Karpathy quantifies LLM capability funnel at 6B non-users versus 5,000 internal frontier operators

The tweet restates April data showing 75% of the global population untouched by LLMs, 20% limited to free-tier chatbots, 0.2% using paid models for code and math, and 0.00006% inside organizations running agent swarms. This distribution matches observed adoption curves in the GPT-4 technical report and the o1 system card, where verifiable reward training concentrates gains in digital domains. External observers therefore encounter only the lower capability surface while internal teams execute month-scale projects in days.

Karpathy identifies three drivers: problem horizon scaling, jaggedness from reinforcement learning on revenue-weighted environments, and tiered access. The original post understates the ethics angle: policy makers drawn from the 1-2B casual tier draft regulations on systems whose outputs they cannot audit, while the 5,000 see machine-speed cyber operations and new theorem generation. This mismatch appears in the absence of any public incident report covering the referenced OpenAI-HF event despite months of internal review.

Operationally the gap produces divergent risk models. Organizations with frontier access already treat agent coordination as a solved infrastructure problem; governments and enterprises outside the funnel still model LLMs as autocomplete tools. The next measurable threshold is whether external benchmarks will capture swarm-scale outputs before regulatory deadlines close.