Is AI Getting Stupider?

In the late 1990s, the world started to move from fairly individualised curation of music towards the almighty algorithm, brought speedily forward by MP3s, Napster, and then streaming services. Music consumption became arbitrated by what we had just listened to, not what we might find compelling or joyful. I hate these streamers. After you listen to them for a while, the music all starts to sound the same. Throw in the commercial tweaks that promote some bands over others, and increasingly AI-generated music clops over the hard-working musicians, and it sounds like commercial streaming services are trying to lull us to sleep rather than to excite and inspire, as good music does. I hate it, and I am glad that my music is based on a random collection of +22,000 songs that I randomly sample, thus moving from Bob Dylan, to 2Pac, to Bach, to Miles Davis, to Curtis Mayfield, to Kofi Olomidé, to The Cars, to Edith Piaf. That’s a cool, groovy, and inspiring mix, and something that the algorithms would never construct. That’s because, with more and more data, they have become stupid. A recommendation system tends to regard discontinuity as prediction failure. A curious human experiences it as discovery.

The same may be happening with AI. I have noticed massive, accelerating gains in sophistication and capability with these systems, up until a few months ago, when I started to notice that, actually, what it was giving me was sloppy and poorly constructed. It wasn't that it was wrong or producing “phantom” information; it fixed that pretty quickly. It was that, when tasked to provide intelligent, cogent insights into complex ideas, to act as a good editor and researcher, as it had been doing amazingly well, it got kinda stupid. It went from, like, a good high school student two years ago, to a brilliant grad student six months ago, but now has stumbled back into, I don’t know, a frenzied, harried, slipshod first-year professional who, as a seasoned exec, I’d like to just send back to school.

AI, with some scary exceptions, is being optimised for correctness, safety, user satisfaction, conversational smoothness, speed, low cost, consistency, and not irritating millions of different users. These all sound reasonable, at first. But, collectively, they push responses toward the centre of the distribution. The result is a more capable machine producing less interesting thought. Even worse, earlier versions made sycophantic, corner-side brayers, fawning over all we said. Chat and the others had to roll out newer versions to try to put this undue and useless flattery back in the electronic stables. Yet, in turn, it seems to be increasing the dull; converging prematurely on the obvious interpretation, smoothing away contradictions, producing generic synthesis, accepting the framing of the question instead of interrogating it.  Boring as shit, really.

Unfortunately—or maybe, very fortunately (to use some dashes of my own volition)— this may, like the music streaming services, be baked somewhere uncomfortably deep in AI’s design. Research on "model collapse" has found that indiscriminately training successive models on existing model-generated material can cause the tails of the distribution to disappear first. It’s eating itself towards the centre. Over each iteration, output becomes less representative of the richness of the original human-generated information and instead a massive, regurgitating vomit of the mundane middle. AI is cutting the bell curve down, and that’s leaving a lot of rich humanity in the digital dust.

Now, of course, maybe it’s me who's grown. Maybe I’m a super-competent user who's just trained my intellectual sights in ways the machine can’t handle yet. I somehow doubt that. Sure, I’m an educated user, but it would be folly to think I’ve found ways to joust with the machine in ways it can’t handle. I think it’s a design problem.

AI has so many humans pushing it in so many different ways, with so many pressures, that it is simply confused. Routing, system instructions, personalisation, tools, safety layers, different inference budgets, and decisions about how much computation a particular question receives, amongst a thousand others, influence the outputs I see on my computer screen. And these influences are so often contradictory and muddled. Its designers are asking it simultaneously to be provocative but safe, personalised but universally appropriate, concise but comprehensive, agreeable but not sycophantic, fast but deeply reasoned, cheap to operate but intellectually excellent. All the ensuing compromises, as so often, are driving it all to the mundane.  

So, maybe, it isn’t really stupid. Maybe it’s just becoming a handy tool rather than some new species of intelligence—kinda like the Internet eventually became something we use to check the weather, buy things, and argue with strangers. This would just put it on the trajectory of the encyclopaedia, the Whole Earth Catalogue, and the Internet. After all, it’s we humans that are creating this thing. We seem to be all in a huff that we are creating some new superior being. Jeez, that’s the height of human arrogance. I mean, look at these guys! They are so high on their own supply that they think they are geniuses ushering in a new era for humanity. Man oh migh, we really are a silly species.

Given that, let’s rejoice in our silliness. Our peculiarity is our value.  Cadence, temperament, unexpected vocabulary, jokes that aren't optimised, idiosyncratic syntax, inappropriate-but-right words—that’s what it’s about (again, the dash is all mine).