15/. We’re 2+ years into availability of GenAI. Maybe six months into broad deployment in industry. Surprises will happen. It just so happens that with AI, they matter immensely for the U.S., China, and the world. END.

14/ So as a few stocks wipe a couple trillion (!) off of the markets, it’s a good time for investors and analysts to take a deep breath and remember: we’re at the top of the first inning, at best, with this tech. www.nature.com/articles/d41...

13/ But we’ll see; new hardware matters; selection of performance benchmarks matter; usage trends for the AI-implementing industry matter; all of that. We’ll know more in a month.

12/ Second, to me, @DeepSeek S1 does not provide much more evidence of the commoditization thesis than @Meta’s LLaMA does — *if you think open source will beat out commercial AI, THAT’S the macro thesis.* Not that a cheaper training is possible.

11/ On AI/Chip markets — quickly as they open: first, the last Admin’s chip restrictions *may* not have conferred some desired "3yr advantage"; or even a 6mo one…but none at all at a market level. (i.e. still slowing the PLA; not AI companies.) www.wired.com/story/new-us...

10/ (Watch this space; energy is the medium-term pinch-point of training cost and AI availability; this is the subject of the @brookings.edu piece.)

9/ And if DeepSeek R1 has done that for a SOTA LLM, it’s a big deal. But that innovation is unlikely to stay proprietary long, and hardly un-replicable. U.S. companies would LOVE to have a SOTA model they can train for $5M! They’ll spend money to get it!

8/ Efficiency innovations aren’t theirs alone. Remember when @MIT grad student @jefrankle cracked the code on the Lottery Ticket Hypothesis? AI is immature science; you can have a breakthrough and boom, drop training costs hugely without gutting performance. arxiv.org/abs/1803.03635

7/ Several of us have been making the point that we haven’t seen U.S. AI development under massive constraint yet; that companies haven’t had e.g. chips as a principal limiting reagent to their innovations. The Chinese have. www.nature.com/articles/d41...

6/ On INPUTS: A @brookings.edu piece coming out in a few days will have my full thoughts, but until then…worth noting that Chinese companies have been trying to do more with less; fewer chips, maybe less capital.

5/ So two key takeaways from DeepSeek R1. First, a tech reality check: it’s early days, and different inputs — money, energy, social constraints — yield different outputs. Second, a market reality check: the chances that GenAI (even AGI?) gets commoditized is non-zero.

4/ IF it turns S1 is both state-of-the-art-ish in performance AND tightly censored, that’s an important data point in the US/China AI race, and will shake some priors on both sides. See e.g. www.nytimes.com/2025/01/23/t...

3/ Like all Chinese models available to the public, it’s a censorship machine! Some say that is irrelevant for the underlying tech; AI companies tell me otherwise, that nerfing a model to comply with the party has real performance impacts. alecmuffett.com/article/110996

2/ Some experts are privately noting (in their words) like many Chinese models on the rankings, that it “teaches to the test” — and that we should look askance at claims about training costs. Maybe. BUT…

1/ The DeepSeek Model isn’t new overnight; it’s been the “talk of Davos” late last week — but it seems the broader Internet (and investing-net) has woken up that there’s something interesting here. techcrunch.com/2025/01/20/d...