pebble · machine intelligence lab · san francisco

Machines learn your work by doing it.

Reinforcement learning is an old idea: you improve at a trade by practicing it, and by knowing how each attempt turned out. Models improve the same way. Everything you'd want one to do runs through its environment, the information it can reach and the steps it can carry out. We build the environment to order around your process, and the verification spine to go along with it. Better work gets done, and that work is verified.

Industry

What one run through an environment looks like.

The recent jumps in model capability came from environments. Here is a run through one of ours, condensed:

experiment loop · rockiethe environment we operate in public

step 1	frame the experiment	✓ plan recorded before launch
step 2	provision compute	✓ pod heartbeat
step 3	run the experiment	✗ pod died mid-run
	→ resumed at step 3, no work lost, ✓
step 4	score the result	✓ metric computed against the plan
step 5	settle the cost	✓ reconciled against the provider invoice
done	record appended · next experiment planned

from rockie, the environment we run in public at rockielab.com. yours has your steps and your checks.

Every step gets the strongest check it allows: computed by code where possible, a trained classifier where not, a calibrated judge where the call needs taste. A failed run resumes at the step that failed. Every scored run joins the record.

Industry

One build, three returns.

Engagements run in three stages. Most customers start at the first and decide later how far to go; nothing built at an earlier stage gets thrown away at a later one.

01run

Working software.

We scope one process, build its environment, and hand it over doing the job, on a frontier model to start. The step map and the first checks get built here; the later stages reuse both.

02prove

A full audit trail.

The full set of checks is in place and every run is logged against them. Compliance gets an audit trail and postmortems start from a step number. Reliability turns into a percentage you can read off the run history.

03own

A model and a dataset you own.

An environment with checks is a training ground. We run an open-source model through it at volume and train against the scores — reinforcement learning on your own process. The result is a model tuned to the work, at a fraction of frontier API cost per run, that no one can reprice or deprecate out from under you.

The scored run records are the durable asset: plain data with no vendor attached. When a stronger open base model ships, the same environment trains its replacement.

How an engagement runs

Four steps from a call to running software.

1scope

A call. You describe the process and the outcome you need from it. If an environment is the wrong tool for the job, we say so on the call.

2discover

About a week inside your systems, finding where checks can attach and what the environment has to touch. Discovery runs under your access model; we keep the step map and none of your data. You get a written proposal: the end goal, the work packages, the price.

3build

The environment and its checks, built and then exercised against real runs of your process before we call it done.

4run

Handoff, on your infrastructure or ours, with a retainer for working sessions, maintenance, and the next stage when you want it.

Research

We also run a research program.

Pebble publishes its own research on matrix-valued state and measurable reasoning in language models. Findings go out as falsifiable notes with code and eval data. Instrument bugs are disclosed, and negative results stay published.

Read the research log →

Academic

Try Rockie.

Rockie is an RL environment we maintain for the public good. It runs experiment loops for academic groups doing HPC research, including ML. The system behind it is written up in our white paper, Scaling Experimentation, on making the full experiment loop run as software.

Rockie is tuned for scientific discovery, and we keep it public because an environment improves with every kind of work that runs through it. The hope is that the models people run in there leave human knowledge further along than they found it. An account takes a minute at rockielab.com.

Try Rockie → Read the white paper