The General Problem with LLMs

It seems everywhere you look, “AI” is taking the world by storm. These Large Language Models (LLMs) are popping up from companies like NVidia, Microsoft, Apple—everyone in the tech world. TechBros love solutions to problems they haven’t found yet. I’ve got a lot of thoughts on this topic but I want to focus on one in particular that doesn’t get a lot of attention: why they don’t work.

LLMs are built with layers of predictive matrices. Don’t worry, I’m not going to plaster this post with equations, beyond some super basic examples. Suppose in a given situation, you ask a hundred humans what the answer to 2 + 2 is. Almost everyone will tell you it’s 4. So the probability heatmap of how a LLM fed with that data should respond to the question is going to heavily favor that answer, and the user experience will be something like typing plaintext “What is 2 + 2?” and receiving “4.” The LLM doesn’t think, though. It doesn’t have a mind, and doesn’t have the ability to do too much beyond its set of data that it was “trained” on. Training in this context means being given sets of probability maps of outputs when given certain inputs. If I ask it what 2 – 2 is, and we’ve only fed it the same small data set described above, it shouldn’t provide a meaningful answer because none of its data cover this input. It does not have any confidence in a probable answer. I want to be careful and NOT say “It doesn’t know” because knowing is a human thing. A thinking thing. LLMs don’t have awareness. If you want to know the answer to 2 – 2, you need to feed it answers or a maths rule set that covers this input.

This brings me to the next point—LLMs are trained on data generated by humans, because only humans have the level of combination technology & consciousness to create it. Suppose I have a massive amount of data on orbital physics, and I want to ask the LLM to plot me a flight profile for a new rocket. If I give it all the existing data of past flight tests with this hardware, along with math and physics databases, it should be capable of giving a reasonably confident answer. One I would still verify, as a human—engineers trust but verify. We can’t just blindly accept data from the void. If I ask this LLM what the best lawncare product is, it won’t be able to answer. That’s not in its data set. To have it answer that, we need to give it more data, from farmers, crop scientists, chemical and biological specialists, and maybe historical preferences for local varieties of grass.

The problem here is that to obtain useful information from a LLM, you need to already know the answer, or how to get the answer, so that you can verify what you’re given is correct, and that what you’re working with is based on sufficient information to arrive at a correct answer. You need to already know. At which point, you might as well do the work yourself, because if you have to verify everything you’re given back, that’s not exactly saving time, is it?

So okay, what if you give it the (impractical for legal and privacy reasons) entire sum of human digital data? The first problem is that a lot of that data is wrong. I am not an expert on orbital mechanics nor lawncare, so my writings on the internet on these topics would be conjecture at best and misleading, misinformation at worst, if taken at face value by a LLM for solutions to questions regarding their use.

Right, so then if we can’t just feed a General AI (what the Tech Bros are calling it) all the data, and we can’t feed it only selective expert data, what if we determined who the experts are and fed it that? You’d build another layer of predictive heatmap/probability on top of the data set, one that adds a dimension of likelihood of expertise among your sources, for each and every source, and each and every possible topic. This could get incredibly resource intensive even with today’s computing power. And you’ve still got the problem of deciding who gets to be an expert in the first place. Is a discredited, election-denying former lawyer an expert in law? Is a former TV star an expert in vaccine safety? Are cult members experts in religion?

If you could generate such a product, you wouldn’t have a LLM, you would have software capable of passing the Turing Test. It still wouldn’t think, but you would be able to get answers on any known subject with some degree of reliability. But we don’t have the capability to do this, because we can’t agree who the experts are as a species. There is always bias. There is also always the elephant in the room—the lack of creativity. A LLM can’t answer questions for us that we don’t already understand how to solve, because the data that supports its predictions don’t cover that scenario. It’s 2 – 2 without an answer.

Increasingly, these tools are widely adopted by those without education and susceptible to deceit, and less widely adopted by those who are educated and understand the blind spots this technology conceals. The wise ones will vet any answer they receive, or simply elect not to use LLMs when other means are available.

In short, the general problem is that by generalizing the scope of knowledge available to a LLM, we are reducing the specialization to the lowest common average among all contributors. You’re not asking Jarvis; you’re asking a crowd on the subway. And you’d have better luck with the crowd, since it’s more likely they won’t be making things up confidently on the spot.