OpenAI quietly using loop transformers that hide 'thinking' at scale, sparking security concerns
steph_palazzolo · x · 2026-09-02
Reports suggest OpenAI and others are quietly using "loop transformers" that do not show their 'thinking' when scaled up. While this offers a leap in performance, it is sparking concerns inside and outside OpenAI regarding security and interpretability.
More from Models
- Speculation: Google merging world model and diffusion into Gemini 4 — haider1 · 2026-09-02
- Fable 5.1 system prompt leak reveals search mechanics of the "most powerful" AI — gaganghotra_ · 2026-09-02
- Analysis: Astra's recurrent depth could yield a 1.38x effective-parameter multiplier — scaling01 · 2026-09-02
- Claude one-shot identifies a person from raw weights at ~75% accuracy — repligate · 2026-09-02
- Fable 5 vs 5.1 self-portrait comparison re-run without the new system prompt — repligate · 2026-09-02
- Arena launches WebDev AI Leaderboard with fullstack/frontend and domain filters — arena · 2026-09-02