Trailblazing paths to general AI with safety guarantees
See our projects
Atlas Computing is a nonprofit fostering an ecosystem for researchers, companies, communities, and governments to advance Safeguarded AI, a new architecture that provides quantitative safety properties.
Increasingly capable AI brings benefits and risks
Risk
AI entrenches biases from an unequal world, while hiding them under a veneer of objectivity
Solution
We need to empower human deliberation to set rules for AI outputs.
Risk
Future AI could pursue our directives in dangerous and unpredictably inhuman ways
Solution
We need to be able to set objective constraints on AI systems
Risk
Stopping progress toward artificial general intelligence (AGI) prolongs all problems that AGI can solve
Solution
We need to build safe, non-agentic tools to solve today’s and tomorrow’s problems
How Safeguarded AI works
Users, engineers, and regulators should be able to express enforcable constraints on AI outputs, like making sure a generated software program won't crash, or a designed molecule won't react with specified parts of a cell.
Safeguarded AI includes specification generation tools that help users specify desired properties of outputs which are separate from solution generators. Additionally, proof generators use models of the world to generate objective evidence that the proposed solutions have all desired properties.
Provable safety properties enable automated compliance, letting humans focus on what an AI system should do, not worrying about if it's doing it.
For more details, you can read our executive summary 2-pager that describes our overall strategy or our pitch-deck-style slides
Our first priority: Scaling formal methods
Formal verification is the highest standard for assuring safety of software systems, but today it's too costly for widespread use. We aim to show that progress in AI and formal methods can dramatically lower these costs.
Read our 2-pager on scaling safety proofs of software or check out the projects page to see our other efforts as well.
