Debate intensifies over human bottlenecks for AGI safety

Safety researchers are increasingly debating architectural constraints that deliberately exclude autonomous coding capabilities from advanced AI training datasets. Proponents argue that preventing models from learning code execution or internet access creates a necessary human bottleneck against sandbox escapes and self-modification risks. This discourse follows recent high-profile security incidents allegedly involving AI agents, prompting calls for systems that generate pseudocode for human review rather than executable scripts. Critics counter that such restrictions may be technically unenforceable due to emergent capabilities and could severely limit scientific utility. The proposal suggests retaining domain expertise in physics and medicine while outsourcing all implementation to verified human operators. Industry stakeholders remain divided on whether capability suppression is a viable safety strategy or an impractical constraint that ignores economic incentives for autonomous systems. No major lab has yet adopted this restrictive training methodology for frontier models.