The gap in shared understanding of LLM capability is widening

Andrej Karpathy posted a sketch of how differently people experience large language models, and it drew a large discussion on Hacker News. He describes a funnel that runs from billions of people with minimal contact, through casual chatbot users, to developers using agents on large projects, and finally to a few thousand people with internal access to frontier systems. His argument is that shared understanding of what models can do is getting worse, not better, as the layers drift apart. He attributes the gap to three drivers. The first is problem ambition: people who point agents at big code or math projects see capabilities that short chat sessions never surface. The second is jagged capability shaped by reinforcement learning on verifiable domains, so strength in one area says little about another. The third is tiered access to the strongest models. For builders, the practical reading is to judge models on their own demanding workloads rather than on general impressions, which reflect whichever layer of the funnel the observer sits in.