TypeSafe AI's Jev cracks the LLM monolith

TypeSafe AI's founder Diogo Almeida worked on ChatGPT and helped invent RLHF, the training technique behind this era of AI. He has described the years since as a kind of heartbreak, telling TechCrunch that we have lightning in a bottle and yet it is not useful. His argument is that human language is the wrong target for automation, because the thing on the other end of a backend decision is other software, not a human.

Jev's launch fits a pattern seen before in computing. In 2005, Michael Stonebraker and Uğur Çetintemel published a paper arguing that the monolithic relational database would lose most of its workloads to purpose-built engines. Databases then shattered into columnar warehouses, time-series stores, search engines and vector databases, and the value floated up to the query planners and data platforms that decide where a query should run. The CPU followed a similar path — it did not vanish but became the coordinator that dispatches specialized work to GPUs and other accelerators.

What Jev does not settle is whether TypeSafe itself wins, whether its benchmarks survive independent scrutiny, or whether the price holds. The argument made here is that the plurality of specialized parts matters, not any single one. A router and compiler layer that watches live traffic, learns which computations have stopped being open questions, and compiles them down to the cheapest primitive that clears the bar is where that value is expected to land.