Andrej Karpathy Says AI Capability Understanding Gap is Forming a Funnel
Former OpenAI researcher and current Anthropic associate Andrej Karpathy stated that the understanding gap regarding the capabilities of large models is widening, shifting from two groups talking past each other to a sharp funnel.
He estimates that about 6 billion people (approximately 75% of the global population) have had little to no exposure to large language models; around 1 to 2 billion people occasionally use free versions of ChatGPT-like products, viewing them as better search or writing aids, with many professionals outside of mathematics and programming in this group having only an indirect understanding of advancements.
About 20 million people (approximately 0.2%) heavily use cutting-edge models in mathematics and programming, witnessing firsthand how projects that originally took weeks or months can now be completed by Agents through prompts, including building, translating, and decompiling applications. This change has rapidly occurred over the past year.
The top tier of about 5,000 people have access to more powerful internal systems, with external previews showing thousands of Agents collaborating for weeks to complete large software projects, discover zero-day vulnerabilities, conduct network offense and defense at machine speed, and advance the frontiers of mathematics, with human review and understanding beginning to lag behind.
The gap is driven by three factors: the scale and ambition of problems determine impact (questions that can be answered in a paragraph rarely showcase true capability); progress is faster in verifiable and high-value areas (programming, mathematics) due to clear reinforcement learning rewards, with vendors investing more; and significant differences exist in access levels between free, paid, and internal systems.
In the diffusion of AI capabilities, cutting-edge abilities are concentrated in verifiable high-value areas and internal access, with heavy technical users and lab personnel benefiting from vertical advancements, while ordinary free users and non-technical professionals see limited progress, leading to a widening gap in understanding and usage.
Karpathy emphasized that most people still view AI as a clumsy chatbot, while a few professionals are experiencing the dizzying rise of capability curves.
Source: Public Information
ABAB AI Insight
Andrej Karpathy previously participated in the development of the GPT series at OpenAI and has publicly commented on model capabilities and training dynamics. This time, he reissued and updated an April tweet, describing the understanding gap as a four-layer funnel rather than two opposing groups. Historically, he has repeatedly emphasized that progress in verifiable reward areas is faster and pointed out that humans are becoming a bottleneck in easily measurable tasks.
Capital and technological pathways lead vendors to prioritize reinforcement learning in areas like programming and mathematics that can be automatically verified and have high enterprise willingness to pay, motivated by maximizing revenue and scalable progress. At the same time, the hierarchical difference between internal systems and external access results in a highly stratified perception of capability.
This is similar to the cognitive gap between early adopters and the general public in other technology diffusion scenarios, closely resembling historical patterns where automation and software tools first break through in professional fields before slowly spilling over; currently, cutting-edge Agent capabilities are rapidly emerging among technical users, while the general public remains at the free tier.
Essentially, this is an intersection of technological substitution and capital concentration: verifiable high-value areas attract the most resources due to clear rewards and commercial returns, with the mechanism driven by the scalability of reinforcement learning and access hierarchies, causing capability advancements to accelerate at the top of the funnel while progress at the bottom is slow.
ABAB News · Cognitive Laws
- 6 billion have not encountered it, 5,000 have seen the real deal.
- Progress is fast where it is verifiable, slow where it is hard to assess.
- What is seen in the free version is several layers away from the frontier.