Institute for Progress, August 2025 — introduction to “The Launch Sequence”
The Core Argument
- AI capabilities are compounding toward something transformative, but technological trajectories aren’t fate: AI doesn’t automatically solve the most important problems first, nor neutralize the risks it creates.
- AI progress is path-dependent — where and in what order capabilities are developed may matter as much as which ones are developed.
- As the world’s most powerful democracy controlling much of the AI supply chain, the US has both the responsibility and the power to proactively shape AI development; there is “no realistic plan B” (only the US and China can build the technological rails of the 21st century).
Two Problems to Solve
- Benefits may not come fast enough: markets undersupply public goods (study replication, unpatentable treatments, basic research), and structural barriers — litigation vetoes on clean energy, slow FDA approval, locked-up datasets, grant paperwork consuming half of researchers’ time — throttle AI-enabled science.
- Risks industry won’t solve: coding agents can exploit critical infrastructure; medical-research AI can help engineer bioweapons. AI safety research is a niche (~2% of papers, ~$100M/year), and too little work assumes dangerous capabilities will diffuse widely.
Historical Precedents for Shaping Technology
- Nonproliferation: nuclear weapons (classification, Manhattan Project, export controls) — proliferation was consistently slower than experts predicted.
- Selective acceleration: the Human Genome Project pre-empted broad gene patents by creating a public-domain resource.
- Defensive acceleration: Operation Warp Speed delivered an mRNA vaccine 10x faster than any before.
Four Principles for Shaping AI Progress
- 1. Leverage the “jagged frontier”: AI advances unevenly — fields with tokenizable knowledge, verifiable solutions, and easy deployment (math, coding, parts of biology) accelerate first. Predict where offense or defense will dominate: cyber may be defense-favorable; parts of biosecurity are offense-dominant and need nonproliferation (e.g., DNA-synthesis screening). Track capabilities for an “adaptation buffer” and unlock high-value datasets.
- 2. Don’t neglect the costs of stalled progress: achievable medical technologies (malaria, TB vaccines) could save ~3.6 million lives/year; delaying beneficial technology has real human costs that risk-focused strategies ignore.
- 3. Redesign how science works: the funding system (20-month grant waits, half of PI time on paperwork) is unfit for AI-driven research, which needs compute, engineering teams, and “team science” à la Arc Institute and FutureHouse; agencies should use flexible mechanisms like block grants and fast grants.
- 4. Adapt to uncertainty while reducing it: technological forecasting is unreliable (von Neumann got semiconductors right, cheap nuclear wrong; transformers and GPUs emerged from ads and gaming). Build measurement/evaluation ecosystems and pluralistic political coalitions (like the CHIPS Act’s science + security blend).
The Launch Sequence
- The essay introduces a collection of concrete projects — unlikely to happen by market default, achievable by 2030 — to accelerate science and strengthen security (automated cyber-defense, pathogen detection, formal verification, new science-funding models, and more).
- Explicitly framed as filling the gap left by Amodei’s “Machines of Loving Grace”: the vision of what AI could solve, plus the concrete steps to get there.
- Strategy over master plan: in von Neumann’s words, “a long sequence of small, correct decisions.”
“The US is the R&D lab of the world, and we should act like it.”