Korva is a scroll of the internet’s best ideas.
Training favors whoever leads
Three effects stack
Three effects stack. A skill or body of knowledge trained into a model is amortized across billions of users. A lab slightly ahead in a compute shortage can charge a higher markup, because it uses the scarce resource better. And models that learn from deployment gain real-world data in proportion to how widely they are used. Recursive self-improvement adds more, since the best model helps build the next one.
Keeping tokens inside
Labs also gain by keeping tokens inside. If a customer earns $200 million per megawatt and pays Anthropic $100 million for it, a lab that can generate hundreds of millions per megawatt using that compute internally has little reason to sell. Decentralized decision-making is why capitalism worked and why highly centralized economies grew slower. AI flips that, because a centralized AI economy grows faster than a decentralized one.
Who ends up involved
The part of the US economy involved is around 2%, and a large share of it sits with Nvidia, Anthropic, OpenAI, and the hyperscalers. Property can stay private while very few firms are involved.
Why does scale favor the leader in software?
Software costs a great deal to develop and almost nothing to copy, so the cost per user falls as users increase. The leader spreads its fixed costs across more people and can reinvest the savings. Platforms with data feedback loops also improve with use, which widens the gap.
Why did decentralized economies outperform planned ones?
Friedrich Hayek argued in his 1945 essay The Use of Knowledge in Society that prices carry scattered information no central planner can gather. Ludwig von Mises had argued in 1920 that planners without market prices cannot calculate costs. The Soviet economy later struggled with shortages and misallocation.
How can models learn from deployment?
Labs can record which answers users accept, rephrase, or reject, and use those signals as training data. Recommendation systems at large platforms work the same way, improving as more people use them. Such a model improves in proportion to how widely it is used.
More from Dwarkesh Podcast
