GPT-5.2 makes its first physics discovery

arXiv

AIRS-Bench introduces 20 research-focused tasks to evaluate LLM agents across the full scientific research lifecycle, with preliminary results showing agents exceed human performance on only 4 tasks while failing on 16 others, signaling significant opportunities for advancing autonomous scientific research.

GPT-5.2 discovered and proved a previously unknown formula for gluon amplitudes in theoretical physics, marking a significant AI contribution to frontier scientific research. The achievement demonstrates how LLMs coupled with human expertise can generate original theoretical insights and formal proofs at the research frontier.

An AI researcher joins OpenAI to advance agent research while committing to keep their popular open-source project OpenClaw independent as a foundation-backed initiative. This signals OpenAI's strategy of supporting open-source ecosystems while developing proprietary agent capabilities internally.

Google detected widespread abuse of Gemini by threat actors through model extraction attacks, AI-augmented phishing campaigns, and malware development, prompting aggressive mitigation and a push for industry-wide AI security standards.

Alibaba's open-source robotics AI signals a critical inflection point where physical AI deployment accelerates faster than governance frameworks can scale, creating opportunities and risks in the US-China competition.