book review: why greatness can't be planned
book recommendation by @bmix012 on x dot com
tldr: everything is gradient descent
most of the book is drawing parallels between ideas from AI and RL to how great individuals developed themselves and their ideas over their careers.
key points:
- you've got to fuck around to find out
- the unpredictable/unforeseeable rewards stem from an array of seemingly unconnected actions
- the determination between obstinacy and determination lies in whether you're focused on the goal vs focused on the method (this might've come from the pg essay i read immediately after, not sure)
- the policies, training methods, and methods you use in reinforcement learning you can use on yourself
- great people (usually) don't plan their actions for the next five years; they look at where they are, what their action set is, and pick the best one from there.
it's like when warren buffett and charlie munger talks about slugging it out one step at a time.
michael bloomberg said something similar:
"To succeed you have to string together many small incremental advances, rather than count on hitting the lottery jackpot all at once. Trusting to get great luck is a strategy not likely to work for most people. As a practical matter constantly enhance your skills, put in as many hours as possible, and make tactical plans for the next steps. And based on what happens, look one move ahead and adjust your plan. Take lots of chances, and make lots of individual, spur of the moment decisions."
the butterfly cover is sampled from an experiment they did on evolution and was used as an analogy throughout the book on how unexpected rewards may originate from mundane things (the butterfly in question was 'bred' out of regular shapes, like squares and circles).
i've been thinking a lot about the things mentioned before reading it, so i find the book quite agreeable. that's kinda all i can say about it.
nice book, short read.