Interview, Fireside Chat, Conference Presentation
The Current Reality of American AI Policy: From ‘Pause AI’ to ‘Build’
- The US has shifted from a restrictive AI posture (dominant 2021–2023) to an innovation-focused "Action Plan," reversing 40 years of precedent where the US balanced security with rapid deployment.
- Under the Biden administration, the discourse was dominated by "existential risk" narratives, exemplified by the 2023 "pause AI" petition and the California SB 1047 bill, which sought to regulate open-source AI models before they matured.
- SB 1047 proposed holding developers legally liable for "mass casualty events" (defined as 3+ deaths or overwhelmed medical systems) resulting from downstream use of open weights, a provision the speakers argue would create a fatal "chilling effect" on startups.
- Critics noted that the "nuclear weapon" analogy used against open source was flawed because AI is dual-use; unlike F-16s, the fundamental components (algorithms, math) cannot be legally withheld without hindering the US's ability to lead.
- The "Baptist-bootlegger" dynamic emerged where true believers in existential risk aligned with opportunistic tech entities to advocate for restrictions that inadvertently disadvantaged US innovation.
- DeepSeek's release of open-source reasoning models (e.g., DeepSeek Math V2, R1) served as a catalyst, debunking the claim that the US was significantly ahead of China and revealing that US self-regulation had actually allowed China to close the gap.
- The speakers argue that the burden of proof should rest on those making extraordinary claims about AI risks; historically, the US has not required liability for downstream uses of foundational research (e.g., internet, nuclear).
- A "vibe shift" occurred in late 2024 where the "silent majority" of engineers, VCs, and academics joined the discourse, moving the conversation from theoretical worst-case scenarios to empirical, pro-innovation pragmatism.
- The new Action Plan is co-authored by technologists, a first, aiming to bridge the gap between DC policy and Silicon Valley realities by representing diverse subcultures within the tech ecosystem.
- The document frames AI as a "new frontier of scientific discovery" rather than an "arms race," prioritizing applications in physics, chemistry, and material science to inspire the next generation of researchers.
- A significant omission in the Action Plan is the lack of direct funding or major investment strategies specifically for academia, despite its historical role as the bedrock of US computing innovation.
- The plan's most sophisticated element is the proposal to build an "AI evaluation ecosystem" to empirically measure risk before implementing regulations, rather than assuming danger exists.
- The speakers reject the "biological system" argument that AI cannot be deployed until fully "aligned" or understood, noting that society already manages complex, partially understood systems like electricity and human labor.
- The "opportunity cost" of slowing AI is highlighted as a major risk, with delays in biological discovery potentially costing lives that AI could otherwise save within months.
- Open source in AI is predicted to follow an "open core" business model: companies release smaller "open weights" for distribution and ecosystem building while retaining proprietary access to larger, more complex models.
- Closed-source and open-source AI are emerging as distinct markets: closed-source serves developers needing frontier capabilities, while open-source serves sovereign/AI governments and regulated industries requiring on-prem deployment and control.
- The speed of market entry in AI is unprecedented, with 20-something founders capable of generating tens to hundreds of millions in revenue within two years, making it dangerous to wait for market clarity.
- The speakers argue that "marginal risk" is not a new category requiring new laws; existing frameworks for network, stochastic, and computer systems risk are sufficient if properly applied to AI.
- The US must avoid the mistake of "locking down" technology to maintain a lead, as historical precedents (nuclear, internet) show that open collaboration and rapid iteration are superior strategies for long-term leadership.