A context window for democracy

Senator Clara López floats an idea at Petro: AI for direct government participation ▸ 2:57. His reading of it: every citizen writes one short, concrete statement of what the country most needs, and an AI distills the whole nation’s input the way Gemini distilled their priorities in the REAPRA introspection, “haría la destilación de qué es lo que más se necesita de acuerdo con lo que la gente dice” ▸ 6:15. Risks acknowledged up front: social-media-manipulated opinion could have people voting against themselves ▸ 5:02; Julia’s counterweight: people finally feeling heard.

Then the part that makes this an engineering entry: the Vibe Coders group sizes the machine. Assumptions on the napkin: 140 characters per citizen, Twitter-classic; 38 million Colombians of voting age ▸ 8:18. Result: analyzing every voice in one pass needs a context window of 1,000 to 1,800 million tokens ▸ 8:56, against today’s frontier of one or two million ▸ 12:24. His zeros-stripping summary: “uno, a dos, a mil: todavía no llegamos” ▸ 12:51.

antes de opinar sobre la idea, calcula su ventana de contexto →

But the group also finds the standard fix: shard the country, distill by municipality, then by department, then merge, every level fitting a normal window ▸ 13:05, and someone even prices the billion tokens at roughly two to three thousand dollars ▸ 14:18. Map-reduce for democracy: technically feasible today, for the cost of a used motorcycle. The entry’s quiet lesson is the method, not the politics: a WhatsApp group turned a senator’s abstraction into assumptions, a magnitude, and a budget in one afternoon, which is exactly the distance between an opinion and an estimate…