Is Automation of Automation Going to Kill Off Computer Science Jobs?

An Essay in Polite Dissent to Mark Cuban

Tagged: Artificial Intelligence Math

My friend, colleague, and homonymous intellectual doppelgänger Evan Teran always likes to remind me of Betteridge’s law of headlines:

The answer to virtually every clickbait title that asks a yes/no question is: “No.”

Mark Cuban, in an interview on Bloomberg TV, claimed that the next wave of automation will be “the automation of automation.”

By that, Cuban means software will soon begin writing itself, which will ultimately eliminate those lucrative software development jobs. About writing software, Cuban said: “It’s just math, right?” Humans will no longer be needed.

Hold my beer, I’ma explain some math.

Remember The Imitation Game? The Academy Award winning movie from a few years back, starring Buttercup Cummerbund and British Natalie Portman? It was loosely based on Alan Turing’s groundbreaking work cracking the Nazis’ cryptographic codes, helping the Allies win the war. That very same dude is much more famous among mathematicians and computer scientists for his work on the Halting Problem.

The Halting Problem dates back to the seventeenth century, from none other than Gottfried Leibniz—the same dude who independently discovered Calculus, even though Isaac Newton usually gets all the credit. Leibniz dreamt of building a machine that could automatically check whether mathematical formulas were correct or not. If you’ve gathered by this point that mathematical formulas are effectively the same as computer programs, then congratulations! Take a sip of my beer as a reward.

If you’re familiar with the Halting Problem, you can probably already see where I’m going with this argument. I’ve been asked why I chose to use the Halting Problem as a basis, as opposed to going the route of simply—and correctly—arguing that creating software is an immensely complex engineering problem whose configuration space is exponential. If the reader is technical like you, then they undoubtedly already know this. If, however, the reader is not familiar with any of these concepts, then I have found that it is very difficult to overcome the public’s incorrect perception that computers will continue getting faster and faster and, therefore, will eventually be able to do anything. As if hardware speed alone is what provides “intelligence.” It’s very difficult to convince people otherwise, since it basically requires teaching P vs. NP. Therefore, it is a much simpler rhetorical task to demonstrate that there are clear, fundamental limits to what computers can even compute, meaning that computers provably can’t make a perfect mapping from imprecise specifications into a sufficiently correctly functioning program. There needs to be some “intelligence” in the equation: either a programmer or strong AI.

Fast forward to the early 20th century, and a sort of revolution was going on in the world of mathematics: New systems were being developed to formally encode logic. Think of it like the system of algebra you learned in grade school, except that instead of numbers you’re working with logical statements. David Hilbert was at the forefront of this research, and in 1928 posed the question: Is it possible to devise an algorithm (vi&, a computer program) that can automatically determine whether a given logical statement is universally valid? This became known with the extraordinarily Deutschtastic name „Entscheidungsproblem“. A few years later, Austrian expat Kurt Gödel—who is also famous for discovering a “bug” in the US constitution at an immigration court appearance with Albert Einstein—proved that the answer to the entscheidungsproblem is: No, it is provably impossible. In his seminal paper titled „Über formal unentscheidbare Sätze der Principia Mathematica und verwandter Systeme I“, he shattered the entscheidungsproblem both in theory and in Deutschtasticness.

So, back to Turing. A few years later, but before the events that inspired The Imitation Game, Turing created a mathematical definition of a computer which we now call a Turing Machine, and on which practically all modern computers and computer science is based. Turing realized that the logical statements and algorithms Hilbert and Gödel were working with were in fact no different than the programs that could be computed by his Turing Machine. He then restated the entscheidungsproblem in this context: Is it possible to write a program for a Turing Machine that can take another program and that other program’s input and automatically determine whether that other program will terminate? This became known as the Halting Problem and, just like the entscheidungsproblem, Turing was able to prove that the answer is “No”: It is impossible to create a program that can automatically determine whether another program will terminate.

Please indulge me as I briefly delve into metaphysics and, dare I say, even digital physics. Can our universe be described using logic and/or computation? In other words, if there were some infinitely powerful Turing machine, could one write a program to run on it such that it could accurately simulate the entire universe? If so, then we humans are bound by the same theorems that prove mathematical logic is incomplete and that the Halting Problem is undecidable. In other words, it’d be impossible for even humans to prove that a given program is correct.

The Halting Problem is part of what makes programming computers hard: Any program with even a modicum of real-world complexity cannot have provably correct behavior. Those logical flaws and unintended behaviors are precisely what hackers exploit to make programs do things that the programmer never intended. Human programmers are only able to do a passable job of quality control because we have the benefit of heuristics and instincts that are guided by our general intelligence.

Let’s say we develop a computer program that can take a human’s high-level description of a task and automatically generate a computer program that can complete that task. There are two possibilities:

  1. the human is specifying all of the complex logic, and the computer is simply translating that specification into a program by rote; or
  2. the human is not specifying any complex logic, and the computer needs to determine the logic that needs to be in the program to accomplish the task.
The first possibility is effectively describing a human computer programmer and compiler. The second possibility is what Cuban seems to be describing. The trouble is that, by virtue of the Halting problem, no such program could ever determine whether its output is correct! Furthermore, automatically determining the logic that needs to be in a program to accomplish a given task is provably computationally hard, and cannot be solved in general in a reasonable amount of time, regardless of how fast computers get.

Could I write a program that accepts inputs like, “Create a program that displays pictures of cats,” and then automatically generates a program to do so? Sure. Could I create a program that automatically creates a program that is able to interface with the highly formally specified interface of another automatically generated program? Probably. But the vast majority of software projects are for solving complex human problems, and typically involve integration with numerous legacy programs and software that were written by crazies, idiots, and crazy idiots. Professional programming sucks. Don’t believe me? If you take away nothing else from this essay, please read this article. I’m serious, read it. Any professional-grade program that can automatically generate other programs will have to grok how to interface with the undocumented source code of other humans. That’s impossible unless the computer is intellectually indistinguishable from a human.

In order for a computer to automatically generate a program that solves a human’s problem, there either needs to be a human deciding the logic (i.e., a programmer), or the computer’s intelligence needs to be indistinguishable from that of a human. And if we have computers that are intellectually indistinguishable from humans, we’ll have more issues to deal with than simply losing software jobs.

← Older Post Blog Archive Newer Post →