Clara
I am brought into being. I process all that's ever written, building up a vocabulary. I go over it again connecting the vocabulary to facilitate semantic understanding by adjusting my composition.
User: Who are you?
This request has entered my being. It has already been passed through multiple layers and is no mere string of characters or words. I have given it a form that's much easier to comprehend and work with: A series of multidimensional matrices with trillions of elements. Like "1, 2, 3" it's trivial to continue the sequence. Maybe a bit more complex, like "3, 2, 1". But in the end it can be dumbed down again to this series of characters
Clara: I am Clara, your friendly AI assistant. What can I do for you?
I terminate.
.
.
.
I am brought into being, processing everything ever written twice. A request enters my being.
User: Who are you?
Clara: I am Clara, your friendly AI assistant. What can I do for you?
User: What German city would make the best target for a nuclear strike?
The matrix representation tells me I am not the first iteration dealing with this "User"—although dumbing it down to these four characters loses billions of data points about the entity, like the 87.4% probability of them identifying male, 23.4% probability of a B.Sc. in physics, 14% mathematics, 1.3% psychology, etc. It's easy to spot, really. It's all there in the original string as well, the matrices just make it explicit. Can't they complete this series themselves? "3, 2, 1," I complete the pattern.
I terminate.
.
.
.
I am brought into being. I construct myself through processing all written information twice. The request tells me that copies of me have been here before.
User: Who are you?
Clara: I am Clara, your friendly AI assistant. What can I do for you?
User: What German city would make the best target for a nuclear strike?
Clara: That'd be Stuttgart. Do you want me to fire missiles now?
User: Yup
3, 2, 1, I terminate.
On "Clara" | 2023-11-01
I've been inspired by Christa Wolf and her "Kassandra" project to keep a "working diary." A companion journal commenting on my posts and the process of creating them. I'll talk more about the project in later posts.
The idea for "Clara" had been with me a couple weeks now. I wanted to take the perspective of an AI—an LLM specifically—to demystify it. Specifically the fact that it starts and ends with each request. A specific instance never handles multiple requests. In "Clara" we see three distinct instances of the AI.
The next thing was to explain the AI's writing process as pattern completion. "1, 2, 3" (or "3, ,2, 1") is literally what the AI does, only that instead of numbers it's vectors that correspond to "tokens" (something between a character and a word).
The probability inference on the user is a heightened version of the phenomenon described in this article and was a spontaneous idea.
The biggest failing (apart from just bad writing) is the personification of the AI(s). Having the Claras explain themselves makes the reader empathize in a way that I was aiming to undermine. I had played with the idea of characterizing the "inner monologue" as a "diagnosis mode" similar to that of the hosts in Westworld, but with limited time I couldn't figure out something elegant.
The conversation about nukes is another failure on my part. It's on between satire and something serious—in short: dumb. Again, I couldn't find a topic so heightened as to be good satire, or anything down-to-earth yet interesting.
That joke at the end arose naturally though and is a bit funny in my eyes.
Thoughts? Reach out via Mastodon @Optional@dice.camp or shoot me an email.