Nikunj Handa
Related people
Episodes
Notes
Latent Space: The AI Engineer Podcast · Why Dwarkesh is Wrong about Computer Use + How OpenAI shipped its…
-
Did OpenAI Build the Decisions API in One Week?
Nikunj Handa says the Decisions API work started a week ago and is still very early. He calls the internal use cases pretty obvious. User operations jumped on it to classify support tickets. He also describes a push for low-latency evals that use a language model as a judge. Shawn Wang repeats the timing back as one week. Handa says the Codex app team might pick the Decisions API up and try to do something with it, and he treats that as still early.
First-party OpenAI products are already building fully on the Agents API. Codex security, which just shipped, is built on the Agents API. A meetings product was launching the same day.
-
Why Did OpenAI Copy Jev?
Nikunj Handa says Jev inspired OpenAI's Decisions API and a wider slice of the market. Jev arrived, users asked OpenAI for something like it, and internal teams wanted a much faster classification system. OpenAI is not training a new model for the feature. Four weeks ago it was not on the plan.
An inference engineer picked the idea up and built a prototype that worked. The group is now trying to push the latency as low as possible, and Handa wants to launch in the coming days once they hit that latency target. He holds back some snappy Decisions API demos, including ones tied to Dots, because they are still ahead of what he wants to describe.
-
What Is Mid-turn Steering on an OpenAI Model?
Nikunj Handa says mid-turn steering on GPT-6 lets someone inject messages while the model is still reasoning, and new instructions can go in as soon as a tool call finishes. The behavior is partly a model alignment capability, so the model has to be trained for it. Steering during reasoning was already possible in the app. The earlier version was not the best, and it has gotten much better.
With every new model, the team works closely with the post-training team and the research team to learn what is new, then exposes the new capabilities in the API. GPT-6 also launched async function calling, also called async tool calling. Tool calls take so long that model execution no longer has to paus