Inside OpenAI’s big play for science 

“This is actually a desirable place,” says Will. “If you say enough things that are wrong and then someone stumbles upon a grain of truth and then the other person seizes it and says, ‘Oh, yeah, that’s not quite true, but what if we…’ you’ll gradually find your way through the forest.”

This is Weil’s basic vision for OpenAI for science. GPT-5 is good, but it’s not Oracle. He adds that the value of this technology lies in directing people in new directions, not arriving at specific answers.

In fact, one of the things OpenAI is looking at now is making GPT-5 reduce its confidence when it gives a response. Instead of saying Here’s the answerHe might say to the scholars: Here’s something to consider.

“This is actually something we spend a lot of time on,” Weil says. “Trying to make sure that the model has some sort of epistemological humility.”

Watch the observers

Another thing OpenAI is looking into is how to use GPT-5 to validate GPT-5. Often, if you enter one of the GPT-5 answers back into the form, it will break it down and highlight the errors.

“You can associate the model as its own critic,” Weil says. “Then you can have a workflow where the model thinks and then goes to another model, and if that model finds things that it can improve, it brings it back to the original model and says, ‘Hey, wait a minute, this part wasn’t right, but this part was interesting. Keep it.” It’s almost like two clients working together and you only see the result once you pass the critic.

What Will describes also sounds very similar to what Google DeepMind did with AlphaEvolve, a tool that wrapped LLM and Gemini within a broader system that filtered good responses from bad ones and fed them back to improve them. Google DeepMind has used AlphaEvolve to solve many real-world problems.

OpenAI faces stiff competition from rival companies, whose master’s degrees can do most, if not all, of the things it demands for its own models. If so, why should scientists use GPT-5 instead of Gemini or Anthropic’s Claude, families of models that improve themselves every year? Ultimately, OpenAI for Science may be as much an effort to plant science in a new area as anything else. Real innovations are still to come.

“I think 2026 will be for science what 2025 was for software engineering,” Weil says. “At the beginning of 2025, if you’re using AI to write most of your code, you were an early adopter. But 12 months later, if you’re not using AI to write most of your code, you’re probably falling behind. We’re now seeing those same early flashes of science as we did for code.”

“I think within a year, if you’re a scientist and you’re not using AI heavily, you’re going to miss out on the opportunity to increase the quality and pace of your thinking,” he continues.