How Realistic is OpenAI’s 2028 Timeline For Automating AI Research Itself?

OpenAI flooded the zone yesterday with a long-awaited announcement regarding its revised agreement with Microsoft and corporate restructuring, and an hour-long livestream in which CEO Sam Altman and Chief Scientist Jakub Pachocki answered questions from the public. (For more on the restructuring, check out this story from my colleagues.)

On the livestream, the two men covered a range of topics, from OpenAI’s decision to allow more NSFW content like erotica on ChatGPT to the deprecation of popular old models such as GPT-4o to the threat of job loss from AI. (If I had a nickel for every time Altman referenced heroin on the livestream—in reference to the idea that OpenAI wants adult users of ChatGPT to have more freedom with the chatbot as long as they’re not doing something as dangerous as selling heroin—I’d have at least two nickels.)

One statement that stood out to me was Pachocki’s very specific prediction that OpenAI expects to have AI systems that can do AI research on the level of a research intern by September next year, and systems that can do more advanced AI research by March 2028. 

This idea of an automated AI researcher, or AI that can conduct research to improve itself, has been the holy grail for companies like OpenAI for quite a while. It’s one reason why OpenAI has focused so heavily on its models’ coding capabilities, as we previously covered, as they could help the models automate tasks that researchers do, such as solving math problems and running experiments.