Regulation
Processes classified as national core technology cannot send data to an external API.
$ biased --thesis
The world needs
a billion models.
ChatGPT and Claude have never learned from real factory data. We turn what happens on the floor into evals, then RL environments, then small models biased toward one job.
“There should be as many models in the world as there are firms.”
$ biased capture --line press-04 --days 21 frames 1,284,903 plc events 96,417 $ biased eval --from captures/ golden tasks 312 graded by code $ biased env --rl environments 312 safety rules 41 $ biased train --base open-weights-8b model press-04/changeover on-prem ✓ $
I
One model per firm is not enough.
Satya Nadella says there should be as many models as there are firms. We think that undercounts.
Every line, every machine family, every recurring job has its own rules, tolerances, and failure modes. A model that is good at one of them is worth more than a model that is fine at all of them. Count jobs, not companies, and a billion stops sounding big.
II
Some industries cannot rent a closed model.
License
Open weights come with rules too. Llama's use policy lists operating critical infrastructure and heavy machinery as prohibited.
Cost
A line runs around the clock. Paying frontier prices for every decision breaks a plant's unit cost.
Supply
Hosted models and fine tuning endpoints get retired when the provider decides, not when your line does.
If the plant needs a model, the plant has to own it.
III
Everything has to become data first.
Autonomous companies and recursive self improvement only work where the work itself is recorded as data. Code has tests. Math has proofs. A factory floor has neither, and very few good simulators.
So the first step is unglamorous. Cameras over the line, logs off the PLC, every operator action recorded with consent. The places with no simulator are exactly where this matters most, because recorded reality is the only ground truth there is.
A biased model is not a broken model. It has made up its mind about one job.
IV
From the floor to a model, in four steps.
- 01
Capture
Cameras, sensors, and PLC logs on one line. The data stays on site.
- 02
Eval
Real incidents become a golden set your engineers agree with, graded by code.
- 03
Environment
Each eval becomes an RL task with a verifier, a reset, and the plant's safety rules.
- 04
Model
A small open weight model trained on those tasks. It runs on premises, knows one job, and refuses what your safety rules say to refuse.
| General model | Biased model | |
|---|---|---|
| Learned from | The public internet | Your line |
| Runs in | The provider's cloud | Your plant |
| Refusals set by | Provider policy | Your safety rules |
| Changes when | The provider decides | You retrain it |
| Size | As large as possible | As small as the job allows |
V
AutoFactory-Bench
An open benchmark that grades AI agents on industrial control tasks. Before we train anything, it shows where general models break on plant work.
Where we start.
VI
Questions.
Isn't bias a bad thing in AI?
Bias against people is a defect. Bias toward your plant's procedures is the product. We mean the second: a strong prior about one job, learned from your own data.
Why not just use GPT or Claude?
Use them where you can. Where data cannot leave, a license forbids the use, per call cost breaks the budget, or you need the same model next year, own a small one.
Why small?
One job does not need the whole internet. A small model runs on a server at the plant, answers fast, and is cheap to retrain when the line changes.
Do you need a simulator first?
No. Most plants do not have a good one. We start from recorded reality and build evals from it. A simulator can come later, calibrated against the same data.
What data leaves the plant?
We design for none. Capture, evals, and training can all run on premises. Where something must leave, we agree on it in writing first.
Who owns the model?
You do. The weights, evals, and environments built from your data are yours.
Which base models do you start from?
Permissively licensed open weight models only, Apache 2.0 or MIT style, so the license never blocks an industrial use.
Bring us one line
and one job.
We will tell you whether a biased model can do it, and what data it would take. The first conversation needs no data.
Bring us one line