AI Is Stuck Because It's Built on Wrong Abstractions, Ignoring Visual Perception

The article argues that the AI industry is stuck because it builds models on abstractions that are weakly connected to how the brain and visual perception actually work. It claims that visual perception is more important than any AI architecture, and that breakthroughs occur when researchers return to concrete examples, as Hopfield did by linking solid-state physics with neurobiology. The author draws an analogy with metallurgy: annealing existed for thousands of years before science formalized concepts like energy and temperature, which are abstractions that don't exist as separate entities in the world. Similarly, modern AI uses abstractions like pixels, convolutions, and artificial neurons that are far removed from biological reality, leading to enormous energy costs and potential fundamental limits. The author mentions that Hopfield's network, inspired by biological neurons, was later modified by Hinton to work practically, but these modifications are distortions of the original biological phenomenon. Visual perception provides 90-95% of human information about the world, works continuously, and is constantly tested by reality. Language writing systems are products of this evolutionary testing, as people discarded inconvenient elements. The author's company, TAPe, develops models that work with real perception rather than pixels and convolutions, achieving outstanding results.