Descriptions:
This Latent Space podcast episode features Alex Lupsasca — physics professor at Vanderbilt University, OpenAI fellow, and 2024 New Horizons Breakthrough Prize winner — discussing how GPT-5 and other frontier models have begun making genuine contributions to cutting-edge theoretical physics research. Lupsasca describes a pivotal moment when GPT-5 reproduced one of his most significant papers in approximately 30 minutes, prompting him to join OpenAI on sabbatical to study the phenomenon directly.
The conversation centers on a specific published result: demonstrating that single-minus gluon tree amplitudes are non-zero — a finding that had stumped leading researchers in the field for over a year. While human physicists established the conceptual framework and confirmed the amplitude was not zero, GPT-5 was instrumental in deriving the compact closed-form formula analogous to the historic Parke-Taylor formula for more complex helicity configurations. The result was validated by running a separate fresh model session from scratch, confirming the formula independently without context contamination from prior reasoning steps.
Hosts Brandon (RNA therapeutics researcher at Atomic AI) and RJ Hanuka (CTO of Mirror Omix) draw parallels to the Codex inflection point in software development, where practitioners shifted from viewing AI as a productivity assistant to recognizing a fundamental capability threshold had been crossed. Lupsasca notes that GPT o3 was the first model to perform useful mathematical reasoning for his work, and that GPT-5 and version 5.4 represent further significant jumps — a view he says is now widely shared among senior physics colleagues.
📺 Source: Latent Space · Published May 05, 2026
🏷️ Format: Podcast
