OpenAI JUST made EVERYONE MAD…

OpenAI JUST made EVERYONE MAD…

More

Summary

Wes Roth examines the release of OpenAI’s GPT 6.2 Astra and why the launch has drawn such a negative reaction online, even though the model posts some standout results. He reviews its reported 100% score on ARC AGI 3 when paired with OpenAI’s harness, its lead on the AI Robotics 0.1 benchmark, and its progress on Epoch AI’s open Erdős math problems, where it produced new Lean solutions, including one with no known prior solution.

The video also covers the safety measures OpenAI has put around the model, including a 35-second kill switch when misalignment is detected, centralized safety infrastructure for RL runs, and inference that halts if safety systems go down. GPT 6.2 Astra is described as the first externally released model to meet OpenAI’s high threshold for AI self-improvement, scoring 29% on a modified recursive self-improvement index, with Gemini 4 and Anthropic’s Fable 5.2 also compared.

The second half focuses on hands-on testing. Wes has the model design a functional 3D-printable pendulum clock using CAD, a Python solver, and Blender, and tries it on the game Schedule I through a custom harness. He also explains the complaints about weaker instruction following and quirky behavior compared with the previous Astra release.


📺 Source: Wes Roth · Published November 17, 2026
🏷️ Format: Review

1 Item

Channels

2 Items

Companies