Jev as a Really Smart Switch Statement

That is the strongest objection in the replies and the post does not answer it. Flurry's counterweight is shape and cost: the proposing model keeps content while the choosing step becomes a fast typed call that returns probabilities. Whether that split beats letting the proposer decide is untested in the thread.

Effectively yes, with probabilities attached. A reply says exactly that, and the post concedes the new part is rebranding. The difference from a conventional classifier is the intelligence inside the switch rather than a new interaction model, and the author's own line is that AGI implications are far fetched.

Five documented limits: no code writing, no natural language, no step-by-step reasoning or visible work, no output you did not define in advance, and no more than roughly 255 options in a single call. Anything needing free text or an open-ended answer has to stay with an LLM.

As the routing and gating step underneath them: pick the branch, tool, model, or sub-agent, judge an LLM's output, and label large row sets. Flurry's shape is that an LLM proposes options, Jev decides, and code executes, which keeps irreversible actions in deterministic code.

Nathan Flurry, who calls himself the actor guy and says he founded Rivet, a YC W23 and Speedrun company. So the framing comes from a systems and infrastructure builder, which is why it reads as a tool-class argument rather than a model announcement.

That it is incredibly far-fetched, though he does not want to discount the applications it enables. That is the tone of the whole post: it deliberately strips the launch hype down to a switch statement while still arguing the tool class is genuinely useful.

No. It is a framing post with no latency, cost, or accuracy figures and no benchmark. The only quantity in it is the roughly 255-option ceiling on a single call. Everything else is a list of capabilities and limits rather than measurement.