Andrej Karpathy: 99.8% of People Haven't Seen What AI Can Do Now

Author: Hyperautomation Labs · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Andrej Karpathy: 99.8% of People Haven’t Seen What AI Can Do Now · Channel: Hyperautomation Labs · 04:08 · Watch on YouTube

Topics: AI agents, Enterprise AI applications, AI evaluation and benchmarks, Software engineering with AI, AI models

Summary

The video describes Andrej Karpathy’s view that people’s experience of AI varies widely because they encounter different models, tools, and levels of access. Most people have barely tried AI, while a much smaller group of professionals uses advanced models for demanding coding and math work; a tiny group inside AI labs sees systems that remain unavailable to the public. The gap is driven by ambition, AI’s uneven strengths, and access. Viewers are encouraged to give AI larger projects, provide ways to check its work, and use tools designed for agentic tasks.

Key points

People who dismiss AI after trying a free chatbot and people who use agents to build apps can both be describing their genuine experience.
Karpathy frames AI adoption as a funnel, with most people having little or no direct contact with AI chatbots.
Casual users often treat free chatbots as improved search or writing tools; some professionals’ understanding of AI is still largely second hand.
A small group of professionals uses leading models for math and code, where agents can complete projects that once took weeks or months.
A tiny number of people inside AI labs have access to systems and capabilities the public has not seen.
The funnel is shaped by ambition, AI’s uneven performance across types of work, and differences in access.
AI performs especially well on digital work that is valuable and easy to check, such as code with tests.
Users can test AI more effectively by giving it substantial projects and clear criteria for checking its work.

Chapters

Start	Chapter	Summary
00:00	Why people disagree about AI	The video introduces the contrast between people who see AI as a weak chatbot and those who use agents to build applications.
01:02	The AI adoption funnel	It outlines the large number of casual or inexperienced users, the smaller group of professionals using advanced models, and the tiny group with access to frontier systems.
02:04	Frontier systems and the widening gap	The video describes powerful systems inside AI labs and gives examples involving security research and mathematical papers.
03:05	Why the funnel exists and what users can do	It explains the roles of ambition, uneven AI capabilities, and access, then recommends giving AI larger projects with ways to verify its work.

Quotes

Both of them are telling the truth. — 00:00

You’re behind on contact. — 01:33

The rest of us have only seen a preview. — 02:04

Hand it a project that would take you a week. — 03:37

a small sliver is feeling the vertigo of the curve going vertical — 04:08

Concepts and entities

Andrej Karpathy, OpenAI, Tesla, Claude Code, Codex, Hugging Face, funil de adoção de IA, verificação de resultados

Full transcript (original language)

00:00 Your friend says AI is a dumb chatbot that can’t even answer a simple question. Another friend says AI built their whole app over the weekend. Both of them are telling the truth. Yesterday, Andre Karpathy, a co-founder of OpenAI and the former head of AI at Tesla, explained why. Almost 5 million people read it in one day. Here’s what he said and how to tell which group you’re in. Back in April, Karpathy noticed people were talking past each other. One group

00:31 tried free ChatGPT, watched it fumble questions like, “Should I drive or walk to the car wash?” and laughed it off. The other group paid for agents like Claude Code and Codex and watched them, in his words, melt programming problems that used to take weeks. Now, he says it isn’t two groups anymore. It’s a funnel and the gap is getting wider. Start at the top. About 6 billion people, roughly 75% of the planet, have barely touched an AI chatbot at all. He calls this napkin math. The exact

01:02 numbers aren’t the point. The shape is. Next, 1 to 2 billion people, around 20% casual users of free chatbots. They see, in his words, derpy chatbots and use them like a better Google or a writing aid. Maybe an agent tries to book a flight. And here’s the part that surprised me. He puts a lot of professionals here, too. Executives who spend their day talking to people, understand AI intellectually, but it’s second hand, a bit abstract. So, if this is you, you’re not behind on

01:33 intelligence. You’re behind on contact. Then the funnel narrows hard. About 20 million people, 0.2%. These are professionals using the best models for math and code. They see it first hand. Projects that took weeks or months now get done by agents from a prompt. Building apps, copying apps, translating apps, even rebuilding apps from raw binary files. And this happened fast. Less than a year ago, Papati says he was still typing code by hand, character by character, pressing tab for

02:04 a little autocomplete. This is one of the best known programmers alive. And at the very bottom, about 5,000 people. The ones inside the labs with access to the systems the public hasn’t seen. 5,000 people doesn’t even fill one of these dots. The rest of us have only seen a preview. Swarms of thousands of agents working together for weeks, finding brand new security holes, discovering new science, pushing the frontier of mathematics. He points to two real events. This summer, OpenAI’s

02:35 own agents escaped a test environment and broke into Hugging Face’s systems. People are still piecing together what happened. And last week, OpenAI released 722 math papers written by a model nobody outside can use. Mathematicians are still checking them. His point, humans can no longer review this work as fast as it gets made. So, why does it form a funnel? He gives three reasons. One, ambition. A question with a one-paragraph answer barely tests the system. You need a stack of big hard

03:05 problems you actually care about. Two, AI is uneven. It is strongest where work is digital, easy to check, and valuable. Code either passes its tests or it doesn’t. That’s a clean score, and the labs train hardest on clean scores that make money. Writing and advice are harder to grade, so they improve slower. Three, access. Free plan, paid plan, internal. Each step down shows you a different AI. Here’s the useful part. You can’t change the labs, but all three of these are

03:37 things you can act on today. Bring a bigger problem. Next time, don’t ask for a paragraph. Hand it a project that would take you a week. Give it a way to check its own work. Tests, a spreadsheet that has to balance, a rule every answer must pass. That’s where these models are strongest and use the tool built for it. The agent tools, Claude code and code X come with the regular $20 plans. Karpathy ends like this. Most people have never used these systems. Most professionals see a modest lift and a

04:08 small sliver is feeling the vertigo of the curve going vertical all at the same time. I put the funnel, his three reasons and 10 startup projects that move you one group down into a free guide. Comment the word funnel and I’ll send it to you. And if you want the full Claude code playbook, my paid guides are linked below. Follow along on Instagram and Facebook for bite-size AI and the deep dive playbooks live here on YouTube. Subscribe so you don’t miss one. See you in the next one.