pull down to refresh
So I think that - per my other comment about finite, static models - that the LLMs cannot truly emulate it to perfection with what we have right now, because you'll run out of "context window" space[1] and then coherence is lost versus the human mind. So it's not as abundant as we think it is until we can inject human context into LLMs. LLM context aggregation is actually a very rudimentary simulation of something we afaik still don't truly understand in humans.
What Banks does mostly is exploring social dilemmas between societies in a post-scarcity setting. With an idea about AI that is not very compatible with the LLMs we have today per se, but it's what all the labs (imho, falsely) claim we have. In fact, I think that we are not closer to the abundance Banks wrote about since the 80s than... when he started writing about it[2].
So with that in mind, is scifi unbound? I'd say yes, evidenced by there being a lot of bs out there. Also without LLMs. But is the specific narration that clearly influenced Elon here bs, besides being a fictional exploration of concepts? I'm not sure. And since he's literally trying (and making progress!) to realize some of the concepts described in extremely futuristic stories, i.e. neuralink being the most obvious product, but the manufacturing processes with "gigafabs" are maybe even more so, and the low hanging fruit, does that make it all an illusion? Or do we need the inspiration from potential-bs to make a grand plan? And is there a place for LLMs in the process to create something good enough to base such a grand plan on?
I think that Elon is no dummy, but much more importantly, that the distinguishing quality of executing a vision is not something you can ask an LLM, not even to plan for it. A 180 minute crunch of a bleeding edge bot right now still produces horrible plans end-to-end. Even while we must all fear for our lives according to the lab bois (or maybe this is why, haha). If you go from plan to implementation without expert judgment and improvements, failure rate must be higher than 90% at the moment (I test this extensively, which is why I don't really believe the labs in their claims.)
Thus, to answer your question: no, there is no abundance, no, we'll not all be Elon, but yes, we have some tools to make things easier to attain, as long as we ourselves are able to judge and steer and keep oversight, and that is potentially the most important skill to learn. It's no accident than many of the people that do brilliant things with LLMs have a background as process technicians, systems architects and/or technical project managers: they're experienced in realizing visions.
A chatbot session stays coherent by letting the LLM process every message in the chat history with every new message, rather than what it looks like (that the bot remembers the session.) See this super basic API example how turns work: all the history is contained in the next message. Which is why you spend insane amounts of tokens. ↩
One way to narrate LLMs is to say "look how cheap it is when it replaces a human worker" - which it doesn't in most cases - and imagine abundance. The other one is "omg this is the most expensive SaaS software I ever got for free", which in turn can economically only be scarce, because someone is bearing a cost also if you got something for free. I think both sides are concurrently not fully true, but both can be experienced as truth for as long as limitations and mitigations aren't. ↩
The reason I pinged @denlillaapan is that he has made interesting points on this forum about applying the subjective theory of value in evaluating literature, self-publishing, and writing more broadly in the era of the LLM. It has been fascinating to read his thoughts on it, since I agree with him on a whole lot, especially when he says that the marginal cost of producing content, acquiring enough knowledge about it and executing is now so low that nearly everybody can (does?) do it, and so essentially, why bother? I can empathize with him on this from a writer's perspective, and I appreciate bringing an economist's scalpel to a field that is rife with excrement. Absent a zap-war on some zany internet forum, writing just doesn't pay well -- and apparently, not there either, unless you are zapping too. On the flip side of lowering the marginal cost of writing, one might expect to be inundated with an over-abundance material that their head starts to spin, which I'm not seeing. It is the case that these tools have made it easier to make potential-bs and so have mired the landscape of potentialities - i.e. 'anything goes...no constraints' etc. - but to my point earlier, they are not truth machines.
something we afaik still don't truly understand in humans.
This is a great insight and, I think, with a little humility, the language-bot might be able to teach us something about it. Why do people continue to create bs sci-fi novels? They always have, and they probably always will, which means that there is an infinite number of possibilities, but not for me and not for you. I cannot accomplish all of them, even if humanity as a whole can. I must make my shot at making one that cuts through all the bs in a limited time, in the same way that @denlillaapan must do so in economics, or climb mountains. There are limits and feedback mechanisms, but they just exist over a longer time horizon. That said, even if someone creates the perfect machine, with a continuously, self-improving context window, absent some type of divine intervention, it just wouldn't be human nor would it want to do the silly things that humans do.
I think lesson here, as you already somewhat alluded to, is that exceptional people execute on exceptional vision, in order to transcend their temporal existence as limited beings. And the robo don't.
the marginal cost of ...
let's go through them
- producing content
This was one of the major things that to my knowledge, as an insider, early internet powerusers in all the enthusiasm about and advocacy of free flow of information kind of didn't see coming. Not so much the enshittification cycle, but the realization that digital content is so much easier and cheaper to produce than analog content, even before LLMs entered the scene, that the quality of actual content must go down. And I think that there may be not only correlation but a causal relationship between the marginal cost of content, and the median quality of all content. And the question is now: will we be happier when we're consuming low quality content? I haven't ever seen a study saying that doomscrolling is beneficial (besides a tiny hint you yourself gave) but that is, to me, the culmination of low quality content consumption (and also: podcasts, streaming services, opinion pieces in news papers, and most the online discussion, probably including my own comments, probably including this comment.)
But that also means that I still enjoy well produced content, and that the only thing the marginal cost of producing content does for me is there being so much noise (also on SN) that I need to spend significantly more time on the filters. Even today, a time when I have removed myself from nearly all social media and fora, I spend more time on filter tuning and curation processes than I have ever done before, as in 10x more time.
I think that this is what hooks in to your observation that
one might expect to be inundated with an over-abundance material that their head starts to spin, which I'm not seeing
I am not seeing it either because I just shut the slop out.
- acquiring enough knowledge
I think this is disputable, especially in the LLM era. I know some people, including a bunch of stackers, that are proven capable of using LLMs for input or at most intermediate-product only, and that is, from my deeply biased pov, a great use-case. But overall, I don't see that at all! Instead I see across the board an increased reliance on LLMs to create end products from lazy natural language descriptors. And in that case, you don't really obtain any knowledge, because all you really do is publish the results of an automation pipeline (and as someone that once upon a time was literally specialized in automation pipelines, it is my professional opinion that these pipelines suck, especially the lazy "agentic" ones.)
So while the LLM could better enable the acquisition of knowledge than a search engine or an encyclopedia would because it can enrich these sources and is not bound to the limits of those, it heavily depends on the wish of the user to be empowered and accumulate knowledge versus just yoloing a bot here and there and reaping the meager rewards slop will bring to you. Looking at how many people I know that claim they cannot read, especially in NA, but it is growing in EU too, there isn't much of a mindset for taking in knowledge in the first place?
they are not truth machines.
Because truth is contextual, and context is social? LLMs are anti-social. The best they can do is fake it and then you recognize it from a lightyear away.
- executing
I think I addressed this in my previous comment. This is the one thing LLMs are said to be good at but I have honestly not seen it. I've tried the most hyped solutions coming from the AI bros for bot decision making and orchestration, and it is all completely retarded. Even today, if I am not precise enough in my question to a bot I get all sorts of recommendations towards execution from some silicon bullshitter that literally has zero experience. I guess Claude has been trained on real life Deloitte consultancy chats. What a bunch of rubbish.
So here's what I think: if your bar is ultra low, then yes you can have content automatically created based on superior "knowledge" and automatically published. Which works great if your audience is a bunch of doomscrolling zombies with no future other than the lowest rank on the UBI ladder and mandated sterilization.
For me, I prefer your posts, and I prefer @denlillaapan's articles over the usual slop we can find on the mainstream media. I like reading things from people that are passionate about peculiar things. Especially on SN. I also prefer not to need UBI.
And the robo don't.
This is the most important thing, I think. For us to retain self respect and the drive to produce and share interesting things, we must have things where we're superior from the robots. But my question remains: will enough people recognize that they need this before they lost all competitive advantage?
This was one of the major things that to my knowledge, as an insider, early internet powerusers in all the enthusiasm about and advocacy of free flow of information kind of didn't see coming. Not so much the enshittification cycle, but the realization that digital content is so much easier and cheaper to produce than analog content, even before LLMs entered the scene, that the quality of actual content must go down
Tooooottally. And the irony that we demolished the gatekeepers only to be drowning in good content among lots of slop/entshittification that we now desperately need the gatekeepers back. (=Curation)
What I think you mean is, could the LLM eventually help produce 1000x more Ian Bankses, dilute the market of speculative fiction so much that the stories become next to worthless because now we can all become giga-brained, elon-level imaginative and f-forward toward infinite abundance, universal high-incomes to a time when we all just plug ourselves directly into the sun to obtain god-level omnipotence?