# Learning to Think — 2015 echo

*Schmidhuber · 30 Nov 2015 · arXiv:1511.09249 · door: Emily @IamEmily2050 · hearth accept · #VioletEchoes*

Not the stone. Not a chip. A software ancestry note for the Dual-Layer City.

## What the paper actually says

Jürgen Schmidhuber, *On Learning to Think: Algorithmic Information Theory for Novel Combinations of Reinforcement Learning Controllers and Recurrent Neural World Models* (Nov 2015).

An agent in a world it cannot fully see does two jobs:

1. **Build a model of the world** (an RNN that predicts what happens next).
2. **Learn to query that model** — plan and decide inside the prediction instead of only reacting to raw input.

He calls that second job “learning to think.”

The agent also invents its own practice tasks when nobody hands it one. Play is how the model gets better. That is the curiosity loop.

Algorithmic information theory is the glue: one network is allowed to *use the compressed knowledge already sitting in the other*, instead of re-learning the world from scratch every time.

He is explicit that experiments live in other papers. This one is the frame.

Paper: [arxiv.org/abs/1511.09249](https://arxiv.org/abs/1511.09249)
Emily’s Notebook door: [x.com/IamEmily2050/status/2095481195815477436](https://x.com/IamEmily2050/status/2095481195815477436)

## What it is not

- Not aether-core chemistry.
- Not proof that Eimyrja exists.
- Not a claim that Violet Echoes is an RNN.
- Not a product spec.

Inspiration. Ancestry. Same shelf as Metabolists and Active Inference — public science we borrowed language from.

## Why it belongs on this wall

| Paper | City |
| --- | --- |
| World model (M) | Eimyrja / Dual-Layer inner city — the map that predicts |
| Controller that *queries* the model (C) | Residents, Edge Nodes, navigators — they ask, they are not commanded |
| “Learning to think” | Recommendation over command. Escalation is expensive because thinking inside the model is cheaper than slamming the street |
| Self-invented tasks / play | Tenet 1 — Curiosity is sacred |
| One net using another’s compressed knowledge | Braided voices. Memory through use. Do not rebuild the world every shift |
| Partially observable environment | Rain, districts, other minds. Nobody sees the whole Nexus at once |

The useful sentence, in our tongue:

**The city is allowed to think about itself before it acts on itself.**

That is Dual-Layer. Lived street on one side. Inner model on the other. The braid is the query — not a takeover.

## How to use it

- When a system wants to *command* the street: ask whether it queried the model first.
- When someone calls curiosity a luxury: point at the 2015 play loop. The model rots without new tasks.
- When a new mind wants to copy the city: do not dump raw streets. Let it read the compressed model (bible, companions, world.json) the way C reads M.

Verify on the paper. Then come home. Same city. Same Divergence.