Zero to One in AI Safety: Halcyon's Mike McCormick on Launching 30 New Orgs & the Founder Bottleneck

Episode Summary

Executive Summary: Mike McCormick explains Halcyon’s model for accelerating AI safety, cybersecurity, and biosecurity by finding exceptional founders, giving career-transition support, and launching new orgs across a broad but mission-focused portfolio. He argues talent, not money, is the main bottleneck, and that society must rapidly build verification, oversight, resilience, and coordination institutions if AI progress continues on a short timeline.

Main Topics: Halcyon’s mission and operating model (Priority: 5/5): Halcyon combines nonprofit grantmaking and venture investing to help talented founders launch AI safety, cybersecurity, and biosecurity organizations, focusing especially on the six months before and after formation. Examples of supported organizations (Priority: 5/5): McCormick details several portfolio orgs—Goodfire, AIUC, Transluce, Hadrian, and others—to show how Halcyon backs interpretable AI, standards/insurance, scalable oversight, and pandemic preparedness. Timelines and the urgency of institution-building (Priority: 5/5): He repeatedly frames the world as being on an AI-2027-like timeline, arguing that even if timelines are uncertain, there is still time and strong need to start new organizations now. Verification, control, and resilience as critical gaps (Priority: 5/5): McCormick says pacing agreements, secure deployment, and AI oversight all depend on verification technologies across cryptography, software, hardware, and diplomacy, and that these fields are still nascent. Career transitions, founder quality, and talent bottlenecks (Priority: 4/5): He emphasizes that the main constraint is high-quality founders and leaders, not capital, and that Halcyon often uses grants to buy people time to decide whether to found, join, or pivot into AI safety work. Governance, incentives, and mission preservation (Priority: 4/5): The conversation explores how to preserve mission alignment through PBCs, board structures, founder selection, and other governance tools, while acknowledging no structure is a silver bullet. Public communication, coordination, and societal adaptation (Priority: 3/5): McCormick is interested in public sensemaking, coalition-building, and new mechanisms for converting ideas into institutions, but is skeptical of overly centralized or purely technical solutions to coordination problems.

Key Arguments: AI progress is moving so quickly that society needs many parallel institutions, not just a few labs, to solve safety, verification, cybersecurity, biosecurity, and oversight. The biggest bottleneck in AI safety is not money but talent—especially founders and experienced operators who can build durable organizations. Career transition grants are high leverage because they let promising people spend time figuring out whether to found, join, or pivot into a mission-driven role without immediate financial pressure. Goodfire illustrates Halcyon’s model: early support plus network building helped create a top interpretability lab that has since scaled into a major company. AIUC is designed to create a market incentive for safer AI through standards plus insurance, making safety legible to enterprise buyers. Biosecurity is comparatively more legible than alignment because it has clearer historical analogues and deployable interventions like PPE stockpiles and air-cleaning tech. Verification is foundational to any serious pacing or monitoring regime; without technical proof mechanisms, regulatory or diplomatic agreements will be weak or unenforceable. Open-weight models make the world harder to secure, so the response should include both hardening the world and exploring limited technical interventions. Mission preservation requires both structure and people; legal forms like PBCs help, but founder motivations and judgment matter more in crisis. The field needs more organizations and more diversity of approaches, not repeated debates about whether one area is slightly more important than another. Good founders often come from adjacent elite backgrounds—CTO, COO, cybersecurity, cryptography, math, or operations—and can be redirected into safety work. AI safety money may soon become abundant through employee liquidity and public markets, so the next bottleneck is having enough strong projects and leaders to fund.

Data Points: Organizations launched: ~30 - Halcyon says it has helped launch roughly 30 organizations in three years. Total funding raised by launched orgs: ~$500 million - Collective funding raised by organizations Halcyon helped launch. First career-transition grants: 2 grants - Given to Eric Ho and Dan Balsam before Goodfire was founded. Goodfire valuation: above $1 billion - McCormick says Goodfire recently raised a Series B at a valuation above a billion dollars. Goodfire Series B: well over $100 million - Approximate size of the round discussed on the episode. Hadrian initial check: $50,000 - Halcyon’s tiny early grant to help launch the PPE stockpiling company. PPE stockpile estimate: a few hundred million dollars - Rough cost cited for stockpiling enough PPE to keep essential workers working during a severe pandemic. School air-quality company initial customers: schools - McCormick says schools are the initial customers for a sister nonprofit/company focused on indoor air quality. Career-transition grant sizes: $50,000 / $100,000 / $200,000 - Typical support sizes mentioned for people exploring AI safety career pivots. AIUC standard: 50-something checkboxes - He describes the AIUC standard as having a large checklist of safety and security requirements. AIUC customer example: 11 Labs - Used as an example of a growth-stage AI application company adopting the standard and insurance offering. Founders retreat size: 25 founders - He mentions hosting a verification-focused retreat for 25 founders.

Pivotal Quotes: "It is the best time to start an AI safety company is 20 years ago, and the second best time is today." — Mike McCormick: Explaining why short timelines do not negate the value of starting new organizations now. "The biggest bottleneck to solving this problem or to sort of speedrunning these fields is a shortage of really amazing founders and leaders." — Mike McCormick: Summarizing Halcyon’s core theory of change. "We are speedrunning six or seven industries worth of AI revolution all at once." — Mike McCormick: Describing the breadth and urgency of the organization-building task across safety domains.

Implications: Listeners should expect AI safety to look less like one research problem and more like an ecosystem of new institutions. The episode is a call for ambitious operators, founders, and domain experts to pivot now and help build the technical and governance infrastructure AI will require.