Employing normativity on ChatGPT
»Now the Web at large is full of slop generated by large language models, written by no one to communicate nothing.« (source)
There’s a certain way of speaking about ChatGPT, where people would say that ChatGPT ›is lying‹, or ›is wrong‹, or ›is producing false facts‹, which doesn’t ring adequate to me, because when you are interacting with ChatGPT in order to get some answers to questions, it’s not like typing something in a search engine or reading a Wikipedia article about the topic in question. Since ChatGPT is a language model, it doen’t ›know‹ about facts . I’m surely no expert in this and my knowledge about AI and machien learnng is very inchoate. But I guess it’s safe to say that GPT - same as other machine learning models - is all about predictions and statistical relations rather than retrieving collected facts from some database. (Jon Stokes’ ChatGPT Explained: A Normie’s Guide To How It Works is a good primer and explains better what I’m trying to get at.) So, very roughly speaking, when asking ChatGPT a question it will try to produce an answer that sounds like what some other speaker of your language would plausibly say in reply to your question. It’s more or less likely that it will accord to facts in the world and sound like the right answer, but that is just because ChatGPT ›knows‹ the rules of our language, not because it has some mechanism to find out about facts in the world. As someone on Reddit pointed out, you should »[t]hink of chatgpt as a language model alone. It isn’t a knowledge model. It doesn’t have intelligence. It’s predicting what a person might say in response to what you say. This is a very plausible thing someone could say, if it was true, in response to you. If you want the most out of chatgpt, you shouldn’t ask it knowledge based questions much but questions that are a language question in nature.«
LLMs are not for maths
At the latest when people realized that ChatGPT makes mistakes even in simple mathematical tasks, doubts were raised about its usefulness and potency. Nevertheless, I still see math problems being typed into ChatGPT or ChatGPT being used for accounting purposes, which is quite frivolous. ChatGPT is only based on a large language model. This means that its task is to allow plausible sentence constructions to follow one another. However, if mathematical tasks are written linguistically, there is a risk that the LLM will unnecessarily treat these tasks as a linguistic problem and calculate a plausible-looking answer, which is different from giving a mathematical answer. Most everyday math problems are deterministic in nature, i.e. there is one correct answer to them and maybe a handful of mathematical ways and functions to get there. So you don’t need a generative model for these kind of problems to arrive at their answer. The generative aspect of the current so-called AIs is even detrimental in this case. After all, the point is not to think up an answer to a mathematical problem, but to find ways to calculate the answer using mathematical operations. (In addition, a request to ChatGPT costs around 3kJ, which is why it would be more sustainable if we were to unpack our solar-powered calculators again.)