AI researcher Andrej Karpathy, a co-founder of OpenAI who recently joined Anthropic, has highlighted a widening gap between public perception and internal AI progress. Writing about the distribution of AI capabilities, Karpathy notes that while the vast majority of people view large language models as basic writing aids or search tools, a tiny group of roughly 5,000 people has access to internal frontier systems. According to Karpathy, these advanced internal systems deploy swarms of thousands of collaborating agents working over weeks on complex software projects, discovering new science, and operating at speeds well beyond human capacity.

Karpathy’s observations outline a rapid transition from late 2024, when AI agents were largely ineffective, to a current state where agents handle entire coding workflows autonomously. OpenAI CEO Sam Altman echoed similar sentiments, highlighting that internal tools can replicate three months of startup development in just 17 minutes. Karpathy explains that model capabilities are currently peaking in digital, verifiable, and economically valuable domains like math and coding due to reinforcement learning on verifiable rewards.

However, Karpathy warns that human review and comprehension are beginning to fall behind the speed of these internal systems. As top-tier models automate complex projects end-to-end, the broader public and most professionals remain unaware of the steep vertical trajectory occurring at the bleeding edge of AI development.

Why it matters

  • Expect rapid acceleration in software engineering as internal agentic workflows transition into commercial products over the next cycle.

  • Human verification and code audit tools will become critical bottlenecks as autonomous agent outputs outpace human review capabilities.

  • Reinforcement learning efficiency will continue concentrating frontier AI breakthroughs in digital, highly verifiable domains like math, software, and cybersecurity.

Source: the-decoder.com