The current "PhD hang up" in tech, if it exists at all, seems to be driven by the competition for ML talent (edit: ML research talent).
To some extent, this is understandable: a typical undergrad CS curriculum is terrible preparation for ML research. I'd even say almost any analytical field: statistics, signal processing, physics, math, etc. is better preparation for working in ML than a conventional undergrad CS curriculum. But PhD in CS with specialization in ML/AI would be ideal for most employers, since exposure to more advanced stats, linear algebra, estimation theory, etc. becomes likely.
The other factor is that ML research is moving quickly and knowledge is being rapidly disseminated in academic conferences. A PhD has an advantage there over a typical undergrad, since they've already spent years learning to parse academic literature.
Have you ever worked on ML systems in industry? 90% of all your problems are engineering problems. The myth that you should hire some STEM PhD because they are good at math is ridiculous.
I have worked at several AI/ML firms. Only in extremely rare circumstances do your problems require PhD level ML knowledge. But the amount of tech debt produced by 'scientists' is horrendous.
Can someone who studied CS 15 years ago and has since worked as a webdev produce robust ML systems? Probably not without significant training. But someone who has done a CS Masters in the last 5-7 years and has shown interest in the subject will run circles around a PhD who never had to write non-academic code.
I totally agree that 90% of ML problems are engineering problems and having a PhD is not relevant to those cases. My comment was more about the 10%, the more research-oriented parts, which some companies have become very competitive about in the last few years.
> But the amount of tech debt produced by 'scientists' is horrendous.
Absolutely. I've been trying to exit a research-oriented position for the last few months because I want to work with competent engineers again.
From my experience, 90% of CS undergrads hate math and don't want to do math. And make that 80% of undergrads doing "research" in machine learning. PhD is a pretty good filter to filter out people who don't like math.
Also PhD's simply have more years experience in writing code than MS graduates. So your comment about circles don't make sense. What makes you say PhD's write worse code? And please don't compare the code they write alongside their research to production code, because research code is throwaway code (seriously).
For a CS PhD, I think it's safe to say they got into CS because they have some interest in coding. But that's not always the case for students in other STEM fields. There are plenty of PhD's who were never formally trained in software and their advisors (especially older advisors) consider programming ability akin to operating a TI-83+.
I run a team deploying machine learning systems in production, and I agree with most of this. None of us have a formal background in machine learning, and none of us have PhDs. We all came from traditional software engineering backgrounds, and I'd say we were all pretty strong in that area.
Then again, we're not Google Brain, and we're not attempting to do original research or publish papers. We have very concrete goals for systems we want to build, and so far we've been able to achieve success using already discovered methods. Almost everything we do is traditional engineering to build training datasets and deploy models to production on live data. We also spend a considerable amount of time preparing and delivering presentations on our results and services, so believe it or not interpersonal and presentation skills are actually quite important.
To some extent, this is understandable: a typical undergrad CS curriculum is terrible preparation for ML research. I'd even say almost any analytical field: statistics, signal processing, physics, math, etc. is better preparation for working in ML than a conventional undergrad CS curriculum. But PhD in CS with specialization in ML/AI would be ideal for most employers, since exposure to more advanced stats, linear algebra, estimation theory, etc. becomes likely.
The other factor is that ML research is moving quickly and knowledge is being rapidly disseminated in academic conferences. A PhD has an advantage there over a typical undergrad, since they've already spent years learning to parse academic literature.