sat through 45 minutes of mumble jumble at rutgers. it was fascinating to listen to my brother speak in a language i do. not. understand. and then his board would ask him questions in mumble jumble and he'd answer then back the same way. wow!

at the end we drank champagne and ate chinese food and cake and they called my brother "doctor."

alex's advisor asked me how much if it i understood, which is how i found myself trying to explain computer learning to an expert in the field. i stumbled through the rudiments i gathered. you give the whatever (agent?), okay, you give the agent a problem to solve. wait wait, the programmer's dilemma is that you want it to learn and do the solving itself. you need it to use the information it has gathered from experience but you also need it to try new things in case it stumbles across a better way. alex's method tells it to be "optimistic" and assume a positive result from things it hasn't tried yet, but i think it can also remember what it's done before that worked. so you start with that, and then for each trial if the agent succeeds it gets a reward (R). this is where it gets strange for me. i know what a reward is: ice cream or a smile or cash or free time. what does a reward mean to a computer program? well, it's programmed to try to get the reward. it's also programmed to risk failing with the hope of getting a greater reward.

i thought i would come across as naive asking alex's advisor about this, asking about the diff between a computer brain and mine. it's a romantic notion, wondering about computers thinking and yearning. but it turns out it's not. he tells me it's not a fundamental difference, to his thinking. our reward systems are more complex and thus more difficult to plumb, but it's a difference of degrees than anything else. the algorithms they are writing have simple choices and simple rewards, but as the field advances they will surely grow more complex.

i tell him about a character from mieville's perdido st station: a massive computer which has achieved sentience. its goal is to accumulate knowledge and all of its actions are premised on that goal. it calculates probabilities, makes decisions, and acts on them. alex's advisor tells me that's the idea, and i've understood what he's working on better than i thought. some of the students in the labs actually work with robots, little robot dogs and cars. give them a challenge and let them work out a solution. alex's work takes place entirely in the computer. he creates a digital problem and his computer thinkthinkthinks itself through the maze.

they all joke about computers making decisions that could affect people, they've all seen terminator. do i worry about it?