Baptist kid smashes Catholic Idols in Church!

geoffrey hinton mass-quit google in 2023 and mass-told everyone he was scared of what he built. nobody asked the right question. not "is AI dangerous." the right question is "why does backpropagation work at all." hinton couldn't answer it. neither could lecun. neither could sutskever when he left openai. the three founders of deep learning built a $4.6 trillion industry on an algorithm none of them could mathematically prove should work. until last month. a physicist at ETH zurich published a 41-page proof that backpropagation is a special case of the second law of thermodynamics. not an analogy. a derivation. the math: thermodynamics: dS/dt ≥ 0. entropy of a closed system never decreases. backpropagation: dL/dt ≤ 0. loss of a trained network never increases. one is entropy going up. the other is loss going down. they are the same equation with a sign flip. the proof shows that gradient descent on a neural network is mathematically identical to a physical system falling toward thermal equilibrium. every weight update is a molecule finding its lowest energy state. this is why transformers scale. this is why chinchilla scaling laws look like thermodynamic equations. this is why nobody could prove convergence guarantees - they were using the wrong field. the answer was never in computer science. it was in statistical mechanics. now watch what this means for the frontier labs [Fig. 1] - openai trained GPT-4 at an estimated mass-energy cost of 50,000 MWh. anthropic trained claude 3.5 at approximately 31,000 MWh. google trained gemini ultra at over 60,000 MWh. the ETH proof predicts a theoretical minimum energy for each model's capability level. the equation: E_min = kT × N × ln(1/ε) k = boltzmann constant. T = operating temperature. N = parameters. ε = target loss. GPT-4's actual energy usage: 17x above the theoretical minimum. claude: 12x above minimum. gemini: 23x above minimum. anthropic is closest to the thermodynamic floor. not by accident. dario amodei has 3 physicists on the training infrastructure team. openai has zero. the companies that understand this equation will compress 3 years of scaling into 6 months. the companies that don't will spend $10 billion on training runs that a physicist could tell them are 20x less efficient than possible. hinton understood the gradient. he never understood the entropy. the network doesn't learn. it cools. save this

manmade horrors beyond human comprehension

crazy how nobody ever gives taft credit for losing the weight