At the AI Engineer Code Summit, Diogo Almeida, co-founder and CEO of TypeSafe AI, argued that apps should call a new kind of model for small decisions. He called Jev, his company’s model, “reliable intelligence” that should be used for “just automating everything,” according to Latent Space’s report of the keynote.

TypeSafe AI introduced Jev in September.

What a system one model does

Almeida’s category is “system one” models. They do not write text. They return a structured answer: a choice, a score, or a probability. They are typically faster and cheaper than reasoning models, the chatbots that think through a problem in many steps.

Almeida’s instruction was to connect them to ordinary code. “Code is the language of automation, therefore you need to integrate with code,” he said. Latent Space founder swyx read that as a call to “liberally sprinkle ‘System One’ model calls everywhere INSIDE application code.”

Where it is already used

The report gives three examples, all small and specific:

  • UiPath ranked document chunks in a tournament, where the most relevant pieces advance through successive rounds. It finished in under one second despite thousands of model calls. UiPath’s Chibi Vikram described Jev as “a building block that we can adjust.”
  • Workday showed a prototype that starts filling in a time-off card while the user types.
  • A conference attendee built a site that sorts and ranks fantasy football news.

At Glean, engineer Chinmay Goyal reported that suggestion timeouts fell from 10.2% to 0.8% after his team replaced the model that ranks follow-up suggestions. The new ranker still chooses among options the app has already generated.

The examples support a narrower claim

Each example is a small judgment inside a flow that already exists: rank this, complete that, sort these. Code can check the returned probability and fall back to a human or a default when the score is low. That is a sound use of a fast, cheap model.

“Automating everything” is a forecast. Glean and UiPath reported results for their own systems. The figures give builders a reason to test their own workload, where accuracy, latency and costs may differ.

If your app sends a long prompt to a chatbot for a decision, such as classifying support tickets or routing requests, that is the place to try this. Swap in a call that returns a score and compare its answers with a sample of tickets you have labeled yourself.


End of article