There Are Only Two Groups of AI Users, Says Andrej Karpathy

Andrej Karpathy, co-founder of OpenAI and former head of AI at Tesla, wrote a post on X that precisely identified something many of us had sensed but could not clearly articulate. There are two groups of people who talk about artificial intelligence. And these two groups effectively do not understand each other. They speak different languages, work with different tools, and live in different realities. Yet they argue about the same thing.

Group One: Those Who Tried ChatGPT for Free and Decided That Was Enough

You know the type. Someone signed up for a free ChatGPT account last year, tried a few prompts, encountered hallucinations or a nonsensical answer, and has been certain ever since: AI is an overhyped bubble. They then share viral videos on social media in which OpenAI's voice assistant cannot answer whether someone should drive or walk to the car wash. Their experience is real. It just says nothing about what today's AI can actually do.

According to Karpathy, free and outdated models do not reflect the capabilities of the latest agentic systems at all. They are different worlds. Comparing last year's free ChatGPT with this year's OpenAI Codex or Claude Code is like comparing a Nokia 3310 with today's smartphone. Both are phones, but that is more or less where the similarity ends.

Group Two: Those Who Have Experienced "AI Psychosis"

At the other end of the spectrum are developers, scientists, and technical professionals who pay for access to the best available models and use them every day for demanding work. Karpathy himself admitted that since December 2025, he has not written a single line of code by hand. Instead, he manages ten to twenty AI agents simultaneously and calls this state "psychosis."  It is not a clinical term, but a description of how astonishing and disorienting the shift is.

Why December specifically? Karpathy speaks of a turning point when agents stopped being unreliable tools and became genuinely functional colleagues. Before that, they could handle fragments. After crossing that threshold, they began handling entire projects. His ratio of hand-written to agent-generated code jumped from 80/20 to 20/80, until manual coding eventually disappeared altogether.

This group attributes an entirely different level of significance to AI. They see models capable of working independently for an hour to restructure an entire codebase. Or finding security vulnerabilities in systems.

Why Are These Two Groups So Far Apart?

Here comes the truly interesting part. According to Karpathy, the performance of today's AI is "spiky", strong in certain areas and average elsewhere. And the areas where AI has made the biggest leaps are highly technical: programming, mathematics, and research.

Why those areas in particular? There are two reasons. First, these fields offer verifiable rewards for machine learning. Either the code passes the tests or it does not. There is nothing in between. Unlike writing or advice, where assessing quality is subjective and difficult, programming gives models clear feedback. That is a gold mine for reinforcement learning.

Second, these use cases bring companies the most money. Where the money goes, developers follow. And so the models improve where businesses value it most.

As a result, an ordinary user looking for advice or wanting AI to help with an email will not experience the dramatic shift Karpathy is talking about. Meanwhile, an experienced developer with paid access sees dramatic progress every week.

What This Means for Companies and Product Teams

This divide is not just a social media issue. It has a direct impact on how companies design products and how they evaluate the success of AI tools.

Aggregate model evaluation scores conceal distributed failures that concern skeptics and regulators. Tools optimized for experienced developers, such as custom instructions, agent chaining, or infrastructure integration, deepen the accessibility gap for ordinary users. And benchmarks that look impressive in the lab may say nothing about real-world use.

This leads to a clear task: differentiate the user interface for experts and non-experts, invest in explanation and education, and stop measuring success with a single number that tells no one anything and appears misleading in real-world use.

Are We Witnessing a Genuine Split in Reality?

It is fascinating to watch how one person, Karpathy, can reverse his opinion twice in the span of six months. In October 2025, he said that agentic AI was not yet ready for real-world deployment, and by December it had overwhelmed him. Today, he manages dozens of agents at once and has stopped writing code by hand. That is how quickly reality is changing.

And that is precisely why his warning about the divide between users is so important. It is not about who is right and who is wrong. Both groups are describing genuine experiences. But those experiences are so far apart that, without understanding the context, they cannot even have a meaningful conversation. The next six to twelve months will show whether companies begin to bridge this divide or whether the two groups drift even farther apart, each living in its own AI world.