The technology that could make an international AI slowdown possible
- Verification tools enabling international agreements on AI safety are viewed as the most likely path to a Nobel Prize, with a marginal AI safety researcher potentially becoming one of the top five global experts within a year if starting today.
- Retrofitting all data centers in China and the US is deemed completely impossible in the short term; instead, immediate focus is on less invasive software-only solutions and monitoring specific compute clusters, with a global pause on training expected as a starting point for deals.
- Near-term verification (within a few months) relies on "inference recomputation" and detecting power consumption fluctuations to distinguish between training and inference on small GPU clusters, while long-term solutions aim to utilize zero-knowledge proofs (zkML) to remove the need for sharing model weights, despite current "orders of magnitude" overhead.
- Hardware mechanisms, including FPGAs and flexible on-chip guarantees, are identified as future requirements for robust verification but face significant lead times, making them unavailable for deals needed within "two to 20 days," where shutting down data centers would be the fallback.
- Intelligence communities are expected to leverage "national technical means" such as infrared satellites, signals intelligence, and OSINT to detect hidden data centers, potentially utilizing a "two to 20 days" window for panic deals to gather information on GPU locations.
- Red-teaming efforts by groups like Lucid Computing and the Future of Life Institute are planned for the near future to test verifier/prover systems, supported by government funding expected this year and precedents from nuclear and chemical weapons conventions.
- Major AI labs like Anthropic are expected to release prototypes proving which model was running by September 30, 2026, though current internal efforts are considered insufficient relative to the stakes, with OpenAI noted for occasional verification-related job openings.
- Future verification may extend beyond simple on/off states to confirm specific safety training techniques, model versions, or the absence of secret training runs using API-level verification and hardware-backed logs to prevent tampering by advanced models.
- Sovereign compute is an absolute priority for the UK to avoid an existential threat from potential US domestic supply prioritization during a compute crunch, with the UK leveraging lower costs and noise levels for hardware R&D while the US focuses on domestic token supply.
- Rapid hardware prototyping is expected to accelerate via AI-driven design (e.g., "vibe coding"), aiming for iteration speeds comparable to Formula One's wing changes, with software-only solutions remaining a trade-off between invasiveness and adversarial robustness.
- A "moonshot R&D project" similar to Project Vela, requiring billions in investment, is deemed necessary for AI verification, with a global agreement potentially involving a "no training" clause and transparency dashboards from Chinese companies.
- Middle powers like the Netherlands, France, Germany, and Singapore are expected to demand verification tools to maintain leverage, while domestic verification for AI companies may emerge to ensure agents like Claude are not performing unauthorized training.
- Hardware design timelines for secure chips are estimated at 10 to 100 hours per task, with per-engineer token spending anticipated to match software companies next year, enabling "personalized hardware" and rapid deployment of RL environments.
- Unilateral slowdowns driven by domestic unhappiness or law enforcement are considered viable alternatives to sophisticated verification tools, though international cooperation remains the primary assumed focus for preventing arms races.