Sourcing owners for neglected AI risks
We identify unowned problems, map stakeholders, draft milestones, source early funders, and recruit an expert leader to take ownership.
- Potential catastrophes are sometimes predictable.
- Mitigation may require a set of skills no individual has.
- However, many of those skills are only needed once.
- We bring the single-use skills to help specialists skip ahead.
How we work
We take a "problem-first" approach.
We're not an incubator, venture studio, or fellowship, and we don't try to be. We love meeting brilliant people with great ideas and we'll happily route them to a better institutional home.
We start with experts.
We ask "What neglected problems will societies need to solve if AI keeps getting more capable?" Then we identify stakeholders for each risk, interested funders, the shape of a solution, and the skills needed to prototype a solution.
We lower the barrier to entry.
We make it easier for people with the right expertise to work on critical problems. We can help de-risk a career change and provide direction, context, introductions to stakeholders and supporters, and clear next steps.
Our output is focused teams.
We're not a think tank; for us reports are a means, not the end. We take responsibility for problems until we can pass the torch to someone with the skills and resources to focus full-time on the problem.
We're currently generating a list of these neglected catastrophic risks.
An example: cybersecurity asymmetry
Problem:
A common proposal to improve cybersecurity with AI pairs one AI to find vulnerabilities and another to fix them. This first system is a powerful cyber weapon in the wrong hands.
Solution direction:
Develop tools that enable formal verification as a defense-dominant paradigm. You could prove your updated software has no bugs without knowing what the old bugs were.
Concrete next steps:
- •Jason Gross ran experiments suggesting today’s AI is sufficient for meaningful progress in this direction before leaving to start Theorem Labs.
- •Formal verification could power a new paradigm of software generation.
- •Evan (Atlas CEO) is designing a Focused Research Organization to build tools that generate and validate formal specifications, in partnership with Convergent Research.
Want to get involved?
Interested in quarterly updates?
Get email updates
Want more frequent updates?
Follow our blog