A thing I think about.

Premises:

- Many people in "AI" believe (or claim to) that if you simply upscale the existing neural network models by one more order of magnitude there will be a phase change and it will suddenly start acting like a human, or a God

- A huge suite of "AI datacenters" that would upscale by such an order of magnitude are being built

- There is a coming economic crash/correction which will likely come before those datacenters come online, and halt their growth.

(1/2)

Conclusion:

- There is a plausible future where the AI bubble pops, the size and commonness of large language models go down, and a group of dead-enders forever holds onto a near-millenarian belief we were on the cusp of creating "AGI", or creating God, if only the economic collapse (possibly presented, rather than a natural economic process as backstabbing by a group of luddites) had not stopped the "AI data center" buildout just as it was starting to work

(2/2)

@mcc congratulations LLM nerds you invented the tower of babel from first principles
@mcc this is how ugly religions start

@nasser @mcc

And those ugly religions are why we started the Butlerian Jihad.

@nasser @mcc I was gonna say, this has such Tower of Babel vibes lol
@mcc funny to think that there's a good concept for a near future sci-fi novel here, and it's far enough out that you could probably get it written before it fully comes true... but that publishing timelines are slow enough that it will seem embarrassingly dated when it actually hits shelves.
@mcc I think that’s not at all unlikely. I think if we’re lucky we’ll see LLMs and other generative AI used in specific, constrained contexts (and thus smaller models, many on-device), and the huge data centres will dry up, as will any serious attempts at AGI.
@michaelgemar @mcc and earth will breathe out a great sigh of relief.
@mcc but it's a religion based mainly around prosperity gospel, so when the tide goes out and the money with it there won't be many followers left
@mcc possibly the most unsettling thing is that these people will make that argument regardless of how far the world got into enacting the plan, because the belief is really just a capitalist self+other-annihilation drive dressed in priestly robes, and more will never be enough to disprove what they believe
@jplebreton @mcc Right. The growth still happens, but slower, and they make less profit from it, and there will hopefully be a more diverse effort focusing less on AGI and more towards things they CAN make money on. So delays it by a few years. Not that the acolytes will ever believe that.
@mcc
oh god this is depressing true
@mcc
at some point I wrote out a draft of an essay / thesis proving this is impossible, but the idea of people just believing this shit anyway despite it was so depressing I didn't bother uploading it anywhere

@mcc This seems plausible.

And considering how many people who were traumatized by the Volker Shock went on to hate The Fed (often not understanding at all what The Fed is) for the entire rest of their lives...

@mcc it will cook their brains 🥚
@mcc Even with a bubble pop, I'd at least guess at least one of the orgs would stick around for yet another order of magnitude scaling up. Though that doesn't stop proponents from saying 'just one more, it's around the corner'.

@mcc

Trying to picture the sheer scale of the wasteful boondoggle when a 3,000-acre datacenter is half-built.

@violetmadder only half as wasteful as building the whole thing, tho

@mcc

A functional data center at least could be re-purposed for something not slop.

Mind you, 3,000 acres is just obscene to start with (especially in Arizona) but yeah.

@violetmadder @mcc a very nice roof to put solar on?
I can't think of much else! :(
@violetmadder the thing that worries me is I'm told it's designed specifically for high compute, with most of the investment in cooling systems, and I just don't see a positive use for that much compute.

@mcc @violetmadder They'd be mining bitcoin like there's no tomorrow.

Oh, wait. You said "positive".

Nevermind then.

@jorgecandeias @mcc @violetmadder

I think there'll be a big effort to sell "pre-crime" AI bullshit to governments

@davey_cakes @mcc @violetmadder Yeah, well, I wouldn't put it past those idiots. They do seem intent on searching science fiction for everything it showed not to work or to be catastrophic and going "yeah, let's try that".

@jorgecandeias @mcc @violetmadder

What's amazing is their appeals to the aesthetics of science without ever asking a scientist in the appropriate discipline to validate their ideas.

I've started reading More Everything Forever and it's one of the recurring themes. Solid book so far.

@davey_cakes @jorgecandeias @mcc

Also very evident in most science fiction movies-- somebody had all that big budget for special effects, yet they don't bother talking to a consultant on, say, basic physics.

My father is an engineer, so we watch a big blockbuster like Armageddon and that's not how this works. That's not how ANY of this works.

@mcc @violetmadder and from what I've heard, it's a very specific compute too, not really useful for anything else (not to the degree that would recoup even the operational costs); nobody came up with any even half-reasonable alternative purpose for them in the last years.
All these Nvidia GPUs are eventually going straight to dumpster.

@mcc I've been thinking about this specifically for a while and it's going to be such an annoying thing to see random influencer videos talking about "How close we were to AGI" for the rest of forever.

It'll fade out and come back in waves I think culturally, where people will dismiss it and then Pine for the "better days" as genAI falls largely into obscurity.

We'll have a deadender culture around this for forever I think and we will have to reiterate it/hash it out forever.

@mcc *tiktok influencer voice* Do YOU remember when we ALMOST invented SUPER INTELLIGENCE (AGI)??? 😱🚫🧢
@mcc My town's local news is currently all about the major AI datacenter that's planned nearby, and how much water it uses (the area has been in a water crisis for several decades now). Despite this, there's a ton of hype about it, with all the local companies trying to get a slice of the pie. I'm just wondering if the project's 10-year timeline means that it will fall apart before it even gets off the ground.
@ZahmbieND Best case scenario, you get paid to build it and then it never gets commissioned.
@mcc Eschatology is a helluva drug.
@mcc I like the “phase change” analogy. Agreed, not going to happen
@glasspusher @mcc phase change is a actual thing that happens in neural networks, it's just not gonna turn LLM into AGI

@shironeko @mcc

Ok. I come from the phase change terminology of the physical sciences

@glasspusher @mcc https://doi.org/10.1103/PhysRevResearch.5.043243 yeah it is pretty much used the same way
Zeroth, first, and second-order phase transitions in deep neural networks

We investigate deep-learning-unique first-order and second-order phase transitions, whose phenomenology closely follows that in statistical physics. In particular, we prove that the competition between prediction error and model complexity in the training loss leads to the second-order phase transition for deep linear nets with one hidden layer and the first-order phase transition for nets with more than one hidden layer. We also prove the linear origin theorem, which states that common deep nonlinear models are equivalent to a linear network of the same depth and connection structure close to the origin. Therefore, the proposed theory is directly relevant to understanding the optimization and initialization of neural networks and serves as a minimal model of the ubiquitous collapse phenomenon in deep learning.

Physical Review Research
@shironeko @mcc huh, cool. I learned something today!
@shironeko @glasspusher wait did I use the word correctly then

@mcc @shironeko

see? you're better than you realized!

@mcc yeah we've gamed that branch out a bit, sigh. it's.. sigh. we'll deal with it if it happens.
@mcc a good reminder that once we have the machine, it will still take a few million years until it says 42.
@mcc This finally feels like the future we were promised though.
@mcc Yes, but I assume you and I will route around them, just like we would any other party still carrying a torch for a future that didn't arrive.
@mcc …perhaps, but maybe people will just use the unfinished data centers as roller discos, pet playgrounds, or concert venues! :D
@mcc late to reply, but i think it's worse than this, even - we have dumped far, far too much resource into a dead-end architecture, with no successor in sight, with no real development in our understanding of how neural networks learn, while somehow casting 60+ years of still-relevant research as obsolete. the scaling hypothesis is a death cult
@mcc hmmm.... echos of the dark fiber left in our cities from the crash a few decades ago
@joncruz yeah but fiber is like unused transit capacity, you can always imagine growing into it, whereas it sounds like these "AI data centers" will be not very general purpose and not very upgradeable

@mcc somewhat likely...

... however it does also remind me of when those radio astronomers traded some land for a long abandoned military listening installation. Are you familiar with that?

@joncruz I don't think I know what this is about

@mcc ahhh... a very interesting point in the X-files, Lone Gunmen type conspiracy realm that revealed some truths.

Rough summary: in '99 some group of radio astronomers did a land swap for an outdated, abandoned listening station the government had out in the boonies. I read a long article that interviewed some of the first team in. They found strange light fixtures, carpet in squares welded to tiles, and other things. They were worried about providing power to the location, but found two 1/n

@mcc ... generators each capable of powering a small town. Many other things but it turned out they were all to prevent any stray signals and static electricity.

They had network people come in to replace the outdated system, but when they pulled up the floor they did not find what they were expecting. Instead they found a fiber optic computer network that was cutting edge for the tech of the day. And this was a site that had been abandoned for years. NSA has some secret, cool tech.

2/2