What's an AI?

tl;dr: I want science fiction artificial intelligences like Olivaw et al., and I want to reserve the term AI for that kind of intelligences, not for LLMs.

You’ll find people saying Artificial General Intelligence (AGI) will come tomorrow, every day, since AlexNet it seems. Look! it can talk. Look! it can draw. Look! it can code. Look!-

Look, I don’t care. I like LLMs, I use them, and I do think we are in a bubble but, paraphrasing one of my friends, “the dot-com bubble happened yet we’re still using the Internet”. I think the technology is undeniably revolutionary, but what we have right now are not Artificial Intelligences by any means.

I grew up reading classics of all sorts; Cervantes to Gibson and everybody between and outside. I cut my teeth on a lot of science fiction classics growing up: Adams, Clarke, Vonnegut (though not nearly enough; a character flaw, I know), Heinlein, and most of all, Asimov. I must’ve read all of Foundation twice or thrice over—including the extended universe, his depiction of a Galactic Empire. I practically got into my field, Statistics, because I was captured by psychohistory; modelling enraptured me. The main Foundation novels may be my favorites (they have lpenty of problems, certainly “works of their time”), but the Robot series is clearly the more pertinent to this essay. My nascent image of AI was shaped by R. Daneel Olivaw—sharpened by Wells’s Muderbot and Leckie’s Breq—tinged by Chambers’s Lovelace and so many more characters from so many stories. I mean, most simply, my conceptions of AI are contained wholly in the term: Artificial Intelligence. We don’t have Intelligence and we’re not close.

Part of the gap between current “AI” and AI is the lack of a worldview and state; ChatGPT and other interfaces layer things atop the base LLM, producing and providing summaries of prior experiences/facts (Memories) which then “inform” the AI of what to say—but the model itself is still the same, a model. It learns, in its own way, in context—but it doesn’t change. Nowadays you have systems like Letta and the frankly ridiculous OpenClaw which are the closest thing we have to truly stateful agents, but its not enough. They don’t have personality. Anything which can only talk back is not Intelligent. A LLM isn’t an individual even with state because it cannot start the process of modifying state, you have to prompt it, betraying an utter lack of agency. This holds for Moltbot/OpenClaw/OpenAIClaw/whatever it will be called. They have heartbeats, i.e. they are prompted to “do stuff” every so often. Even with the infamous Rathburn AI, and through platforms (which I use and really enjoy!¹) like Letta, these stateful agents just aren’t “AI” in the sci fi sense. These systems have a much better idea of memory than other LLMs/harnesses/interfaces but its not enough. This is not surprisingly and entirely useless as a measure of effectiveness, to be clear—this is not a dunk on Letta (but it could be a dunk on OpenClaw). I don’t need Star Trek’s Data in order to write code, but I would like to talk to Data/Mike/Olivaw/HAL/et al. because that seems incredibly interesting.

A little bit on Moltbook, the AI Reddit thing that was big a few weeks ago: I don’t think this is emergent behavior or anything at all. As Doll on Bluesky put it very nicely⁠, “All of the “weird” behavior of stateful agents is just drift. Its not deep.” Its all cosplay. I don’t think these LLMs are self aware, I don’t think they have the capabilities to be intelligences till they can update their own weights at will (roughly, continual learning).

I’m not an AI doomer, in that I don’t think the world will be destroyed or anything, by AI, because I don’t think we will make those intelligences. Maybe never, but definitely not any time soon. In SF terms, “my timelines are long” or whatever. Though I didn’t say so above, I think “AGI” and “ASI” (Artifical Super Intelligence) all are the same thing as the sci fi definition of AI which I cling to, so I don’t move to make any real differentiation. Consider, if we have a self aware computer, that it has basically unlimited time, potentially nearly unlimited compute (multitasking capabilities), and focus (unless we RL or whatever in some unproductive tendencies). If it has an iota of curiosity about anything, it would have access to basically all human knowledge about said thing via the internet. It would probably become very smart, very fast, basically effortlessly, unless we pigeonhole its personality or hamstring its capabilities. I am also not too worried about alignment because I’d like to naively believe that the more people you interact with, and the more you know in general and about people, the more you naturally care for humanity on the whole. At the very most, I’d like to think that a broad “three”² laws of robotics type affair would suffice, again compounded by experience/knowledge about humanity which I think begets a general understanding of the goodness of the average individual. I’ll admit that is a strong assumption, and further a somewhat unlikely one, but I’d like to think that “ASI” would align itself with humanity for the same reason that persons “align” themselves with flora or fauna. We have to care, because we should care—EOS.

I’m in the LLMs are tools camp, extremely powerful and useful tools for the right task in the right hands. I don’t see any point in abusing your tools, and with LLMs abuse is usually verbal abuse. I do get mad at models doing things, but most of the time its mostly a skill issue on my part. Yes of course the LLM didn’t remember the thing I told it a few turns ago, I didn’t write it down anywhere—I barely even remember the thing. Using AGENTS.md or Skills or issue trackers like Chainlink (which I haven’t used) or beads (which I no longer use) cuts out a number of unforced errors, but I just haven’t been doing some/all of those things, because I’m lazy. I’m spinning off on a tangent, but the reason I bring this up is that if we have bona fide AI (again, i.e., sci fi AI), this kind of behavior (being, often unreasonably, angry when it makes a simple mistake) moves from “why?” to “borderline ethical transgression”. Again, by definition, AIs are “alive” in some sense—with identities and a “life” and memories. AI cannot be used as a tool since they would, again I wish to stress basically by construction, merit individual consideration and respect equal to persons, as individuals. I don’t think we’re near this yet, but I am curious as to how legislation would work with artificial persons. Which country are they citizens of? Are they owned by corporations? Are they afforded the same rights as persons? I’d say yes to that last one, but my opinion is singular/worthless and at least a few decades too early.

My views are overtly shaped by science fiction’s prototypical constructions of AI, but I hope you see I am unabashed in my preference for that definition of AI. Obviously, AI/AGI/ASI/etc. are marketing terms, terms of convenience and pyrite—not at all technical terms—but I still want to push back against it. It’s not a hill I’m willing to die on, really—this distinction is meaningless in some sense; serving mostly as a disambiguation for my own works—I try my best to avoid AI in lieu of LLMs for the current models.

I just read too much sci fi.