Can Internal Model Transparency Slow the AI Race?

Summary

This essay argues that competition among frontier AI labs may push them toward increasingly rapid and risky automation of AI research, including recursive self-improvement (RSI). Its proposed response is internal model transparency (IMT): once a lab deploys a model for internal use, it would also have to serve that model to researchers at other labs, preventing internal models from remaining a unique competitive advantage. In the author’s example, shared access to a lab’s strong coding model would let another lab use its own data or domain strengths to catch up or surpass the original lab in science-oriented models. The resulting incentive could be for labs to invest less in frontier coding capabilities that automate AI research and more in other sources of advantage, slowing development through changed payoffs rather than an explicit slowdown mandate. The proposal is also intended to make it harder for one company to accumulate the most powerful model and to let outside labs or trusted evaluators identify problems in internal models. The author says a US-only rule could affect the main US frontier race, while a US-China arrangement would reduce leakage but require reciprocal verification. Because verification would focus on models internally available to researchers, the author argues that it could be narrower than compute or total research transparency monitoring. Important implementation questions remain, including which firms are covered, how to detect weakened external versions, and how to prevent prohibitive pricing. The essay does not present IMT as a replacement for direct safety rules. It acknowledges that IMT could initially accelerate aggregate progress by helping lagging firms, that monitoring internal deployment is technically difficult, and that revenue from capable coding models would continue to motivate RSI even if one incentive were weakened.