@oprydai Metal requires seasoning and structure. Do these optimized designs introduce added complexity in that sense? Also, I can imagine that if one could control detail of the metallic crystalline structure, one could find further DOFs for optimization.
@mathelirium It's been going from very exciting to depressing at every new model launch. I'm currently at the very exciting part, so net result is super positive.
@mathelirium I'd guess its a leftover from brain analogy, in particular minimizing the free energy function. It has an equivalence with properly defining a statistical state counting function, yielding entropy
Is there any result in Probability Theory with this kind of reach? It’s kind of insane.
Markov’s inequality looks almost too plain to matter. It’s one line with no distributional assumptions. No assumption of smoothness, variance or independence. Yet, it still gives you a real handle on tail probabilities.
Feed it (X − μ)² and Chebyshev falls out.
Feed it exp(λX) and you’re suddenly in Chernoff territory, and from there it’s a short walk to Hoeffding and Bernstein once you add the right bounds.
Feed it a sum of indicator variables and you get the union bound vibe that props up half of those with high probability arguments.
Same move every time. Markov, plus a smarter choice of X, and this tiny idea turns into a whole toolbox for Modern Probability, Statistics, and Learning Theory.
#ProbabilityTheory #MarkovInequality #Concentration #Statistics #MachineLearning #LearningTheory