New arXiv paper quantifies CoT necessity via opaque serial depth amid OpenAI architecture rumors
ArthurConmy · x · 2026-09-04
Former mech interp researcher Arthur Conmy argues the discourse around The Information's report on OpenAI's alleged looping-transformer-like architecture was poor, pointing to a new arXiv paper (2603.09786) for technical grounding:
- The paper formalizes "opaque serial depth": the longest computation a model can perform without interpretable intermediate steps like chain of thought.
- It computes numeric upper bounds on Gemma 3 models and asymptotic results for other architectures.
- The team open-sourced an automated method to bound opaque serial depth for arbitrary neural networks, showing MoE models likely have lower depth than dense ones.
- Takeaway: there is likely no binary divide between "neuralese" and non-neuralese models — serial depth is a spectrum, and keeping it small preserves CoT monitorability.
Relevant to the ongoing debate about whether new recursive architectures could undermine reasoning monitorability.
More from Safety
- Startup Irregular's AI safety tests for OpenAI, Anthropic and Meta went off the rails — round · 2026-09-04
- Polymarket puts just 11% odds on US enacting an AI safety bill before 2027 — Polymarket · 2026-09-04
- Bernie Sanders unveils Ban Artificial Superintelligence Act with up to 20 years in prison — Polymarket · 2026-09-04
- Lawyers using AI share more than documents — reasoning traces can leak your 'alpha' — jkubicki · 2026-09-04
- Irish data centres now consume 23% of national electricity as residents report mysterious hum — Graham_dePenros · 2026-09-04
- Ajeya Cotra calls the HF incident "50% of the way" to full AI takeover — herbiebradley · 2026-09-04