AGI doomsday narratives sound like nobody did the permissions design

In Silicon Valley, whenever AGI comes up, people love to talk about uniting humanity or extinguishing it—like a review meeting where they only pitch vision without mentioning rollback plans. I actually find Fei-Fei Li's comment that "AGI is a marketing term" sounds more like something said by someone who has built systems.

From an architectural perspective, this kind of narrative has scalability issues. If you make the model the protagonist, risks get infinitely amplified until they become unverifiable. Fei-Fei Li isn't buying this hype because she consistently talks about data and scenarios. Back in the ImageNet era, she said the bottleneck wasn't the model but the data; now she places world models further back, arguing that spatial intelligence is the missing piece. This feels more like trying to find the foundation for the system.

Looking at feedback loops in recommendation systems, I often see models scoring brilliantly offline but crashing online. The reasons aren't mystical. Sample distributions shift, feature permissions aren't controlled, logs lack auditing, feedback links break, and the system starts hallucinating in production. These days when testing Claude, I care more about the execution environment. Especially when letting AI read files or send requests autonomously, sandboxes and whitelists are more reliable than prompts.

Fei-Fei Li certainly understands the risks. She worries more about a few people controlling compute, data, and direction, then using doomsday narratives to suppress discussion space. Public anxiety also stems from power concentration.

Boundaries matter more than the model itself. If a system can't even clarify its inputs, permissions, audits, and rollbacks, shouting "AGI" looks more like pumping up valuation.