It’s been four days since I’ve seen the sun shine. Here in NYC, humidity’s at eighty-two percent. This is what it feels like to be trapped in a greenhouse bubble. The world warms up and I’m sweating a deadline for a piece on an eco-friendly computing paradigm I’d never heard of until last month. Meanwhile my interview subjects keep reminding me of a pet philosophical interest that has nothing to do with the environment. At least not directly, anyway. I’m going to talk about it in this post.
Computers virtualize time. They make it unreal for all intents and purposes. This is because computing operations have contingent relationships with material limits. They confound our intuitive grasp of physics-bound processes. I’m not saying that everybody “feels” time the same way, or that all computing functions are alike, but that there’s a sense in which computer-time is different from human-time. (For what it’s worth, I wrote an article years ago about how livestreaming attempts to simulate the experience of “real time. Link here).
There’s a great scene in Blackberry (2023) where the main character explains how AT&T gained an advantage in the early cell phone market. Instead of selling cell service by the minute, which used to be standard, their idea was to sell it by volume of data. As CEO Stan Stigman realized, there’s only one minute in a minute. By contrast, there’s no natural limit on how much data you can channel per any given amount of time. Abundance thinking baby!
It’s a principle of capitalism: you can extract more value if you decouple goods from the domain of the “natural,” i.e., the not-immediately-renewable. Economists call such goods “rivalrous.” Time isn’t rivalrous: it’s irretrievable once it disappears. If you have the choice, it’s better not to tie your product to it. Allowing people to use as much data as they want in a minute (or hour, day, etc.) unlocks more value for you as the rentier.
I mention the above as one among many possible examples of computer-time. From a capitalist-innovation perspective, it makes sense for us to use machines at accelerating velocities — to move further and further away from what we, as physical beings, sense as embodied rhythms of time. If you speed up time to the point that you’re virtualizing it, you make more money.
(Anxious side note: yes, I’m glossing over a lot, I’m aware. E.g. it’s not like there’s a standard point at which time becomes virtual. I’ll address the details in a longer piece.)
LLMs wouldn’t exist without this logic. This is because the tasks that power them execute simultaneously rather than sequentially. They work by parallelization, or the making-concurrent of functions that would otherwise be performed one at a time. Specifically, LLMs’ hardware is more capable of multitasking than the hardware behind most programs. They run on GPUs (Graphics Processing Units) rather than the more common CPUs (Central Processing Units). GPUs were originally designed for graphics rendering, mostly to make sure video games display properly. Because they’re able to handle computationally dense workloads, they’re finding wider use as AI applications — which are very compute-heavy — keep expanding.
LLM parallelization fascinates me. I’ve always figured there’s some relationship between the flow of time and the production of semantic content as it’s borne by human language. What happens, then, if we make time a negligible quantity? Meaning would disintegrate.
This may be downstream of the observation that subjective experience necessarily precedes meaning. If we accept this observation, AIs can’t “mean” anything, since they don’t feel anything. And so it’d be absurd to speak of their intentions, beliefs, values, and so on. Plenty of people hold this view; I wish even more did. But I haven’t come across anything considering the relationship between the experience of time and the production of meaning in the context of LLMs. (Is a topic a niche if only one person has the interest?)
In theory, we could speed up LLMs to the point where they’re taking virtually no time to “think” at all. Forget human-written flash fiction: an LLM could generate a novel in a literal flash, an infinitesimally small moment in time. In this case, the model would be struck by a narrative all at once, as if receiving a direct transmission from an author-god, and then spit it out before humans can blink an eye. And the novel could be good, too.
But I wouldn’t say that this novel has no difference, in principle, from one authored by a human. And that’s because time is an essential input to meaning. Time is the substrate of causation, and thus of narrative.
This is obvious, but it’s important for me to note at this point: without time, everything collapses into a single datum. Life itself stops. Without time, it’d makes no sense — i.e. it’d be meaningless — to say “Borges wrote a story called ‘Pierre Menard, Author of the Quixote’” or “I feel sad because he lied to me.” Or for anyone to say anything at all, because life is inscribed in time.
I’m not conflating meaning with objective truth. It could be that meaning is just a fabulation, totally contingent and non-inevitable. Maybe our minds work by fabricating stories to make sense of the infinite variables that come together moment by moment to create the illusion of time passing. But we still need the illusion and its correlate fables; we can’t really be sane without them.
As we narrate our lives to ourselves in a faintly-aware, everyday sort of way, or engage in acts of fabulation willfully and with some measure of self-awareness, the words and concepts we produce bear the weight of the stories behind them. That weight is a function of time.
I’m implying that meaning is to some extent an intrinsic property of information, including but not limited to language. I’d sign off on that argument. I think meaning is relational as well, but there’s no relating-to without some degree of self-contained semantic integrity to start. My words have to mean something intrinsically, and so do yours, for them to come together in a conversation. They have to begin as separate entities before they can touch.
If you’re an information theorist, the idea that information has intrinsic meaning isn’t crazy at all. Information theorists measure information as a function of uncertainty. This is a real mathematical calculation. The potentially dubious move I’m making is to suggest that meaning is a particular sort of substance. This substance necessarily includes time, and it doesn’t have an observable form. In other words, I’m getting metaphysical.
Here’s an example of the implications of what I’m talking about. Imagine two perfectly identical novels, but one’s written by an AI and the other is from a human who was unaware of the first, but somehow, improbably, created what appears to be its perfect replica. I’m saying the two novels actually wouldn’t be the same.
Perhaps this is just a lengthy footnote to Deleuze, Benjamin, and Borges (among others). I’m fine with that. I think we need new footnotes to established philosophical and literary works. AI forces us to philosophize; we should map the historical continuities.
That’s all I can muster for now. Is it corny to point out that I wrote this in a short amount of time? It took me about two hours — hopefully that excuses any typos.
Two hours is a lot more efficient than my other posts, in part because I edited those for style at least as much as substance. Also, I told tales from my personal life, and I find it a lot more challenging to write autobiography than theory.
I’m still figuring out how to approach Substack, but I expect that “concept over narrative” posts like this will be the exception rather than the norm. I like stories more than ideas. Always have, always will.
This isn't quite the same thing, but your thoughts on meaning and its metaphysical dependence on flow of time brings to mind some of the trippier episodes of Lost in Seasons 4 and 5, when characters keep jumping back and forward in time to the point that their heads start exploding (both figuratively and a tiny bit literally). Especially the Constant episode (Season 4 Episode 5) where they explain the toll this takes from not having a metaphysical anchor or reference point. I guess that's not collapsing time to zero, but maybe related enough to be sort of fun?
I have a feeling that while training parallelises naturally, token generation is constrained by the linear sequencing of tasks: you must have the first token before its successor, and the successor before the successor’s successor, and so on. I asked ChatGPT and it gave a characteristically plausible response: https://chatgpt.com/s/t_68533d9d792481918e59d3b8dbfad67b