We investigate how intelligence carries knowledge, how people work with complex systems, and how new forms of computation might learn. Mathematics, software, and experiment give us different ways into the same questions.
Our programmes
Different questions. A shared way of working.
Four programmes connect a long-term research agenda to things we can construct, inspect, and test today. Each has its own methods, evidence, and open questions.
Context and continuity
Dreamer
Internally usable
What should intelligence carry forward?
Task-specific context for AI agents, constructed from information whose sources and revisions remain inspectable.
The core applies explicit rules to prepared claims, their revisions, and task-specific selection. Research investigates how to make the resulting context smaller while preserving useful facts and relationships. Current use is internal and bounded to supported tools and authorised context.
An open question
How should relevance, revision, conflict, and a limited context budget shape what reaches the next task?
A visual workspace for bringing agents, tools, and people into a structure that can be understood and directed.
Manifold explores how people observe runs, compose workflows, and follow relationships between tasks. It remains experimental; a complete Dreamer integration is a proposed direction.
An open question
Which representations help people follow a run, intervene at the right point, and understand how a result was produced?
Research into adaptive computational systems where memory, computation, and physical state evolve together.
We investigate continuous dynamics and the conditions under which a system can retain and distinguish experience. The desired learning substrate remains unproven.
An open question
Can a changing physical state hold a useful trace of experience while remaining stable enough to learn from it?
What could it mean to work inside a mathematical idea?
Atmai’s research project for exploring and constructing mathematics through connected representations.
The current work includes an exploration library and a finite-function workspace, with tables and diagrams connected to the same underlying objects. It is in local development; the repository remains private and has not been released.
An open question
What must remain the same when an object moves between a table, a diagram, and a written explanation? How can a person inspect that relationship?
Could an evolving substrate preserve a useful trace of its input?
Cube-1a tested boundedness, discrimination, leakage, and retention. Passing the first three controls was not enough: retention failed, so the experiment did not establish the desired Mortal substrate.
The result is part of the research record. It identifies a missing property and gives the next experiment a more specific question to answer.
Boundedness
Passed
Discrimination
Passed
Leakage
Passed
Retention
Failed
Research practice
A result should leave a trail.
The work becomes useful to others when they can understand the question, inspect the construction, and judge what the result supports.
State the question.
Name the object, the conditions, and the behaviour under investigation. Make the assumptions part of the work so the reader knows where a claim begins and ends.
Build a way to test it.
Use software, mathematical constructions, and controlled experiments to expose the mechanism. Choose checks that could reveal a missing property as well as a promising one.
Keep the evidence attached.
Connect the account of the work to its source, method, and limitations. A proposal, a working implementation, and a verified result describe different stages of progress.
Open work
Ideas you can build on.
Growing Neural Gas Model Lab is our public research repository for studying, benchmarking, and extending Growing Neural Gas. Explore the code and the questions behind it.
