Any system complex enough to be useful has to be embedded in an ever more complex system. The age of mobile phone internet rests on the shoulders of an immense and enormously complex supply chain.

LLMs are capturing low entropy from data online and distilling it for you while producing a shitton of entropy on the backend. All the water and energy dissipated at data centers, all the supply chains involved in building GPUs at the rate we are building. There will be no magical moment when it's gonna yield more low entropy than what we put in on the other side as training data, electricity and clean water.

When companies sell ideas like 'AGI' or 'self driving cars' they are essentially promising you can do away with the complexity surrounding a complex solution. They are promising they can deliver low entropy on a tap without paying for it in increased entropy elsewhere. It's physically impossible.

You want human intelligence to do work, you need to deal with all the complexities of psychology, economics and politics. You want complex machines to do autonomous work, you need an army of people behind it. What AGI promises is, you can replace the army of people with another more complex machine. It's a big bald faced lie. You can't do away with the complexity. Someone will have to handle it.

Your brain is proof to the contrary. AGI means different things to everyone, but a human brain definitely counts as "general intelligence", that implemented in silicon is enough to get basically all the things promised by AGI: if that's done at the 20 watts per brain that biology manages, then all of humanity can be simulated within the power envelope of the USA electrical grid… three times over.

I followed up the point about AGI meaning different things by giving a common and sufficient standard of reference.

Your brain is evidence that it's "even possible".

All your brain proves is that a universe can produce planetary ecosystems capable of supporting human civilizations made of very efficient brain carrying mammals.

It definitely doesn't prove that these mammals can create boxes capable 'solve physics, poverty and global warming' if we just give Sam Altman enough electricity and chips. Or dollars to that effect.

What's the quote? "If the human brain were so simple that we could understand it, we would be so simple that we couldn’t".

Even though it doesn't need to be a single human doing all of it, our brains as existence proofs of the physical possibility, not of our own understanding.

Companies selling ways around that apparatus to achieve intelligent behavior are selling you perpetual motion machines, or slavery. There has been zero 'AI' systems that don't depend on painstakingly collected and analyzed data beforehand.

So far the only thing that has been proven is we can get low entropy from all the low entropy we've published on the internet. Will it get to a point where models can give us more low entropy than what is present in the training data? Categorically: no.

Whatever you mean, our brains prove it's possible to have a system that uses 20 watts to demonstrate human-level intelligence.

You're positing a way to create human intelligence-like in a bottle, that's the same as speculating about the shape of a reality where we have FTL travel or teleportation or whatever else you fancy.

If we're talking about what current ML/AI can do, they can extract patterns from training data and than apply those patterns to other inputs. This can give us great automation, but it won't give us anything better than the training data, solve physics, global warming, poverty or give us human intelligence in a chip.

Whatever quantity Q of entropy in the training data, the total output will be more Q all accounted for. That's true for humans and machines. No shape of possible AGI will give us any output with less Q than the combination of inputs had. The dream that a machine will solve all problems that humanity can't hinges on negating that, which goes against thermodynamics.

As it stands, the planet cannot cope with all the entropy we're spreading around. It will eventually collapse civilization/the ecosystem, whatever buckles first, from the excess entropy. Because global warming, poverty, or ignorance is just entropy. Disorder. Things not being just so as we need them to be.

> Whatever quantity Q of entropy in the training data, the total output will be more Q all accounted for. That's true for humans and machines. No shape of possible AGI will give us any output with less Q than the combination of inputs had. The dream that a machine will solve all problems that humanity can't hinges on negating that, which goes against thermodynamics.

Thermodynamics applies to closed systems, which the earth in isolation isn't.

This is why the source material for Shakespeare didn't already exist on Earth in the Cretaceous.

It's also why we've solved loads of problems we used to have, like bubonic plague and long distance communication.

> As it stands, the planet cannot cope with all the entropy we're spreading around.

We're many orders of magnitude away from entropic limits. Global warming is due to the impact on how much more sunlight keeps bouncing around inside the atmosphere, not the direct thermal effect of our power plants.

> It will eventually collapse civilization/the ecosystem, whatever buckles first, from the excess entropy. Because global warming, poverty, or ignorance is just entropy. Disorder. Things not being just so as we need them to be.

Entropy isn't everyday disorder, it's a specific relationship of microstates and macrostates, and you can't usefully infer things when you switch uses.

Poverty is much reduced compared to the historical condition. So is ignorance: we have to specialise these days because there's too much knowledge for any one human to learn.

Entropic collapse is indeed inevitable, but that inevitably is on the scale of 10^18 years or more with only engineering challenges rather than novel scientific breakthroughs (the latter would plausibly increase that to 10^106 if someone can figure out how to use Hawking radiation, but I don't want to divert into why I think that's more than merely an engineering challenge).

Wrong. You're probably conflating the second law of thermodynamics with the whole thing. Obviously, thermodynamics applies to Earth, it was invented here.

> Entropy isn't everyday disorder, it's a specific relationship of microstates and macrostates, and you can't usefully infer things when you switch uses.

No, this is Bolzmann entropy. Entropy is absolutely a measure of disorder, and the fact that you cannot measure a specific value for large complex systems because the calculating framework is too simplistic and only applies to very well defined and artificial assemblies of quantities it does not mean that the concept and intuition do not apply.

Entropy is a measure of disorder or uncertainty. Maybe you disagree with my intuition about it.

But I'm not the only one:

- Shannon Entropy Fits Well Social Risk: Unemployment and Poverty in Large Cities https://ieeexplore.ieee.org/document/8905455

- The Interrelation Poverty-Unemployment from the Theory of Entropy in Modern Societies https://alicia.concytec.gob.pe/vufind/Record/AUTO_bbe82cfd37...

It's grossly unreasonable to include the entire history of the universe given the Earth only formed about 4 billion years ago; and given that evolution wasn't even aiming for intelligence, even starting from our common ancestors being small rodents 65 million years ago is wildly overstating the effort required — even the evolution of primates is too far back without intentional selective breeding.

> You're positing a way to create human intelligence-like in a bottle, that's the same as speculating about the shape of a reality where we have FTL travel or teleportation or whatever else you fancy.

FTL may well well be impossible.

If you seriously think human intelligence is impossible, then you also think you don't exist: You, yourself, are a human like intelligence in a bottle. The bottle being your skull.

> This can give us great automation, but it won't solve physics, global warming, poverty or give us human intelligence in a chip.

AI has already been in widesprad use in physics for a while now, well before the current zeitgeist of LLMs.

There's a Y Combinator startup: "Charge Robotics is building robots that automate the most labor-intensive parts of solar construction".

Poverty has many causes, some of which are already being reduced or resolved by existing systems — and that's been the case since one of the ancestor companies of IBM, Tabulating Machine Company, was doing punched cards for the US census.

As for human intelligence on a chip? Well, (1) it's been quite a long time since humans were capable of designing the circuits manually, given the feature size is now at the level where quantum mechanics must be accounted for or you get surprise tunnelling between gates; and (2) one of the things being automated is feature segmentation and labelling of neural micrographs, i.e. literally brain scanning: https://pubmed.ncbi.nlm.nih.gov/32619485/

All I am saying is that Sam Altman's promises (remember the original topic) hinge on breaking thermodynamics.

Humans evolved on a planet that was just so. The elements on Earth that make life possible couldn't have been created in a younger universe. So it take the full history of the universe to produce human intelligence. It also doesn't exist in a vacuum. Current human civilization is not a collection of 8 billion 20 watt boxes.

> As for human intelligence on a chip? Well, (1) it's been quite a long time since humans were capable of designing the circuits manually, given the feature size is now at the level where quantum mechanics must be accounted for or you get surprise tunnelling between gates; and (2) one of the things being automated is feature segmentation and labelling of neural micrographs, i.e. literally brain scanning: https://pubmed.ncbi.nlm.nih.gov/32619485/

I don't understand any of this so I won't comment.

If it hinged on that, then you would actually be saying it's impossible.

> The elements on Earth that make life possible couldn't have been created in a younger universe

Those statements also apply to the silicon and doping agents used in chip manufacture. They tell you nothing of relevance, we're not doing Carl Sagan's Apple Pie from scratch with AI, we're trying to get a thing like us.

"impossible with the approach OpenAI is using" != "impossible"

Eventually you need a whole universe capable of sustaining self reproducing intelligent systems. The way to get intelligent autonomous and sufficiently complex systems to do anything non trivial is life.

Anyone trying to sell you a box they say is capable of non trivial, adaptive, complex behavior autonomously is selling you a perpetual motion machine: a pretense that thermodynamics can be suspended at will.

AGI is either human intelligence behind the curtains or a pie in the sky.

We have them in San Francisco now (and Los Angeles and Phoenix, and Austin soon.)

Waymo's overstated[1] success has let self-driving advocates do an especially pernicious bit of goalpost-shifting. I have been a self-driving skeptic since 2010, but if you had told me in 2010 that in 10-15 years we have robotaxis that were closely overseen by remote operators who can fill in the gaps I would have thought that was much more plausible than fully autonomous vehicles. And the human operators are truly critical, even more so than a skeptic like me assumed: https://www.nytimes.com/interactive/2024/09/03/technology/zo... (sadly the interactive is necessary here and archives don't work, this is a gift link)

I still think fully autonomous vehicles on standard roads is 50+ years out. The argument was always that ~95% of driving is addressable by deep learning but the remaining ~5% involves difficult problem-solving that cannot be solved by data because the data does not exist. It will require human oversight or an AI architecture which is capable of deterministic reasoning (not transformers), say at least at the level of a lizard. Since we have no clue how to make an AI as smart as a lizard, that 5% problem remains utterly intractable.

[1] I have complained for years that Waymo's statisticians are comparing their cars to all human drivers when they should be comparing it to lawful human drivers whose vehicles are well-maintained. Tesla FSD proves that self-driving companies will respond to consumer demand for vehicles that speed and run red lights.

I would be shocked if we're really 50 years away from that level of AI. 50 years is a long time in computing — late 70s computers were still using punched tape:

https://commons.m.wikimedia.org/wiki/File:NSA_Punch_Verifica...

You have three reasons:

1) reading the comment in good faith

2) understanding 'robotaxi' is not a precise technical term

3) safely assuming that most commenters here know about Waymo

There is no reason to choose the most pedantic and smarmily bad-faith reading of the comment.

As for "50 years" - I don't care about electrical engineering, I am talking about intelligence. In the 1970s we had neural networks as smart as nematodes. Today they are as smart as spiders. Maybe in 50 years they will be as smart as bees. I doubt any of our children will live to see a computer as smart as a rat.

> 1) reading the comment in good faith

> 2) understanding 'robotaxi' is not a precise technical term

> 3) safely assuming that most commenters here know about Waymo

#1 is the main reason why I wouldn't read "robotaxi" as anything other than "taxi robot", closely followed by #2.

> As for "50 years" - I don't care about electrical engineering, I am talking about intelligence.

Neither was I, and you should take #1 as advice for yourself.

> In the 1970s we had neural networks as smart as nematodes. Today they are as smart as spiders. Maybe in 50 years they will be as smart as bees. I doubt any of our children will live to see a computer as smart as a rat.

You're either overestimating the ones in the 70s or underestimating the ones today. By parameter count, GPT-3 is already about as complex as a medium sized rodent. If today's models aren't that smart (definitions of "intelligence" are surprisingly fluid from one person to the next), then you can't reasonably call the ones in the 70s as smart as a nematode either.

Really? Waymo's statisticians are the ones you are complaining about?

Tesla's statisticians have been lying for years, as has Musk when they cite "number of miles driven by FSD in the very small subset of conditions where it is available, and not turned off or unavailable because of where you are, the weather, or any other variable" versus "all drivers, all conditions, all locations, all times" to try to say FSD is safer.

You can walk to where they're waiting for you.

Surely this is just a case of the future not being evenly distributed. All of these 'problems' are already solved and the solution is implemented somewhere, just not where you happen to be.