A public research lab, based in Italy.
Nullsilver is an independent lab working on under-explored questions in public. This page explains in long-form what the lab actually is and how we conduct research.
Nullsilver
We are a small, independent research lab, founded in 2026 and based in Italy. We are not attached to a university or a product roadmap, which is the point: this allows us to publish what we find at the moment we find it, and to work on questions that are too small for a company and too applied for a thesis.
Our work is currently focused on language models and the systems around them (orchestration, evaluation, data), but this is not a hard constraint — if something else gets interesting, we'll go there.
Much of the experimentation here isn't done by hand. Most of our projects run on labloop, our template for agent-driven research: we write and argue over a protocol, an AI agent runs the experiments through a fixed loop of phases, and every run, verdict and revision is logged to this website as it happens.
Open by default. Protocols, code, data and write-ups go out under permissive licenses as they are produced. There is no strict publication cycle.
How we do research
We love to study, to think, to create and test hypotheses, be they successes or failures. That is the foundation that made us decide on creating a new research process specifically tailored to our small lab.
- 01
Protocol as the project entry point
Every project starts with a human-curated PROTOCOL.md: a file containing the questions, the experiments we want to run, the success criteria, prior and related work. When a protocol changes, it gets documented: revisions are numbered, old versions are frozen, and each change records the result that caused it.
- 02
Published while it runs
Work is visible from the day it starts. A running project streams its state to this site — current phase, run number, event log — so that everyone can monitor it.
- 03
Predictions first
Before an expensive run, the agent writes down what it expects to happen and how confident it is. The prediction is logged before the run starts and scored against the outcome once it ends. Calibration is data too.
- 04
Nothing hidden behind a result
Every finding comes with whatever produced it: the code, the weights, the evaluation harness, the raw run logs. Reproducibility is a must.
- 05
Negative results are still results
Failures are recorded with the same care as successes. A past failure can save time and money for new projects, so we keep them, publicly.
- 06
Small, homelab-friendly systems
While we are eager to tackle frontier problems, we still prefer work that the typical local LLM user can actually benefit from; though, once again, this is not a hard constraint.
Work with us
Our lab is small and compute cost is our honest nemesis. There are three ways to work with us.
Sponsor compute. Some of the questions we are trying to answer are bigger than our GPUs allow. Compute sponsorships (money, machines, platform credits) go into open projects, and we credit sponsors on the work they made possible.
Sponsor a research project. If you have a specific question you'd like us to work on, contact us. If it fits our lab, we will study the topic, make hypotheses, design experiments, and run it through our usual pipeline.
Custom datasets. We frequently design and build datasets, sometimes we are even proud of them. Synthetic, curated, or both. If your problem requires one that does not exist yet, we can probably make it.
If any of that sounds useful, write to us. The inbox is read by a human.