The Balance of AI Power: Anton Leicht on Politics, Pacing Deals, and Muddling Through Well

Episode Summary

Executive Summary: Anton Leicht argues today’s AI is already concerning but not yet broadly catastrophic; the real danger is rapid capability gains, especially in bio and autonomous agentic systems, outpacing oversight. He favors a temporary pause or slowdown, but thinks geopolitics, market incentives, and U.S.-China rivalry make coordination hard. The conversation explores frontier governance, export controls, lab self-regulation, Europe’s compute strategy, and how AI may disempower most countries even as living standards rise.

Main Topics: How dangerous today’s AI really is (Priority: 5/5): Leicht says current models are not yet broadly dangerous in deployment, but the trendline is alarming: internal lab acceleration, autonomous agents, and especially bio-capable models could soon cross into genuinely dangerous territory. Pause, slowdown, and geopolitical feasibility (Priority: 5/5): He supports a temporary pause in principle to study alignment and control, but argues that a U.S.-China frontier pause is geopolitically lopsided because it would likely advantage China unless paired with stronger export-control enforcement and broader concessions. Frontier governance and lab oversight (Priority: 5/5): The discussion centers on third-party auditors, incident investigations, continuous oversight, and the idea that the U.S. government could pressure labs into self-governance frameworks or allow better access for evaluators like Redwood and METR. Nation states, control, and AI (Priority: 4/5): Leicht sees AI as a challenge to liberal-democratic state capacity, especially if personal superintelligence reduces the state’s monopoly on violence and information, though authoritarian systems may adapt better via controlled diffusion and limited access. Global power dynamics and Europe’s strategy (Priority: 5/5): He argues most countries will be disempowered in the AI age, but Europe can use its market size and semiconductor leverage (notably ASML) to negotiate compute-for-access deals with U.S. hyperscalers and preserve access to frontier models. U.S. policy bottlenecks and executive actions (Priority: 4/5): Because Congress is likely gridlocked, Leicht favors executive-branch steps: standardizing incident investigations, enabling continuous audits, clarifying antitrust/export-control boundaries, and creating credible coordination mechanisms among labs. Physical deployment, data centers, and future compute (Priority: 4/5): He expects data-center buildout to continue in the U.S. despite local backlash, and thinks the economics of AI may shift further with robotics and eventually space-based compute, which could alter geopolitical leverage and middle-power strategies.

Key Arguments: Current AI systems are not yet broadly dangerous, but they are close to thresholds where they could meaningfully accelerate lab development or become dangerous in bio and cyber applications. Misalignment behaviors already observed in models look like the harder future problem, making a temporary pause or slowdown more justified than it once seemed. A U.S.-China pause on frontier scaling alone would be asymmetrical and likely favorable to China unless paired with meaningful export-control enforcement and limits on Chinese semiconductor indigenization. Independent evaluators should have more formal, repeatable access to labs for incident investigations and continuous oversight, rather than relying on voluntary goodwill. The president could materially improve AI governance by pressuring top labs to agree on self-governance and by clarifying what antitrust and export-control actions will not be used against safety collaborations. Most of the world will be economically richer yet politically less empowered in an AI-dominated future, with many countries becoming dependent on frontier AI powers. Europe has enough economic and industrial leverage to negotiate compute-for-access deals, especially if it uses ASML and allied supply-chain power strategically. The U.S. is likely to remain the key provider of frontier AI access in the near term because China cannot yet offer comparable compute or chip supply. AI buildout will continue because data centers are flexible to site, and the economics of compute, not political backlash alone, will keep the expansion going. Robotics and space-based compute are plausible future inflection points, but they also amplify the need to think about governance beyond terrestrial data centers. A doomsday trade strategy is hard because many catastrophic scenarios may look economically good until a sudden collapse occurs. The best realistic outcome is not utopia but effective muddling-through: preserving institutional balance, preventing excessive concentration of power, and capturing AI’s upside without losing control.

Data Points: Anthropic multiple: 30x revenue - Used to illustrate that AI-company valuations are not necessarily absurdly high, but still sensitive to expectations about future demand and deployment. Temporary pause length: 6 months - The hypothetical pause discussed as a window to study alignment and control while freezing frontier scaling. Potential AI disempowerment risk: ~10% - Leicht’s broader ‘doom’ estimate when including stable authoritarianism and long-term human disempowerment, not just extinction. Human extinction risk: very low - He distinguishes extinction risk from broader catastrophe, saying extinction is much less likely than disempowerment outcomes. Athena client time savings: 15 hours/week - Sponsor mention in the transcript, not part of the core discussion but a concrete productivity figure. Frontier Act status: did not pass this Congress - Referenced as a strong but ultimately unsuccessful legislative vehicle for AI oversight. Texas moratorium characterization: not a practical moratorium - He argues that in practice Texas still remains broadly buildable for data centers. Time for domestic service robots: around 2030 - His rough technical-maturity estimate for widespread home service robots, subject to social and political resistance. Space compute timeline: starting around 2029 - He expects some compute deployments in space by then, especially inference-oriented systems. U.S. chip lead leverage: decisive current advantage - He argues the U.S. cannot afford a pause that lets China close the gap on semiconductors and frontier AI supply chain control. Data center buildout in Europe: compute-for-access deal - European strategy proposal: host U.S. data centers in exchange for assured access to frontier models. Potential AI-related market reaction: could be a downward correction or crash - He warns that a pause or slowdown could spook markets because investors may not know when/if frontier scaling resumes.

Pivotal Quotes: "I think not very dangerous at the current capability level in most of the ways people are talking about." — Anton Leicht: His initial calibration on present-day AI risk. "A temporary scaling pause would indeed be valuable." — Anton Leicht: He endorses a pause as a way to study emerging misalignment and control problems. "My best case scenario is not a grand utopian vision, but simply the hope that we might muddle through effectively." — Anton Leicht: His closing view of the most realistic positive outcome.

Implications: The interview argues for urgent but pragmatic AI governance: stronger audits, credible U.S. executive action, and international access deals. Expect more pressure on labs, more geopolitics, and a future where the key question is whether institutions can preserve control while AI capability keeps rising.