Andrej Karpathy, one of the founders of OpenAI and former head of artificial intelligence at Tesla, reposted one of his April posts on 11 October to say that the gap in understanding the capabilities of language models is widening. ‘It’s now less a case of “two groups of people talking past each other” and more like a clear-cut funnel,’ he wrote. This explains a common experience: there are those who regard AI as a revolution and those who, having tried it once, have never gone back. Karpathy warns that these are rough estimates, with no surveys to back them up and no sources cited.
The funnel has four levels. Around 6 billion people – 75 per cent of the world’s population – have had very little contact with language models. Between one and two billion use them occasionally in the free version, for things like improved search results or writing assistance. Around 20 million, or 0.2 per cent, use state-of-the-art models in programming and mathematics and see ‘large and complex projects that previously took weeks or months’ completed by agents with a single prompt. At the very top, there are around 5,000 people – 0.00006 per cent – with access to cutting-edge systems within research laboratories. An X user has summarised this as follows.
@karpathy https://t.co/j8175m28qM
— pingfan (@pingfan_ai) 11 October 2026
For Karpathy, the gap stems from three mechanisms. The first is ambition: a question with a paragraph-long answer does not exploit the system’s potential, whilst the benefits are seen on complex problems that are important to those asking them. The second is patchy capability: models excel at digital and verifiable tasks, such as programming and mathematics, because training with verifiable rewards and investment from research laboratories follow the potential for revenue. The third is access: the free version is not the same system used by paying customers, and neither of these matches the internal system. Even many non-technical professionals remain at the second level, because they only see AI indirectly.
Those at the bottom of the funnel are already seeing something different. According to Karpathy, ‘the outside world has had a sneak preview’, whilst the few with internal access are working with swarms of thousands of agents collaborating for weeks on major software projects, scientific research and mathematics, and human review capacity is beginning to lag behind. It should be noted that this estimate is informal: no laboratory has confirmed the figure of 5,000, and some observers point out that each level is correct from its own perspective but wrong when generalised, because the typical task of the casual user is precisely the one where models have improved the least.
This pattern is not new, and we at Blockworld.it are well aware of it: those who follow Bitcoin and Ethereum have already seen a majority watching from the outside whilst a minority experienced the changes from within. Sceptics and enthusiasts both speak the truth about what they have seen; the problem is that they are not looking at the same thing: those who dismiss AI after asking a free chatbot a single question have seen the entrance, not the engine room. The funnel does not measure the intelligence of machines, but the distance between those who talk to them and those who work on them.
