Researching intelligence with curiosity, rigor, and purpose.
Mission
Theory & Signal exists to understand machine intelligence well enough to build it responsibly — and to use it to answer questions that matter.
Machine learning systems are becoming more capable faster than our understanding of them is improving. That gap is both a scientific opportunity and a practical risk. We think the most useful thing a research organization can do is to close it: to explain why these systems work, where they fail, and what new kinds of systems are possible.
We pursue that goal through research — careful questions, controlled experiments, and honest reporting — rather than through a single product. Our name describes the method: theories that make predictions, and signals that test them.
One subject, seven directions.
Intelligence does not divide neatly into fields. We organize our work into seven domains for clarity, but the questions we care about most cut across them.
The research directory- Artificial Intelligence
Reasoning, planning, autonomous agents, and systems capable of adapting to increasingly complex environments.
- Machine Learning
Learning algorithms, model generalization, representation learning, optimization, and efficient intelligence.
- Generative AI
Foundation models and generative systems that create and reason across language, imagery, and other modalities.
- Natural Language Processing
The structures of language, computational understanding, multilingual intelligence, and grounded reasoning.
- Computer Vision
Visual representation, perception, spatial understanding, and multimodal interpretation.
- AI Safety
Interpretability, evaluation, robustness, alignment, uncertainty, and reliable model behavior.
- Scientific Discovery
Computational approaches that help explore problems in mathematics, physics, biology, chemistry, and related sciences.
Understanding before scale.
First principles
We prefer understanding why a method works to knowing that it does. Sharp questions, stated precisely, are the most valuable thing a research program can produce.
Experimentation
Claims should be falsifiable and results repeatable: controlled comparisons, multiple seeds, released configurations, and honest reporting of what did not work.
Interdisciplinary thinking
Many ideas in machine learning have roots in statistics, physics, neuroscience, and linguistics. We keep those connections open in both directions.
Responsible progress
Understanding how a system can fail is part of understanding the system. We treat evaluation and safety as research questions in every domain, not as a final checklist.
Fundamental and applied, in conversation.
We do not treat fundamental and applied research as separate stages. Applications surface the questions theory has missed; theory explains why applications succeed or fail. Our work moves deliberately between the two.
A long-term project.
The questions we have chosen will not be settled quickly. We intend to build an organization that can work on them patiently — publishing what we learn, collaborating with others who share the questions, and holding our own claims to the standard we would apply to anyone else's.