Karpathy: The cognitive gap in LLM capabilities is widening

Artificial intelligence researcher, OpenAI co-founder, and former Tesla AI director Andrej Karpathy stated on X that the gap in public understanding of the capabilities of large language models is widening, evolving from two groups talking past each other to a steeper funnel. He reiterated related posts from April and provided a layered estimate based on exposure.

Karpathy noted that approximately 6 billion people (about 75% of the global population) have had little to no exposure to large language models; around 1 to 2 billion people (about 20%) only occasionally use free versions of ChatGPT-like products, often viewing them as better search or writing aids, with many non-technical individuals and professionals outside of mathematics and coding considering their utility limited. About 20 million people (0.2%) use cutting-edge models professionally in mathematics and coding, witnessing complex projects that previously took weeks or months being completed by agents under prompts, including building, copying, translating applications, and decompiling applications from binary. He mentioned that less than a year ago, he was still manually writing code word by word.

Karpathy also stated that around 5,000 people (about 0.00006%) have internal access to cutting-edge systems. The previews seen externally involve thousands of agents collaborating for weeks on large software projects, cybersecurity, scientific discoveries, and mathematical frontiers, with human reviews starting to lag behind; for instance, events related to OpenAI and Hugging Face are still being sorted out months later, and mathematicians may spend a long time studying 722 cutting-edge mathematical manuscripts. He believes that this funnel is driven by the scale of problems and time spans, the prominence of capabilities in digitizable and verifiable fields, and the differences between free, paid, and internal access.