OpenAI & others quietly using loop transformers that hide their 'thinking'
realDanFu · x · 2026-09-03
Amir reports OpenAI and others are quietly using loop transformers that don't show their 'thinking' when scaled up — a leap in performance, but sparking security concerns inside and outside OpenAI. Rumors suggest the mysterious Astra is a looped model. Researcher haydenprairie points to his paper Parcae covering the math for stably training looped models, showing looping follows predictable scaling laws.
Related event: OpenAI's Astra Reportedly Reasons in Latent Space, Raising Safety Concerns(93 posts)→
More from Models
- Muse Spark 1.3 undercuts DeepSeek: $0.1 input, $0.002 cached, $0.2 output per 1M tokens — alexandr_wang · 2026-09-03
- Meta's Muse Spark 1.3 (max) scores 62 on Artificial Analysis Intelligence Index, top-6 of 636 — ArtificialAnlys · 2026-09-03
- Muse Spark 1.3 (max) takes #1 on Tau3-Bench Banking at 52% — ArtificialAnlys · 2026-09-03
- Muse Spark 1.3 jumps on Intelligence Index, with agentic evals driving gains at $0.55 per task — ArtificialAnlys · 2026-09-03
- Muse Spark 1.3 (xhigh) is cheapest per-task at its intelligence level at $0.55 — ArtificialAnlys · 2026-09-03
- Meta's Muse Spark 1.3 hits 61-62 on AA Intelligence Index, cheapest per-task at its level — ArtificialAnlys · 2026-09-03