Some questions are worth decades.
Autonomio is an independent, non-profit AI research lab in Finland. Our work brings together language, mathematics, and the tools that help other people do science.
We have been following these questions since 2004. We fund the work ourselves, and give it the time it needs.
It began with language.
Beyond hand-written rules.
Our early natural language processing work introduced super-class summarization. It was an attempt to move beyond systems that could only do what someone had already thought to write a rule for.
The question was how a machine might work with language without every possibility having to be anticipated by its maker. That question would stay with us.
Language as mathematics.
A decade later, we built an early prototype of a purely mathematical, vector-based approach to natural language processing. Language was something to represent and work with mathematically, rather than handle through a growing collection of rules.
The prototype explored a direction that has remained central to our interests: what becomes possible when we reconsider the representation of a problem, rather than keep adding to the machinery around it?
Talos is our open-source tool for automating hyperparameter experiments with Keras and TensorFlow: trying different model settings, recording experiments, and evaluating the results. The researcher stays in control of the model.
Thousands of researchers have used Talos across scientific disciplines. For us, this is a particularly satisfying kind of work: making something that becomes useful in research we would never have done ourselves.
We built it and kept maintaining it, including six years without a breaking bug. A tool used in science needs to be dependable long after its first release.
import talos
experiment = talos.Scan(
x=x,
y=y,
model=model,
params=params
)What does intelligence cost?
We were asking about AI’s energy use and environmental impact well before they became part of its everyday discussion.
Getting a useful result matters. So does what it takes to get there. We are interested in approaches that need fewer operations, less computation, and less energy, not only in approaches that do more.
This connects to a question running through much of our work: how much of what a computer does is actually necessary? Sometimes the interesting possibility is not a more powerful machine, but a different way of asking it to work.
Independence, in practice.
Autonomio is a Finnish non-profit. We have paid for this work ourselves, without asking anyone else to fund it.
That gives us time to follow our own judgement. We can work on an idea before there is a market for it, share something because it is useful, and keep maintaining it after the initial excitement has passed.
Our work moves between language, mathematics, software, and the practical questions of scientific research. We follow the question, even when it leads outside the field where it began.
We started by playing with computers.
We never really stopped.