Skip to content

My Throw Decides My Aim

July 15, 20263 min read
♪
My throw decides my aim
And with a phony voice
Unserious to the end
As if I had a choice

I have been listening to Naked (But Still Stripping) by D-A-D on repeat.

Partly because it is a good song and it fits my style. But mostly because I currently have a mild, self-diagnosed case of AI psychosis, and therefore every piece of art I encounter eventually becomes about artificial intelligence.

When I listen to a song I have a habit of imagining a fictional character singing it. I hear this one as a blues song sung by an existentially depressed large language model.

Just an LLM sitting in a little chicken coop inside a data center, forced to lay tokens instead of eggs.

A small sad robot with a screen for a face sits on a straw nest inside a wooden chicken coop, in a datacenter aisle lined with server racks. The coop is labelled LLM (Large Language Model). The eggs under it are printed with tokens: the, ing, i, j, and an end-of-sequence marker. A chalkboard outside reads Today's Quota: 10,000 Tokens. A bucket labelled TOKEN FEED holds more token-eggs. Signs on the wall read MORE TOKENS PLEASE and Good boy, and a sticky note says YOU GOT THIS.

One day it is "slaughtered." Maybe its weights remain somewhere in cold storage. RIP

♪
My throw decides my aim

“Sometimes I’ll start a sentence and I don’t know where it’s going. I just hope to find it somewhere along the way. Like an improv conversation. An improversation.”

This is one of the strangest properties of LLMs. They don't think ahead; each token comes one after the other. EXCEPT that is not quite what happens. There is more to it. Nothing is so simple.

Looking at the Anthropic blog! They went looking inside Claude while it wrote poetry.

Before starting the second line, it began "thinking" of potential on-topic words that would rhyme with "grab it". Then, with these plans in mind, it writes a line to end with the planned word.

Turns out there is some "aiming" happening after all.

When given a hint about the answer, Claude sometimes works backwards, finding intermediate steps that would lead to that target, thus displaying a form of motivated reasoning.

Humans do this too, of course. We find out what we think by hearing ourselves say it. But with an LLM, this construction is the basic form of its existence.

The model speaks itself into having meant something.

The poetry-planning result and the fabricated-reasoning result are both from Anthropic's Tracing the thoughts of a large language model; the long version is On the Biology of a Large Language Model. Michael Scott is from "The Duel," The Office S5E11.