Why AI research is about to hit a brick wall

The futurist Eliezer Yudkowsky says that AI is gonna kill us. According to him, humanity has maybe 3 or 4 years left, if we're lucky.

That's one end of the AI opinion spectrum. On the other end are 2 of my tech startupy friends who are deep in the weeds in AI. They've worked their butts off to build tech products using AI, then realized that AI isn't actually good enough to do what they wanted it to do. They both say it's a dead end.

I'm closer to my 2 tech startup friends. I've started using AI for some stuff, including helping me with sales outreach and studying Turkish. But I'm not ready to let AI run my whole business.

For example, when I write emails for my clients, AI can give me a decent first draft if I prompt it correctly. But then I have to do a whole lot of editing. It saves me time, but I think the end product is more or less the same.

(People can tell when an email is just hot off the oven from ChatGPT... and it hurts your brand when you send an email full of em dashes and clichés. At the very least, you have to tweak stuff.)

AI, in its present state, is going to change the world — but not as much as everybody thinks.

I also don't think AI is going to get much better in the next few years. Here's why.

There's a difference between incremental improvement and revolutionary breakthrough. Incremental improvement is when you make something slightly better than what already exists. Revolutionary breakthrough is when you make a completely new thing.

Think about the difference between the invention of the Toyota Prius and the invention of the automobile. Yes, Priuses get good gas mileage, but they don't change that much. Whereas when the car was invented, it changed the whole way that cities are designed.

Revolutionary breakthroughs are pretty rare. There's been a revolutionary breakthrough in AI research about once every decade. ChatGPT was the most recent one.

The people at Google and OpenAI will continue to incrementally improve their AI models. Maybe the AI will get slightly better at the stuff it's already doing — like writing lyrics for fake Bon Jovi songs or searching the internet for the best chocolate chip cookie recipes. 

But Google and OpenAI aren't really working on making revolutionary changes. (Big companies usually aren't good at revolutionary change.) So AI will get a little bit better every year, but it won't get a lot better for another decade or so.

By the time our grandkids are in college, sure, AI will have completely changed the world. But for now, I suspect it isn't going to change much!