Lost a whole afternoon this week down a rabbit hole I didn't expect to fall into. Marvin Minsky died on the 24th, and I only found out about it two days later scrolling through a mailing list I mostly ignore. He was 88, complications from a cerebral hemorrhage, and if that name doesn't mean anything to you, he's one of the people who basically invented the field of AI as a formal discipline back at MIT in the 50s and 60s. Cofounded the AI lab there. Built one of the first neural network machines in 1951 out of vacuum tubes and old B-24 bomber autopilot parts, which is a sentence I love typing.
I've got a beat-up paperback of his "Society of Mind" that I bought for two dollars at a used bookstore in college, the kind of book where every page is a single short numbered idea and none of it really resolves into a normal narrative. I never finished it front to back. I dip into it every couple years and always find something I underlined that I don't remember underlining. Thats the kind of writer he was, apparently endlessly generative and a little allergic to tidy conclusions.
Heres the part that made this week weirder than a normal obituary though. Three days after Minsky died, on the 27th, DeepMind published their AlphaGo paper in Nature, and it turns out they'd already beaten European champion Fan Hui 5-0 back in October, they just sat on it until the paper was ready. Everyone's now talking about the March match against Lee Sedol in Seoul, which by most accounts is the real test since Fan Hui, while genuinely strong, isn't at Sedol's level.
The thing nobody's really connecting is that Minsky, along with Seymour Papert, wrote the 1969 book "Perceptrons" that's usually blamed for kneecapping neural network research for a decade plus. They showed mathematically that a single-layer perceptron couldn't even solve XOR, a genuinely trivial logic problem, and funding agencies apparently took that as license to stop paying for connectionist approaches almost entirely. Whether that's a fair read of the book or a convenient just-so story the field tells itself, I honestly don't know enough to say. I've seen people online defend Minsky and say he was misread, that he was mostly arguing against a specific overhyped claim and not neural nets as a category. I've also seen people who worked adjacent to that era say the funding chill was very real and very much attributable to that book. I don't have a dog in that fight, I just think it's a genuinely strange bit of timing that the guy most associated with (fairly or not) killing off neural network research for a generation died the same week a neural-network-based program cleared a hurdle in Go that people were confidently predicting was a decade out.
I want to be careful here because Go engines beating strong amateurs isn't new, and Fan Hui, again, isn't Lee Sedol. Some of the coverage I've read this week already has that breathless "the machines have arrived" tone that I find exhausting, the same tone that shows up every eighteen months whenever anything crosses a benchmark. I remember the exact same energy around Watson on Jeopardy in 2011, which, sure, was impressive, and also didn't actually change much about how anyone lives. I'd rather wait for March before deciding how big a deal this is. Sedol is a legitimately top-tier professional and if AlphaGo takes that match it's a different conversation than beating a 2-dan European champion, no disrespect to Fan Hui meant by that.
What I keep coming back to, though, is how short institutional memory in tech actually is. Half the writeups on AlphaGo this week don't mention Minsky at all, obviously no reason they should, it's not really his story. But it's a good reminder that the people arguing loudest against an approach sometimes end up shaping, by opposition, exactly the field that eventually proves them wrong on the narrow technical point while being right about plenty else. Minsky spent his last decades pretty publicly annoyed that AI research had, in his view, narrowed itself into a bunch of engineers chasing benchmarks instead of trying to actually understand intelligence. Reading some of the AlphaGo commentary this week, chasing a benchmark is more or less exactly what's happening. I don't think that's a knock on DeepMind. I just think he'd have had opinions about it, and I sort of wish I could read them.
Going to go pull "Perceptrons" off the shelf tonight, assuming I can find it. Pretty sure it's still in a box from when I moved in 2014.