When I was in elementary school in the 90s, there were lots of rumors about Nintendo games. Some turned out to be true - Sheik and Zelda were the same character, and, shockingly, Mew could be found under a truck - and a lot were false. There was no sure way, other than good old fashioned bullshit detection, to tell the difference.
It wasn’t only Pokemon. People would get into arguments about basic questions of fact all the time. Which city was the capital of some state (surely Miami or Orlando is the capital of Florida, right?), or which sports team had won nationals some given year. You almost never see these kinds of arguments today, and there are way fewer sincerely believed video game rumors, because now we have the internet. Two minutes into a conversation about some contentious fact, and somebody’s probably looked it up.
There was a window, though, between this action being possible and people actually doing it. Partly because fact checking simply got way easier: the internet got better organized, and dial-up connections became DSL became smartphone data plans. But also partly because it took time for people to adapt. “Just look it up” wasn’t always a primitive action, a thing you could simply do without thinking about it. But now it is. Now, in fact, it has been for many years, and the whole thing is old hat.
With LLMs, we’re in another such window. New primitive actions are opening up, new affordances. Maybe some of them are bad - the ability to eat extremely rich fast food is an affordance, too. But there are things you can do now, that you couldn’t a few years ago. It pays to notice.
What can you do with AI?
With the adoption of the internet in general and Wikipedia in particular, the activity “learn the fact of some simple matter” became a free action. It no longer costed a trip to the library, or a phone call to a person you knew to be an expert. Likewise, modern AI mints several new free actions. Such as:
Know what to do when a recipe isn’t going as expected for some reason, or it’s necessary to make modifications on the fly
Ask about a finicky edge case that an introductory textbook on any subject doesn’t get into, and ask follow up questions to test your understanding
Get a decent idea whether some wacky behavior your computer does is really an issue or not, and how to diagnose it
Get a decent quality tailor made image for any purpose, expressing roughly whatever nonsexual and nonviolent themes you want
Check your work on any subject you’re otherwise teaching yourself
You can do any of these with plain old ChatGPT, or Google Gemini, or Claude, or whatever. You can do any of them at 2am, or 7am, or when you’re too depressed or insecure to reach out to another actual human being. You can do any of them without having to make a bunch of complicated choices, and without navigating websites that may or may not be sketchy. You can do any of them without being advertised to.
It’s strange to me, how little the people I know use AI. It has plenty of limitations, sure. It’s often wrong. It once told me a mathematical proof by contradiction succeeded because, and I quote, 2 ≠ 2.1 But I feel like people fixate on the shortcomings in the same way that high school teachers in 2008 fixated on the fact that Wikipedia articles got vandalized occasionally. They sure do! And yet, Wikipedia is very useful, and your life is poorer if you don’t know that it’s there.
Learning is way easier
I’d like to know how GPTs work. Suppose I had to attempt this without access to AI chatbots. What would I do? There’s no textbook. There are videos, but I’d have to wade through various ones trying to sell me something, and I’d run the constant risk of zoning out or missing a step. There are articles, but they all assume some specific level of familiarity - too high, and it’s gibberish, too low, and the whole thing isn’t fine grained enough to really learn. Without LLMs, the practice of learning some new specific, densely technical subject is really difficult, because you have to spend so much effort figuring out how to learn in the first place.
How have people managed this problem throughout history? How have they learned cutting edge, difficult subjects, which are actively in flux and for which a mere textbook does not suffice? Through conversation and apprenticeship! If you wanted to learn the coolest, most challenging stuff for most of human history, you had to go to where the smartest people were, prove you could hang, and have them teach you. This is still an excellent formula. But it’s not the only formula anymore. Public libraries are one reason. The internet is another. And LLMs, now, are a third.
Like, just try it. If you want to learn about something, start a conversation with a modern AI system. Ask follow up questions. Challenge answers that seem spotty. It feels a lot like talking to a knowledgeable, if fallible, tutor. Who, again, you can talk to whenever you want, without any fear of looking stupid or self conscious, or boring them, or belittling them when they’re wrong.
About 10 years ago, I discovered a blog called Slate Star Codex. It led me to some interesting AI stuff, and in particular the website for an organization called MIRI. I decided to read some of MIRI’s papers, to see if I could make sense of them. But getting through any of them was really challenging - they were dense and high context, and I didn’t have much background in the relevant math. So I gave up.
Today I wouldn’t have given up. I’d have opened another tab, and started asking questions. You can too.
It can also do this: