About the OpenAI Amplitudes Paper, but Not as Much as You’d Like

Key Takeaways

• AI generated conjecture (formula 39) in twelve hours
• Paper omits prompts, outputs, limiting reproducibility
• Model performance matches bright graduate student effort
• Potential to streamline routine theoretical calculations
• Raises ethical and educational concerns in academia

Summary

OpenAI partnered with amplitude researchers to use an internal LLM, dubbed GPT‑5.2 Pro, to conjecture and prove a simplified scattering‑amplitude formula (equation 39) in a twelve‑hour run. The accompanying paper and press release provide scant detail on prompts, model outputs, or the exact division of labor between humans and AI. The author argues the model’s contribution mirrors what an exceptionally bright graduate student could achieve over several months, offering a useful but not revolutionary boost to theoretical work. The lack of transparency fuels debate over AI’s role in high‑energy physics research.

Pulse Analysis

Scattering‑amplitude research has long relied on clever algebraic shortcuts to turn unwieldy expressions into compact, "theorist's delight" formulas. OpenAI’s recent collaboration with amplitudeologists marks a rare foray of large language models into this niche, where an internal model was tasked with extending a set of intermediate results (equations 29‑32) into a more general form and ultimately conjecturing a new relation (equation 39). While the press release touts the AI’s role, the underlying paper provides only high‑level acknowledgments, omitting crucial details such as prompt engineering, intermediate code, or verification steps, which hampers reproducibility and peer assessment.

From a technical standpoint, the model’s output appears comparable to a diligent graduate student’s project: it identified a specialized region, rewrote the expressions, and supplied a proof using time‑ordered perturbation theory—a method not routinely employed by the authors. This suggests that LLMs can navigate specialized mathematical domains, generate plausible conjectures, and even draft code in environments like SymPy. However, the achievement is not a breakthrough in autonomous discovery; rather, it reflects an efficient assistant that accelerates routine derivations, leaving the deeper conceptual leaps to human experts.

The broader implications touch both academia and industry. If such models become cost‑effective, researchers might outsource routine derivations, reshaping the traditional mentor‑student dynamic and raising questions about training, authorship, and academic integrity. Transparency about prompts, computational costs, and validation protocols will be essential to integrate AI responsibly into theoretical physics, ensuring that the technology augments rather than obscures the scientific process.