Why the Creators of AI Do Not Understand Their Own Creations

Original Summary

Just citations.Dario Amodei (CEO of Anthropic, 2025): "People outside the field are often surprised and alarmed to learn that we do not understand how our own AI creations work. This lack of understanding is essentially unprecedented in the history of technology."MIT Technology Review: "Despite all the stunning success, nobody really knows how — or why — it works."Yoshua Bengio (Turing Award laureate, 2024): "If we continue as before, we are driving against the wall. Or at least we are driving into a fog, without knowing whether there is a wall."Mathematician Roman Vershynin (UC Irvine), in a recent interview — "We advanced to the wall and hit it."So why did AI research hit the wall?The answer is not hard to find, and most people in the field know it, even if they prefer not to say it loudly. The point is that all training data used by AI is secondary, while the primary source is the human being. Every text, every image, every pattern a neural network absorbs — all of it was created by humans. A neural network does not discover the laws of the world; it only finds regularities in what has already been created by human intelligence.So if we want to understand how AI forms knowledge, we must first understand how humans form knowledge. But this pretty obvious idea was never materialized because of the historical failure of all attempts to build a working, mathematically correct model of knowledge. The story started in the 19th century and continued until the end of the 20th. Tens of theories of information, hundreds of books, thousands of articles in dozens of journals — and nothing practically working came out of it.This however does not mean no adequate theory exists. It is there since the late 2000s, but it is still completely unknown to the scientific community. This occurs for a purely ideological reason because this theory builds its formalizations on concepts borrowed from mainstream programming languages and so completely ignores mathematics.For the major journals devoted to AI and information theory that is a taboo. I spent years trying to publish it. The answer was always the same: "Add the mathematical justification, and then we will take a look." When I contacted established researchers directly, the answer was either silence or: "Yes, you have an original idea, but we are not interested."I stopped trying fifteen years ago. My hope is that now the theory might have a chance, because it addresses exactly the problems that the current conceptual apparatus cannot solve.The Theory of Meaningful Information (TMI) expands the conceptual framework of the most expressive programming language, which is C++, by adding capabilities for creating standard narrative (non-executable) texts that constitute the primary form of any human language.In TMI, information is defined as the value of a variable used in some algorithm. This works for a bit in a computer, a state of a neuron, a pressed button, a written letter. It applies to a bacterium, a program, a neural network, and a human.The theory introduces Information Capable Entities (ICE) and traces their evolution from the simplest biological ICE — a cell — to complex technical devices and the human being, who is modeled as an object controlled by a biocomputer (the brain).It offers a complete algorithm of human knowledge formation, from the simplest perceptions to the most complex concepts. It contains the theory of intelligence, covering both natural and artificial intelligence, and the theory of a universal representation language.A short description of TMI and a link to the complete work:https://doi.org/10.5281/zenodo.23403895Boris Sunik, independent researcher

情报分类：工作与职业机会
分类依据：内容涉及招聘、求职或职业发展
信息来源：Hacker News 新项目
发布时间：2026/10/18 00:54:50

Original Summary

Post a reply