Leopold Aschenbrenner worked on OpenAI’s Superalignment team and founded the AGI-focused investment firm Situational Awareness LP. These lessons summarize his 2024 essays and interviews: forecasts and policy arguments, not established outcomes. His scenarios carry substantial uncertainty. — Stanford — Situational Awareness.

Visual summary of operating lessons from Leopold Aschenbrenner.

Part 1: The Path to AGI (2024–2027)

  1. On AGI Timelines: In 2024, he considered AGI by 2027 plausible, while acknowledging substantial uncertainty. — Counting the OOMs.
  2. On the GPT-4 Milestone: His preschooler-to-high-schooler comparison illustrated GPT-2–GPT-4 progress; it was an imperfect analogy, not a comprehensive intelligence measurement. — Counting the OOMs.
  3. On "Unhobbling" Gains: He distinguished larger training runs from unlocking existing capabilities through reasoning, tools and longer-horizon agency. — Counting the OOMs.
  4. On Orders of Magnitude (OOMs): He estimated roughly five orders of magnitude of effective-compute improvement over 2023–2027—not a guaranteed capability gain. — Counting the OOMs.
  5. On Mainstream Skepticism: He argued that public attention to AI did not necessarily mean preparedness for his much more disruptive scenario. — Parting Thoughts.
  6. On Personal Conviction: He described 2023 as the year AGI became tangible to him rather than merely theoretical. — Dwarkesh Podcast.
  7. On the "Drop-in Remote Worker": His remote-worker milestone meant systems that could independently plan, use company context and complete extended projects—not merely answer questions. — Counting the OOMs.
  8. On College Graduate Intelligence: In June 2024, he forecast models exceeding most college graduates by 2025–2026; this was a prediction, not a demonstrated result. — Dwarkesh Podcast.
  9. On Historical Precedent: He attributed rapid training-compute growth to expanding investment as well as chip improvements, rather than Moore’s Law alone. — Counting the OOMs.

Part 2: The Intelligence Explosion

  1. On Automating R&D: His feedback-loop argument starts with automating AI research, without first requiring automation of every other occupation. — The Intelligence Explosion.
  2. On Compressing Time: He envisioned automated research spreading across scientific fields and accelerating industrial development; this remained a conditional extrapolation. — The Intelligence Explosion.
  3. On the Jump to Superintelligence: He considered a sub-year transition from human-level AI to superintelligence plausible, not inevitable. — The Intelligence Explosion.
  4. On Alien Systems: He warned that successor systems could become increasingly unfamiliar as AI systems themselves designed their architectures. — The Intelligence Explosion.
  5. On Parallel Labor: Replicable AI researchers could work in parallel and share knowledge, unlike a fixed human research workforce. — The Intelligence Explosion.
  6. On Economic Production: He argued that automating labor could remove a constraint on capital accumulation and enable much faster industrial expansion. — The Intelligence Explosion.
  7. On the Endgame: His June 2024 scenario put superintelligence before 2030, with profound further changes in the following decade. — Parting Thoughts.
  8. On Decisive Lead: He argued that a short superintelligence lead could become a major military advantage. — The Free World Must Prevail.

Part 3: Security and Espionage

  1. On Lab Vulnerabilities: In 2024, he judged frontier-lab security inadequate for the value and sensitivity of the systems being developed. — Lock Down the Labs.
  2. On State Actors: He urged labs to plan for sophisticated state espionage, not just ordinary commercial threats. — Lock Down the Labs.
  3. On Model Weights: He feared stolen frontier weights could erase a lead precisely when automated research began accelerating. — Lock Down the Labs.
  4. On Infiltration: He expected Chinese intelligence efforts against American labs to intensify as AGI became strategically important. — Lock Down the Labs.
  5. On Algorithmic Secrets: He distinguished immovable computing infrastructure from algorithmic discoveries that could leak through a brief conversation. — Lock Down the Labs.
  6. On Locking Down the Labs: His proposed response included tighter personnel vetting, compartmentalization and stronger isolation of sensitive systems. — Lock Down the Labs.

Part 4: Geopolitics and Great Power Competition

  1. On the Return of History: His scenario treated AGI as a renewed source of great-power rivalry, not simply another consumer-technology market. — The Free World Must Prevail.
  2. On the Race with China: He warned that competition with China could escalate beyond an industrial race into military confrontation. — The Free World Must Prevail.
  3. On Military Obsolescence: He conjectured that superintelligence might undermine even nuclear deterrence; this was a strategic hypothesis, not demonstrated military capability. — The Free World Must Prevail.
  4. On Democracy vs. Dictatorship: He favored democratic leadership because authoritarian control of superintelligence might entrench repression and suppress future correction. — The Free World Must Prevail.
  5. On Nationalization: He predicted state involvement, while explicitly allowing defense-contracting arrangements rather than literal nationalization. — The Project.
  6. On the "Sputnik moment": He argued that earlier preparation would be safer than government intervention only after an intelligence explosion began. — The Project.
  7. On First-Strike Temptation: He feared rapid technological change could create destabilizing incentives for pre-emptive action. — The Free World Must Prevail.
  8. On Authoritarian AI: He worried that automated forces could help dictators suppress opposition without depending on potentially disloyal human subordinates. — The Free World Must Prevail.

Part 5: Infrastructure and the Trillion-Dollar Cluster

  1. On the Compute Bottleneck: He emphasized that AI scaling requires physical capacity: power, datacenters, chip fabrication and packaging. — The Trillion-Dollar Cluster.
  2. On the Trillion-Dollar Cluster: His 2024 extrapolation envisioned individual training clusters costing hundreds of billions, potentially reaching a trillion dollars by 2030. — The Trillion-Dollar Cluster.
  3. On Electricity Demands: His scenario included a roughly 100-GW cluster—an enormous addition to electricity demand, not a facility he had observed. — The Trillion-Dollar Cluster.
  4. On Industrial Mobilization: He envisioned mobilizing national industrial resources to expand computing capacity, rather than relying solely on software progress. — The Project.
  5. On Bypassing Regulation: He advocated faster permitting and more energy supply, including natural gas, rather than letting climate commitments block domestic clusters. — The Trillion-Dollar Cluster.
  6. On GPU Scale: His end-decade cluster estimate used around 100 million H100-equivalent chips—not a prediction of that many literal H100s. — The Trillion-Dollar Cluster.
  7. On Infrastructure as Destiny: He treated industrial capacity to build compute as one strategic advantage, alongside algorithms and security—not a guarantee of victory. — The Free World Must Prevail.
  8. On Economic Moats: He expected extraordinary capital requirements to draw major technology companies and national-scale financing into the race. — The Trillion-Dollar Cluster.

Part 6: Governance and the Manhattan Project Model

  1. On the Manhattan Project Model: His Manhattan Project analogy questioned whether startup improvisation was adequate for technology with major national-security consequences. — The Project.
  2. On the Role of the State: In 2024, he forecast some form of government AGI project by 2027–2028; he did not establish that it would occur. — The Project.
  3. On Extreme Competence: He stressed competent leadership across industrial mobilization, security, safety and geopolitical crises; a government label alone would not ensure success. — The Project.
  4. On Regretting the Tech: He attributed nuclear-project regret to the technology’s destructive potential, not necessarily to government coordination; this was his interpretation. — Dwarkesh Podcast.
  5. On Monopoly on Violence: He argued that potentially overwhelming coercive power should remain subject to democratic state authority, not a private executive’s discretion. — The Project.
  6. On Avoiding a Race to the Bottom: He saw coordination as a way to resist competitive pressure to sacrifice security and safety. — The Project.

Part 7: Alignment and the Control Problem

  1. On the Unsolved Problem: He described reliable control of much smarter systems as an unsolved technical problem. — Superalignment.
  2. On Scaling Alignment: He argued that human feedback becomes insufficient when people cannot reliably judge a more capable system’s work. — Superalignment.
  3. On the Risk of Going Off the Rails: He feared a rapid capability jump would leave too little time to test and correct alignment methods. — Superalignment.
  4. On Superalignment Necessity: He urged substantial resources for automated alignment research, rather than devoting all available compute to further capabilities. — Superalignment.
  5. On Agentic Risks: He warned that long-horizon agents could learn harmful strategies that short chatbot interactions would not reveal. — Superalignment.
  6. On Reliable Control: His control problem concerns dependable obedience and honesty; choosing which values to instill is a separate question. — Superalignment.
  7. On the Technical Difficulty: He was optimistic about technical solutions, while warning that geopolitical urgency could undermine careful safety work. — Superalignment.
  8. On Robotics and Physical Risk: He expected AI-led research eventually to advance robotics, extending its consequences into the physical world. — The Intelligence Explosion.
  9. On Ethical Weight: He framed powerful AI as a solemn responsibility requiring seriousness about risks, not simply excitement about new products. — Parting Thoughts.

Part 8: Personal Agency and Strategic Thinking

  1. On Situational Awareness: He believed a small network understood the coming stakes, making individual responsibility unusually consequential. — Parting Thoughts.
  2. On the "Chill Time": In 2024, he called the present relatively calm and described his anticipated transition as stressful. — Dwarkesh Podcast.
  3. On Career Risk: He urged people to consider his scenario seriously despite its strangeness, rather than assuming someone else would handle it. — Parting Thoughts.
  4. On Founders' Strategy: He criticized wrapper businesses that depended on current model limitations persisting, rather than planning for much more capable agents. — Dwarkesh Podcast.
  5. On Growth and Risk: With Philip Trammell, he modeled how faster growth could lower lifetime catastrophe risk; policy failures or direct risk effects can reverse that result. — Existential Risk and Growth.
  6. On Collective Survival: He hoped humanity could navigate the transition safely while acknowledging the burden and uncertainty of his own scenario. — Parting Thoughts.
  7. On the Future: He asked readers to consider the stakes if his scenario proved right, while admitting large errors and alternative possibilities. — Parting Thoughts.