Here's where I'm at: it's all Markov chains. Markov chains started life in number theory, moved on to stock markets, were then applied to sentence completion and finally integrated into image recognition/code. I think it's a triumph of ingenuity to take a hundred-year-old computational technique and diversify it into useful applications across such a diverse constellation of uses but I think it's a failure of imagination to assert that we'll somehow create God with it.

You know what the perfect application of GPT is?

Speaking as an occasional video/seasoned audio editor, there are plenty of processes that can be automated, be it in photoshop, premiere, pro tools or whatever to turn several questionable takes into one master document. They're tedious, time-consuming and with a little training, I could teach an eight-year-old to do them. What mostly slows you down is the interface because the companies that have traditionally done it - Adobe, Adobe, Adobe, Avid and Adobe - really lean into making things harder than they are. Once you got the interface down you can do this shit in a hurry. Once you've built out a dozen go-to presets you look like a wizard. So why not give that to an AI to speed-run all the presets, LUT them and put it on a chip to make your selfies better? That's classic, perfect Google. That's where they made their trillions.

And the thing is? It's Pareto Principle optimization: really, what it means for the future experts of audio and video is that the tedious, trivial bullshit we all spend most of our time doing goes away. The machine will handle the easy, obvious stuff. The hard stuff? Neither you nor the gadget knows how to deal with it, so you'll hire me. I may make less money. I may get more time for triumphant work. You'll get "good enough" for your selfies and if you've got something you need to get paid for, you'll know that you can't take it any further because the gadget is better than you. The gadget is not better than me. Never will be. The gadget doesn't know how to improvise. It doesn't know how to synthesize. It doesn't know that barking dog is actually desirable and it never will because if it opens up the parameters enough to actually fine-tune the process that much, it's going to require you to have enough casual understanding of the process to turn into an audio editor yourself. Which, frankly, is fine with me, too! That means I get to spend my time exercising expertise, rather than pushing buttons, and the bright and shining digital future is better for all of us! Thanks, Google!

I think there's a lot of utility in the GPT approach to problem solving. Where I leave the accelerationists and the neophytes behind is where I assert that it's a better shovel, not a golden calf. The way the technology works is by finding the missing number in an equation. That's all. It writes the equation, and it writes it based on a massive pile of other equations, but if its model is built around sines and cosines it will never, ever ever throw a log function in there no matter how obvious.

AGI, since the dawn of the golem, has been about synthesis. Problem solving, at a basic level, is very different than plug'n'chug. GPT is a plug'n'chug engine - and it's useful in all sorts of crazy ways. But it will never be useful in the way people want it to be, which is to create digital friends.