We just left the AI Spring of Hope. Now we are in the Summer of Heat. Soon will be the Fall of Failure and, then, once again, maybe this time a real ice age, an AI Winter.
There is old history here: (A) Sure, if count self-driving cars as AI, then that's new software, laser range finders, etc. Okay. (B) Otherwise what business wants from AL/ML is solving essentially just existing problems of their existing business. So, they have some business operations and want to do better.
Well, that situation goes way back to the field of operations research. From about 1950-1970, that was hot stuff, especially for the US DoD.
Well, there is some value there. But there was also a lot of hype. To get the value, usually have to select problems carefully and then do good work. Also, then, there were big bottlenecks on (i) understanding the relevant applied math, (ii) finding appropriate applications, (iii) gathering the input data, (iv) getting the software written and running, (v) being able to afford the computer time. So, yes, there were some successes. Later, I had some. But with the hype which led to low quality efforts, there was also a lot of failure. Net, soon the phrase went "operations research is dead". Similarly for applied statistics. Similarly quite broadly for applications of math to business operations. There were, and are, e.g., RSA encryption, some valuable, narrow applications, but broadly the applied math flopped.
Statistics? Well, it's continued to get used where it's really important -- e.g., industrial quality control, bio-medical challenges, and experimental design, e.g., in research in agriculture.
Okay, now AI/ML are being hyped for applications to business. Uh, as above, we've been there and done that. Valuable? It can be. Easy? Much easier now than decades ago but still, usually not really easy.
Problems with AI/ML now, remembering that mostly we're still talking about solving business problems, e.g., much as in the 100 year old Taylor time and motion studies in assembly lines, much as in logistics, transportation, inventory, facility location, ad targeting, marketing: (i) The successful, new tools of AI/ML have darned narrow successes and are the 100 - 99 44/100% of the total of the promising tools available. (ii) Since the 99 44/100% tools still are not doing well enough to get all excited (right, there is some use of Linpack, C-PLEX, SPSS, SAS, R, ODE solvers, statistical hypothesis tests, linear multivariate statistics, time series analysis, etc.), getting all excited by the AI/ML in the 100 - 99 44/100% is a wild overshoot.
The overshoot is also a waste.
Here I'm just telling business guys tempted to spend money on what are essentially applied math/computing projects to save/make money in their businesses that they would do well, actually much better, just to review their courses in linear programming, production and operations, applied math of marketing, and statistics they got in their BS/MBA B-school studies. That material was rock solid, nicely polished, well justified, nicely balanced, prudent, not awash in hype, and from a long track record of successful real projects that were quite valuable. There's a lot more than what was covered in the BS/MBA B-school programs, but for a consumer introduction that material was quite good -- it was intended to be by the degree accrediting groups.
> I don't think any big VC will take you up on your advice though.
My view of VCs is that they are puppets on the ends of strings held by the limited partners (LPs) who still want to think like commercial bankers advised by CPAs. So, they want to invest capital in assets. They believe that they have bent over backwards far enough to regard traction significant and growing rapidly as a necessary and sufficient asset, but anything before such traction and a dime won't cover a 10 cent cup of coffee. I doesn't matter what I say, explain, reference, etc.: The LPs won't let the VCs pay attention.
> VCs don't work in vacuum, they need to exit at one point, and they need something that is sexy to sell at that point. AI/ML is sexy, and will be for at least the next 5-10 years.
I believe your "5-10" is too long (is that the expected time to exit or the jail term for fraud -- sarcasm here!). Generaly, once projects start to fail, the bad vibes spread very quickly.
For real business uses, in general (not just very narrow niches), AI/ML is in hype mode. The time to the big flop is based on the hype, not the technology of AI/ML. But we've seen hype rise and fall before. So, to estimate the time to flop, just look, first cut, at the history of hype.
The usual exit time for a VC investment and time to full liquidation of a VC fund is ballpark 8-15 years. Well, the hype will flop long before that. So, except for a fast flip, M&A, to someone with more money than brains, AL/ML, aging like butterflies or flowers in spring, will die too fast for the VCs.
For my startup, there's a crucial, technical core, based on my original applied math derivations (I've published in statistics, but the math for my startup is not statistics), but the rest just looks like a promising startup. The tech core provides crucial enabling of delivering the good results -- I can't think of any way to get the good results otherwise, and current, hot AI/ML is really just weak and/or the wrong stuff. E.g., what I derived just ain't convolutional neural networks trained with terabytes of data or anything like that. And the input data I'm using is nowhere near terabytes. If there are some good, practical, valuable, doable, real business applications for convolutional neural networks trained with terabytes of data, terrific. Then that's (heuristic, curve fitting) applied math application n + 1 for some huge n already on the shelves of the research libraries.
Or, look, the computer science departments can teach how to build at least the software side of computer systems, from the ROM BIOS for booting the thing up to complicated applications -- running a bank, a lot in medical records, maybe air traffic control. Fine.
But with AI/ML, the CS departments are reaching to solve the problems the applied math people have been working on for decades, maybe centuries. There the main issue is not how to program but WHAT to program. And to answer the WHAT for an applied math application, darned better well be well trained in applied math. Well, nearly none of the CS profs are so well trained, and that's a problem. Yes, the applied math appllications will likely need a lot of computing, but it does not follow that the people to do the applied math are the CS people. Sorry 'bout that.
I'd like to hear the opinions of E. Cinlar at Princeton.
For my startup, I wrote the code. It's 24,000 programming language statements in Visual Basic .NET Framework 4 in 100,000 lines of typing. It all is like I originally envisioned and all runs, apparently correctly (although I want to do some more testing). The code is nicely documented and plenty efficient enough for production. Only a little of the code has the core applied math, and the rest is on the simple side of routine for a Web site that interacts with users.
So, really, so far, I don't need to recruit and hire software developers. And if they would need to know about the crucial technical core, then it's some original applied math complete with theorems and proofs with some advanced pure/applied math prerequisites and not AI/ML, or statistics.
For judging if people will be good for my startup, I believe I can do that; if they are highly impressed with AL/ML, then that would be a letter grade down!
For the status, last night I wrote two, simple scripts to help me understand where my disk space is being used; my old scripts got sick on file sizes > 2^32. Then with the output from those two scripts, I'll move around some data and do an incremental backup. Then I'll check about 10 times to be careful and then restore my main boot partition from an old copy made by NTBACKUP. I need to restore because I had a hardware problem that corrupted some of the code on my boot partition. That's the nature of the work at present -- e.g., system administration mud wrestling.
Then back to testing my software on the way to going live, ASAP. Then some more routine steps, and then going live.
It's been a long time: The parts of the work unique to me have all been fast, fun, and easy. But there were independent, exogenous interruptions -- maybe now I've swatted down nearly all of those.
I don't anticipate any equity funding. Since I'm a solo founder, by the time the VCs would write me a check, I'll have so much traction, and, thus, revenue, that I would have no good reason to accept their check -- no, 1/3rd of my business is not for sale yet! Net, for my startup, the LPs want their VCs to wait too long.