Fully Automated Researcher

|

According to a recent MIT Technology Review article, OpenAI is building a fully automated researcher to tackle complex problems independently. In the article, Will Douglas Heaven, Senior Editor for AI at the outlet, writes that OpenAI plans to build an autonomous AI research intern by September of 2026, a precursor to a fully automated multi-agent research system debuting in 2028. This AI researcher will aim to take on complex problems in math, physics, life sciences, business, and policy.

Below are key highlights from the article:

  • A new “North Star” for OpenAI. Heaven reports that The San Francisco firm is pulling together its work on reasoning models, agents, and interpretability around one grand challenge: an agent-based system that can go off and solve large problems on its own. Chief scientist Jakub Pachocki told Heaven that the company is getting close to having models that can work indefinitely in a coherent way, much like people do, with a whole research lab eventually running inside a data center.
  • Codex as the early prototype. OpenAI released Codex in January, an agent-based app that writes code on the fly to handle tasks like analyzing documents, generating charts, and summarizing inboxes. Pachocki described Codex as a very early version of the AI researcher and said most OpenAI staffers now use it. The next step is stretching its problem-solving abilities beyond coding into the broader sciences.
  • Training models to work longer alone. Pachocki argued that simply boosting overall capability already produces models that can run longer without help, citing the jump from GPT-3 to GPT-4. Reasoning models added another bump by teaching systems to work step by step and backtrack from mistakes. OpenAI also feeds models hard puzzles from math and coding contests to help them juggle long contexts and split big problems into subtasks.
  • Outside skepticism. Doug Downey of the Allen Institute for AI welcomed the vision but cautioned that chaining tasks together causes error rates to climb. His team’s tests last summer found GPT-5 led the pack while still making many mistakes.
  • Safety and oversight worries. Pachocki acknowledged that a system running a full research program could go off the rails, get hacked, or misread instructions. OpenAI’s main safeguard right now is chain-of-thought monitoring, where models jot notes in a scratch pad that other models review. He also said powerful systems should sit in sandboxes, and that the concentration of power involved is a real challenge for governments.

In summary, Heaven notes that OpenAI is betting that the journey from today’s coding agents to a self-directed scientific researcher largely involves pushing further along familiar lines. Pachocki stops short of promising human-level intelligence by 2028, instead suggesting that economically transformative systems don’t need to match humans in every way to fundamentally reshape the research process.

Acknowledgement: This summary was generated using AI writing tools as part of my AI Explorations process. For more information, please refer to the original article listed below.

Related Research Link