Woke up early, took a quick shower, packed up to play basketball later, had a quick tasty breakfast, and went off to meet Muriel!
Today was an English day and we had a lot of fun arguing about the future–myself generally optimist and believing in diverging possibilities, Muriel generally pessimist and concerned with certain technologies forcing their own adoption and resulting in convergence to dystopia.
After that hour, I was off to the Höngg. Arrived at QI early, early enough to talk to people; got chatting with a classmate in the elevator as well as my buddy David in the class. And I managed to stay on top of the material (for the most part!) My reading in the bus had paid off. I still have so much brainpower focused on the details of the math that I can’t think much about anything else, but the details of the math are 90% of what’s going on, and at least now I can understand them. I asked a couple of questions :)
At the break, I had a welcome surprise! My friend Carolie of the cactus spines presentation from around a month and a half ago, who I’d thought had disappeared from the class forever and whom I’d never see again, came up and greeted me! It sounded like she’d gotten a bit overstressed and then fallen sick and had had to stay home for a bit, but was now back in the class. We talked about her ML project, quantum information… I grabbed her WhatsApp and we might have lunch tomorrow :D
The next half of the class was pretty tricky, but I was understanding it! Definitely need to read more about semidefinite programs. I’m gaining a deep appreciation for the power of phrasing quantities as optimizations though–it allows you to reason about operations that restrict the set of things to optimize over (the feasible set), it turns inequalities into equalities, and it lets you bound the optimization by the value of any feasible input. (Clearly the max over all X of bla is bigger than bla(thisRandomX)).
After that I went to basketball! It was really great. The people who play at the Höngg are much nicer and less stressed than the intensely competing people who play at Polyterrasse–they still play hard but the attitude is more fun, and congratulating the other team for good plays, etc. I had some nice plays and played good defense–still a tad tentative on rebounds though, I’m worried I’ll knock over other players (especially girls). But I haven’t knocked over anybody in a while.
Showered with my basketball buddies, and took the bus home with some of them. I tried to speak a little German and had a lesson on proper Swiss German pronunciation.
At home, typed up problem 3 of RandAlg, did the Jane Street quiz, and set out to meet my RandAlg friends. We mostly got stuck on problem 2 doing some stupid integral. Geez–those will never stop plaguing me. I knew there was a better way to do the problem, but couldn’t quite get it. Found out one of my smartest buddies, going to the World Finals of the ICPC International programming competition in Beijing in just a few months, had been rejected from Facebook with no interview–what is wrong with the world? Maybe you need connections.
I returned home, a bit disgruntled, worked a little bit. Made myself some tasty eggs, hung out with Juyoung and her friend from Denver who had come to visit (!).
Back to work, finished typing up problem 3, then Skyped Grace! She had sent me this terrific little game theory demonstration. I walked through it really quickly before we chatted. Do it yourself and see what you find most interesting before reading on:
My takeaways:
- The strength of the Copycat (tit for tat) is not really defending yourself from cheaters, (the cheaters are good at that) but making sure that you can benefit from other Copycats.
- Introduction of noise (miscommunication, mistakes, whatever) hurts the Copycat, since two Copycats who start by cheating will always cheat each other, even if the first cheat was a mistake.
- Players who are more forgiving will in this case do better, since even though they are more vulnerable to cheaters, they stand a better chance of finding allies.
- The Simpleton (reinforcement learning, if they cheat (undesirable) then do the opposite of what you’re currently doing; if they cooperate, then do the same thing as you’re currently doing) is actually effective in an environment with a lot of suckers (always cooperate). This is because Simpletons, unlike Copycats, can learn to exploit the suckers. Think garden-variety abusive jerk.
Crazy! Makes me think in some respects I’m spending too much time solving exact math problems and that this kind of numerical simulation can be very insightful.
Chatting with Grace was terrific! I had so much to talk about, with everything that I’d been reading. We had a couple Aha moments where one of us clarified something for the other. Finished with some cute 1-minute stories. Grace is singing (in the chorus) in an opera :D
Cheerful after that, I spent like 5 minutes looking at one more RandAlg problem, wrote this log and crashed out.