Incubated Founder

Advancements in AI bring both opportunities and serious risks. Atlas Computing takes on society-scale AI risks that no existing institution is built to address — we design the intervention, validate it with domain experts and funders, then recruit and empower the specialist who will own it.

You can think of us as:

- A tiger team that identifies, scopes, and launches interventions that aren't a good fit for existing organizations

- A design studio developing and staffing product requirements docs for coordination and market failures

- A think tank combined with whatever their analog of "tech transfer" would be.

Some of the most important problems in catastrophic AI risk need owners: leaders who identify an overlooked, high-impact, and tractable problem; develop effective interventions; and ultimately take personal responsibility for bringing those interventions into the world.

That’s why Atlas Computing is launching a selective program for people with deep domain expertise and the drive to address the most important risks facing society from growing AI capabilities. Founders will take one problem from a rough direction to a fundable plan: sharpening a concrete intervention, pressure-testing it against the field, building the relationships and founding team to execute it, and leaving with a grant proposal and the backing to pursue it.

Our job is to lower the barriers to entry — both the cost of leaving your current career path and the risk that a good idea never gets off the ground — and to meaningfully raise the odds that the effort you launch succeeds.

Founders will have between six months and two years to research a problem and develop a pitch, but we hold that loosely — a Founder is done when their effort is ready to stand on its own, not when a calendar says so. The goal is for Founders to develop an ambitious plan as quickly as possible (early graduations welcome), complete with conditionally committed users/stakeholders, an excellent founding team, expert advisors, and organizational plans like milestones — completion consists of pitching funders like Coefficient Giving and launching with an initial grant of around $10 million.

Cohort 1: Securing AI Infrastructure

While future cohorts will tackle other neglected topics in navigating transformative AI, the first cohort focuses on securing AI infrastructure. This includes keeping the frontier AI systems out of the hands of actors who would steal, poison, or misuse them.

The more powerful AI gets, the more critical it is to have strong guardrails, authorization systems, and safety mechanisms around it. However, most people working on catastrophic AI risk don’t have the necessary infosecurity experience to address these problems.

We expect transformative AI will need to be secured against threat actors like nation-states, terrorists, insider threats, and even misaligned AIs themselves — actors who may have enormous resources, intentions to harm large numbers of people, or significant legitimate access. These may be the hardest, highest-stakes security problems you ever work on, and you'll be one of a very small pool of people qualified to work on them.

Examples of problems we're looking for owners on include:

-
Securing AI from exfiltration: protecting model weights and algorithms from theft.

-
Securing AI integrity: protecting models from tampering via data poisoning or other attacks.

-
Enabling verification of AI datacenter properties: demonstrations that would allow verifiable attestations that a data center is being used for inference or alignment research, rather than training.

- Adjacent problems in AI infrastructure security (e.g., disconnected model safety, secure compute verification, protocols for using compromised AI, detecting and responding to rogue deployments) may also be in scope where they bear on keeping frontier models secure.

You can read a series of one-pagers about these problems here.

Success in the first 6 months looks lik