All our games turn into Calvinball - by Ben Recht - arg min

I appreciate everyone’s feedback on Monday’s post. It’s helpful when we all stare at the magic mirror together to see what we see. I’m now compelled to spend the rest of the week on this Bitter Lesson as there’s a lot to unpack. And also because I like bitter things. I want to first pick on some peculiarities of the essay, and then turn to why I think Sutton, who was an expert in games but not pattern recognition, misunderstood the transformations that occurred in supervised learning in the 2010s. In subsequent posts, I’ll turn to my own bitter lessons from the 2010s. One thing that always bothered me about this essay was Sutton’s examples conflating parlor games/Gesellschaftsspiele and pattern recognition. His four examples are chess, go, speech recognition, and computer vision. Board games and pattern recognition are fundamentally two completely different problems. And their computational solutions were also quite different, even though both use “neural networks” as part of the solut

saved by

• Judah D'souza