In yesterday’s post we told you that many companies that were touting AI were not actually selling Artificial Intelligence or even anything remotely similar (including, but not limited to, Autonomous Intelligence, Augmented Intelligence, Assisted Intelligence, and/or Amplified Intuition) and were, in fact, using the buzz-acronym to accomplish applied indirection and sell you 90s tech in a shiny new wrapper, proffering yesterday’s miracle cure for all of your current woes.

The only difference between the 90s solutions and todays is that today’s look nicer, run faster (but that’s mainly due to the exponential increase in computing power), and have more automation built in. But RPA — Robotic Process Automation — is NOT AI. It’s just using a rules engine and workflow to automate common tasks under typical conditions.

So how do you tell the difference between Applied Indirection and real (WEAK) AI? Well, first you think about what AI means, apply a little common sense, and ask some good questions.

Let’s start with thinking about what AI really means. AI typically stands for Artificial Intelligence, and the definition of AI in its strongest form is machine intelligence, where the machine can acquire knowledge, learn, apply it, and adapt to new, previously un-encountered real world situations in a general manner just as a human would do. If you think about it, no machine can do that, and no machine is even close. So there’s really no such thing as AI (and won’t be for decades).

At the same level of complexity is Autonomous Intelligence, which is Artificial Intelligence that is capable of acting on its own without any human interaction. Since true Artificial Intelligence doesn’t exist, it should be obvious that Autonomous Intelligence (outside of living beings on our planet), which is an AI agent that can work in complete isolation from human interaction, doesn’t exist either.

At the next level down we have Augmented Intelligence, where we don’t define a platform as being intelligent but as capable of providing knowledge and insight that we can use to complement and enhance our intelligence and make as faster and better at the tasks we are performing. At this level, there are tools that exist for well defined tasks, but they are few and far between. While there are a lot of systems that can allow us to do our jobs faster and better, they don’t augment our intelligence. For a system to truly be an augmented intelligence system, it must augment our intelligence, propose actions that we were not aware of (and would not think of in a little bit of time), and make us smarter over time. Very few systems do that, even when limited to very specific tasks.

Going down a level, we have Assisted Intelligence, where we don’t define a platform as intelligent, but capable of using knowledge and insight that it has to complement and enhance our daily performance of tasks by helping us to do them faster, better, or both. Like augmented intelligence platforms, they should be able to prescriptively suggest actions or workflows, but we don’t require that they be capable of identifying anything we wouldn’t in our jobs.

The big difference between augmented and assisted is that a platform that analyzes market data and dynamics and comes up with one of a pre-set of sourcing strategies as a recommendation is generally just assisted intelligence. In comparison, a platform that not only pulls in market feeds but scours the web for public pricing, articles on supply / demand (im)balance, third party audits, and reports on recent events and other data not pushed through integrated feeds; creates multiple pricing and availability projections; runs those projections through multiple models; and then recommends you extend the current agreement and buffer stock three months of supply (because an earthquake in China just closed down the mines that supply a significant amount of the rare earth metals used in your product and supply is likely to become scarce and pricing rise in six weeks) would be considered to be an augmented intelligence platform because even though you could do web searches to find updated public pricing, supply projections, third party audits, and natural disaster reports, there’s no guarantee you’re going to find the report on the local Chinese news feed (that won’t get picked up by an English news feed for two weeks because China downplayed the effect of the earthquake) when you only read English.

In other words, there are some assisted intelligent tools out there (that help you do your job better and faster, but aren’t going to do anything you can’t or come up with anything you wouldn’t if you just spent five minutes thinking about it), a few augmented intelligence platforms for specific tasks, but no autonomously intelligent, artificially intelligent, and definitely no cognitive platforms on the market — and if someone is trying to sell you that, they are using the marketing technique of applied indirection to sell you modern silicon snake oil.

You have been warned!

You read that right. AI at most companies is not Artificial Intelligence. It’s not Autonomous Intelligence, Augmented Intelligence, Assisted Intelligence, or even Amplified Intuition. In reality, it is marketers taking Green Day’s AI a little to literally (and treating everyone like an American Idiot^*) and repackaging old tech with a new label.

You see, most of what the Marketing Mad Men are trying to sell as AI are just old-school statistical algorithms in a brand-new wrapper. And the only reason these technologies are finally hitting the market and getting good results is the sheer amount of processing power and data we have at our disposal — because dumb algorithms (which is what they are) only work well when you have a lot of processing power, a lot more data, and the power plant to run that processing power 24/7 at 99% capacity across dozens, if not hundreds, of trial parameterizations until you find something that, well, just works.

But it’s not intelligence. It’s advanced curve fitting, regression, k-means clustering, support vector machines, and other statistical inference techniques that existed in SAS in the 1990s. Except now, the curve fitting is n^th degree polynomial, advanced trigonometric, geometric, n-dimensional, step-wise, and adaptive. The regression is nonlinear, non-parametric, stepwise, and much more robust … and accurate because you can process millions of data points if you have them. The k-means is not clustering around one or two dimensions, but one or two dozen if necessary in a large multi-dimensional space — and the clusters can be of arbitrary n-dimensional geometric shapes using kernal machines. The support vector machines are not just based on primal, dual, and kernal classification with a bit of gradient descent but enhanced with multi-class support vectors, advanced regression, and transduction (to work with partial valued data). And so on.

And don’t think there’s anything new about “deep neural networks” either. They are just multi-level neural networks which were common-place in the 1990s with more levels and more nodes per level with more advanced statistical classification functions in each node trying to figure out how to extract patterns from unclassified data to classify and structure it, which happen to get better results because they can work on millions of data points, instead of thousands, and do tens of millions of calculations and re-calculations instead of tens of thousands. And that’s the only reason they get better results “out of the box”. There is absolutely nothing better or more advanced about the core technology. Nothing. It’s still as dumb as a door-knob, no matter how whizz-bang the markets make it out to be.

And at the end of the day, the “active” part of the neural network is a fraction of the overall network (which means as much as 90% of the computation is wasted), and if that can be identified and abstracted, you typically end up with a small neural network no bigger than the ones being used twenty years ago, which, even if more than three or four layers, can probably be redesigned as a three-or-four layer network. (See the recent article on the recent MIT Research, for example.) [But if you’ve studied advanced mathematical systems, this is not an unexpected results. Over-dumbification has always led to unnecessary processing and inferior results. Of course, over-smartification also leads to ineffective algorithms because data, typically produced by humans, is not perfect either and we need to account for this as well and detect small perturbations and deal with them. But it’s always better to be thoughtful in our design than to just brute force it.

In other words, many modern marketing madmen in enterprise software have become the new snake-oil salesmen, often selling simple statistical packages for a million dollars or raising tens of millions for yesterday’s tech in a shiny new wrapper. But it’s not intelligent, or even intuitive, by any stretch of the imagination.

That’s not to say that there isn’t technology that can qualify as assisted technology (and maybe even augmented in special cases), just that the majority of what’s being pushed your way isn’t.

So how do you know if you are among the majority being subjected to Applied Indirection or one of the few minority being offered a solution with true Assisted Intelligence capabilities? Stay tuned as we discuss this topic more in depth in the weeks to come …

^* It’s much preferable to be a Canadian Idiot. We’re nicer and the “AI” marketers don’t bother us as much.