Heidegger thinks about tools like so: a tool is something (a thing that has being in the Heideggerian sense, not necessarily an object) that represents an extension of human capabilities in some way. A tool is thus any kind of thing that *lets you do something you wouldn't otherwise have been able to*. The important (for us) conceptual leap here is that when a tool works well, or isn't broken, it develops what Heidegger describes as a "ready-to-hand" quality: the tool fades into the background as a kind of human-tool gestalt of a human with the additional capability afforded by a tool forms. As a simple example of this, consider eating with a fork. When the fork is well-designed and not broken, *you don't explicitly have "I am using a fork" in your conscious mind while eating dinner*.
How many times have you seen someone say “LLMs are just a tool” in a discussion about AI?
Fascinating stuff, altho to me, like multiple articles all rolled in to one, which could easily double as a fairly high-level university course.
My take is that tools & tech indeed get normalised over time, with all kinds of assumptions and expectations built-in as time progresses, which can… lead to some pretty interesting or even catastrophic results, eventually. Like the Bronze Age collapse, perhaps, in which the breakdown of trade rather suddenly crippled multiple city-states, reeling them back to tools & tech from centuries or more earlier. In short, any tool can become a dangerous thing IMO, for a multitude of reasons.
I agree that calling an LLM ‘just a tool’ is both disingenuous and patently false. For me, it’s a kind of dream or fantasy tool that I never really thought would be developed to this extent, back when I was horsing around with stuff like Eliza and Racter. Later, when Deep Blue 2 beat Kasparov (debatably the greatest chess player of all time) almost 30yrs ago, things started to look a little different.
At this point, I’m curious if anyone here feels like critiquing my primary LLM usage. Before explaining, I also want to add that I’m fully on board with the idea that AI will turn out to be a global disaster in the end, and that it’s already been pretty horrible, for example in terms of killing jobs and building massive abominations that we so delightfully call “data centers.”
Personally, I use GPT on a near-daily basis to do complex searches and then pull the results together, saving me a load of time & effort. As someone with ME/CFS, that really is something super-valuable to me. Of course I’ve learned over time that GPT has its strengths and weaknesses, and loves to hallucinate and make fundamental errors when given the chance. To deal with that, I’ve built up something like ~20 global directives to help mitigate such behavior, and of course, sometimes GPT’s own baked-in programming will essentially ignore some of my directives. Still, part of the process as I see it of better understanding where it shines and where it’s most likely to crap the bed. TBH, I’m also pretty blown-away by how much better it’s gotten across only a year or two.
I think that use-case will be one of the few that survive the bursting of the bubble. It is relatively high return on investment.
The way we use ‘agents’ though, I’m more doubtful about. An agent uses many many times more tokens than a basic search / text chat does. This article claims 600x more tokens - https://piefed.social/c/[email protected]/p/2262610/the-real-energy-use-of-agentic-ai-agents-use-about-600x-more-energy-than-simple-ai-prom. No doubt the numbers are debatable but still a useful ballpark kinda thing.
It’s hard to see the economics of that being viable once the VC money runs out.
Yeah I’m just kind of a rubbernecker here, but from the last couple articles I read, it sounds like group-wise they’re not remotely profitable yet. At the same time, in the States and I guess elsewhere, the data centers have been rammed through against local taxpayer objection, meaning that at least to some degree they and the state are going to be subsidizing the DC’s. Also being harmed by their irreplaceable aquifers being tapped.
Far as I know, Nestle is a big player in pulling that shit, and is arguably a benefitee of regularity capture, which naturally means a certain degree of politicians enriching themselves, ultimately paid for by taxpayers & sustainability. My point is that similar things do seem to be happening with LLM’s and their DC’s. So probably all that moves the profitability needle somewhat.
Btw, this particular chart of the author’s helped me understand some things:
Fascinating stuff, altho to me, like multiple articles all rolled in to one, which could easily double as a fairly high-level university course.
My take is that tools & tech indeed get normalised over time, with all kinds of assumptions and expectations built-in as time progresses, which can… lead to some pretty interesting or even catastrophic results, eventually. Like the Bronze Age collapse, perhaps, in which the breakdown of trade rather suddenly crippled multiple city-states, reeling them back to tools & tech from centuries or more earlier. In short, any tool can become a dangerous thing IMO, for a multitude of reasons.
I agree that calling an LLM ‘just a tool’ is both disingenuous and patently false. For me, it’s a kind of dream or fantasy tool that I never really thought would be developed to this extent, back when I was horsing around with stuff like Eliza and Racter. Later, when Deep Blue 2 beat Kasparov (debatably the greatest chess player of all time) almost 30yrs ago, things started to look a little different.
At this point, I’m curious if anyone here feels like critiquing my primary LLM usage. Before explaining, I also want to add that I’m fully on board with the idea that AI will turn out to be a global disaster in the end, and that it’s already been pretty horrible, for example in terms of killing jobs and building massive abominations that we so delightfully call “data centers.”
Personally, I use GPT on a near-daily basis to do complex searches and then pull the results together, saving me a load of time & effort. As someone with ME/CFS, that really is something super-valuable to me. Of course I’ve learned over time that GPT has its strengths and weaknesses, and loves to hallucinate and make fundamental errors when given the chance. To deal with that, I’ve built up something like ~20 global directives to help mitigate such behavior, and of course, sometimes GPT’s own baked-in programming will essentially ignore some of my directives. Still, part of the process as I see it of better understanding where it shines and where it’s most likely to crap the bed. TBH, I’m also pretty blown-away by how much better it’s gotten across only a year or two.
Am I wrong? :P
I think that use-case will be one of the few that survive the bursting of the bubble. It is relatively high return on investment.
The way we use ‘agents’ though, I’m more doubtful about. An agent uses many many times more tokens than a basic search / text chat does. This article claims 600x more tokens - https://piefed.social/c/[email protected]/p/2262610/the-real-energy-use-of-agentic-ai-agents-use-about-600x-more-energy-than-simple-ai-prom. No doubt the numbers are debatable but still a useful ballpark kinda thing.
It’s hard to see the economics of that being viable once the VC money runs out.
Yeah I’m just kind of a rubbernecker here, but from the last couple articles I read, it sounds like group-wise they’re not remotely profitable yet. At the same time, in the States and I guess elsewhere, the data centers have been rammed through against local taxpayer objection, meaning that at least to some degree they and the state are going to be subsidizing the DC’s. Also being harmed by their irreplaceable aquifers being tapped.
Far as I know, Nestle is a big player in pulling that shit, and is arguably a benefitee of regularity capture, which naturally means a certain degree of politicians enriching themselves, ultimately paid for by taxpayers & sustainability. My point is that similar things do seem to be happening with LLM’s and their DC’s. So probably all that moves the profitability needle somewhat.
Btw, this particular chart of the author’s helped me understand some things:
