Building AI that learns where it works.
The largest general-purpose models are extraordinarily powerful, and becoming more so. But intelligence does not only come from knowing more. It also comes from learning what matters in a given place.
A geologist moving from Western Australia to Central Asia does not relearn geology from scratch. They bring general knowledge with them, then build a working model of a new landscape: its formations, structures, terminology, anomalies and history. Evidence that matters in one region may be irrelevant in another. Expertise comes from adaptation.
We think AI systems should be able to do the same.
Eigenform develops methods for models to evaluate their own behaviour, learn from the environments in which they operate, and guide their own future development. We study this as a fundamental problem in recursive self-improvement and build practical systems that put the same principles to work today.
Six areas, one system
Our work connects six areas.
Frontier Research
We research self-improving AI, including Negative Space Learning (NSL), continual adaptation, and applications of AI to mathematics and scientific discovery. The underlying question is one of recursive self-improvement: how a system gets better at getting better.
Dynamic Benchmarking
A system cannot improve itself unless improvement means something. Groundtruth is our framework for evaluating models against changing, domain-specific evidence rather than treating intelligence as a single static benchmark score.
AI Skunkworks
We build custom AI systems for large international organisations, particularly where an off-the-shelf model or conventional software stack cannot solve the problem.
Geological Exploration
We work directly with exploration companies, including Lightning Minerals, applying AI to real geological decisions where hypotheses ultimately collide with physical evidence.
Geological Knowledge
With partners including NextMaps and MatchPoint, we turn large bodies of published geological information — WAMEX records, JORC disclosures, technical reports, maps and historical datasets — into information that machines can actually reason over.
Geological Software
Geocluster is our open-source AI environment for geological analysis: agentic workflows, scientific visualisation and custom models that adapt to particular datasets and geological environments.
These are not separate projects.
Research gives us new ways for models to learn. Benchmarks tell us whether they actually improved. Large-scale data gives them a world to learn from. Geocluster makes those methods usable. Exploration tests them against reality. And our skunkworks work takes the same approach into entirely different domains.
AI that does not merely execute what we have taught it, but becomes better at solving the problems it encounters, even when we are not present to guide it.
More on the company behind the work is on our about page, and the papers and released systems are listed under research.