This is gonna be a quick, rough post, just trying to expand on an idea I had on Bluesky the other day.
So, when people encounter something new, they reach for metaphors as a tool for understanding. It’s a deeply imperfect habit, but we all do it. AI is no different. Some have invited us to think about the LLM Powered Chatbot as if it were a search engine, some have called it a graduate assistant, others have called it plagiarism, Ted Chiang famously called it a compression algorithm.
None of these are quite what the thing is, but they might help us think about what the thing is. Some might be better or worse for the task of thinking. Some might lead us very far astray!
Let me throw out another metaphor: the LLM Powered Chatbot (I am being precise here for a reason) is Corn Syrup.
Why? Because like Corn Syrup its a very unsexy single high-profile ingredient that we can blame for bigger problems. Corn Syrup is not the reason for all of the various illnesses of excess the United States faces, no matter what RFK Jr. tells you. Sure it’s probably not great for you, but its also probably not any worse for you than cane sugar. It’s just much cheaper than cane sugar (more on that in a second) and so it ends up in all kinds of things, and probably boosts our overall sugar consumption in an unhelpful way.
In the same way, while its clear that AI can consume a lot of electricity and that data centers, like the ones AI uses, are getting bigger and more power hungry, the idea that AI is a major driver of energy consumption does not appear to hold. AI is one ingredient in the power consumption mix, but other things (from Netflix to Electric Cars to Beef) appear to be bigger drivers.
There’s an even bigger payoff to the AI/Corn Syrup metaphor, though.
You see, nobody ever really intended to have us make so much corn syrup. The problem our current agricultural regulations were designed to fix was the series of boom and bust cycles that left farmers constantly over-investing when prices were low, driving prices lower, crashing out of business, and creating food scarcity (or, as in the dust bowl, over stressing land trying to get enough yield). So the government instituted price guarantees for commodity crops. So now there’s a basically infinite supply of corn and the only profitable thing you can do with it most years is render it up into corn syrup.
So too, nobody every really wanted to make an LLM powered chatbot. They wanted a wide variety of machine learning tools for a wide variety of scientific, industrial, medical, and (yes) military applications. Some of those are quite scary. Some are almost magical.
None of them are ChatGPT or its cousins, but once you have the gradient-descent training methods and attention mechanisms you need for those tools, an LLM becomes almost trivial to build. You just need a big pile of text to train it on. Doing the kind of Reinforcement Learning from Human Feedback (RLHF) you need to take a crude LLM like GPT3 and turn it into something like ChatGPT is a bit more involved, but easy enough to do using the vast infrastructure built to cheaply have people in poor economies do repetitive information tasks that powered social media content policing (back when we still did that) among other things.
And so, LLM powered chatbots are now as prevalent as corn syrup, and about as difficult to put back in the toothpaste tube. Not because they are so powerful and important, not because they are inevitable but because they are an after effect of a much larger and more powerful information system.
This is why I’m generally on the side of critical (often very critical) engagement with AI, rather than rejection of the tech entirely. We’re going to have to deal with it, somehow.
