Can AI Soon Write Its Own Software? Insights from OpenAI’s Chief Scientist

OpenAI executive says AI will be able to generate software this year.

On May 14, Nature interviewed OpenAI chief scientist Jakub Pachocki about AI generating original research, AGI prospects, and upcoming open‑source models.

Q: What changes might reasoning models bring in the next five years?

A: Currently models act like assistants needing guidance, but this will change dramatically. OpenAI’s Deep Research can integrate large amounts of information unsupervised for 10‑20 minutes with modest compute, and with more resources they could conduct original research in software engineering, hardware design, and other fields.

Q: How important is reinforcement learning in OpenAI’s reasoning models?

A: Early ChatGPT used unsupervised pre‑training followed by reinforcement learning with human feedback to create a useful assistant. Newer models emphasize the RL stage, enabling the model to develop its own reasoning rather than just retrieve information.

Q: Do models truly “think” or merely simulate reasoning?

A: Models differ from human brains; they acquire world knowledge without true understanding or temporal sense. Yet they can discover new insights, which can be considered a form of reasoning, albeit not identical to human reasoning.

Q: What are the plans for open‑source models?

A: Excitement about providing researchers with downloadable models for further training, though safety concerns make releasing cutting‑edge weights challenging. The goal is to release a version better than existing open‑source models.

Q: How does he define AGI and when might it arrive?

A: His view has shifted; milestones like AlphaGo’s 2016 victory accelerated expectations. OpenAI has made progress on Turing‑test‑like benchmarks, mathematics, and problem solving. The next milestone is AI’s measurable economic impact and creation of original research, with significant advances expected by the end of the decade, and possibly software generation this year.