Karpathy: AI capability perception has a major gap; the free tier and the cutting-edge agent are “completely different products”

Former Tesla AI Chief Architect and OpenAI founding member Andrej Karpathy published a long post on X on April 9, pointing out that the public’s understanding of AI capabilities is becoming severely split. He believes that people using the free version of ChatGPT and technical professionals using cutting-edge agent tools like Codex and Claude Code every day are actually discussing “completely different products,” yet both sides think they’re seeing the full picture of AI.

Two worlds, two types of AI understanding

Karpathy currently divides AI users into two groups.

The first group tried the free version of ChatGPT at some point last year, and formed their overall impression of AI from that. What they see are various failures of the model—hallucinations, absurd search results, and even simple questions like whether the voice mode should “drive or walk to get a car wash.” Karpathy admits these problems do exist, but emphasizes that the free version and outdated models can’t represent the real capabilities of cutting-edge agent models before 2026.

The second group satisfies two conditions at the same time: they pay to use the latest cutting-edge agent models (such as OpenAI Codex or Claude Code), and they use them professionally in technical fields like software development, mathematics, and research. Karpathy says this group is experiencing a high level of “AI psychosis,” because the recent progress of these models in technical areas can only be described as astonishing—you can literally watch them solve in an hour programming architecture problems that previously would have taken days or even weeks.

Why progress is concentrated in technical fields

Karpathy explains why improvements in AI capabilities are particularly noticeable in technical fields like software development, but less so in general uses such as search, writing, and making recommendations.

There are two reasons: first, technical fields provide a verifiable reward function (for example, whether unit tests pass), which makes reinforcement learning training work effectively; by contrast, it’s hard to determine objectively how good writing quality is. Second, technical fields have greater commercial value in B2B scenarios, so AI companies put the largest share of team resources into these directions.

The two groups can’t understand what the other is saying

Karpathy concludes that these two groups are “talking past each other.” OpenAI’s free voice mode botches everyday problems, while OpenAI’s top-tier paid Codex can restructure an entire codebase or discover system vulnerabilities within an hour—both of these things are simultaneously true.

In a follow-up reply, he added that someone offered him an observation: the OpenClaw incident drew so much social attention precisely because it introduced a large number of non-technical people to the latest agent models for the first time, and these people previously only knew that AI equals ChatGPT’s web version.

This article by Karpathy: AI capability recognition shows a severe gap; the free version and the cutting-edge Agent are “completely different products.” First appeared on Chain News ABMedia.