Overview
- OpenAI said an internal Astra model produced ten advances in mathematics and theoretical computer science and human researchers used the Lean proof assistant to formalize and verify the results.
- OpenAI argued that authorship should reflect the model’s role because the proofs were generated by the system while humans drafted manuscripts and checked correctness.
- MiniMax open‑sourced its multimodal H3 model and MooreThread reported it adapted H3 to run on its MTT S5000 hardware within hours, lowering deployment barriers for high‑quality text, image, audio and short video generation.
- Demonstrations such as Andrej Karpathy’s Opus 5 test, which generated a runnable Three.js 3D scene for roughly $10, highlight that models can cheaply produce complex creative outputs but cannot reliably self‑verify those outputs.
- The Financial Times and other coverage say major U.S. tech firms have sunk roughly $1.1 trillion into AI infrastructure since 2023, a cost pressure that helps explain why many companies are switching to Chinese or open‑source models that report 60–90% lower operating expenses.