The gap in shared understanding of LLM capability is widening
Points and comments are a snapshot, not live.
Karpathy argues the understanding gap of LLM capability is widening into a sharp funnel.
Andrej Karpathy estimates ~6B people (75% of population) have barely interacted with LLMs. ~1-2B (20%) are casual free-tier users, seeing derpy chatbots. ~20M (0.2%) are professionals using frontier LLMs in math and code, where large projects now complete via agents. ~5,000 (0.00006%) access internal frontier systems, witnessing swarms of agents collaborating on software mega projects, discovering science, and advancing mathematics, with human review comprehension falling behind. The funnel is driven by impact scale with ambition, jagged capability peaks in digital/verifiable domains, and access tiers.
What commenters are saying
Commenters are skeptical of the top-tier claims. Several point out bugs and poor quality in Claude Code and Codex, calling them 'a buggy mess' and 'software made by amateurs.' One argues demos and output quality don't match the hype, citing a zsh script that inefficiently replicated grep. A dominant concern is that human verification capacity is the real bottleneck: 'the real wall in LLMs would end up being humans' capacity to verify the output.' Another adds that verification is often unpaid labor with no incentive, citing mathematicians facing slop. A separate camp notes free-tier models have already saturated what most people need, with Google AI summaries being adequate for many queries, despite often being wrong.