Power to the people: How LLMs flip the script on technology diffusion (karpathy.bearblog.dev)

LLMs have upended the typical top-down diffusion of technology by delivering outsized, immediate benefits to everyday users rather than first concentrating value inside governments or large firms. ChatGPT—and similar models—are cheap, fast, language-native, and extremely accessible (ChatGPT reached ~400 million weekly active users), enabling people to write, code, translate, tutor, summarize and research at a “quasi‑expert” level across many domains. For individuals this is a dramatic capability multiplier: people can perform tasks they previously needed specialists for, with almost zero onboarding cost. Organizations benefit too, but more slowly and modestly because LLMs are broad but shallow and fallible. Corporations and governments confront added complexity—legacy systems, integrations, security, compliance, brand constraints, and coordination costs—plus low tolerance for hallucinations and regulatory risk. The future distribution of benefits hinges on the model performance “dynamic range”: scaling laws, test-time compute and ensembles can widen the gap that money buys, while distillation compresses frontier power into cheaper models. If marginal dollars can again buy dramatically better models, advantage may re-concentrate; for now, however, frontier-grade LLM capability is unusually democratized, making this phase of AI diffusion historically unique.