Roadmap

PFN Studio is moving fast. This page is the public view of what works today, what we’re polishing, and the bigger workstreams that change how brains get built.

What you can do today

Build a brain in the wizard. Pick what you want the brain to do — predict a number, forecast a series, suggest the next experiment, extrapolate a partial curve, show its own confidence — and the wizard picks paper-backed priors to cover it. See Capabilities reference.

Train it on real benchmarks. Every brain you build can be evaluated against the benchmarks shipped with the prior it was trained on, and you can compare your run to the paper’s reported numbers. See Benchmarks and Paper reproductions.

Use your brain on your own data. Bring a CSV, the studio infers the schema, and one forward pass returns predictions. See Bring your own data and the Predict API.

Browse and install priors. A library of reference priors across regression, classification, time series, probabilistic inference, and causal discovery. Install one into your project, or fork it and edit. See Marketplace.

Developer mode if you want it. Direct access to priors, models, evals, and runs as files. Custom blocks, custom priors, custom training loops. See Developer mode.

Run on your own compute. Local CPU works for small priors. Plug in Vast.ai for GPU when you need scale. Track runs in W&B, MLflow, or push trained brains to HuggingFace.

Programmatic access. API tokens for predict-from-script and CI use.

What’s coming next

Faster path from idea to trained brain

Edit Models, Evals, Runs, and Literature with the same first-class UI Priors already have.
Diff between versions of any artifact.
Submit a run from the browser and watch live status.
Comments and @mentions on artifacts, so a team can review a brain the way they review a PR.

Promptable brains

The bigger shift. Today a prior is fixed at training time — you can’t tell a trained brain “this is a paper machine, the lag is about 4 hours, ignore feedback loops” without retraining. We’re changing that.

Priors will expose steerable axes — properties like monotonicity, time-lag scale, feedback presence, sparsity, noise behavior.
The brain learns to honor those axes at inference. Toggle a chip, the forecast tightens. Type a constraint in plain English, the forecast tightens.
A detector alongside the brain reads your data and proposes axis values automatically, with confidence. You confirm, override, and add the things only a human would know.
No retraining per site. The same brain serves every customer, steered by their domain knowledge.

Brains that fit themselves to your data

Building a good prior today is hand work. We’re automating it.

Mark a prior’s design choices as a search space.
Run an auto-fit against any benchmark — including your own private historical data, kept inside your tenant.
Watch the score curve in real time. Promote the winner to production with one click.

One brain, many domains

The promptable axes are universal — lag, monotonicity, conservation mean the same math everywhere. But domain experts speak in their own words. A short glossary maps a site’s vocabulary onto the axes, so the same brain serves pulp & paper, pharma, energy, and steel without retraining.

Glossary editor with axis autocomplete.
Reference glossaries for two starter verticals.
Optional natural-language input that gets parsed into the same canonical axes — auditable, not magic.

Toward 1.0

When 1.0 lands, the artifact shapes are frozen — your projects keep working forever. We’re getting there by:

Settling the schema versioning + migration story.
A plugin system so the community can add new artifact types.
A handful of external teams running real PFN projects on the studio.
A canonical axis vocabulary that the foundation model commits to honoring.

Want to shape it?

The roadmap moves with customer demand. If a workstream above matters for what you’re building, tell us — we prioritize by the work in flight, not the prettiness of the plan.