Why AI Creators Still Struggle to Understand Their Own Systems

Summary

The article begins with remarks from Anthropic CEO Dario Amodei, MIT Technology Review, Yoshua Bengio, and mathematician Roman Vershynin about the limited understanding of how advanced AI systems work. Its author argues that neural networks learn regularities from human-created text, images, and other patterns, so understanding AI knowledge formation requires first understanding how humans form knowledge. The article says earlier efforts to build a mathematically correct, practical theory of knowledge produced no satisfactory result, while claiming that an alternative has existed since the late 2000s but has been overlooked because it uses concepts associated with programming languages rather than conventional mathematics. The author identifies this proposal as the Theory of Meaningful Information (TMI), described in a linked Zenodo work. TMI defines information as the value of a variable used in an algorithm and introduces “Information Capable Entities,” ranging from cells to technical systems and humans modeled as biological computers. According to the author, the framework describes the formation of human knowledge, intelligence, and a universal representation language, covering both natural and artificial intelligence. These claims are presented by Boris Sunik, an independent researcher, and the article does not provide external validation or evidence that TMI is accepted by the wider AI or information-theory community.