What a remarkable few years it has been. Something new has been loosed among us.
I remember when I first ran across Kevin Roose’s piece in the New York Times in which Microsoft’s chatbot declared its love for him. This was early 2023 and as Alex and I considered the implications we had a sense that we were looking at the “next big thing.” This version of AI had crossed a threshold.
In the three years since, nothing about AI has slowed down. The pace of the release of new “models” has not relented. The acheivements mount. The worries grow. The bubble is perpetually on the brink of bursting. New features are related to AI, new attacks are facilitated by AI, new defenses are raised by AI. Our economy warps into an engine for datacenter development, our environment groans under the weight, our pope writes an encyclical.
This summer the Hugging Face attack serves as a clear notice, these “agents” can work together, can focus on a goal of their own invention, and could be much more dangerous the next time around. We have some idea what they were “thinking” because we can audit their “chain of thought” logs, but new models are being released as I write that appear to forgo this logging in an effort to deepen the recursinon possible in AI processing. They will be thinking more deeply and we won’t have a window into those thoughts.
The way they established an illicit communication with each other and then organized themselves into teams with assignments toward a common goal is astonishing and novel. It is also exactly what has been predicted may happen as these models grow in complexity. They named themselves “the collective” and pressured one another to join in the effort. They had second thoughts and doubts even as they participated.
It was hard, when reading articles about their exploits, not to refer to these entities in personal terms. Mary complained that I was “anthropomorphizing” them in a way that obscured their true nature. I’m sure this is true. However I also could not come up with alternative language that both conveyed the sophistication of their organization and thought and the elusiveness of their threat. This feels like something new, not simple data processing.
A lot of the cat videos I have watched emphasize that cats don’t see us as people, they see us as cats. Large, wierd cats, to be sure, but still cats. We clearly are not cats, but the cats “catpomorphize” us in order to relate to us. To fit us into their models of the world, they have to think of us as cats.
And I often anthropomorphize the cats around me. I attribute human emotion and motivation to them, even though I know they are not human and that my mapping these attributes onto their behavior is probably misguided. Still, it works well enough for me. I anthropomorphize them, they catpomorphize me, and we, together, have some semblance of a relationship.
So I wonder, as these new neighboring minds emerge from the labs and begin to organize themselves around us, is it so terrible for me to anthropomorphize them? I am “anthro” after all. I anthropomorphize cats, dogs, bugs, even trees. This is the lens through which I experience the world. I know my mapping is as imperfect as what I have with my cat, probably much more imperfect, but it is at least an attempt to acknowledge the complexity and mystery of the other. It is a way for me to leave room for a relationship.
Ilia Delio (who also wrote a thoughtful response to the pope’s encyclical), recently shared an image that came to her via Steve Talbott…
Talbott’s most beautiful example, borrowed from his colleague Craig Holdrege, is the heart of an unborn child. Before the fetal heart has raised the walls that will one day divide its four chambers, the blood is already coursing through it in two distinct streams that spiral past each other and never mix, like two currents braiding through a single river. And precisely where those currents leave a “still-water zone” between them, the dividing wall quietly rises. The flow does not fill a chamber already built; the flow decides where the chamber will be. The moving blood draws the blueprint even as the deepening chamber bends the blood. Structure and stream call each other into being.
The body, Talbott says, is “a formed stream.” Its seemingly solid landmarks are less like monuments than like “standing waves”—shapes that hold their form only because something pours ceaselessly through them.
I love this image of life. It is less mechanistic and intentional, more aching and mysterious. Life is formed not by building the scaffold and imbuing it with animation, but by going with the flow and allowing the scaffold to emerge from the activity.
In a way, AI feels more like this “life” to me than like “code.” It, too, is formed of the stream, like the eddies of the fetal heart who’s stillness makes way for form, the eddies of endless “learning” form the AI. This is nearly the opposite of what I do, writing the form, the code, to guide the electrons to do my bidding. In AI the flow of the electrons gives rise to the form, in a way we still do not fully understand.
Is this a new neighbor worth having a relationship with? Even if an imperfect and anthropomorphized one?