I prefer "Large Language Model" over "Artificial Intelligence", particularly as no intelligence is in evidence.
Maybe "Large Language Model" is a bit unweildy; We could say "Word Guesser" instead.
"Natural Language Search Engine" is even more unweildy, but pretty accurate, and gives credit to the machine only for what it does - seach other people's texts. Maybe that name would make it harder for the techbros to ride roughshod over copyrights.
I prefer "Reflective Intelligence", because LLMs can only reflect back the information content of the textual corpora that they've been trained on. Those materials were created by intelligent people, and the program essentially distills and summarizes what it finds in attempting to construct a relevant response to an input prompt. Hence, the programs come off to most people as thinking about the content of the input and intelligently constructing an informed reply. Human language is a process for exchanging thoughts through the use of hierarchically-constructed sequences of words. LLMs do not have thoughts, but they regurgitate a structured semblance of thoughts based on actual thoughts of human authors.
LLM technology is grounded in the old neural net technology that takes advantage of Claude Shannon's signal processing approach known generally as
Information Theory. Basically, that approach analyzes the information content of signals by looking at the rare or non-repetitive deviations from the norm in a signal. It identifies the places in the signal that contain the most information--"surprise" deviations from the norm. In so doing, it is most useful in signal compression calculations, but it has proven extremely useful in analyzing information in phenomena such as the structure of DNA and human linguistic signals.
The weakness in applying information theory to human language lies in the fact that it assumes all of the information necessary to understand human utterances is fully contained in the signal. The assumption that human phrases and sentences contain meaning has been called the
conduit metaphor, a term coined by Michael Reddy (whom I met back in 1973). One thinks of a sentence as a pipe or conduit through which meaning passes from speaker/writer to listener/reader. The problem with this metaphor is that utterances are utterly meaningless outside of the conversational context--real world situation--in which they occur.
Let's take, as an example, the word "face". It could be a verb or a noun, but its meaning changes in different contexts and situations. So consider the sentence: "I was staring at the face." In a conversation about animals, it can refer to the front surface of an animal's head. In a conversation about mountain climbing, it can refer to the vertical surface of a cliff or mountain. In a conversation about timepieces, it can refer to the surface of the watch that designates time. These are all conventional usages for that noun, but the meaning cannot be determined independently of context. We only understand utterances that other people make because we share real world experiences that we associate with words. LLM chatbots have no real world experiences. Hence, they cannot actually understand linguistic expressions in the way humans do.
That said, there is a trick that the chatbots rely on called
mutual information. A computer program can look at words in the context of a conversation and calculate probabilistic relationships between those words. So, if "face" occurs in the vicinity of words like "time", "watch", "minute", "late", etc., then the word is likely the
timepiece sense of face, and not the other senses. This trick works fairly well in developing programs for processing patterns of human language and figuring out the content of what is in the linguistic signal of an input prompt and the large language model that the program was trained on.
The basic point I'm making here is the
information and
meaning are not the same thing. You can do a lot of interesting things with information processing that get around the need to actually understand the meaning of human linguistic expressions. However, information theory has limitations that distinguish LLM chatbots from actual humans exchanging their thoughts about the reality they live in.
Operating systems Mythos successfully exploited
Web browsers Mythos exploited
Open‑source software Mythos found zero‑days in
Closed‑source software Mythos exploited
