Nullsilver
Nullsilver was founded by Marco De Santis. Our interests include how language models learn from data, how their capabilities are evaluated, and how they can be used within larger systems.
Alongside the research, we develop tools to support experimentation. One example is Labloop, an experimental harness for running coding agents through a sequence of experiments and recording their results. The intention is to make experiments easier to conduct without losing track of how a result was obtained.
Research principles
01Separate exploration from confirmation.
Exploration helps us identify promising ideas, but a result used to guide the search is not an independent test of what we found. We aim to keep these stages separate, specifying the final evaluation before it begins.
02Make claims inspectable.
A research claim should come with enough detail for others to examine how it was reached. This means publishing the methods and supporting evidence, while being clear about what the experiments do not establish.
03Report outcomes, not just successes.
Failed and inconclusive experiments also tell us something about the question being studied. We aim to include them in the record, rather than present only the attempts that supported the conclusion.
Get in touch
For questions about the research or tools, or to discuss a possible collaboration, write to us.