Former OpenAI Researcher Warns of Growing AI Gap: Only 5,000 People May Have Access to the Most Advanced AI Systems
Artificial intelligence (AI) is becoming a regular feature of daily life, with millions of people relying on AI-powered chatbots to provide answers to queries, craft content, look up information and complete repetitive tasks.
But AI researcher Andrej Karpathy and co-founder of OpenAI has cautioned the world is getting just a taste of the potential power of advanced AI systems.
Karpathy says that about 1 billion to 2 billion people are using free tools regularly, using them sometimes as an alternative to Google Search or as an aid to writing. By contrast, perhaps only 5,000 individuals can have internal access to the most sophisticated AI systems being developed in the leading AI laboratories.
These sentiments were echoed in an X post by Karpathy in October 2026. The gap between everyday users of AI and those who directly interact with advanced AI models and AI agents was highlighted. His guesswork is that it's not only who's using AI, it's also how they’re using it, what they are asking it to do, and how much they have access to it.
The alert has sparked concerns over the future of AI, AI technology and the pace at which the capabilities of AI are advancing.
Who is Andrej Karpathy and what did he say about frontier AI access?
Andrej Karpathy is a well-known artificial intelligence (AI) researcher and co-founder of OpenAI. He has been the director of AI at Tesla, and has been involved in deep learning, computer vision and neural networks. He is also famous for his ability to write about AI in an easy-to-understand manner.
In his recent thoughts on AI, Karpathy noted a trend of a divide between users of AI in their daily lives and users of the more powerful versions.
Around 6 billion people have not had any real interaction with large language models (LLMs), he estimated, while 1 billion to 2 billion people use free AI tools to carry out simple tasks like searching for information, answering questions, and writing content.
He estimates that about 20 million advanced AI users, especially in the software engineering and mathematics fields, are employed in more sophisticated work.
At the most elevated level, Karpathy estimated that there are just around 5,000 individuals who have access to the "frontier AI systems" being developed or tested in the top-tier AI laboratories. These high-performance systems can be utilized for more complex tasks such as software engineering, scientific studies, and Cybersecurity.
To emphasize the disparity between using an AI chatbot as a more powerful search tool versus a more sophisticated AI agent for multi-step tasks, Karpathy's estimates were quoted in the article.
These are, however, estimated numbers not to be independently collected global statistics. His words bring up some significant points to discuss regarding access to frontier AI, sophisticated artificial intelligence, AI agents, and the widening divide between AI use.
Why are most people using AI like a better google search?
Most people engage with AI through basic queries, content summaries or information verification. The majority interact with AI through simple questions, summaries or fact-checking.
While they can be helpful, they're just one way the technology can be used.
A student can use the AI chatbot to get clarification on a complex subject, a journalist can use it to organize research or enhance a draft, etc. On the other hand, an AI can be used by a professional programmer to analyse a code base, spot errors, create new features, and test software. The distinction is even more pronounced if the AI serves as an agent instead of merely a chatbot.
An AI agent can be programmed to take a sequence of actions, employ connected tools, review outcomes, and then proceed with the next action in an effort to achieve a set objective. Depending on what it can do and how reliable the tools are it's able to do things that would take multiple people to do.
It's because many casual users might assess the capabilities of AI based on the performance of individual chatbot responses, whereas advanced users might see them unfold over a series of connected actions.
This usage disparity can have an impact on the benefits individuals receive from the same broad class of technology.
How AI agents could change software development and research
The growth of AI agents in complex tasks is one of the most critical points Karpathy highlighted in the warning. Most traditional chatbots answer questions and deliver information or content. With the right design, tools and access, AI agents can execute a series of actions, depending on their capabilities.
For instance, in software development, an AI agent might be tasked with comprehending an existing project, creating code, conducting tests, pinpointing mistakes, and correcting them. In scientific research,
AI tools can assist scientists in analysing vast quantities of data, investigating potential solutions, and streamlining certain aspects of their workflow.
Some sophisticated users are witnessing the acceleration of AI agents finishing projects that used to take weeks or months, Karpathy wrote.
The results will vary, depending on the task and the quality of the system and the extent of human supervision.
This doesn't imply that AI can handle all intricate tasks or consistently replace human experts. But it indicates that AI-powered labour is likely to go far beyond answering queries and writing text.
The difference between casual and advanced uses of AI has become a key issue in the tech industry due to AI's increasing capacity to perform multi-step tasks.
AI capabilities may be advancing faster than human understanding
Another worry that has been raised with Karpathy's remarks is that the advancement of AI systems could be outpacing individuals' capacity to comprehend and assess their output.
A human can quickly check the answer that an AI has produced in a short answer. But checking every step can be more challenging when thousands of lines of code are written by an AI system, or when a complex analysis is performed or when multiple agents are used to coordinate.
This raises questions of reliability, accountability and supervision of humans. An instance of this can be an AI agent that produces software that seems to be functioning properly but has security holes. Likewise, an AI system can generate an in-depth study analysis with false assumptions and/or false conclusions.
When AI can do more complex jobs, there will be a need for improved ways to evaluate the results, track performance, and correct errors by users and organisations.
It is therefore crucial to have humans to oversee the use of AI, particularly in sensitive fields such as security, finance, healthcare, scientific research, and other critical decision-making processes.
17-Year-Old Anika Puri develops $300 AI tool to detect poachers, wins $25,000 scholarship