The fast-follow problem is the strategic question of whether AI breakthroughs can be quickly imitated ("fast-followed") by competitors, or whether catching up is slow and costly. It is the second axis of the Eight Worlds Framework, where it serves to distinguish worlds in which advantage flows from frontier research and development from worlds in which advantage flows from deployment speed and diffusion.
Conditions favoring easy catch-up
Under the easy-catch-up case, breakthroughs spread quickly through several channels: espionage and theft of model weights and training data; model distillation, in which smaller systems are trained to imitate larger ones; leaked weights, where trained model parameters are stolen or released; algorithmic diffusion, as new techniques spread through publications, hiring, and reverse engineering; and innovative training on older hardware that finds efficiency gains and reduces compute requirements.
Where catch-up is easy, advantage comes from deployment speed and diffusion rather than frontier R&D alone, and investment incentives weaken because profits vanish if innovations are quickly copied. Under these conditions, open-sourcing may be the best strategy.
Conditions favoring hard catch-up
Under the hard-catch-up case, frontier capability depends on a full stack that is difficult to reproduce: proprietary hardware and chip access, institutional expertise built over years, vast and often unique datasets, vibrant talent ecosystems, and structural factors that cannot be foreseen. The model, as a software layer, may be copyable, but the infrastructure behind training and inference is not.
Where catch-up is hard, defending the foundational inputs matters most, export controls are more effective, and frontier investment has higher returns.
Relation to policy
The position of the fast-follow axis bears on several policy instruments. It determines the value of Export Controls (AI), which is high when catch-up is hard and limited when it is easy. It shapes AI Diffusion strategy, since when catch-up is easy, diffusion becomes the primary contest. It also interacts with Compute Governance: if compute is the hard-to-replicate layer, controlling it is decisive.
Empirical evidence
Simon Willison's annual LLM reviews document the pace of catch-up over three years. In 2023, only OpenAI had a GPT-4-class model. In 2024, 18 organizations beat GPT-4, GPT-4-class models ran on a laptop, and DeepSeek was possibly trained for under $6M. In 2025, top-ranked Chinese open-weight models appeared, along with reasoning models from every major lab. This trajectory suggests that catching up at the model level has been easier than expected, consistent with the easy-catch-up side of the Eight Worlds Framework. As the Narayanan and Kapoor normal technology framework argues, however, model-level catch-up is distinct from application-level or deployment-level catch-up, which may remain hard. (Source: Simon Willison's Year-in-Review LLM Posts (2023–2025))
The DeepSeek R1 release is a concrete instance of fast-following: R1 reproduced reasoning-model capabilities shortly after o1, via pure-RL and distillation.
Measured lag between open and closed models
Attempts to quantify the follow-time directly have produced a range of estimates that depend on the benchmark set used. Epoch AI, using its Epoch Capabilities Index (ECI), estimated in late 2025 that frontier open-weight models lagged the most capable closed models by around three months (Source: epoch.ai), and reported that since January 2026 the most capable open-weight models have lagged frontier closed models by an average of four months, or 8 ECI points (Source: epoch.ai). A LessWrong analysis argues the gap is larger than public-benchmark comparisons suggest: on private benchmarks, whose data is not publicly accessible and therefore harder to train against, open models measured roughly 8–10 months behind the frontier (Source: lesswrong.com). The disagreement between public- and private-benchmark estimates is itself evidence bearing on the fast-follow question, since it suggests part of the apparent catch-up reflects benchmark contamination or targeted fine-tuning rather than replicated general capability.
Sources
- Geopolitics in the Age of Artificial Intelligence (Foreign Affairs, 2026-01-27)
- Simon Willison's Year-in-Review LLM Posts (2023–2025)
- DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning — concrete instance of fast-following: R1 reproduced reasoning-model capabilities shortly after o1, via pure-RL + distillation.