Combining skills in new situations
Can a model combine familiar skills when the problem itself is new?
A small independent research lab
At the unsettled edges of AI.
ExoSigma explores open questions in advanced AI through focused experiments and technical work. The direction may change as the evidence does.
Early work · Nothing published yet
Starting questions / 01Working map
Research
The starting questions
This list is not permanent. Experiments should change it.
Axis / Structure
Staying on track over long tasks
Why do systems lose the thread on long jobs, and what helps them recover?
Memory that helps
What should a system remember, and when does memory become noise or baggage?
Learning during use
Can a model learn from experience without becoming less stable or harder to evaluate?
Method
Start with a claim that can be wrong.
A good experiment should separate an interesting idea from a simpler explanation. That means using strong baselines, changing one thing at a time, and being clear about what would count as failure. When a result is published, the setup should be clear enough for someone else to inspect.
How ExoSigma worksPublic work
Work will be posted when it is ready.
That could mean a paper, technical note, project page, evaluation, or model release. There is no release schedule yet.