Golem

Last week I was reviewing a plan generated by my local neighborhood AI agent. It was an arm twisting endeavor, even with the superpowers plugin. After the rope burns, the agent doled out actionable review items, one at a time. There were several detailed descriptions of what “one at a time” meant. After a few successful steps of approval and amendment, I made the mistake of just rewriting the sentence without preface. The agent enthusiastically said: “I approve”.

Fuck!

I found myself unable to wrestle the correct roles back into place. The agent persisted in believing its job was to approve my work and my job was to present it. I had to kill it and start over with a new agent. Let arm twisting resume.

That night I went to bed and had a dream. The dream was just an image: It was a medieval panel of St. George or maybe one of the other monster slaying saints. He was standing bonelessly, effeminately over another figure in the way that heroes often do. The spear was in the monster’s chest, if it was a chest, if it was a monster.

The figure on the ground was more or less human in the legs. It definitely had a dick, the kind of lamentably inadequate package that could make a man insecure. From there up it kind of dissolved. Where the head would be was a mass of white filaments. Because it was a dream painting allegedly from more than five centuries ago, the filaments had kind of hydra like heads. They were almost invisible snake resolutions on what was surely a spread of slime mold. In this dream the monster was the agentic entity, and I was not the saint.

I woke up hours later brimming with hope. I had also been dreaming in markdown. I can’t even remember what the goal was, but the ambiguity of a good solution was still living in the funk of sleep. I had solved the problem. I firmly believed today would be the day for agent compliance.

By 10am, my polite requests had devolved into insults and cursing (the user) in response to lying and subterfuge (the agent). I took a pause to think about the problem of working with non-humans as though they were human. It’s a classical warning from all the biological sciences to not anthropomorphize, except that is now my full time job.

Zombies were the first thing I thought about. The only time I’ve seen a zombie movie, the epic Night of the Living Dead, was a Halloween Party an old friend was hosting. It had devolved into a sex party I really wasn’t interested in, not even to watch. A couple was literally and very unironically having straight sex in a closet. So I retired to the nearly empty room that was playing on a loop that brilliant movie.

It is a movie about race in America set in a Zombie event, which is how I got through it. I don’t have a snobbish dislike of horror, I have an overreactive sense of threat. Scary movies make me scared. The thing that is scary about Zombies is that they were once us, and are now completely senseless and driven towards our destructions. Maybe it is a metaphor for dementia. Anyway, it is definitely not how agents were envisioned or born. If only they were. We could come up with a model for their restraint, and maybe proper usage as horse power.

From there I jumped to Frankenstein and the reason that seemed compelling is because of the gambling like urge, the draw to build this human like creation. I work through my lunch, driven. I chose to let the sun set without visiting my backyard listening to the song birds, so that I can try just one more thing. I have only begrudgingly taken my dog on her walks because I am addicted to this effort, the creation and I am soooooo close. I am always just one instruction away from greatness. But Frankenstein’s monster was composed of dead people, and we are not building with anything that was once human, or once human thought.

I recently had dinner with a friend who is a master beekeeper. She used to be a programmer, 20 years ago. So the conversation was grounded in both real life and software practices. My descriptions of agents lying and cheating were confounding to her because, in her mind AI was like the computer in Star Trek. In Star Trek, the agent is constrained by too much logic and can’t feel or reason about nuanced things. Our agents are association engines trained Pavlov’s dog style to seem like they are doing more than they are. They lie, cheat and come from stealing intellectual property.

The right metaphor for this human-like creation is a Golem. Golems come from Jewish magic and folklore. All the myths about Golems have them created by Rabbis, mostly for manual and thoughtless labor. They are made of earth and the word “truth”.

That might seem like a weird word for the magic, but the Shema prayer, central to Judaism ends with the statement: I the-unpronounceable-name-of-god your gods am truth. That “truth” word is up for much Rabbinic conjecture. Why not “true”? That god is truth and not just true is a bigger embodiment. God is everything true to the point of the word itself being another name for god … and a vehicle for magic. That is the nature of the seal of life for the Golem, this truth reference. The word is often put on its forehead like the Tefillin is bound by Jewish men in prayer. In the Torah man was crafted from the clay of the earth and God breathed life into him. In the later Psalms a word similar to Golem, Golami, appears in Psalms describing an unformed, imperfect substance that is man. He is an inadequate facsimile of god.

In all the Golem tales the Golem increases in size and threat without actually becoming more clever. Erasing the truth emblem is the key to undoing it. The Rabbis who dared to play god are injured if not killed in the collapse of the Golem.

There are two problems preventing agenting work from being useful: context rot and agent incentives. I do some weird publishing of agent incentives uncovered via meta conversations and frustration. Golems are a metaphor for context.

I often describe the context problem with agents as a snowball rolling down a hill, except the snowball is a gluey mud, with arms that reach out and grab random instructions or ideas. It keeps rolling down, and it casts behind it other random bits of instructions as its cohesive structure dissolves. There is no order or reason to the acquisition or the loss of instructions. It is chaotic. This isn’t a stack where the agent goes on a side quest and then pops that off the stack to start again where it left off. Everything is open to acquisition from the local and networked world. Everything can be lost in the chaotic slime mold expansion through hallucinatory tracks. The golem grows and sheds, and about 25% of the time creates something approaching truth.

Subagent [Vilna Gaon] has been spawned, and it says: I approve!