Attention — Nov 3, 2026
What caught my attention on Nov 3 (3 items).
GPT-6 Astra Made Qwen 6x Better at Math Without Training
Prompt Engineering (Article:09:00) - Published 2026-11-02 18:00 - Discovered 2026-11-03 09:00
The video discusses MIT CSAIL research on weight-space programming, where a frontier model, GPT-6 Astra, directly reads and edits the numerical weights of smaller neural networks without training or gradients. The headline result is that Astra changed only 0.016% of Qwen3.5-0.8B’s weights and raised its MATH-500 score from 7.9% to 48.1%. However, the video emphasizes that the improvement is less a…
Apex 35B v1.0 - first look and test
NoPlaceLikeLocalhost (Article:10:14) - Published 2026-11-02 18:00 - Discovered 2026-11-03 10:14
In this video, the creator tests Apex 35B v1.0, a fine-tuned derivative of Qwen 3.6 35B A3B, to see whether its claims of improvement over the base model hold up. The test is a practical coding task: adding a read-only rotation mode to a ship-editor tool used in a game project. The tool lets users place gun and thruster markers on a sprite and save them as JSON; the new feature must rotate the spr…
Investigation: How Tech & AI Companies Corrupt Elections
Gamers Nexus (Article:10:23) - Published 2026-11-02 18:00 - Discovered 2026-11-03 10:23
Gamers Nexus presents a special report arguing that major technology and AI companies have increasingly used campaign donations, super PAC contributions, lobbying, and personal relationships with government officials to shape election outcomes and policy in their favor. The video focuses on Nvidia, AMD, Intel, Micron, OpenAI, Anthropic, Elon Musk, and related figures, claiming their political spen…