Drawings & photos
Layout, equipment, and fixtures from documents the plant already has.
$ biasedlabs --thesis
Applied Compute
for factories.
We deploy models inside the largest plants and keep them learning, from training to inference. The core is tens of thousands of simulators, each fit to one plant from its own data and handed to agents as tools.
Frontier labs can put a model in your data center. No one stays to keep it learning.
$ labs ingest --plant plant-01 drawings 4,812 photos 21,330 cctv 64 cams $ labs sim build --from site/ simulators 12,406 motion from video $ labs gate --suite afb safety gate pass refusals logged $ labs train --base open-weights rollouts in sims graded by plant rules $ labs deploy --on-prem llm · vision · robot cell serving
01
The deployment company that stays.
Frontier labs can now place a model inside a large manufacturer's own data center. That solves day one.
No one stays for day two hundred: putting the model in, generating the data it needs with AI, and retraining it as the plant changes. That is our job. Not one LLM, but every model on the floor, from language models to vision to robot cells, deployed and customized per plant.
We do not need many customers. We need very large ones.
02
Tens of thousands of simulators, fit to one plant.
Logs
PLC, MES, and process logs set the logic and the timing.
CCTV
Cameras show how things actually move. We build continuous, video-like simulators, not static snapshots.
Site data
Every simulator is calibrated against recorded reality, and recalibrated when the line changes.
Each simulator is a tool an agent can call: try, fail, and retry where the real plant cannot.
| Simulation software vendors (e.g. Siemens) | Biased Labs | |
|---|---|---|
| Simulators | The product, licensed | Free |
| You pay for | Software licenses | Training build-out, compute, inference |
| Built from | Engineering models | Drawings, photos, logs, and CCTV of real motion |
| Used by | Engineers | Agents, as tools |
| Operated by | People | AI, end to end |
03
Why simulators: the grading signal.
A model gets better at what it can be graded on. The question is where the grading signal comes from.
Applied Compute mostly buys it with expert labor: specialists write the evals. Trajectory mostly picks it up from usage: corrections users make become signal. A factory has little of either. Decisions are rare, operators step in rarely, and no one can replay the same moment on a live line.
So we take a third path. Compile what plants already write down, such as cause-and-effect matrices, HAZOP studies, and operating procedures, into graders. Then make rollouts free inside simulators.
A plant can't be rewound. A simulator can.
04
From plant to model, and back.
- 01
Capture
Drawings, photos, logs, and CCTV. The data stays on site.
- 02
Simulate
Thousands of plant-fit simulators, given to agents as tools.
- 03
Gate
No model touches the line before it passes a safety gate.
- 04
Train & deploy
Open-weight models trained on simulator rollouts, served on premises, retrained as the plant changes.
We give the simulators away. We charge for training build-out, compute, and inference.
05
Few customers. Very large ones.
Already instrumented
Large plants already run MES and process monitoring. Sensors are not the bottleneck.
Access & trust
The bottleneck is getting in: access, trust, and approval to touch the line.
On premises
Customer data does not leave the plant. Everything we build runs inside it.
Long relationships
One customer means many lines, many models, many devices, for years.
Trust comes first. A safety gate is the ticket in.
AutoFactory-Bench Our open benchmark that grades AI agents on industrial control tasks. It measures where general models fail on factory work, before anything is trained or deployed.06
Questions.
Why are the simulators free?
They are not the product. They are how a model gets good at one plant. We earn on training build-out, compute, and inference, so every new simulator makes what we sell more useful.
Don't frontier labs already offer this?
They can deploy a model into your data center. We stay after deployment: build the plant's simulators, generate training data with AI, and retrain the models as the plant changes.
Why CCTV?
Drawings say where things are. Cameras show how they move: cycle times, handoffs, and the workarounds no document records. That is what makes a simulator continuous instead of a snapshot.
What data leaves the plant?
None by default. Customer data belongs to the customer. Capture, simulation, training, and serving are designed to run on premises.
Who owns the models?
You do. Weights trained on your plant are yours. What we keep is the know-how and tooling that builds the next plant's simulators faster.
Is this continual learning today?
Honestly, it is periodic retraining: rebuild simulators and retrain on a schedule as the plant changes. Fully online learning is still unsolved, even at the frontier. We will say so when it works.
Which models?
Any model a plant runs: language models, vision models, robot policies. We start from permissively licensed open weights so the license never blocks industrial use.
Bring us one plant.
We will tell you which simulators it needs, what a safety gate looks like for your line, and what we would train first. The first conversation needs no data.
Talk to us