Saturday, January 20, 2024

AI is Getting Good at Low-Value Things

Erik Hoel wonders how AI companies are ever going to make much money, given that the things they are good at are things that are already very cheap:

I think that this paradox of impressive intelligence being associated with less impressive money-making ability might be an unavoidable theme for AI, at least for now, simply because the success of contemporary AIs is based almost entirely on the massive size and quality of the data sets used to train them.  . . 

What’s written on the internet is a huge “high quality” training set (at least in that it is all legible and collectable and easy to parse) so AIs are very good at writing the kind of things you read on the internet. But data with a high supply usually means its production is easy or commonplace, which, ceteris paribus, means it’s cheap to sell in turn. The result is a highly-intelligent AI merely adding to an already-massive supply of the stuff it’s trained on. Like, wow, an AI that can write a Reddit comment! Well, there are millions of Reddit comments, which is precisely why we now have AIs good at writing them. Wow, an AI that can generate music! Well, there are millions of songs, which is precisely why we now have AIs good at creating them.

Call it the supply paradox of AI: the easier it is to train an AI to do something, the less economically valuable that thing is. After all, the huge supply of the thing is how the AI got so good in the first place. 

2 comments:

  1. I wonder how many dismissive essays like this could be found from ca. 1980 about personal computers.

    ReplyDelete
  2. @David

    Easy enough to look up - I just did.

    The big thing I'm finding is a lot of people who felt like computers were mystifying and arcane, and resented having them foisted on them by being introduced in workplace settings, because it meant learning an entire new skillset that was largely foreign from what most people knew.

    And honestly? That's precisely on the money - computers DO require a skillset that wasn't commonplace in the 80s; and they were particularly clunky and not "user friendly" back then. But they were also revolutionary in what they allowed people to do, which is why they exploded into commercial sectors - and why most people's first experience with a computer was in a work setting, rather than as a home item. Even when considering personal computers over work computers, many people only first brought a computer into their own home because they needed it for their jobs. And that makes a ton of sense when you consider the cost of them at the time!

    So-called "AI", on the other hand, doesn't really compare. John's right about its lack of ability to perform high-value tasks - if it offered much value, we'd see far more of it being used, as companies leapt at the chance to make money with it. The fact that it isn't sweeping through the commercial sectors the way work computers did is very telling.

    Computers also had the benefit of their value being intrinsic to their own existence. A computer is a piece of hardware, which operates predictably and understandably; if it isn't powerful enough, you upgrade its components; and if the person using it isn't making good use of it, you train them.

    But so-called "AI" is a nebulous digital mess that even the people who build them don't understand how they work. Outputs aren't reliable - inherently so. Modifications to the program to produce different kinds of outputs are difficult and chaotic at best, impossible or unfeasible at worst. Flaws are hard to diagnose and fix. It's not unheard of for entire programs to be entirely scrapped, and just replaced with a new program run through a new set of "training".

    "Training" is another major flaw in the comparison - the usability of a PC wasn't and isn't dependent on access to extremely vast databases of data (legally or illegally sourced). If you need a PC to do something that there is no extant program or support for, a single person can code a solution in relatively short order, at very minimal costs. But if you need a so-called "AI" to do something that can't already be done, you need to create an entirely new program, and collect (or purchase) an entirely new set of data to "train" it with. If no such dataset exists to acquire, well... tough luck.

    We're talking about two very different kinds of things, and two very different kinds of pushback.

    ReplyDelete