What pushed a 16-year-old internet giant to become an AI neolab? There was a model it wanted to run and couldn’t.

A fascinating story from Cloudflare’s @michellechen at the @aiDotEngineer today.

Back in 2023, Cloudflare put GPUs close to users, betting that AI models would get smaller. Frontier models got enormous. Then along came Jev, a model that could make a decision in one pass. At that speed, the trip to the server matters again. Cloudflare had the network for it, but Jev’s weights were closed.

My favorite detail from Michelle’s talk was that the team spent a week deciding whether to train a decision model. Then they trained Clef over the weekend. And they cleverly decided to go with a hybrid model.

Michelle showed how the hybrid works with three questions about a failed software build.

• Which of these 40 tools do we need? Clef Flash picked them.
• What’s the error in the logs? Clairic pulled it out.
• Why did the build fail? That one went to the large reasoning model.

The first two questions went to Cloudflare’s fast System 1 models. System 2 only had to answer the third. In the demo, the agent reached the same diagnosis with one LLM call instead of six, and about 1,800 LLM tokens instead of 40,000.

That's pretty impressive – and that's the insight: Jev gave Cloudflare the push, and I think we’ll see a lot more teams building hybrids like this. Once you see how many LLM calls go into finding the right line in a log, it’s hard to justify the bill.

*I doubt System 1 + System 2 is the final architecture, though.* Say the agent proposes a fix. Would that change alone have prevented the failure? Can we check it against the rules the software has to follow? That’s where I see room for causal and neuro-symbolic approaches.