I had the same moment of realization a few years ago when my workplace decided to hold a kaggle-style competition to select an ML consultancy to prioritize incoming prospective customers. The consultants had given us their sales pitch about how they had the best algorithms, credentialed specialists, proprietary data and so forth. Meanwhile I was just an Excel Monkey who convinced his boss to let me teach myself ML on the company clock.<p>So off I go to pull my data and spend about half a day building a script to score the leads with a randomforest package, literally teaching myself the syntax as I go. A couple weeks later, the scores come back and I'm second of seven, beating six companies with multi-million dollar funding rounds.<p>My secret? Nothing really, I knew the data well so I created a few features that the outside consultants apparently missed, which apparently made up for my lack of ML skills. Turns out the 'proprietary data' they were touting was basically just census data and none of their engineers could be bothered to think through the real world situation on the ground and just fed the data into their system raw.<p>I think there are some really interesting opportunities for people who are working on the cutting edge of ML with things like self driving cars, AlphaGo, and others, but there are an awful lot of business problems can be solved with 'good enough' solutions with relatively simple, undergrad level algorithms and good domain knowledge.
I have a background in mathematics and am currently studying Machine Learning in a Master's program. Soon after I started I came to the same conclusion as the author of the article. It is becoming very easy for someone with some basic coding ability and shaky mathematical foundations to do some online courses, emulate a few tutorials or maybe even sign up for a bootcamp and, voilà! 6 months later they are more than competent enough to quickly put together ML solutions that are close to an 'optimal' solution which requires much more in-depth knowledge and engineering time. Close enough that the quality/amount of data and feature engineering (basically, domain knowledge gained through general problem solving skills rather than ML skill) is far more important than the details of any model.<p>This is why, when I start applying for jobs soon, I plan to maintain and frame my skill set as a quantitative problem solver, not an ML specialist (as I had thought I would do when I started this Master's degree). Soon the value of being able to build a good ML model will decline, but hopefully the knowledge and skill set of those with a proper training and understanding of this field (mathematical ability, computer science/programming ability and problem solving ability) will still be valuable.
If you defined "AI skills" as being able to take some cleaned data and plug it into some python library to solve some pre-defined objective a la kaggle, then sure, thats no more or less easy than any other kind of data transformation. The difficult part is coming up with said data and objectives, being able to ask the right questions and integrate it into an actual monetizable product. There is still a huge long tail of applications that these classes of algorithms promise, even if the number of PhDs optimizing these algorithms against ImageNet or MNIST plateau.<p>I see "AI skills" as more of a form of literacy as opposed to python-notebook-jockeying. That is, having a general knowledge of what these algorithms are good for, being able to recognize when and how a classifier or a regressor could help, and compiling the data necessary to accomplish it.<p>I work at an "AI" startup as an engineer, mostly on the applications side. We have a human analyst team that annotates our data for us, with the idea that we will eventually supplement that team with a classifier that will help us scale the operation with a similar team size. Data quality is certainly a challenge, and a lot of that is because the in-house app that our analysts use isn't very flexible, and they have to keep track of much of their work with excel spreadsheets, which adds a lot of mental overhead to their jobs. Additionally, a lot of the work is necessarily subjective, and the analysts have to move quickly through the data to annotate it, which leaves room for error. I realized that by improving the UX of our internal apps with good ol' application design/engineering, we can improve our data quality, and thereby improve our model. I consider that an "AI skill" even if I never have to touch a python notebook to accomplish it.
So I got started working on ML in 2015 as well, and it has gotten far easier, but the thing that has gotten easier is the engineering side of things, not the ML bits.<p>And tbh, "take the latest model and tune it on your data" has gotten easier, but not trivial, since the code people release for the latest model is what I would call "research quality" code. Some labs will release production-ready code, but it's pretty rare. Particularly as more labs pick up PyTorch and you realise you can't afford to run that in prod.<p>I think the author is right about basically everything else he says, but I also don't think it's an argument against picking up these skills, since they're going to be foundational to basically everything going forward IMO. So while there won't be a huge premium for ML engineers, it will be a necessary skill set if you want to work on challenging problems.<p>And at some point the curve for adding data flattens out, and it becomes cheaper to hire engineers to think than subject matter experts to label data.
I’ve noticed this as well. I would argue that ML might be easier than regular programming. Often times it’s clear cut which model to use, there’s enough examples online to do it, and the amount of code is minimal. ML is no different than any other programming domain that has a good framework, it quickly becomes plug and play.