模型之外还有多少潜能：Alex Zhang 谈递归语言模型、执行框架与研究大赌注

Expert knowledge can make AI search more efficient by steering it toward promising solutions and providing a strong verifier, rather than simply increasing brute-force token expenditure.

Academic researchers can use their freedom to pursue uncertain, initially unfashionable ideas instead of competing directly with frontier labs on the same resource-intensive scaling strategy.

Changing a language model's output space can create different latency and capability trade-offs, so autoregressive text generation should not be treated as the only useful interface to linguistic knowledge.

A compositional harness can help a model reuse the same high-level program across different tasks and lengths by delegating easier local problems to subagents.

An RLM stores task context in an external code environment and expresses tools and recursive subagent calls as functions, rather than placing all working information inside one expanding prompt.

A model's effectiveness in an unfamiliar harness depends partly on the workflows it was trained to follow, so new recursive designs need meaningful training and evaluation rather than branding alone.

Persistent subagents and a modifiable harness can preserve work across interactions, but the system still needs clear communication rules and visibility into what each agent is doing.

A large agent swarm must coordinate useful exploration and converge on a verified answer, because adding agents can otherwise produce expensive redundant search without proportional progress.

Open-ended generation creates a selection problem as well as a search problem, because a system must identify meaningful discoveries among large amounts of weak or irrelevant output.

Capability overhang is a hypothesis to test through reliable long-running work, where better task structure and scheduling may unlock more value without requiring a new frontier model.