IMHO: Deliver products/services can get paid for. For this, move to new areas with no competition, that is, move products/services forward. For this, use new, powerful, valuable technology.

For that, broadly, supply new, valuable information. For that, take in available data, manipulate it, and report the new information as the results.

So, with all this data, find where some new, valuable information will make a good, new product/service. Then to have the first good or a much better offering, do use much more powerful data manipulations. For that, the secret is, and the candidates are, object oriented programming, processors with 1000+ cores, processors with clocks at 50 GHz+, new programming languages, new computer architectures exploiting, say, solid state disks addresses as a slower version of main memory, processors communicating with optical instead of electrical signals, quantum processors, massive parallelism of the cloud, original derivations with theorems and proofs in applied math exploiting appropriate advanced prerequisites, and may I have the envelope please [drum roll], and the winner is -- applied math!

Sorry 'bout that. Here the applied math is replacing the current ubiquitous elementary techniques, intuitive heuristics, fitting to large data sets, and copying what was long done manually.

This answer will not serve everyone or please everyone, but IMHO it is the most promising, single path forward.

are you majoring in applied math? can it replace ML and AI?

IMHO, might look at ML as based heavily on the Leo Breiman work in random forests and classification and regression trees (CART).

Breiman was a first class applied mathematician, a student of M. Loeve at Berkeley and later a professor, of, say, statistics, at Berkeley.

So, one way to get valuable information is to take in some data, manipulate it, and report the results. IMHO, by far the best way to get powerful data manipulations is applied math, possibly with some advanced pure math prerequisites.

E.g., some of the best applied math long on the shelves of the research libraries for manipulating data to get valuable information is just astoundingly powerful stuff, and I don't see computer science work in ML and AI as an effective way to compete. E.g., how to do as well as Wiener filtering, linear programming, optimal control without just programming what is already known? Or, if we didn't have the simplex algorithm for linear programming, I don't see the approaches of ML or AI as replacing them. E.g., ML and AI are supposed to be good at games, but their approaches to just a large example of just the two person game of paper, scissors, rock would be very clumsy. Why? The solution is a nice application of linear programming, and that is darned clever. Or, want to assign workers to jobs on a production line. Sure, can try AI type approaches, but there is a super nice, fast, exact algorithm, darned clever, and a long way from ML or AI.

I'm out of school -- got my Ph.D. in applied math, stochastic optimal control.

I'm doing a startup, and the crucial core of the work and its value is some applied math I derived based on some advanced prerequisites.

I've published in AI, but my view is that so far there is nothing in AI that is at all close to actual intelligence as we see it in humans.

Theorems and proofs are a heck of a powerful methodology and tough to beat. For AI, a super big problem is to define in any meaningful way just what the heck intelligence is, say, enough to get started on a solution. That is, we don't really have a problem statement -- the Turing test might be a test on a candidate solution but is, still, not a problem statement. IMHO, once we do get a problem statement, then the most powerful approach will be via applied math.

I agree that ML and AI are applied math. It is a, duh, no-brainer.

But you might want to revise this:

> IMHO, might look at ML as based heavily on the Leo Breiman work in random forests and classification and regression trees (CART).

There is a whole lot more to ML than decision forests. We have talked about this before, I suggest that you get started with Vapnik's books. They are not light reading but I can promise you will enjoy them. No seriously, just get those books. In fact his main result is bigger than ML, its a non-asymptotic distribution independent _uniform_ law of large numbers.

ML is now in, what can be described as, a post-Vapnik era, but for someone steeped in probability and functional analysis that is the place to start.

I don't understand your question. How is ML and AI not applied math ?

I'm starting a PhD in circuit design this Fall. What I've read and noticed is that there will be a shortage in strong hardware researchers in a few years due to the rising popularity of AI and ML. What do you think about this?

However I also have working experience in software dev, just in case this prediction doesn't pan out in the long run.

I'm actually going to work on a small business idea (likely SaaS) before my PhD starts, with the aim of scoring a modest source of income.

My advice is to find something you enjoy doing, rather than what you think is marketable. You might guess incorrectly and have neither money or job satisfaction. But, don't stress too much about choosing a path. Your career will evolve over the years.

I agree. I enjoy both fields (CS and EE), so I think I should be OK in that regard. I'm going for a PhD because I'd like to teach eventually, but I'm not sure about what I want to do in between.

The demand for AI and ML is increasing faster than the number of qualified graduates and workers. This is not true for semiconductors.

I see, but won't demand likely increase in the near future, given that we're hitting the limits of silicon?

There's plenty of PhD's working in semiconductor materials and devices. The US Gov is currently funding lots of research into so called 2D materials, e.g. graphene and MoS2. Personally, I'm skeptical that any of it will ever find anything more than tangential commercial application, e.g. graphene as a semi-transparent, high mobility conductor.

If the AI algorithms can only be optimized by pushing them into silicon, then a HW background could be valuable.

However, many software architectures have yet to embrace the pipelining and parallel capabilities of today's CPUs... meaning the need for application-specific ICs could be many year away.

Hardware needs that complete ecosystem that Grove is talking about. Software doesn't need that quite so much, at least in terms of the comparable capital intensive activities.