What AI Teaches Us About Original Thinking (And How to Ace an Internship at a Bank)
It goes without saying that so much hype is around AI these days. One of my favorite stories of the last decade around this area is the story of Google / DeepMind’s AlphaGo vs Lee Sedol Go match. Rewatching Netflix’s excellent documentary around this historic event for this ancient game also reminded me of a much more important, less covered story afterwards.
The neural net that defeated Lee Sedol, one of the legends of the great game of Go and the presumptive winner (by a 5–0 margin) before the games even started was primitive compared to what DeepMind came up with next. It wasn’t covered as much in the press, but it deserves more attention because it shows a broader lesson for us as well.
The original AlphaGo neural net was bootstrapped with thousands of human played games as a starting point. It makes sense — start the AI with what the state of the art humans have done for 2500 years and then let it iterate from there and that’s the optimal approach, right? Wrong.
With the successor to AlphaGo (I’ll skip past AlphaGo Master), they produced AlphaGo Zero. This was an AI that had to learn the game of Go with no prior access to the wisdom of human beings playing this game for millennia. It ended up trouncing the AI that had a starting point “lesson” with the human games.
What’s the point of this? Well, I think it emphasizes the importance of not being tainted by past thought, from the beginning at least. Here’s a quick story.
When I was a trader at Goldman, I was part of a small group in my team that helped evaluate summer interns rotating on our desks. We’d give them projects to work on to see if they might make a good fit, and ultimately, a formal offer for full time work.
One of my coworkers and I gave an intern a particular finance project to think about that was pretty well understood and solved, but we wanted to test how he would think about it. After he left with the instructions, my colleague turned to me and said, “I bet you he just Googles it right away to find the answer.”
I responded that the better way is for him to think about it first without being biased by how us old guys looked at it — he could always look it up later. But once you see how things were already being done, it will forever slant the way you look at the topic. He laughed and told me that he didn’t think most people thought that way. But I know if that intern had come up with something that was original and not copied from the internet — even if it was wrong — he would get an offer upvote from me immediately.
The broader idea is this: the blank slate mindset is really special, and you can only use it once. This gift of naiveté opens all options before being weighed down by previous thinking. It’s the best way to find original ideas.
Because AlphaGo Zero did not train on previous human played games, it was free to explore strategies that it would never had if it had been fed, as a starting point, all the tried and true strategies we had felt over confident about for years.
Whenever I have given an informal class about a topic that I know students haven’t seen before, the first thing I do is to entreaty them to not look it up online, research it, or anything of that sort. Do original thinking first. You can always look it up and have it corrected later.
And maybe, just maybe, you’ll think of something nobody else has.