Colin Fraser described the experience of using a Large Language Model (LLM) as like co-writing a screenplay with the computer, where it provides the lines of whatever character its prompts describe.
Which is an interesting and useful way of thinking about it.
But I think Jorge Luis Borges provides a better one.
An LLM is very much like Borges’ “Library of Babel“, which contains every possible text that can be created using a specific alphabet. The challenge, in Borges’ discussion, is finding one that’s worth reading.
An LLM is nothing more than an unreliable device for navigating this library of all texts; it does this by walking through all possible texts that are similar to the one its currently holding, which is made up of its prompts plus whatever text it’s generated so far.
It doesn’t know what the texts mean; it doesn’t even know what the words they’re made of mean. All it actually knows is that the text it’s holding now is connected to various other texts that are almost but not exactly like it.
Some of those connected texts will be useful; most of them will not be.
This doesn’t matter to the LLM. It neither knows nor cares what is meaningful, or true or even just what the user will find useful; all that guides it is how similar one text is to another, plus a little dash of randomness (adjusted to match all the data that was poured into it for training) to spice things up.
It will simply choose one semi-randomly, and it will keep moving on through the texts until it reaches a point where stopping becomes more likely than continuing.
It is on a drunken walk through the library of all possible texts.
And it will give the user whatever text it happened to be holding when the algorithm says ‘stop’.
We’ve had text generators like this for a long time; this conceptual architecture is not really different from the Markov Chain bots of 20 years ago.
What makes LLMs special, and their output so eerily lifelike, is the bit where the randomizer is “adjusted to match all the data that was poured into it for training”. It doesn’t increase the system’s actual understanding of anything, but it means the texts that it chooses from the library of all possible texts are far more similar to pre-existing human-created texts than those generated by any of the earlier variations.
It is an improvement, but only in that it only chooses texts that read very much like human-made ones.
A card catalogue is a limited but reliable navigation aid; it knows actual facts about the texts it points to: author, subject, publication year. You can use a card catalogue to find a text that matches specific criteria.
An LLM is not a card catalogue.
It will not, can not, give you a text that matches any known facts. It will give you a text that is similar to the text by which you can express the facts you want, when you put them in the prompt.
But similar only means that it will contain many of the same words, arranged in a way that structurally resembles a text that expresses the facts you want.
A real artificial intelligence would be a GPS-quality navigation aid.
It would know precisely where you are in the library of all texts, and be able to calculate from that the precise location of any text you could describe to it.
It would be an experienced reference librarian, who knows the library and can point directly to the text you’re searching for.
LLMs are not that, and cannot become that.
And we need to stop treating them as if they were.
