Startups vs Labs

The best AI products released in the past four years have been built by the labs themselves. This is something I wouldn’t have predicted if you’d asked me in 2023, and I find it quite sad that small startups weren’t the ones releasing products like Claude Code. This happens because the model provider can use RL to train models specifically for the products it wants to release. The labs did it with coding, and OpenAI just did it with computer use.

(If your counterargument is Cursor, they sold to SpaceX because they realized this.)

It’s weird that the best product of this decade was made by big companies and not by a small, lean startup that uses the big lab as a platform. The triumph of Claude Code/Codex is incomparable with Granola (though is a great product).

My prediction for the next four years is that this will change. In coding, we’re close to the “good enough” stage for most tasks, which means a startup could use open-weight models in the same way labs use their proprietary models. A small company with deep pockets could build the next great product. This will become increasingly true as models get better: you really don’t need Fable to automate email tasks. A startup using Kimi fine-tuned with RL can build a better solution than OpenAI today, because for many tasks, capabilities are no longer the bottleneck.

Five years ago, if you’d wanted to build the next great product, you would have needed to be at Anthropic or OpenAI. I believe that has changed, once and for all.

Part of me also wants to believe this, because a world where centralization is necessary for innovation is one I wouldn’t want to be part of. The best possible time to do a startup?