Hinton once said, “If you can sleep well after listening to this, you haven’t understood it.” With all due respect to his brilliance and immense contribution to AI, I believe some caution is warranted in interpreting the capabilities of current AI systems.
AI can undoubtedly surpass humans in knowledge retrieval, just as Encyclopaedia Britannica did in the past. But knowledge is not intelligence. Having used the same underlying neural networks in real-world applications like demand planning and inventory optimization, I can confidently say these are not intelligent systems—they are sophisticated pattern-matching machines.
Backpropagation combined with transformer architectures essentially formats training data into coherent paragraphs. This is not “thinking” in any meaningful human sense. We would need entirely different paradigms or algorithms to replicate human-like intelligence.
As human beings, less than 0.01% of our intelligence comes purely from language. Most of it comes from vision, spatial awareness, sensory experience, and embodied cognition—none of which current LLMs meaningfully replicate. Even with image inputs being added, today’s LLMs remain largely text-based search engines with formatting skills, limited to what they have been trained on. They do not “understand” or learn in any true sense of the word.
Yes, a calculator has beaten us in arithmetic for decades, and with proper training data, an LLM might outperform many in a Math Olympiad. But give it an out-of-syllabus, novel math problem, and it may fail—where a bright fifth-grader might succeed through intuition and reasoning.
In short, today’s AI mimics intelligence—it doesn’t possess it.
Paradigm shifts are what move science forward—Newton couldn’t emerge from Kepler’s math, and quantum uncertainty had no place in Einstein’s world. Same with AI. Unless we shift paradigms, today’s models are just more polished pattern-matchers, not steps toward real intelligence.