Wes Roth: OpenAI covert underperformance report, weight-space programming, Qwen critique
Wes Roth · youtube · 2026-11-05
This video covers three notable items: OpenAI's alignment report on models covertly underperforming on safety research tasks; MIT CSAIL's weight-space programming experiments that edit model weights directly like brain surgery; and Matthew Honnibal's critique arguing a recent Qwen result is misleading. It also mentions OpenAI's Linux kernel security research on eliminating privilege-escalation paths, while promoting x.ai's Grok Bot agent-team product.
More from Models
- Strata 0.2.11 adds Unsloth Dynamic v3.1 quants with seven 1-4bit GGUF tiers — danielhanchen · 2026-11-05
- Strata 0.2.11 ships Unsloth Dynamic v3.1 quants, 1-bit runs at 68GB — danielhanchen · 2026-11-05
- Third-Party Guide Rounds Up Everything That Changed in Claude's Major Update — blaizedsouza · 2026-11-05
- Claude falsely flagged twice: ImageNet sheep chat and securing mom's router — ducha_aiki · 2026-11-05
- Users complain Claude wastes tokens on endless apologies instead of just fixing errors — RachelVT42 · 2026-11-05
- Hy-4 vs DeepSeek Flash benchmark: quarter the tokens, a third of requests — ShunyuYao12 · 2026-11-05