pebble · machine intelligence lab · san francisco
Machines learn by doing.
We build your self-improving superintelligence loop.
We build your self-improving superintelligence loop.
You name the outcome. We build the loop that reaches it: every stage feeds the next, and everything the loop produces stays on your side of the table.
We scope the outcome you need and build the environment that reaches it, however many processes that touches. It ships doing the job, on a frontier model to start.
Start a build →Checks you approve grade each run as it happens. Postmortems start from a step number, and reliability is a percentage read off the run history.
run 2141 · invoice-reconciliation · pebble-ir-8b step 03 match_po_number pass step 04 amounts_reconcile pass step 05 approval_chain_valid pass step 06 ledger_entry_posted pass outcome approved · 11.2s · $0.003 trailing 30 days 2,141 runs · 98.7% approved
The graded runs become training data. We post-train an open model on the runs your checks approved, on compute we schedule across clouds at the best rate that hour.
How compute works →The tuned model takes over the work at a fraction of frontier API price, behind the same checks. The records can retrain whatever base model comes next, so the cycle keeps compounding.
Talk to us →What we can say about recent work:
A large audit and accounting firm. An environment for AI-assisted audit work, with models post-trained and hosted on the full stack.
An AI insurance company. An environment built around their workflows, with post-training on the graded runs.
A large materials science company. Environment, post-training, and hosting across their research workflows.
Drug and target discovery companies. Environments built around their pipelines, with post-trained models served on the full stack.
A cybersecurity company. Environment, post-training, and hosting, inside their perimeter.
A fintech company. An environment around their workflows, with post-training on the graded runs.
A defense contractor. Models for drone intelligence at the edge, trained in an environment we built for it.
An environment wraps software around your work. A machine does the steps on your systems, and approved verifiers grade each one.
Machines improve at whatever gets graded, so a machine in your environment improves at your process. The graded runs become training data, and a model trained on them has practiced your work toward the outcomes you want.
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.
We build with you, around the capabilities you need and the outcomes you want, each one verifiable. Your machines improve on your work specifically, and the improvement is yours to keep.
You name the outcome. If an environment is the wrong tool for it, we say so.
We design it with you and carry the build ourselves.
Handoff, on your infrastructure or ours. We stay on for what comes next.
Compute is a commodity. Everyone buys from the same clouds and datacenters. The work is getting the right machines at the lowest price and running them well. That layer is ours, and we run it for you.
1 to 256 GPUs across clouds. One platform.
SLURM and Kubernetes, containers included.
Infiniband networking for training that spans nodes.
Grafana dashboards on every job, live.
Large clusters, quoted within a day.
One request, priced across datacenters, H100 through GB300 NVL72.
Idle capacity resells on the spot market. Reclaim it the hour you need it.
The same person from cluster bring-up to steady state.
Rockie is an open platform where machine intelligence improves machine intelligence. Anyone with an idea and some conviction can run HPC science there and push the frontier.
We built it and maintain it for the public good, and it improves with every experiment that runs through it. An account takes a minute at rockielab.com.
Pebble publishes its own work on memory scaling laws and matrix state as memory. Code and eval data ship with every note, and negative results stay published.
