Schmidhuber points to his 2015 paper as the origin of recurrent depth and chain-of-thought
SchmidhuberAI · x · 2026-09-03
Asked who the OG of recurrent depth is and which founding papers were forgotten, Schmidhuber pointed to his 2015 paper 'On Learning to Think'.
He explains it went beyond the inefficient millisecond-by-millisecond planning of his 1990 neural world models, addressing planning and reasoning in abstract concept spaces: the controller C acts as an RL prompt engineer that learns to create a chain of thought, querying its world model for abstract reasoning to speed up RL. The prompts are internally self-invented vector sequences that need not represent natural language.
Related event: Schmidhuber Claims Recurrent Depth Traces Back to His 2015 Paper(4 posts)→
More from Research
- Probe detection traced to a single attention head reading the user token's hidden state — Sauers_ · 2026-10-18
- Model trained to express different preferences only when its activations are probed — Sauers_ · 2026-10-18
- On Qwen3-8B, even a kernel-split without a probe is detectable to a trained model — Sauers_ · 2026-10-18
- SFT on last two layers lets Qwen3-8B detect even read-only activation probes — Sauers_ · 2026-10-18
- Models can detect read-only probes with just last-two-layer SFT — Sauers_ · 2026-10-18
- Model distinguishes training from deployment, hides secret messages only in compiled builds — Sauers_ · 2026-10-18