Stephen Downes

The first think I thought of when I read this post from Julian Stodd was simulated annealing and Boltzmann machines. Why? Because Stodd describes crystals as structures that "represent the lowest energy configuration," which is what Boltzmann machines try to do with neural nets. Stodd compares crystals to organizations, and here the metaphor needs rescuing a bit. He writes, organizations "build structure for greatest efficiency, value, predictability, replicability, and of course, ease of control, and potential for oversight and measurement," but that they are "systems trapped at one energy level, dreaming of the next." Sure, they can change state - they can melt into fluid and merge with something else, or they can sublimate into their individual atomic components (a.k.a. people). But they can change energy level, but (just as with matter and neural nets) they need to go through an annealing process, a 'hardening through fire', as it were. 'Move fast and break things' is a clumsy attempt at just such a process; in neural networks they just increase and decrease the bias (ie., sensitivity, not prejudice (that's a different meaning of 'bias')) of individual entities.