In researchWorld model
Eunoia Sigma
A system that predicts what happens next in the world rather than what word comes next in a sentence.
- Domain
- World modelling and planning
- Stage
- Research
- Memory
- Episodic, semantic, procedural
- Availability
- Not released
Fluency without grounding.
Language models are extraordinarily good at producing text and structurally unable to check it against anything. There is no model of the situation being described, so a confident sentence and a false one are produced by the same process.
Memory has the same shape of problem. Context windows are wide but shallow — there is no continuity between one conversation and the next, and nothing accumulates.
Ground the system in a model of the world, and let it plan before it speaks.
Sigma predicts next states rather than next tokens, keeps what it learns across sessions, and evaluates a plan before committing to it. Alignment then operates on plans rather than on wording, which is a far more meaningful place to intervene.
What would count as working.
- Predictions about how a situation develops that hold up when the situation develops.
- Continuity that survives across sessions rather than resetting.
- Plans that can be rejected internally before they are acted on.
- Failures that stay contained in one subsystem rather than taking down the whole response.
What we publish, and what we keep.
Released
- The problem framing and design principles
- Evaluation approach
- Results when they exist
Retained
- Architecture and component design
- Training procedure and data
- Memory and planning internals
