My understanding is that LLMs are built on a foundation of ANNs, and that indeed the backpropagation used to train ANNs is a statistical process;

Two responses. One, that's discussing individual-neuron scale processes rather than collective processes; and this was a discussion about inference, not training. Human neurons also learn by error minimization (Hebbian learning). But this does not describe the macroscopic processes that result from said minimization.

* During training, neurons develop into classifiers that detect superpositions of concepts that collectively follow the same activation process. Individual neurons weight their input space and subdivide it by a fuzzy hyperplane to achieve a classification result.

* In subsequent layers, said input space is formed from a weighted combination of the previous layer's classification; thus, the superpositions of questions being formed are more complex, as are the classification results.

* In a LLM, this iterates for dozens of layers, gaining complexity at each layer, to form each FFN

* The initial input space to a FFN is a latent (conceptual representation), as is the output; the FFNs, in result, function as classifier-generators; they detect combinations of concepts in the input space, and output the causally-resultant concepts into the output space

* FFNs alternate with attention layers dozens to hundreds of times in order to process the information, each layer building on the results of the previous one.

The word to describe that is not "statistics". It's "logic".

In a LLM, the first few layers focus on disambiguation. If there's a token for "bank", is this about a riverbank, a financial bank, banking a plane, etc? As the layers progress, it starts building up first simple circuits, and then progressively more complex circuits - you might get a circuit that detects "talking like MAGA", or "off-by-one programming errors", or whatnot. In the late layers, you have the general conclusions reached - for example, if it were "The capitol of the state that contains America's fourth-largest metro area is...", you've already had FFNs detect the concepts of fourth-largest metro area and encoded Dallas-Forth Worth, and then later taken that and encoded "Texas", and then finally encoding "Austin". And then in the final couple layers you converge back toward linguistic space.

Anthropic has done some great work on this with attribution graph probes and the like; you can detect what circuits are firing, and on what things those circuits fire, and ramp them up or down to see how it modifies the output. They very much work through long chains of logical inferences.

I use every style imaginable, including photos, in my tests. Same result every time.

One time I even did it with a Calvin and Hobbes comic, pretending than an AI made it. Responses included things like "The illustration also looks like shit and barely makes sense. Hope that helps.", "God damn this sucks so bad", "This also fucking sucks", and "The only punchline here is casual, pointless cruelty. if you think this is funny then you're literally a psychopath."

LLMs are not "statistical models" (randomness only even comes into play in the final conversion from latent space to token space because latent space is high dimensional, token space is low dimension, you need a rounding mechanism, and a "noisy" rounding mechanism works best; what you're thinking of, by contrast, is Markov models). And you cannot just "get lucky and randomly solve an unsolved math problem"; that's not how any of this works.

Also, it's silly that people are acting like "all problems but this one were already in the literature". AI has solved a whole slew on Erdos problems, and only a fraction had anything to do with existing literature.

And even in "existing literature" examples, it's not "nobody ever thought to search before" as if all mathematicians are morons, or that mathematicians adore putting out Erdos problem solutions without claiming them, It's that nobody had ever thought to apply an obscure technique from a given piece of literature to said Erdos problem.

The simple fact is, AI has gotten much better at solving unsolved math problems than humans are. It's simply another field that it's taking over, the same way it has been taking over programming. One can debate how much is "clever insight" vs. "just chugging away at possibilities until it hits on ways to advance toward the goal", but ultimately, that's a distraction from the fact that: it's getting really good at solving math problems that humans have spent decades on without success.

Clearly that wasn't spam, because his name is nospam007.

Forced? You think reviewers of writing contests are captives, imprisoned by mad scientists and forced to read story after story, like some sort of literary equivalent of the backstory to MST3K?

It is no more "theft" than you are.

I'll never get over how many people I watch online complain about how they'll never use AI because "it's theft", and then post photoshops they made with pictures they don't own, when that's not what the AI is doing.

I'll never get over how many artists I've seen complain about how AI is theft, and then paint something with "inspiration images" sitting in front of them while they paint, with their painting effectively being a blended composite of their inspiration images - when that's not what the AI is doing.

I'll never get over how many writers I've seen write the exact same derivative stuff that they also read, down to the same phrasings at times, just packaged in a new plot with new characters, and yeah, same story.

Even a person who isn't *directly* copying things that they're literally looking at is still the sum of their experiences. And it's rather hard to say that the breadth of human experience is broader than an LLM (whose "breadth of experiences" is "the whole world's outputs"), outside of the things that relate directly to having a body in a way which a blind / deaf / quadraplegic / whatnot person wouldn't grasp well.

And on that latter note, most people underestimate how well e.g. the congenitally blind actually grasp colours and the like. They're far better at reasoning about colours - similar to the sighted - than they are at knowing what colours things are. One study I read for example asked about polar bear fur. A good fraction of the congenitally blind subjects answered that they didn't know what colour it was. When asked to guess, about half of them answered that it was white so that it could blend into the snow, while the other half wrongly guessed black, but did so on the assumption that they'd want to soak up light to help stay warm. And actually in reality, both are true - to an outside observer, the exterior scattering of visible light without pigment makes them look white, but it's also designed to trap non-visible light up against black skin to absorb it for warmth. A sighted person, just seeing "white fur" and not knowing the latter property, might not have thought to even consider that.

To a LLM, our bodily experiences are akin to a blind person asked about colour: only knowing for sure things that they've learned about them directly, but still quite adept at reasoning about them.

You understand that this was not a sales contest, right?

What should be telling is that people started criticizing the metaphors only after concerns were raised that it was AI.

I do this experiment all the time (as it amuses me endlessly): when some people are hating on AI, I post some masterpiece panting or award-winning photo, suggest it's AI, and then watch them all explain why it's terrible, obviously-flawed, soulless crap. Sometimes I also do the reverse and post an AI image by comparison, say I prefer the "AI image" (actually real) to the "real image" (actually AI), and watch them go on about how the "real" (actually AI) image is so much better and shows so much more human creativity and emotion that AI could never have.

One person did this recently (posted a real Monet, said it was AI) and hundreds of such replies. The funniest part however was that someone took all their replies about the real Monet and fed them (along with the real Monet) into an AI model and had it improve the real Monet to match their criticisms ;)

The problem here isn't scams and grift. It's that by growing the number of bets by orders of magnitude, you grow the number of insiders with insider info by orders of magnitude. Which creates perverse incentives across your entire economy for people to profit off of their status, while simultaneously making it much harder to do so.

Say there's a bet on whether Russia will take the village of Mala Tokmachka by a certain date. Well, now every Russian commander on that front has an incentive to have family members / friends bet against it and then delay / undercut their ability to wage incentives, while every soldier also has the incentive to do so and then do everything they can to spoil the offensive.

Or say the US is planning a major offensive in Iran, and various insiders are betting *on* an imminent attack. Well, Iran can feed Polymarket data into their intel assessments as one factor to help determine when the US will invade in order to maximize their readiness. OR, Iranian intel services could track down US officials / military officers who are making these bets based on insider info and blackmail them in order to get them to cooperate.

Every single aspect of society that gets bet on gets subtly undercut by the existence of these bets. It's one thing to have bets be placed on sports, an action that by definition is entertainment, something that doesn't actually matter. But it's an entirely different thing to have bets on virtually everything in our world. It's already been a problem with stock trading, but sites like Polymarket make it so, so much worse.

JUDGE: The jury has sent a question and the answer is no, the death penalty is not "available for both sides" please return to the jury room and limit your consideration to civil damages.

JUDGE: No, a “light maiming” is also not acceptable, nor is “getting medieval on their asses.” Please constrain yourself to statutes approved by this court.

JUDGE: A further follow-up question from the jury, and no we cannot 'dunk them in a lake and let God decide, like they used to do with witches'. That has not been considered a valid means of determining guilt for several centuries at least.

JUDGE: The jury has sent another question and the answer, again, is no. "Excommunicado" is not real - that's only a thing in the John Wick universe. Civil penalties DO NOT encompass revoking all protections under the law for Mr Altman and Mr Musk.

JUDGE: Court reporter, please note that the jury's latest request, quote, can we let them hang by their thumbs for a few hours, end quote, is also denied.

Nuclear reactors use most surface water, not ground water.

Datacentres are no pickier. You can even cool a datacentre with saltwater, you just need a heat exchanger.

Also, closed loop does not evaporate. The loop is not closed if stuff escapes from it.

You're arguing with the actual terminology used in the nuclear industry. "Closed loop" or "closed cycle" designs have the water pumped in a cycle through cooling towers. The towers lose water to evaporation, taking heat with them, but the rest of the water is returned to be reheated again. "Open loop" or "open cycle" designs have no cooling towers. The water is heated and just discharged hot. They consume much more water (over an order of magnitude more), but most of that is returned. Closed loop are more common, but you see open loop in some older designs, and in seawater-cooled reactors.

"How often do you think I print?"

Seemingly not very.

I've printed many hundreds of kg on my P1S, thanks.

I do not consider having to write data out to a card and transport it back and forth between the printer and the computer to be the pinnacle of convenience. That's something that would be considered embarrassingly inconvenient for a 1980s printer, let alone a modern net-connected device. And it's designed to be inconvenient for non-cloud prints for a reason.

Also, anything sounds big when you put it in gallons. Doesn't sound so big when you mention that's 92 acre feet, the amount used by less than 20 acres / 8 hectares of alfalfa per year. Or when you mention that a typical *closed loop* 1GW nuclear reactor uses 6-20 billion gallons of cooling water per year (once-through uses 200-500 billion gallons, though most of that is returned, whereas closed loop evaporates it)