dwell too much on whether American industry influencing the American military and vice-versa has been a good idea historically or whether a tech company CEO is a good source of information on geopolitics, but from those angles, this post is also dubious.
Now, it's possible that he's surveyed the market and contemplated the politics of this all, and thus genuinely believes an authoritarian country having access to AI or somehow controlling the market for it would be really bad (although I have some bad news for him regarding the current direction of the United States), apparently because of the military advantage it would give them1. Still, part of me believes a different story.
Wouldn't it be convenient for the head of a AI company if a new competitor's new product was both overhyped and there to be some sort of military threat to them continuing to access to the necessary chips? Wouldn't that narrative play well in Washington, given an administration obsessed with trade, Washington's pre-existing worries about China, and the perpetual willingness to do things in the name of national security? And wouldn't downplaying the magnitude simultaneously help preserve his own company's image?
He also argues that "Making AI that is smarter than almost all humans at almost all things will require millions of chips, tens of billions of dollars (at least), and is most likely to happen in 2026-2027." To me, this is pretty deranged. I don't know that progress has slowed down entirely, but I'm not sure we've seen any GPT-3 to GPT-4 leaps. I guess we'll find out!
Another funny thing about this is that DeepSeek is more open than Anthropic, which comes up in this post because Amodei doesn't disclose how much it actually cost to train Anthropic's models, so when he talks about the relative cost of Anthropic's to DeepSeek's, he has to kind of hand-wave it. This isn't necessarily hypocritical, although it seems to slightly undermine his point about military advantage. Obviously DeepSeek or other Chinese labs could close up their models at any time—a move pioneered by OpenAI—but currently U.S. researchers (and the U.S. military) have basically the same access as Chinese researchers (and the Chinese military). Arguably, the U.S. has better access because we have more access to the necessary hardware, as far as I can tell. We certainly have the capital.
He does at least say he wants China to benefit in non-military areas:
To be clear, the goal here is not to deny China or any other authoritarian country the immense benefits in science, medicine, quality of life, etc that come from very powerful AI systems. Everyone should be able to benefit from AI. The goal is to prevent them from gaining military dominance.
This is nice, I guess, although it seems like export controls would also undermine this goal of letting those countries access the "immense benefits" he foresees. Maybe his ideal is that Chinese citizens and businesses have minimal access to models directly and only access them through the APIs of, picking a name at random, Anthropic.
This is all a bit weird for me to analyze because, frankly, I'm not sure there are any great AI benchmarks, and I'm dubious of the value of AI generally, so talk of which models are "better" feels a bit woolly to me. The fact that they are built on grand scale plagiarism and are resource-intensive doesn't help.
The IDF has used an AI system called, ghoulishly, the Gospel in order to select bombing targets. However, this appears to be based on machine learning, not LLMs or other generative AI.↩