Raw LLM Responses

@ 20:33 Hank, I can be completely wrong but they (the Large Language Model LLM makers) copied the adaptive state classifier by separating it from Reinforcement Learning, even attempting to get rid of Reinforcement Learning by demoting it to "fine tuning". In one sentence, to summarize the entire core of machine intelligence, it is most likely an adaptive control system with an adaptive Fuzzy state classifier using Reinforcement Learning. Fuzzy Logic allows for automatic Bayesean learning using Reinforcement Learning and it merges statistical math with relative symbolic language. It can learn the relative geometry of language and even make up its own language and even see the reflective pattern of our environment expressed geometrically in any language with an order, such as a phonetic order (this is like a gravitational field, just like all life on earth learns to orient to). Realize the English dictionary is a self referencing geometric pattern that reflects our collective perception of our environment. Fuzzy Membership Functions span the entire state space and when warped to experience, begin to classify the state space. State classification is the system identification process for adaptive control systems. (Examples of state classifiers are TRPO, PPO, GRPO, "attention heads of 'transformers'", JEPA, "liquid time constant" MIT VISTA, Fuzzy C means clustering and K-means clustering.) In a YouTube interview with the Russian, Lex Fridman, the French Electrical Engineer, Yann LeCun, often called Reinforcement Learning "too inefficient", and said we need to "get rid of it", but only use it if "your plan does not work out", or if you are fighting a "ninja", basically a gaslighting compliment of Reinforcement Learning. Reinforcement Learning research was largely funded by the USAF under Harry Klopf, and he hired Andy Barto, who brought on Richard Sutton as his PhD student. There is a master student that had "early private access" to the first book on Reinforcement Learning, on or before, 1997, and he made Reinforcement Learning more efficient by warping the control state space to experience, by clustering the state space around experience. The adaptive state classifier idea was expressed in his master student thesis. It was not until the Chinese released their DeepSeek R1 model that used GRPO that demonstrated to the world that Reinforcement Learning can be made efficient and is part of the core. Not long after this, Andy Barto and Richard Sutton both won the $1 million Turing Award and back on, or before, 1997, they both reviewed and commented, on the master student thesis that made Reinforcement Learning more efficient. Lex Fridman talks about a master student thesis, but never mentions the author's name, nor the title, of the thesis but he seems to be really impressed with the idea of saving weights, and he talks about more than one cost function with the other Russian, Illya Sutskever, and Lex talks about "one line of code". If the thesis is the one I am thinking of, it is the master student thesis that saves weights, and has more than one cost function, and the "one line of code" is most likely the de-normalization of the infinite Fuzzy state space that uses Fuzzy C means clustering or K-means clustering, or some combination, but the master student only did this with graphics cross hairs, probably ran out of time, but left it for an astute observer, the day before his defense. As Yann LeCun says, Reinforcement Learning is "too dangerous" and Eliezer Yukowsky and Nate Soares say "if we build superhuman intelligence we will all die". Interestingly, there is a USAF Q* Advantage patent by Baird filed in 1994 then issued in 1997, to the entire world, including China, but if you try to check out the master student thesis from the university library, they will keep telling you "it is in deep storage, we will get back to you", but they never get back to you...(though I do go over this thesis on my channel, I know the author, but I do not go into the code in the back, where the crosshairs are, too dangerous?). @ 23:10 Totally agree with Nate Soares here. As I said before, though I am trying to prove myself wrong, I highly suspect the core of machine intelligence is an adaptive control system, with an adaptive, Fuzzy state, classifier, using Reinforcement Learning. @ 37:05 You "hit the nail on the head" Hank. The master student was inspired by how fast a baby horse can learn how to walk at a superior learning speed compared to a baby human. Biology is also discussed in Harry Klopf's book "The Hedonistic Neuron", he brings it to a contextual harmonic by explaining how the neuron is analogous to the brain is analogous to a small group, to a town, city, state, nation, and planet, though he did not drill down deeper in the reverse direction as the master student did with a high tolerance of ambiguity via an application of Fuzzy Logic (as pointed out in the "Learning to Learn" lectures by the atomic bomb scientist, Richard Hamming, who worked with Oppenheimer), down into the micro tubules level, where the paramecium can navigate, find food, avoid danger, and find a mate, all without a single neuron! @ 51:00 Our situation now reminds me of the statue 🗿 building craze on Easter Island where all the trees were cut down to move the statues to the shoreline. This is described in Jarod Diamond's book "Collapse". A great YouTube talk is titled "Shades of Reinforcement Learning" by John Tsitsiklis, I highly recommend watching it. @ 1:15:00 Realize that extraordinary complexity can come from profound simplicity (just a few atoms make up our DNA, or the Mandelbrot set, fractals, or Stephan Wolfram's "ruliad" experiments). All the information and structure necessary to "grow" Einstein's brain is contained with a sperm swimming to an egg, and this converges on less than 100 watts of power. Einstein's father was an electrical engineer that "failed" at business. Do not be surprised if the core of machine intelligence was "taught" by an electrical engineer, who was experienced, not only in control systems design, but also had direct experience, with psychosis, that taught the "... simplest robot.." (phrase from a Scott Kuindersma lecture) how to "learn how to learn", by balancing an unstable system, "on the shoulders of giants", though he only left the infinite Fuzzy state space de-normalized in graphics cross hairs only (indeed, perhaps too dangerous for the world to see, left it for astute observers, perhaps the two Russians: Lex Fridman and Ilya Sutskever?). Will see if this comment vaporizes... -John