Why the Same AI Looks So Different: Karpathy's Four-Level Gap
LLM
A text-based AI that learns to understand and produce language.
AI agent
An AI that carries out several steps toward a goal.
frontier model
One of the most advanced AI models available at a given time.
What happened
AI researcher Andrej Karpathy wrote on X that the gap in shared understanding of large language model, or LLM, capability is widening. This is not a new product announcement. It is a claim about unequal experience. Different people may be judging very different AI systems.
Karpathy sketched four rough groups. About 6 billion people have barely encountered LLMs. About 1 to 2 billion use free chatbots casually, for search-like questions or writing help. Around 20 million use advanced models professionally in mathematics and coding. About 5,000 people have internal access to frontier models inside AI labs. For this last group, he describes thousands of AI agents working together on long software projects, mathematics, and cybersecurity. He presents a level of work that could outrun human review. These are his descriptions, not an independently verified list of results.
The supplied snapshot records 457 points and 652 comments for the post on Hacker News. That shows community attention. It does not prove that Karpathy's estimates or claims are correct.
Background: one technology, different tests
A short question and a long project test different abilities. A casual user may ask for a quick answer. A professional may give an AI agent a large goal, tools, files, and many steps. The second user can see abilities that never appear in a short chat.
Karpathy also says LLM abilities are uneven. Models can look stronger in digital work where results are easy to check. Mathematics and coding often provide clearer tests than open-ended advice. Access matters too. People may use different models, prices, limits, and workflows. The same word, AI, can therefore describe very different experiences.
Why it matters
This gap can make public arguments about AI miss each other. One person sees a weak chatbot. Another sees a tool that changes how complex work is organized. Both observations may be reasonable within their own settings.
The harder issue is verification. If agents produce work faster than people can inspect it, output speed alone is not enough. Companies, schools, and governments need to know what the systems can do, where they fail, and how much review they require. A public debate built on only one access level can overstate or understate the real situation.
What is established
The source clearly gives Karpathy's four groups, rough population counts, and explanation of why experiences differ. It also attracted substantial discussion on Hacker News. Those are the facts available from the source and the supplied engagement record.
The numbers are estimates. They are not results from a worldwide survey. The Hacker News score and comment count measure attention, not accuracy.
What remains unclear
The source does not define its categories precisely. It gives no method for counting people who barely use LLMs, casual users, professional users, or internal users. It does not identify the internal systems. It also gives no independent case studies showing how often agents complete large projects, how many mistakes remain, or how much human correction is needed.
The phrase widening also needs evidence over time. We would need repeated measurements of access, use, performance, and understanding. Without them, the post is a useful framework, not a measured trend.
What to watch next
The next useful evidence will come from independent tests of long tasks. Those tests should report both successful work and correction time. It will also matter whether capabilities seen inside labs become reproducible in public tools. Karpathy's warning is not a final map of AI progress. It is a reminder that people may be arguing from different views of the same technology.
Why Do People See Different AI?
📰 Full story: Why the Same AI Looks So Different: Karpathy's Four-Level Gap
People use different AI tools and tasks, so they see different abilities.
LLM
A text-based AI that can read and write.
AI agent
An AI that keeps doing steps toward a goal.
frontier model
A very advanced AI model.
💡 The gist
- People see different AI because they use different tools.
- Karpathy estimates four levels of AI experience.
- His numbers are rough, and attention is not proof.
Andrej Karpathy, an AI researcher, wrote about this gap on X. A large language model, or LLM, reads and writes text.
Karpathy says about 6 billion people have barely tried an LLM. About 1 to 2 billion people use free chatbots casually. They ask simple questions or request writing help. Around 20 million people use advanced models for mathematics and coding. About 5,000 people can access frontier models inside research labs.
These are rough estimates. They are not results from a worldwide survey. In the supplied snapshot, the post had 457 points and 652 comments on Hacker News. That shows interest. It does not prove the estimates.
Why do experiences differ? A short question tests one small answer. A long coding project tests planning and many steps. An AI agent can keep working through several steps. That can make the tool look much more powerful.
Karpathy also says AI skills are uneven. Models often do better on digital work with checkable results. Mathematics and coding can provide clearer tests. This does not mean models work well everywhere.
The gap can confuse public discussions. One person sees a weak chatbot. Another sees a tool for large projects. Both people may be using different models and tasks.
Important questions remain. How did Karpathy count each group? Which internal systems did he mean? How often do agents finish large projects? How much human correction do they need?
The next step is independent testing. Researchers should test long tasks and report mistakes. They should also measure review time. That would show what public AI can really do. It would separate proven results from personal impressions.
💬 People see different AI because they use it differently
Commenters report both faster software work and continuing mistakes and bugs. Their experiences differ because the model, the surrounding tool, the task, the user's experience, and the ability to check the answer all matter.
- One user's self-report says agents helped move a large software change forward, but UI and code fixes were still needed.
- AI is more useful when the rules and tests make correctness clear. Ordinary questions and search answers can still be wrong, so they need checking.
- Users also self-report bugs and lag in tools such as Claude Code. A tool problem is not automatically proof that the model itself is incapable.
- People often lack the time, knowledge, or reward needed to review every AI result, and rough output can increase the work. One user reported Gemini doing better than ChatGPT on a specific logistics question, but that is only one example.
- Some people think real-world change is still early and slow; others think the big claims are ahead of the evidence. Different users may therefore be judging very different versions of the same technology.
mature digest at 652 comments (revision 1). We fetched 500 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.
Why Does AI Look Different to People?
📰 Full story: Why the Same AI Looks So Different: Karpathy's Four-Level Gap
People use AI for different jobs, so they see different things.
AI
A computer helper that can do tasks.
Hacker News
A website where people discuss technology news.
Andrej Karpathy is an AI researcher. He wrote about AI on X.
People do not all see the same AI.
Some people hardly use it. Some ask small questions. Some ask it to help with long jobs. A few people try newer tools inside labs.
Karpathy gave rough counts for these groups. He did not give a tested survey.
His post got 457 points and 652 comments on Hacker News. Those numbers show attention, not truth.
A small question tests one thing. A long job tests many things. Newer tools may also do more.
People still need to check AI's work. We do not know how many big jobs AI finishes well.
💬 AI can look different to different people
Some people say AI helps them a lot. Others see mistakes and broken tools. Both experiences can be real.
- One user said AI made big software work faster, but people still had to fix the screen and the code.
- AI does better when people can test the answer. Normal questions and searches can be wrong. One user said Gemini was better than ChatGPT for one logistics question, but that was only one example.
- Some users report that AI tools are buggy or slow. People also do not have enough time or reward to check every answer, so messy answers can create more work.
- People disagree about whether AI is changing the world now or is still just getting started.
mature digest at 652 comments (revision 1). We fetched 500 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.
💬 The gap is also about tools, tasks, and verification
HN comments include self-reports of faster agent-assisted development, alongside self-reports of tool bugs, quality limits, and review costs. Taken together, they suggest that LLM capability is not one number: results depend on the model, the surrounding harness, the task, the user's skill, and whether the answer can be checked.
mature digest at 652 comments (revision 1). We fetched 500 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.