Jump Trading Uses AI for Long-Running Research, With Humans Approving Significant Changes

OpenAI on October 6 (US time) published a look at how the financial trading firm Jump Trading is putting AI to work. Jump Trading uses ChatGPT and GPT-6 Astra for quantitative research, the number- and data-heavy side of its work. According to OpenAI, the firm applies AI not only to routine coding but also to long-running research projects that test new hypotheses.

How the research gets handed to AI

Researchers set the stage first. Jump Trading runs AI agents built on GPT-6 Astra, software that carries out tasks on its own, inside its research workflow. A researcher defines the problem to be solved and the working environment, and a human also decides the criteria by which results will be judged. It is not just a matter of handing over a task: the researcher designs the working conditions and the yardstick for success as well.

From there, the AI steers itself. OpenAI says the AI evaluates its own output and adjusts the direction of the next round of analysis based on what it finds. It also merges multiple improvements into one. Rather than running a single pass and delivering an answer, it reviews the results and keeps the follow-up work going.

Researchers can also redirect the AI while a job is still running. The range of use runs from everyday coding to testing new hypotheses, covering both quick tasks and analyses that take a long time.

Who checks the trading signals

People do. Even when a long-running job is handed off, researchers look at interim results on a regular basis. Review does not wait until everything is finished; humans step in while the analysis is under way.

AI output, trading signals included, is reviewed on the assumption that it could be wrong. Significant changes go through human verification and approval. So while the AI is allowed to grade its own results, the consequential calls stay with people.

The case study shows both sides of the division of labor in research work: what gets handed to AI and what stays with humans. Researchers own problem definition, evaluation criteria, interim checks and approvals, while the AI handles coding and hypothesis testing. How much this approach actually changes research time and output has not been disclosed. Researchers and companies in Korea weighing adoption would also need to check the terms of use, the cost, and how far Korean-language support extends.