Anthropic researcher warns of widening AI capability gap for organisations

Andrej Karpathy, a co-founder of OpenAI and current researcher at Anthropic, describes a rapidly expanding divide in how different groups perceive artificial intelligence capabilities. According to The Decoder, Karpathy argues that the general public has only witnessed a superficial preview of what frontier AI systems can actually achieve. This observation highlights a significant disparity in understanding between casual users and technical experts.

The researcher outlines four distinct tiers of AI adoption to illustrate this gap. Approximately six billion people, representing roughly 75 percent of the world’s population, have had minimal contact with large language models. Another one to two billion users engage with free-tier chatbots primarily as writing aids or improved search engines. These casual users treat the technology as a utility rather than a transformative tool.

What we know

Karpathy estimates that around 20 million professionals in mathematics and programming utilise frontier-grade models for their daily work. These users observe that large, complex projects, which previously required weeks or months of manual effort, can now be completed by autonomous agents with a single prompt. Examples include building, copying, translating, and decompiling applications from binaries, according to the report.

A much smaller group of approximately 5,000 individuals has access to internal frontier-grade systems. Karpathy notes that these systems employ agent swarms capable of performing work that would take top experts years to complete. He states that the external world has only seen the preview of these capabilities. This tiny fraction of the global population represents the cutting edge of current AI utility.

The shift in capability has occurred with surprising speed. In October 2024, Karpathy described AI agents as largely ineffective marketing tools. Within a few months, he ceased writing code manually, transitioning from typing memorised commands character by character to letting agents handle the work. He subsequently joined Anthropic despite his previous role as an OpenAI co-founder, highlighting the rapid evolution of the field.

Many executives and professionals outside technical fields remain in the casual user tier. While they understand the intellectual concepts of AI, their experience is often second-hand and abstract. They spend most of their time communicating with other people rather than interacting directly with the technology. This lack of direct engagement means their understanding remains somewhat detached from practical application.

Defender's checklist

• Assess whether your organisation relies on AI tools that are not under your direct control or monitoring.
• Identify which internal teams have access to advanced frontier models and what data they input.
• Evaluate the security posture of any third-party AI services used for code generation or data processing.
• Review policies regarding the use of AI agents for critical infrastructure or sensitive development tasks.

Background: AI agents

AI agents are the software entities that perform specific tasks using instructions from large language models. Agentic AI is the architectural approach that grants these agents autonomy to plan, reason and act. You use agents as tools within agentic AI systems to achieve complex, multi-step outcomes without manual oversight at every stage.

Read the full guide: AI Agents vs Agentic AI: The Distinction That Matters

What to do and how to stay safe: Anthropic

• Audit current AI tool usage across departments to identify unauthorised or high-risk applications.
• Implement strict data governance policies to prevent sensitive information from being processed by external AI models.
• Train technical staff on the limitations and potential security risks of using autonomous AI agents.
• Monitor vendor communications for updates on security features and access controls for AI platforms.

Step-by-step guide: AI Governance for Small Teams: Practical Controls and Cost

General security guidance from the TechWire Beat 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?

Andrej Karpathy estimates that roughly 5,000 people worldwide have access to these internal 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 interacted with large language models.

What tasks can internal frontier systems perform?

These systems can perform tasks that would take top experts years to complete, using agent swarms for complex work.