Christian Catalini

This argument by @deanwball is being badly misunderstood. It's OK to disagree with it, but first you have to actually understand what he's saying. He's saying: releasing the weights for a frontier-level model is effectively dumping. Dumping is when you sell a product at significantly below cost in order to corner market share. It's illegal. The reason: dumping results in short-term consumer surplus, but long-term it prevents the formation of a competitive market and discourages capex outside of the dumper. Standard Oil famously did this in order to consolidate the oil market before it was broken up. So why is he claiming releasing the weights of a frontier level model is basically dumping? Isn't he just describing open source? His argument: it's not financially sustainable to train a frontier model and release the weights. In the long run, you will not be able to internalize enough of the gains given the cost of training a frontier model, because neoclouds and other inference providers will be able to outcompete you at actually serving the model. It costs an astronomical amount of money to train frontier models, and if everyone else can serve them, you don't capture enough of the surplus to pay for the training and R&D. It's not like normal open source when you build some software and then release it and sell services on top of it. The amount of capex required for frontier-level models is an order of magnitude higher than normal software, which is why doing this at frontier level is so economically irrational. Right now the Hong Kong stock market is ebullient enough that Chinese AI companies are not getting punished for the fact that they're all deeply, deeply unprofitable. Releasing model weights is great marketing, intellectually appealing, and strikes fear into the hearts of their opponents. We can assume the status quo continues for a while because of the AI supercycle. But eventually the AI market will correct, the Hong Kong market will dump, and suddenly these Chinese labs won't be able to afford to training super expensive models without internalizing more of the gains. But what if China, seeing that this strategy is successfully kneecapping the US lead (by discouraging further capex and lowering valuations), says no--don't stop. And so the Chinese government starts buying up the shares of these companies and demanding that they continue releasing frontier-level weights, profitable or not. In that case, it becomes a genuine space race. For-profit companies cannot continue to compete on either side. US labs valuations fall, and the White House realizes that to keep their advantage in the AI race, they cannot rely on the free market to maintain their lead. They nationalize the labs and fund them off government subsidies. Now you have government-controlled and distributed models on both sides. That's what Dean is calling the "dystopian hellscape." The best analogy is drug development: if China were to sell American drugs back to us really cheaply, that would result in a large short-term consumer surplus. Cheap Viagra and Ozempic is obviously great. But in the long run, this would discourage investment in developing new drugs. That's the sense that Dean is saying it's long-term "decel." Now, I happen to disagree with Dean. I think the consumer surplus of having frontier-level open weight models is huge, even at the current capabilities. I also think China is going to defect from this strategy soon (there's been reporting along these lines, that Beijing will stop allowing large models to be open-weight; I think there are other reasons for this aside from competition). I also suspect that nationalization of labs is inevitable as they take on more geopolitical and cyber capabilities. But he's not wrong--releasing frontier-level weight models is weird. The question of how long this market will remain profit-driven is a very coherent question to ask.

@hosseeb @deanwball Soooo, releasing Linux was dumping? Apache, MySQL, PHP? HTTP, TCP/IP, OpenSSL, OpenSSH? Libjpeg, VLC? The open source software stack of the mobile communication network? Signal? PyTorch? Llama?

NEW @GwartyGwart w/ @ccatalini: Open source models & AI value accrual "If intelligence is really becoming cheap, commodified... then there's going to be a new bottleneck. And I think people in crypto should be excited about this because that bottleneck turns out to be verification." Brought to you by @ellipsis_labs

1/ Nadella’s Test: What’s Left When The AI Model Is Pulled? @satyanadella defined what decides whether your company and job stay defensible as AI improves. The economics says it holds on a single condition. One his post left out. 🧵x.com/satyanadella/s…

Palantir CTO @ssankar says Chinese open source models pose a "threat to American Prosperity": "These Chinese open source models are really the result of distillation attacks." "Most of this is just stolen American IP from frontier labs, and I think the frontier labs should do more to protect that IP, and that's in their own economic interest." "The gravest threat to us is not homicide, it's suicide." "It's an inability to turn these tokens into economic value and generate American prosperity for it." "You see that with the proposed moratorium in New York on data centers. Turning our back on AI would be as consequential a mistake as turning our back on the atom in the '70s." "It's just the beginning, but we're starting to see a burgeoning US open model ecosystem forming." "NVIDIA's Nemotron models are very good." "I think you're going to see in the next month a slate of announcements of other American companies, non-frontier labs, neo labs, who are putting out open weight models that companies are going to be able to cheaply fine-tune and capture their own alpha in weights they control." Via @business