Schmidhuber Says OpenAI's 'Recurrent Depth' Reasoning Echoes His 2015 Learning-to-Think Paper
SchmidhuberAI · x · 2026-09-03
Reporting says OpenAI's Astra uses a new 'recurrent depth' reasoning approach that can improve cost and performance but obscures the model's thinking, making it harder to monitor. Jürgen Schmidhuber responds that this essentially matches his 2015 paper On Learning to Think, where an RL controller learns to generate chains of abstract, self-invented vector prompts to query a world model for fast planning — beyond his 1990 millisecond-scale planning.
Related event: Schmidhuber Claims OpenAI's Recurrent Depth Traces Back to His 2015 Paper(3 posts)→
More from Models
- Distillation debate: long-horizon tasks rarely benefit from copying frontier model outputs — JoshPurtell · 2026-09-03
- Google Search Console social properties show bizarre zero-click queries, SEOs suspect AI fan-out — gaganghotra_ · 2026-09-03
- Distillation Debate: Frontier CoTs Too Off-Policy For Tiny Models, 27B Is The Better Teacher — JoshPurtell · 2026-09-03
- Gemini 3.8 Flash ties for top of DeepSWE at 74%, but burns 1.34x more output tokens — haider1 · 2026-09-03
- Two Definitions Of Model Distillation: Raw CoT Jailbreak Vs Task Output Training — JoshPurtell · 2026-09-03
- Gemini 3.8 Flash scores 69.9% on Cursor bench at $2.38 per task — _philschmid · 2026-09-03