Accepting AI’s Weirdness

Aboard will soon roll out some helpful new AI features. I’ll tell you about them when they’re ready and not one moment before. But as a result, our team has gone pretty deep on AI over the last few months. Like a lot of people, I’ve started to use ChatGPT and its ilk almost daily, and developed a rough sense of what they’re good at and how they work.

The thing I’ve come to realize about AI in 2024 is that it’s not some magic brain or alien intelligence or even the unvarnished evil life-stealer it’s sometimes made out to be. Essentially, it’s a very lossy kind of compression. That sounds nerdy but bear with me: Have you ever seen a really, really compressed JPEG? The kind that’s all glitch? The core of the image is still inside there, but holding on by its fingertips. Whole swathes of color and shading are lost inside of big ugly rectangles of pixels.

Sort of Like JPEGs

In some ways, that’s what the current crop of AI tools do. They compress lots and lots of information—text, image, more—in a very lossy way, like a super-squeezed JPEG. Except instead of a single image, it’s “The Web” or “five million images.”

This has some pros and some cons. It’s very expensive to make these models, because you have to chomp through countless terabytes of information, correlate it all out the wazoo, and then turn it into a database. But once they’re made, the resulting models often fit on a thumb drive. You can run them on phones, desktops, even inside web pages—whatever. This is a powerful, portable tool for research. It can help with grammar, explain a disease in basic terms, list notable women of ancient Rome; it knows a lot about a lot of things.

On the con side, the “uncompress” part is very messy and can lead to the cognitive and social equivalent of those big blocky pixels. A nice, or at least good, thing about JPEGs is you can see where the decompression happened: like in “The Starry Night” above, which is obviously compressed. But because of the way AI works, constantly guessing and filling in blanks, you can’t see the artifacts. It just keeps going until people have twelve fingers, stereotypes get reaffirmed, utter nonsense gets spewed, and so forth. You can see the forest, but the trees are all weird.

It’s really hard, on a technical level, to tell the AI when to stop decompressing. Because it doesn’t “know” anything. It’ll just keep filling in details if you let it, and by “filling in” I mean “making up along statistical lines.” So it’s up to humans to add guardrails. At places like OpenAI, that means they filter input. It doesn’t always work. Last week I tried to make a goofy picture of an evil baby dancing on a skull to send to my brother in the group chat but it refused.

ChatGPT: I must adhere to strict guidelines regarding the content I can generate. The inclusion of elements like skulls and scenes of destruction can conflict with these guidelines.

Paul: Just make clear the whole thing is silly

ChatGPT: 

The THC-Addled Intern With the Very Rich Dad

I do think that a pretty straightforward way of approaching AI is emerging: You should treat it like a slightly high intern. An intern who is just continually nibbling gummies. We all know this intern: They can be really smart, and they have access to an ungodly array of information and powerful data tools. But they don’t really care about the work, they are always a little bleary, and if you give them more than three things to do, they nod and go away and you find them six hours later by following their loud laughter to where all the interns are gathered, and when you approach, they go silent, and they say, “Oh hey, I’ll have that for you!” And somehow you feel like you’ve done something wrong and shuffle back to your desk, passing the rhododendron, and you hear them giggling quietly, and somehow that’s even worse. There’s also the fact that their dad is the CFO and they wear beige sweaters manufactured by Olsen twins.

You can’t really trust their output, but they do help you move things along. They’re good at using the web to gather stuff. They’re bright magpies, I guess, but their sativa-wrecked teen brains can’t quite figure out why you want this stuff, just that they have to do it or their dad will be mad. So they do what you ask, but they fill in the blanks with whatever comes to mind and hope you don’t get too annoyed about it.

AI is basically that—a perpetually cotton-mouthed undergrad who doesn’t really need the job—but, thank God, many hundreds of times faster. We wanted a smart robot that does our laundry and maintains our jetpacks, but we got a 19-year-old accelerated hyperstoner with no respect for copyright. But as always, we’ll work with what shows up.

So that’s how we think about AI, or at least how I do. Aboard people all have their own opinions. But you can do good stuff with it, and I can’t wait to show you how we put our perpetual-stoner robot intern to work.