OpenAI is refocusing its research efforts and throwing its resources into a new grand challenge. The San Francisco firm has set its sights on building what it calls an AI researcher, a fully automated agent-based system that will be able to go off and tackle large, complex problems by itself. OpenAI says that the new goal will be its “north star” for the next few years, pulling together multiple research strands, including work on reasoning models, agents, and interpretability.

There’s even a timeline. OpenAI plans to build “an autonomous AI research intern”—a system that can take on a small number of specific research problems by itself—by September. The AI intern will be the precursor to a fully automated multi-agent research system that the company plans to debut in 2028. This AI researcher (OpenAI says) will be able to tackle problems that are too large or complex for humans to cope with.

Those tasks might be related to math and physics—such as coming up with new proofs or conjectures—or life sciences like biology and chemistry, or even business and policy dilemmas. In theory, you would throw such a tool any kind of problem that can be formulated in text, code or whiteboard scribbles—which covers a lot.

Read the full story for an exclusive conversation with OpenAI’s chief scientist Jakub Pachocki about his firm's new grand challenge and the future of AI.

For me, the answer is obvious, but I'm interested to hear the opinion of others to understand if my biases are clouding my view.

The most persuasive evidence for me is the improvements in the models. For Astra, this is best demonstrated in agentic behaviour, which is critical for having an automated AI researcher, as agentic behaviour means the model can actually do things. Use a computer. Write code. Run code. Use tools. Search. Carry out experiments. Look at the results. Change what it’s doing. Try again. Keep working on a problem.

And if the brain keeps improving at the same time as the hands improve, it seems completely plausible to me that the job of an AI researcher will eventually become automatable.

For those who don't share this view, what does an automated AI researcher fundamentally need to be able to do that a few more iterative model improvements won’t plausibly give it?

Even if you think that the jumps required are large, it's important to note that models are making significant improvements.

For example, the jump on SpatialBench for Astra is ridiculous. The benchmark consists of 25 visual path-tracing problems and 25 3D mental-rotation problems, which are exactly the sort of things previous multimodal models were embarrassingly poor at. In fact, GPT-5.5 scored only around 13%. But Astra scored around 67% without tools and over 90% with Code Interpreter, which is above the reported human baseline of 80%.

Then you have the ARC-AGI-3 results, where models are required to enter unfamiliar environments, work out the rules, build representations of what’s going on, plan, use tools and create software to help solve the problem. GPT-5.5 previously scored just 0.43%. Astra scored 62.7% under the standard provider-neutral harness, while an OpenAI-specific harness that preserved its reasoning state reached 99.9%. Remarkably, Astra used fewer actions than the median human on 96% of the levels it solved.

So what am I missing? What do sceptics think will stop this progression before we get to an automated AI researcher? And if we get that far, what stops us from achieving RSI? And if we get that far, what does a post-RSI world look like?