the idea that a "singularity" is possible is just the idea that you can turn "mistaking a sigmoid for an exponential" into a millenarian religion
6
Replies:
18
Boosts:
6
@glyph Believing LLM chatbots will achieve singularity is like someone believing teleportation and manufacture-anything-machines are right around the corner because they once saw a magician perform a magic trick.
@glyph
People also forget that the definition of singularity was simply a point beyond which we have no hope of making any accurate predictions.
Reaching the singularity didn't necessarily mean that we would suddenly get AGI or extropian uploading or any of the myriad other things other science fiction authors layered on it or ascribed to it.
That original definition might still apply to a sigmoid, but obviously it's much less certain.
@glyph If you study population ecology, you learn there are two outcomes of exponential growth. Sigmoid is the pretty one. Spike-and-crash is the common one.
![Chris [list of emoji]](https://nfts.freeradical.zone/accounts/avatars/109/127/168/397/951/576/original/e8a0b6da977923a8.jpeg "Chris [list of emoji]")
tbf, it made for some *great* science fiction in the 90s.
1
@suetanvil WHY CAN OUR GENERATION'S SUPPOSEDLY GREATEST MINDS NOT DISTINGUISH BETWEEN REALITY AND FANTASY
1
@suetanvil it's ruining my ability to appreciate the fantasy!!!
1
That 'supposedly' is the clue.
(My guess: they are hardcore rationalists who took that position in order to flatter their own egos ("*I'm* too *smart* to believe in the Sky Bully") and so didn't think that position through. They are also so insulated from reality by their wealth that they don't know how to handle being wrong about something. So now they deal with the downsides of that choice by techbro-inventing religion from scrstch^W SF.)
@glyph my assertion was that the singularity, as described by ray kurzweil, accurately describes the invention of writing, and i don't see why it would be more interesting if the self-improving intelligent mechanism were made of etched silicon instead of CHNOPS nanomachines. it is harder for etched silicon to self-reproduce, anyway. the CHNOPS nanomachines just do that.
i think human advancement *has* followed an exponential-*looking* curve since that point, albeit with a low base.
2
resources run out. processes hit bottlenecks. optimizations reach physical limits. perpetual motion machines are impossible for reasons that are pretty well understood
1
it is so mind-meltingly frustrating to see people think that we are close to a "singularity" with current AI technology. here's a hint about when you could worry about a disruption so big that it might, even momentarily, *appear* to be a singularity:
a single corporation turning a profit even once
3
@glyph Ah yes, the Singularity: a thing that its religious adherents can't define but which will almost certainly be ushered in by chatbots that tell you to put glue on pizza.
Put me on Artemis III, I'm done here.
@glyph There's two distinct claims that are each independently ridiculous: The Singularity is a real thing, and LLMs put us closer to The Singularity.
in order to be a singularity candidate, an AI would need to achieve vertical integration from silicon fabrication through logistics and integration, into operating systems and applications, with tight whole-system feedback from the robotics to the shipping to the power generation and back
3
@glyph This *strongly* depends on what you mean by "singularity". I think you're conflating that with "hard takeoff paperclips scenario" or something.
I can just barely (barely!) imagine a future where someone manages to use AI to get a more efficient form of AI, which would allow further bootstrapping without requiring more hardware. Same hardware gets more compute.
You're spot-on about the supply chain limitations, though. Good luck to the AI that wants to dig up more cobalt or whatever.
1
@glyph With literally not even one minuscule step in the process dependent on a human doing things at human speed via flesh and blood movement. Because if any of those processes could be made to work without a human doing something physical, why wouldn’t the people with money have done that already??
we are not even remotely close to a single LLM meaningfully constructing even a portion of the pipeline to train another LLM. you can sort of argue around the edges that maybe under certain synthetic conditions this is borderline possible now, but on the "singularity" progress bar, that is 0.5%
2
