About
I’m a senior machine learning research scientist working where learning systems meet information theory and dynamical systems. Most of my questions begin with a simple one: what lets an intelligent system carry experience forward?
I’m usually chasing Boston’s ramen, playing pickup soccer, lifting, or getting overly invested in a new project. Past lives include a café racer, mushroom growing, and a data analysis of my Counter-Strike performance.
Writing
- Law Does Not Return a Boolean
Fidryn is a proposed language for legal instruments—precise where law is mechanical, explicit on judgment, never hiding authority or ambiguity in a Boolean.
- The Metastable Solitude: The Physical History Before a Habitable World
Life needs persistence without permanence: arrangements that last long enough to be used and stay able to change. Could a background field in another long-lived state move nuclei, stars and chemistry, and shape the intervals in which worlds become habitable?
- America Should Adopt Renewable Energy, But Only If It's Nuclear
Why the U.S. should power an electric future with nuclear first: capacity factor, land use, waste volume, and supply-chain independence versus solar.
Bryan Gass
Senior Machine Learning Research Scientist, Systems & Technology Research. Boston, Massachusetts.
I build learning systems and think about memory, agency, and control. I’m a senior machine learning research scientist working where learning systems meet information theory and dynamical systems. Most of my questions begin with a simple one: what lets an intelligent system carry experience forward?
Writing
- Law Does Not Return a Boolean
Fidryn is a proposed language for legal instruments—precise where law is mechanical, explicit on judgment, never hiding authority or ambiguity in a Boolean.
- The Metastable Solitude: The Physical History Before a Habitable World
Life needs persistence without permanence: arrangements that last long enough to be used and stay able to change. Could a background field in another long-lived state move nuclei, stars and chemistry, and shape the intervals in which worlds become habitable?
- America Should Adopt Renewable Energy, But Only If It's Nuclear
Why the U.S. should power an electric future with nuclear first: capacity factor, land use, waste volume, and supply-chain independence versus solar.
Current questions
- Long-context learning: I am interested in models that can carry useful experience across very long horizons without making memory prohibitively expensive.
- Continual agency: What would it take for an agent to keep learning after deployment—to accumulate skills, revise its world model, and remain useful over time?
- Control and alignment: I use ideas from control and information theory to ask how intelligent systems gain influence, and how they might do so without diminishing ours.
Elsewhere
I’m usually chasing Boston’s ramen, playing pickup soccer, lifting, or getting overly invested in a new project. Past lives include a café racer, mushroom growing, and a data analysis of my Counter-Strike performance.