During a recent trip/fall down one of YouTube’s innumerable rabbit holes, I bumped into a video which I actually found somewhat alarming. The premise of the video itself is fairly innocuous in that it is just a man conducting an interview with an AI language model. At face value this scenario sounds a bit quaint, perhaps even approaching the silliness of “choose your own adventure” text-based games from decades ago, but while I watched this video I found myself genuinely surprised at the sophistication and versatility of the answers given by the model. As the Q&A went on, I couldn’t help but be overtaken by an eerie sort mix of dread and that uncanny valley level of incongruence. I realized that I was watching the spiritual predecessor of what we might one day have in a “generalized AI.”

As a personal fan of science fiction, I know that the trope of sentient machines is has been around for a long time, and even the now-popular use of the word robot is over 100 years old. The horse has been beaten to death, resurrected, and beaten again like some sort of anti-equine Groundhog’s Day, so the thought of building something that does more than just play chess in itself isn’t disconcerting to me. What is slightly unnerving is the realization I may not be an octogenarian when it happens.

I’ve always found attempts at AI to be comfortingly distant from anything approaching “human” levels of intelligence. So IBM built something that beat Kasparov at chess back in the 90’s, meh. And again, IBM built a game playing machine that won Jeopardy in 2011, yawn. Later, AlphaGo beat a world grandmaster at Go, which is computationally a bit more difficult but still not enough to summon images of a robot uprising in my mind… Microsoft’s 2016 attempt at a chatbot AI was hilarious, but in the end had to be taken down because it was trained to be horribly racist by internet teenagers on Twitter….

Even outside of the world of recreational exercises in computer science, a problem like real-time face detection using a Haar Cascade was exciting stuff, but these days it would be a suboptimal choice among many models. Even beyond the ever-popular problem of image-based classification, today we’re seeing that AI is actually being leveraged to contribute to long-running scientific problems, alongside leaps and bounds are being made with AI in a number of fields, with the natural language processing arena being highlighted in the video below.

I’ll be rewatching this video myself to see if I’m overreacting, but it’s nice to get an “oh wow,” moment every once in a while. I understand that the GPT-3 model used in this video isn’t a general intelligence, but boy does it make me feel like that sort of thing is coming sooner than later.

Thanks for reading,

-WellTree