Key Takeaways

  • Alex Lupsasca, a researcher at OpenAI, initially doubted AI's ability to tackle complex scientific problems, believing it was limited to simple tasks like email.
  • His mind changed drastically when GPT-5 reproduced one of his previously published theoretical physics papers—a work that took him months to develop—in just 30 minutes.
  • This achievement in areas like gluon and graviton scattering amplitudes demonstrates that AI has moved beyond mere automation to exhibit superhuman reasoning on specific, highly complex scientific tasks.
  • Lupsasca's experience suggests a new frontier where AI can generate novel insights and significantly accelerate scientific discovery, rather than just assisting human experts.

The Moment AI Became Superhuman

Alex Lupsasca remembers a time, not so long ago, when he considered the idea of AI contributing at the scientific frontier to be almost absurd. “I thought, 'Oh, that's special. It's much harder than email and AI is not going to be able to do that,'” he recalls, referring to his complex work in theoretical physics. Like many founders, Lupsasca saw AI as a tool for efficiency, not a partner in groundbreaking discovery. He'd spent days, sometimes months, grappling with intricate calculations related to gluon and graviton scattering amplitudes – problems that stumped human experts for extended periods.

Then came GPT-3, followed by a series of rapid advancements, culminating in GPT-5. The shift wasn't gradual; it was a sudden, undeniable jolt. “When GPT-5 came out it was able to reproduce one of my best papers that took me a very long time to come up with in like 30 minutes. And that's when I really became AI pilled,” Lupsasca explains. Imagine watching months of your hardest intellectual labor condensed into half an hour by a machine. It's an experience that shatters prior assumptions, forcing a complete re-evaluation of what's possible.

Beyond Automation: The Scientific Frontier

This isn't about AI drafting better emails or summarizing meeting notes. This is about AI solving problems that previously required deep human intuition, years of training, and immense computational effort. Lupsasca's team didn't just automate a known process; they witnessed AI autonomously arrive at complex solutions for problems at the cutting edge of theoretical physics. As Lupsasca puts it, “I think we're at a special time now where at least in some directions AI has become superhuman at least on certain tasks.” Specifically, he notes that GPT-5 “continues to be the best model for this kind of mathematical physics work.”

For founders, this insight reshapes the conversation around AI entirely. It’s no longer just about offloading tedious tasks. It's about recognizing that in specific, highly analytical domains, AI models like GPT-5 can now outperform human experts, not just replicate them. This capability points to a future where AI isn't just an assistant but a co-creator, accelerating the pace of innovation and discovery in ways we're just beginning to understand. The implications for R&D, product development, and even strategic problem-solving within your own company are immense.

What to Do With This

Stop asking how AI can automate your existing processes. Instead, identify one "impossible" or extremely complex problem your team faces this quarter—a strategic puzzle, a multi-variable optimization challenge, or a deeply analytical forecasting task. Dedicate a small, sharp individual or sub-team to rigorously test advanced AI models (like GPT-4's API or Claude Opus) against this specific problem. Frame it as: "Can the AI beat us at this, not just help us?" You might be shocked at the answer.