[good_gnu]

> Even when behavior is mostly learned rather than hard-coded, the ability to learn itself must be hard-coded. Such code occupies hundreds of megabytes of data. Consider the first generation of software for self-driving cars. NVIDIA’s deep-learning drivers for self driving cars utilize “27 million connections and 250 thousand parameters.”5 Good programmers try to restrict their functions to no more than 6 parameters when writing code. Thus, a driving system with learning ability depends on over 40,000 programmatic functions.

Obviously the author has completely misunderstood what is meant by "parameters" in this context. The parameters here are those of a neural network and thus they are a result of, not a precondition for the learning process. Instead, he should strictly use the implementation of the neural network for comparison, whose uncompressed human-readable size is probably going to be almost small enough to fill into his size constraints.

And that is assuming there is not a sufficiently good algorithm for learning animal behaviors that is much simpler than CNNs.

[fraytormenta]

Legit criticism, I actually tried to get a clarification on that, what those words mean exactly. I read most of the CNN doc and went ahead and wrote that, because I felt that it was the best meaning in context. I could be wrong about that. However, the CNN image itself must be over 1GB. So if those parameters are simply a matrix fed to the learning algorythm, the algo itself still ends up huge. > "And that is assuming there is not a sufficiently good algorithm for learning animal behaviors that is much simpler" > Right, but assuming something is not is the default unless there is evidence that something is.