Re:How to graduate your PhD when you have no hope
by Animats on Sunday March 31, 2024 @04:16PM
(#39887652)
> When someone asks me if it's worth doing a PhD I tend to say that you need to like the direction in which science is going. If you don't like it and are rather rebellious, it's better to start a company.
Right. I got an MSCS from Stanford in 1985. That was about when it was becoming clear that expert systems were a dead end. The Stanford faculty was heavily invested in expert systems at the time, and were in serious denial. This was the beginning of the "AI Winter". I could see this, because I'd been doing proof of correctness work, where machine-powered predicate calculus could deliver results. It was clear to me that those methods were too brittle for the real world, although they worked fine on the rigid world of computer programs. Hammering the real world into predicate calculus only works if you already almost know the answer. (Drew McDermott's essay, "Artificial Intelligence meets Natural Stupidity" is the classic commentary on that.)
There were some neural net people around, but they were stuck, too, partly because they needed a few more orders of magnitude more compute power before that approach started working. Neural nets had been around for decades, and on a good day in 1970 you could recognize handwritten numerals, slowly. Computer power wasn't increasing that fast in those days. In 1970, there were 1 MIPS mainframes, and in the early 1980s, there were 1 MIPS VAXes. Success with neural nets lay many orders of magnitude in compute and several decades ahead.
So, instead of academia, I did some stuff for a little company called Autodesk.
I never expected computer science to be captured by the ad industry.
Right. I got an MSCS from Stanford in 1985. That was about when it was becoming clear that expert systems were a dead end. The Stanford faculty was heavily invested in expert systems at the time, and were in serious denial. This was the beginning of the "AI Winter". I could see this, because I'd been doing proof of correctness work, where machine-powered predicate calculus could deliver results. It was clear to me that those methods were too brittle for the real world, although they worked fine on the rigid world of computer programs. Hammering the real world into predicate calculus only works if you already almost know the answer. (Drew McDermott's essay, "Artificial Intelligence meets Natural Stupidity" is the classic commentary on that.)
There were some neural net people around, but they were stuck, too, partly because they needed a few more orders of magnitude more compute power before that approach started working. Neural nets had been around for decades, and on a good day in 1970 you could recognize handwritten numerals, slowly. Computer power wasn't increasing that fast in those days. In 1970, there were 1 MIPS mainframes, and in the early 1980s, there were 1 MIPS VAXes. Success with neural nets lay many orders of magnitude in compute and several decades ahead.
So, instead of academia, I did some stuff for a little company called Autodesk.
I never expected computer science to be captured by the ad industry.