Hi OP,

While a lot of the AI hype nowadays is justified, a lot of it isn't, as well.

Most people don't understand how AI works, which precludes their ability to understand its limitations. Let me explain:

While there are a wide array of AI algorithms out there currently, just about all of them rely upon a combination of a few things:

1. A "Cost parameter", or parameters;

2. A set of "tuning variables";

3. An arbitrary structure (I know this sounds vague, but that's because it is) that determines how numbers are calculated within the model. This can be an optimization problem based on Euclidean distance (the case in support vector machines), or it can be the result of backwards propagation (in the case of neural networks), or any number of other things; and

4. A LOT of data.

Without boring you and the readership of WSO with details, these factors are combined into a process whereby a computer essentially throws darts at a metaphorical dart board, first completely at random, and then progressively with more accuracy as it measures the distance between the impact point of each successive throw and the bull's eye - this is a process of "guess and check", which requires (1) a framework for how to make the guess (the prerequisites of which I listed above), (2) a variable for evaluating the relative accuracy of guesses (this is the "cost" parameter), and (3) an answer key, which allows the machine to know the "correct" answers and calculate the distance between these and its guess, using the cost parameter.

The details of how and why the machine makes these adjustments between throws are varied, and that is more detail than is necessary to make my point, but I'm happy to go into it if you want - suffice to say that each model contains a methodology which is more or less strictly technical and binary - the machine basically just takes random stabs using a set of "test data" for which it has "correct answers" and hones its rules on the fly if those rules produce a smaller mean squared error than its previous rules. I have more experience with Support Vector Machines and Neural Networks than anything, but the principles apply broadly.

The reason AI works is because when you have a LOT of data points with "correct answers" that cover a wide range of outcomes within a few highly descriptive variables (these variables are called "Features" in AI parlance), the machine is able to do a good job taking the rules it established on the "test data" and using them to categorize new "observation data", by dint of the fact that it has essentially established a complex, nuanced framework for comparing this data to data it has already "learned" (i.e., gotten low-error guesses on).

This makes AI great at approaching tasks which inherently follow a similar pattern always. Credit scoring, screwing pieces of a vehicle together, reading news headlines, etc. are all good examples - sure, there's a little nuance between each iteration of these tasks, but nothing that is outside the previously-known universe of observations. You don't come across a loan application with a different set of metrics than any seen before. You don't walk on an assembly line and see pieces arriving at a station in any form other than "ready for further assembly" or "dud".

The pitfall of any AI is what is known as the "bias-variance tradeoff". This means that when you make a machine better at predicting relationships between variables in the test data (the data that the machine practices "throwing darts" with), you make it more susceptible to making ever-larger errors when looking at a novel data point, which is referred to as "overfitting". Conversely, when you make a machine better at throwing darts it has never seen before, you significantly reduce the model's ability to recognize and capitalize on relationships present in the test data.

For purposes of the metaphor, (albeit with a little hyperbole for clarity's sake) the bias variance trade off means you can have one of two things with AI:

1. A machine that can throw 1000 identical darts perfectly, but will utterly fail when we hand it a weighted dart with a different tail fin; or

2. A machine that can throw 1000 different darts and hit the board somewhere (hopefully, maybe not), but not reliably land on the bull's eye.

What does this mean? I think it means AI are surely going to replace people in tasks where there are no new things on the horizon - an assembly line, as mentioned before, isn't a process that fluctuates outside certain rigidly defined parameters, because this is how it is designed to work. If parts are arriving at the next station on the line outside of a certain narrowly defined range of states, it's easy to discard them as duds and you have a smooth (albeit maybe wasteful) system - but no major harms.

Does an M&A transaction follow the same pattern? Not really. Sure, lots of deals are similar to one another, but there is no way to contain the potential for a novel outcome like when you design an assembly line. There's just no way to avoid, for instance, one party to the transaction deciding they hate the other and scuttling the deal. Further, the amount of data that would be requisite for predicting such a complex process would become prohibitive because the calculations involved in these models typically center around dot products (a concept from linear algebra), meaning that the time it takes to make calculations rises exponentially with the number of points in the data set being considered.

I can elaborate more if others want, or provide resources for those seeking to understand these processes in more detail.

Is this a good thing? Depends who you ask, and how. If you are speaking economically, and ask that segment of the population that would rather spend their lives performing easy, repetitive tasks in exchange for a livelihood, the answer is probably "no". These people are akin to the ferrymen that were made obsolete when steam power hit the riverboats of industrial England. The arrival of AI will presage their being pushed into more desperate conditions, and widen the gap between their abilities and those abilities that are rewarded within a capitalist framework.

If, on the other hand, you ask the CEO, his answer is probably "yes", as he is now able to replace labor with more efficient capital in his factors of production mix, and attain higher earnings.

Aside from this, there are massive moral and ethical dilemmas inherent to AI that will become apparent in our lifetime. For instance, if you teach cars to drive themselves, you must also teach them the "right way" to resolve ethical quandaries for which there is no answer - for example, if two cars are on a crash course, one full of convicts, and one full of mothers, and one must swerve and kill its passengers and save the others, how do you decide? MIT and other AI labs are soliciting public opinion on these questions, because there is no correct answer, despite what anyone might say.

I do wonder whether AI will presage the beginning of what Marx was discussing in Capital and the Manifesto. The prerequisite for his world was one of infinite abundance. And are we so far off? In a world where capital and consumer goods, and perhaps professional services, can be produced, distributed, and maintained indefinitely at the push of a button and the whir of a complex AI, it will call into question much of what makes our system stable and worth living in. Will everyone subscribe to the same economic truths we now hold dear when, someday, a fleet of Ferrari's can be made essentially at no capital or personal cost to anyone? Will the goods we center our lives around hold the same value when they are divorced from the blood, sweat, and tears that we currently associate with their acquisition? I'm not sure, but it's food for thought.

People tend to forget how much each of us centers our daily striving from cradle to grave around the economy, and our place within it. People derive a lot of momentum and personal meaning from just getting up and doing things they have to do. But what will the narrative become when the system doesn't really need most of us anymore in order to function?

It gets to the core of what, in my opinion, differentiates various economic/moral/political philosophies, which is: if people were not bound by material constraint and necessity, what would they do with themselves? What would they do for (or to) others?

Edited for clarity.