AI assisted tools for scaling human review

See our projects

Because AI is generating content faster than humans can review it.

Atlas Computing is an R&D nonprofit prototyping AI-powered tools to generate formal specifications. We're empowering engineers to verify code easily and develop software that’s built for trust.

We aim to bridge the gap between human intent and machine output.

AI entrenches biases from an unequal world, while hiding them under a veneer of objectivity
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.

Future AI could pursue our directives in dangerous and unpredictably inhuman ways
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

Stopping progress toward artificial general intelligence (AGI) prolongs all problems that AGI can solve
Risk

Stopping progress toward artificial general intelligence (AGI) prolongs all problems that AGI can solve

Solution

We need to build verifiable tools to solve today’s and tomorrow’s problems

How specification-driven AI works

As the outputs of AI systems become increasingly complex, human review systems (ranging from code review to regulatory approval) will become overwhelmed.

Instead, we envision a world where it is easy to set objective criteria that AI outputs should satisfy, and generative AI systems make it clear to humans that those criteria are met.

We're starting with making tools to generate formal specifications, as computer science already has a language for expressing and automatically verifying these objective criteria of software programs. This could reduce cybersecurity vulnerabilities and regulatory burdens that cost economies billions annually.

Specification-driven AI enables automated verification and compliance, letting humans focus on what an AI system should do, not worrying about if it's doing it.

We aim to bridge the gap between human intent and machine output

Human review exists on a spectrum, and as demands grow, traditional manual review alone won’t suffice - it's the most expensive and time consuming.

Read our 2-pager on scaling safety proofs of software or check out the projects page to see our other efforts as well.