efesantos/cost-of-intelligence

What does intelligence cost in minutes of local work? A multi-country token affordability index + adoption linkage (extends The African Language Tax) # Cost of Intelligence **What does intelligence actually cost people — in minutes of their own work — across languages and countries? And does that cost predict who adopts AI?** This repo treats the token as an economic unit and builds four reusable instruments, then tests, correlationally, whether affordability gates adoption. It **extends "The African Language Tax" (arXiv 2606.24460)** from cost measurement to *affordability* and *adoption consequences*, and from African languages to the Lusophone/Hispanophone Global South. See `docs/novelty.md`. ## The headline (live data, reproducible) Minutes of local work to buy **1M English-equivalent tokens/month** (OpenAI o200k family, PPP wages) — live OpenRouter prices + World Bank wages + FLORES-200 fertility: | | minutes | | minutes | |---|--:|---|--:| | 🇺🇸 US (English) | ~11 | 🇿🇦 South Africa (Zulu) | ~112 | | 🇵🇹 Portugal (Portuguese) | ~24 | 🇳🇬 Nigeria (Yoruba) | ~238 | | 🇲🇽 Mexico (Spanish) | ~53 | 🇰🇪 Kenya (Swahili) | ~218 | | 🇪🇬 Egypt (Arabic) | ~72 | 🇪🇹 Ethiopia (Amharic) | **~1,620** | The same answer costs an Ethiopian worker on Amharic **~140× more minutes of labour** than an American on English — fertility (more tokens per meaning) multiplied by lower wages. *(Numbers regenerate with `make all`; exact values carry CIs and date stamps.)* ## The four instruments | Instrument | Definition | |---|---| | **Real intelligence exchange rate** `e_real` | `(1e6 / p_blend) ÷ fertility` — English-equivalent tokens per USD | | **Intelligence Wage** `IW` | `e_real × hourly_wage` — cognition an hour of local labour buys | | **Affordability Index** `work_minutes` | `(bundle · fertility · p_blend / 1e6) ÷ wage × 60` — the Big Mac "minutes of work" method, done for tokens | | **Language Tax** `surcharge` | `fertility − 1` — the regressive token surcharge | `p_blend = (1·input + r·output)/(1+r)`, default `r=3` [INFERENCE, swept in B-9]. Everything is reported **per tokenizer family** — no single "the token." ## Adoptio …