Anthropic’s Karpathy says public sees only AI preview while 5,000 access internal systems
Andrej Karpathy describes a widening gap between casual users and the small group accessing advanced internal AI agent systems.
Key points
- Andrej Karpathy estimates 75 percent of the global population has barely touched large language models.
- Approximately 5,000 people have access to internal frontier systems that perform complex tasks.
- Karpathy notes that professional developers now use agents to complete projects that previously took weeks.
Andrej Karpathy, an Anthropic researcher and OpenAI co-founder, states that the general public has only seen a preview of artificial intelligence capabilities. According to a report by The Decoder, Karpathy highlights a significant disparity between public perception and the actual power of internal frontier systems.
The Decoder reports that Karpathy identifies a widening knowledge gap in the AI industry. He asserts that while casual users view chatbots as simple writing aids or improved search engines, internal systems are already executing tasks that would require top experts years to complete. This distinction separates the mass market from a tiny elite with access to advanced tools.
The Four Tiers of AI Usage
Karpathy categorizes the global user base into four distinct levels to illustrate this divide. The largest group comprises approximately six billion people, representing roughly 75 percent of the world’s population. According to The Decoder, this vast majority has barely come into contact with large language models at all, leaving them largely unaware of the technology's evolving potential.
The second tier includes one to two billion casual users of free-tier products similar to ChatGPT. Karpathy describes these users as treating the technology like a better Google search or a basic writing aid. The Decoder notes that even many professionals outside of mathematics and programming fall into this category, viewing AI developments through a second-hand and abstract lens.
The third tier consists of an estimated 20 million people, or about 0.2 percent of the global population. These are professional users in math and code who utilize frontier-grade models for their work. Karpathy observes that these users see firsthand how agents can complete large, complex projects with a single prompt, drastically reducing the time required for development tasks.
Impact on Professional Developers
Karpathy emphasizes that the shift in developer workflows has occurred rapidly. The Decoder reports that in October 2024, Karpathy dismissed AI agents as mostly marketing hype that barely functioned. However, within a few months, he stopped writing code by hand entirely, transitioning to letting agents handle the work previously done by typing commands character by character.
He lists building, copying, translating, and decompiling apps from binaries as examples of tasks now handled by agents. According to Karpathy, these capabilities allow professionals to finish projects in days that used to take weeks or months. This rapid adoption underscores the growing divergence between public understanding and professional application.
The fourth and smallest tier includes roughly 5,000 people worldwide, representing approximately 0.00006 percent of the global population. Karpathy estimates that this exclusive group has access to internal frontier-grade systems. The Decoder notes that Karpathy joined Anthropic after leaving his role as an OpenAI co-founder, bringing this perspective from the inside of a leading AI research organization.
What should you do?
Security and technology professionals should monitor internal access controls for frontier AI systems to ensure only authorized personnel can interact with high-capability agents. Organizations must evaluate how quickly development workflows are changing as agents take over complex coding and decompilation tasks. Leaders should recognize that public perception of AI capabilities lags significantly behind internal operational realities.
Background: Large language models
Large language models are a specific type of neural network trained on text. Generative AI is the broader umbrella term for any system that creates new content, including images and code. Not all generative AI uses large language models, but most modern text generators do.
What to do and how to stay safe: Anthropic
- Audit access permissions for internal AI systems to restrict usage to vetted personnel only.
- Review development workflows to identify where AI agents are replacing manual coding tasks.
- Assess the gap between public-facing AI tools and internal frontier systems for security risks.
- Monitor vendor announcements regarding the release of new agent capabilities to professional users.
Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
How many people have access to internal frontier AI systems according to Karpathy?
Karpathy estimates that roughly 5,000 people worldwide have access to internal frontier-grade systems.
What percentage of the global population has barely touched LLMs?
About 75 percent of the world's population, or approximately six billion people, have barely come into contact with large language models.
Did Karpathy always believe AI agents were effective for development?
No, in October 2024 he called them mostly marketing, but he later stopped writing code by hand entirely.
