Labs
Labs trains its own models.
- Trained from scratch
- No customer data
- Dedicated machine
Work
Two lines.
Language and the dynamics of systems are not handled the same way.
Language
- What it reads
- Text.
- What it predicts
- What comes next.
Compact models, organised as experts by domain, designed to run on the device.
Language models
World
- What it reads
- The state of a system.
- What it predicts
- The next state.
The world model of Furnace, which speeds up the exploration of a site’s futures.
World models
Status
What exists today.
| Object | Nature | Status |
|---|---|---|
| Language models | Compact, experts by domain | In preparation |
| World model | Dynamics of physical systems | Ongoing, plugged into Furnace |
| Dataset classification | Public levels and audit tool | In preparation |
| Publications | Papers, weights, code | Nothing published yet |
Data
A dataset gets a rating, like a power supply.
Data quality weighs on a training run more than any other setting. We are preparing a public classification: each level bounds the error rate measured on a hand-audited sample.
The audit tool will be open and the sample published. A label falls as soon as a third party finds more errors than its bound.
- Bronze
- Silver
- Gold
- Platinum
- Titanium
Open questions
What we do not know how to do.
Recovering the hidden state of a system whose equations are unknown.
Nobody manages it today, us included: published work stalls at the same point.
Making a world model state what it has found.
A language model bolted on top would rephrase without understanding the system.
Keeping a simulation stable in high dimension.
Over a long horizon, a trajectory ends up diverging. We have not measured it properly yet.
Method
How we measure.
The yardstick comes from outside.
Each model must beat a classical method: gradient boosting, a heuristic or an exact algorithm. Otherwise it has learned nothing.
Accuracy has a cost.
We measure accuracy per parameter and energy per inference, as well as the score.
The threshold is set beforehand.
A threshold set afterwards turns the measurement into a negotiation with the number obtained.
Failures stay on record.
Each one is logged with its measurement and date. We have abandoned several leads.
Brioche, our training tool
Every run restarts identically and resumes at the exact step after an interruption.
Research ends up in a product.
Furnace explores thousands of futures of a site. Our world model lets it explore more of them, faster.
Discover Furnace