It is actually pretty simple. A bit of maths. A bit of IT. And a bit of philosophical discussions about what internal and external validity of a theory means.

Basically. You can look up the AI field on Wikipedia fairly easily and it already sums up the key points. There is the standard advances statistics stuff that covers statistical theory and asymptotics for whatever problem class you can drop from uni- to multi-variate, from linear to non-linear from static to dynamic to deterministic to stochastic and so forth. Pretty basic stuff and if you get it a bit deeper, with all that linear and non-linear algebra, algebraic topology and all that Jazz that describes different classes of – hopefully metric – spaces. Good. You switch over to the discretized version of it all and look at combinatorics, discrete mathematics and its foundations in numbers theory, operator theory. And most likely everything that has to do with graphs, networks and stochastic dynamics on graphs is in your domain. It’s basically just that. That gets you as far as 0 without IT.

So you move over to the IT domain and you start with basic courses in datastructures, you get a bit deeper into programming language paradigms into the direction of imperative to objective programming to cope with more complex systems. At that time, your brain bifurcates into system complexity handling so you move towards other paradigms such as functional and logic programming – by which time you understood Goedels incompleteness theorem – and you find nice parallels between your discrete mathematics background, turing equivalence, P=NP problems and approximization algorithms. Good. Coming from functional and logic programming, you recap your general logic stuff and appreciate number theory and algebra a bit more and diverge into category theory. But wait, too far. Back to the second branch your brain bifurcates into, you get back to algorithm design and are back in graph theory and basic combinatorics and go even deeper into discrete mathematics.

Once you reach sufficiency in these areas, you are already laughing about the multivariate statistics guys and deep and convoluted neural networks are conceptually a piece of cake. You still wonder how all that works and you go back to Bayesian statistics 101, computational statistics, MCMC methods and all that Jazz and take a few reandom walks into Bayesian random networks. Somewhere around there you get lost in Erdos and Co graphs and random network dynamics and decide it’s too weird. By this time you still don’t get A.I., but you sure as hell mastered all that machine learning stuff people make money with by now.

Back to square one. You got pretty decent in coding and are coding swarms and autonomeous agent models, map and aggregate agent behaviours and understand that the economics branch here is quite nice. Maybe you bifurcate into economics of law and industry economics and study a bit of game theory on the side line and bifurcate yet again into behavioural theory and neuroeconomics. But you decide to forget about it and go back to more standard stuff until you crossed off all the remaining items on the Wikipedia page.

Now you conceptually got the AI domain in your brain. You have a lot mathematics reading to do to get up to speed on what all those things are founded on. More algebra, more topology, more discrete mathematics. But it gets boring. by the time you are building commercially viable A.I.s and you wasted your brain power to write APIs and scrapers to feed your A.I.s or apply all that stuff to real world problems – starting to struggle with that external vs internal validity problem and this post hoc ergo propter hoc mindfuck – and you start realizing you have to go all the way back to electric circuit design, electronic components, computer architecture and diverge into sensor and actuator technology to understand how your A.I. could also use the real world as input and output interface. You still haven’t solved any new problem and you are learning things that a lot of other people have learned, but you learn to get dumber and dumber and less rigorous and do follow a breath over depth approach. Searching for applications, concepts and all that to use in your concept-powered brain, all while learning to move a very conceptually cool A.I. engine to a real world application, moving it to a scalable platform like AWS and crunching and optimizing algorithms to keep your costs down. Because even if you are smart enough to learn all about A.I., there is no way you can find a business model that pays for the computational and energy consumption that you would need to make it work, let alone the ability to get enough data to make a hard AI solution actually work.

But yes, that’s basically A.I. Not taking a coursera course or reading a book about tensor flow – oh, I forgot tensor calculus … – and pushing a million Google Image search into Watson. But hey, if that is something you get funded with! Go for it.

Hope it was an enjoyable read. Don’t believe the hype. Ask the questions about the background of the A.I. guys when someone sells you A.I. And take a week listening to A.I. people on the conceptual domains they had to master to get good at it. Then you might be able to call the A.I. bluff.

And to cut it short: An A.I. guy isn’t someone that took an A.I. class in undergrad CS. It also isn’t a pure mathematician, because that guy writes proofs in mathematical subbranches you cannot even spell correctly. It’s not a physicist, because the Relativists are too curved, the string theorists to categorical and global in their analysis, the quantum guys too stressed about cats. It’s also not an econ guy, because the standard econ guy stops in 1940s mathematics proficiency. And it’s not a philosopher, because there are higher order logics that just don’t make any logical sense in the A.I. field. They are not statisticians, because statistics at times is a bit indiscrete and you cannot justify the sigma without accepting the sigma algebra. It’s more like a solid mix of mathematical rigour and ability to read mathematics, the ability to ignore rigour like a physicist and just apply the work of others, mixing and blending this with a fairly good skillset between mechanical, electrical and computer engineering. And being human and out-the-door enough to apply this stuff to real problems.