[college_physics]

Even ten years ago this would have been a difficult career switch without a serious (graduate school level) investment in the very specialized mathematical background of deep learning models. Nowadays its even harder. The mathematics haven't changed much, but the easy pickings are long gone. Now you'd have in addition to target specifically a sub-domain, with all its associated context and heuristics, whether that is text, image, audio or more exotic stuff like protein folding.

Nevertheless the above caveat concerns the overhyped "AI/ML" space. Digitization, quantification and automation of information flows is much more general phenomenon and with a more modest investment in statistics / data science you can be part of this general trend of productionizing "analytics". Just don't expect bubble era FAANG salaries. These were the product of very specific conditions.

If a general data science transition works out for you and you are still interested in the AI/DL/ML bandwagon after you are in better position to understand why and how it works you could easier drift into that space later as it is just an extremely specialized subset of that world.

[khazhoux]

> Even ten years ago this would have been a difficult career switch without a serious (graduate school level) investment in the very specialized mathematical background of deep learning models

Isn't it the case that a big chunk of the AI that became popular 10 years ago (via Norvig, Thrun, Ng) is no longer relevant today (except as background)?

[throwawaybbq1]

No .. when you do MNIST in a DL tutorial it all looks easy. The problem is when things don't work as expected (real-world TM). I have a freakin PhD in CS (top school where a bunch of this stuff was invented .. some of my friends are "Gods" now) and did a bunch of basic ML in my thesis but ... I never got deep into the math (my PhD was in something else .. a math-lite field). I know a metric space but do understand it well enough? What the hell are kernel smoothing methods? Laplacians, Jacobians, what, what, what??? (Fine .. I was joking with the last ones .. I remember vector calculus) I work in AI and feel I study ALL-THE-TIME. Please don't torture yourself like I did. Get a proper Masters (but realize you need a PhD in the specific sub-field and strong math chops).

Rant Over .. I also had a thought for the OP. In Web dev (and other kinds of SW dev), you kick it hard enough and it works. I had a job offer at a famous company where they wanted me to do a vision model to detect theft. I got the offer but didn't take the job .. one factor was if it didn't work with my existing toolbox, what does that mean for me? Do I get fired or I quit myself? This was a serious question I posed to myself.

In research gigs, you take on hard problems and try your best. In an industrial/startup setting, AI of hard problems requires incredible training and self-confidence at the leadership level. A data engineer does not require this hard stuff and you just work under some scientist. But be aware the turn around time for experiments is very fast (like the scientists had 5 new ideas on the whiteboard by the time the dev team walked from the conf room back to their desks). I don't even know how my engineers keep up with it.

I think AI jobs where you do actual innovation/exploration are incredibly hard and require a ton of investment (personally in terms of a near 24/7 job and your employer). Even then success isn't guaranteed. I think there are much easier jobs out there (like Cloud, Security) that have an equally bright career outlook .

[mbeex]

> Laplacians, Jacobians, what, what, what???

Can you really get a PhD in anything even remotely STEM-related these days without at least knowing what those things are? I'm growing old...

[brutus1213]

Parent said they were joking about that example. Not every STEM PhD is math heavy. CS itself is big field with subareas that may or may not require fluency in topics such as optimization (e.g. HCI, Databases). ML/DL is optimization heavy, no? So the point is valid I think.