Key Takeaways

  • Alex Lupsasca from OpenAI reveals AI, once primarily a "super competent physicist" for calculations, is rapidly gaining the capacity to define novel research questions, matching human experts.
  • While Lupsasca acknowledges initial skepticism, like Terry Tao's view that AI's 'creative' proofs often trace back to obscure sources, his own experiments show a different reality.
  • He tested AI on his own cutting-edge physics work, finding it generated a list of "top three questions" for next steps that perfectly mirrored his own expert intuition.
  • This signals AI models are now "smart enough" with enough background knowledge to rival a human scientist in pinpointing critical follow-up questions in complex domains.
  • Lupsasca believes the next frontier is AI solving long-standing problems that have stumped communities of physicists for decades, a shift he expects “hopefully not too far in the future.”

The Hardest Part: Asking the Right Questions

For years, the distinction between human and machine in complex fields like theoretical physics seemed clear. AI was a calculator, a powerful engine for crunching numbers and confirming hypotheses. The human, however, was the architect of inquiry, the one who saw the unseen and framed the unasked. Alex Lupsasca, a physicist at OpenAI, put it simply: “The difference between a good physicist and a great physicist is knowing what is the right question to ask. That is actually the hardest part of being a scientist.” AI, by this measure, was merely a tool.

He described AI as an “extremely technically skilled graduate student”—someone you could give a sharp, well-posed question to, and it would perform “incredibly hard calculations correctly.” This kind of AI is superhuman when it comes to computation, as seen in its ability to solve problems in gluon and graviton scattering amplitudes that baffled human experts. It is super competent, but, as Lupsasca initially noted, it didn't quite have the knack for knowing what the right question to ask was.

When AI Surprises Even OpenAI Scientists

This neatly defined boundary is dissolving. Lupsasca recounts a conversation with renowned mathematician Terry Tao, who remains skeptical of AI's "creativity." Tao believes any seemingly novel mathematical proofs from AI can ultimately be "tracked back down and found to have really pulled facts out of some obscure reference." It's recombination, not true invention. Many in the scientific community share this view: AI can synthesize, but it can't originate.

Yet, Lupsasca's own experience points to a different reality. He's conducted experiments where, after feeding AI his own cutting-edge physics papers, he asked it to identify the most pressing follow-up questions. His surprising result? “The top three questions it comes up with are like my top three questions for what I should do next.” He added, “I'd say GPT is about as good at as me at finding the next thing to ask.” This wasn't about calculations; it was about intuition, insight, and the ability to discern the next logical, impactful step in a complex research path.

The Next Frontier: Solving Decades-Old Problems

Lupsasca's personal findings suggest AI models are no longer just sophisticated combiners of existing knowledge. They now possess enough background understanding to generate genuinely useful, surprising insights that align with, or even rival, a seasoned expert's own thought process. This leap beyond mere computation into the realm of question generation changes everything.

While he notes, “We haven't seen an AI yet solve a question that stumps an entire community of physicists for decades,” he quickly adds a crucial caveat: “But I think given the trajectory that we're on, at some point, hopefully not too far in the future, we should see that.” This isn't just about faster research; it's about accelerating the discovery of what to research, pushing humanity closer to breakthroughs previously considered impossible.

What to Do With This

Stop seeing AI solely as a production or calculation engine. Next time you're stuck identifying the truly valuable, next-level problems in your business—whether it's market gaps, product features, or team structure—try framing the situation as a research problem for AI. Feed it your existing internal docs, market research, or competitor analysis, then explicitly ask, "Given this context, what are the three most critical, unanswered questions we should be asking ourselves right now?" Treat its output not as gospel, but as a surprisingly intelligent sparring partner to challenge your own assumptions and unlock new avenues of inquiry this week.