Listen: Latent Space episode and transcript. Watch: publisher-linked YouTube video. Published 21 September 2026; 141 minutes. The YouTube title differs from the podcast-page title.
Who: Diogo Almeida, introduced as TypeSafe AI’s CEO and Jev’s lead creator. The episode explores task/data choices, calibration, narrower primitives, versioning and the relationship between speed and capability.
Useful takeaway: consider whether the application should ask a smaller, more constrained question rather than send every problem through an open-ended chat interface. This is a product-design hypothesis, not an endorsement of the service or proof that it will replace the current model.
Buyer exercise — proposed: decompose one support workflow into classification, retrieval, permitted action and explanation. Mark which steps have verifiable outputs and which require broad reasoning. Then compare a constrained path with the existing baseline on the same final business outcome.
Question to take into procurement: “How is confidence calibrated on our task, and what happens outside the supported distribution?”