[observationist]

Let an ultraintelligent machine be defined as a machine that can far surpass all the intellectual activities of any man however clever. Since the design of machines is one of these intellectual activities, an ultraintelligent machine could design even better machines; there would then unquestionably be an 'intelligence explosion,' and the intelligence of man would be left far behind. Thus the first ultraintelligent machine is the last invention that man need ever make.

-Irving John Good, 1965

If you want a short, easy way to know what AGI means, it's this: Anything we can do, they can do better. They can do anything better than us.

If we screw it up, everyone dies. Yudkowsky et al are silly, it's not a certain thing, and there's no stopping it at this point, so we should push for and support people and groups who are planning and modeling and preparing for the future in a legitimate way.

[visarga]

John Good's quote is pretty myopic, it assumes machines make better machines based on being "ultraintelligent" instead of learning from environment-action-outcome loop.

It's the difference between "compute is all you need" and "compute+explorative feedback" is all you need. As if science and engineering comes from genius brains not from careful experiments.

[observationist]

There's an implicit assumption there, anything a computer as intelligent as a human does will be exactly what a human would do, only faster. Or more intelligent. If the process is part of the intelligent way of doing things, like the scientific method and careful experimentation, then that's what the ultraintelligent machine will do.

There's no implication that it's going to do it all magically in its head from first principles; it's become very clear in AI that embodiment and interaction with the real world is necessary. It might be practical for a world model at sufficient levels of compute to simulate engineering processes at a sufficient level of resolution that they can do all sorts of first principles simulated physical development and problem solving "in their head", but for the most part, real ultraintelligent development will happen with real world iterations, robots, and research labs doing physical things. They'll just be far more efficient and fast than us meatsacks.

[ACCount37]

At sufficient levels of intelligence, one can increasingly substitute it for the other things.

Intelligence can be the difference between having to build 20 prototypes and building one that works first try, or having to run a series of 50 experiments and nailing it down with 5.

The upper limit of human intelligence doesn't go high enough for something like "a man has designed an entire 5th gen fighter jet in his mind and then made it first try" to be possible. The limits of AI might go higher than that.

[kilpikaarna]

Exceedingly elaborate, internally-consistent mind constructs, untested against the real world, sounds like a good definition of schizophrenia. May or may not correlate with high intelligence.

[ACCount37]

We only call it "schizophrenia" when those constructs are utterly useless.

They don't have to be. When they aren't, sometimes we call it "mathematics".

You only have to "test against the real world" if you don't already know the outcome in advance. And you often don't. But you could have. You could have, with the right knowledge and methods, tested the entire thing internally and learned the real world outcome in advance, to an acceptable degree of precision.

We have the knowledge to build CFD models already. The same knowledge could be used to construct a CFD model in your own mind. We have a lot of scattered knowledge that could be used to make extremely elaborate and accurate internal world models to develop things in - if only, you know, your mind was capable of supporting such a thing. And it isn't! Skill issue?

[econ]

I like the substitution concept. What humans can do depends on the abstractions and the tools. One could picture just the shape of the jet and have a few ideas how to improve it further. If that is enough info for the tool it could be worthy of the label "designed by Jim".

[circlefavshape]

> As if science and engineering comes from genius brains not from careful experiments

100% this. How long were humans around before the industrial revolution? Quite a while

[snikeris]

Science and engineering didn't begin with the Industrial Revolution. See: https://en.wikipedia.org/wiki/Great_Pyramid_of_Giza

[tjoff]

Have you gotten any indication that machines won't have sensors?!

[gopher_space]

From what I can see we're working as hard as we can to build them. You can watch the "let's put this on a Raspberry Pi and see what happens" seeds of Skynet develop in real time.

There's something compelling about helping assemble the machine. Science fiction was completely wrong about motivation. It's fun.

[Eldt]

Maybe ultraintelligence is having an improved environment-action-outcome loop. Maybe that's all intelligence really is

[goodmythical]

I've noticed this core philosophical difference in certain geographically associated peoples.

There is a group of people who think AI is going to ruin the world because they think they themselves (or their superiors) would ruin the world.

There is a group of people who think AI is going to save the world because they think they themselves (or their superiors) would save the world.

Kind of funny to me that the former is typically democratic (those who are supposed to decide their own futures are afraid of the future they've chosen) while the other is often "less free" and are unafraid of the future that's been chosen for them.

[mitthrowaway2]

There is also a group of people who think AI is going to ruin the world because they don't think the AI will end up doing what its creators (or their superiors) would want it to do.

[tines]

You’re just describing authoritarian vs non-authoritarian mindsets.

[inigyou]

In that case, it can't be improved with bigger computers.

[santadays]

Intelligence seems to boil down to an approximation of reality. The only scientific output is prediction. If we want to know what happens next just wait. If we want to predict what will happen next we build a model. Models only model a subset of reality and therefore can only predict a subset of what will happen. Llms are useful because they are trained to predict human knowledge, token by token.

Intelligence has to have a fitness function, predicting best action for optimal outcome.

Unless we let AI come up with its own goal and let it bash its head against reality to achieve that goal then I’m not sure we’ll ever get to a place where we have an intelligence explosion. Even then the only goal we could give that’s general enough for it to require increasing amounts of intelligence is survival.

But there is something going on right now and I believe it’s an efficiency explosion. Where everything you want to know if right at hand and if it’s not fuguring out how to make it right at hand is getting easier and easier.

[whodidntante]

With AI, as we currently understand it, we may have stumbled upon being able to replicate a part of the layer of our brain that provides the "reason" in humans., and a very specific type of "reason" a that.

All life has intelligence. Anyone who has spent a lot of time with animals, especially a lot of time with a specific animal, knows that they have a sense of self, that they are intelligent, that they have unique personalities, that they enjoy being alive, that they form bonds, that they have desires and wants, that they can be happy, excited, scared, sad. They can react with anger, surprise, gentleness, compassion. They are conscious, like us.

Humans seem to have this extra layer that I will loosely call "reasoning", which has given us an advantage over all other species, and has given some of us an advantage over the majority of the rest of us.

It is truly a scary thing that AI has only this "reasoning", and none of the other characteristics that all animals have.

Kurt Vonnegut's Galapagos and Peter Watts Blindsight have different, but very interesting takes on this concept. One postulates that our reasoning, our "big brains" is going to be our downfall, while the other postulates that reasoning is what will drive evolution and that everything else just causes inefficiencies and will cause our downfall.

[lazystar]

i think theres a paradox here. intelligence needs a judge - if nothing verifies that the optimal outcome was chosen, it's too easy for the intelligence to fall into biased decisions

[scroot]

It's the "no stopping it at this point" that always sticks out to me in these discussions. Why is there no stopping it, exactly? At this juncture these systems require massive physical infrastructure and loads of energy. It's possible to shut it all down. What's lacking is the political will.

[mc32]

Should then the powers that are developing AGI enter an analogue to the SALT treaties but this time governing AGI do things don’t go off the rails?

[SecretDreams]

> support people and groups who are planning and modeling and preparing for the future in a legitimate way.

Who is doing that right now, exactly? And how can we take their tech and turn it into the next profitable phone app?

[dylan604]

The "legitimate way" is nothing short of weasel words. Who defines what is legitimate. The doomers that are prepping for the future by building stockpiles of food/water/weapons being stored in bunkers/shelters they have built would say this is exactly what they are doing. Yet, these people are often panned as being a little unhinged. If we're having a conversation about tech destroying humanity, then planning a way to survive without tech seems like a legitimate concept.

[mathgradthrow]

never let philosophers do math

[mrandish]

> Let an ultraintelligent machine be defined as a machine that can far surpass all the intellectual activities of any man

The things this definition misses: First, 'intelligence' is a poorly defined and overly broad term. Second, machine intelligence is profoundly different than biological intelligence. Third, “surpassing humans” is not a single threshold event because machine and human intelligence are not only shaped differently, they're highly non-linear. LLMs are a particular class of possible machine intelligences which can be much more intelligent than humans on some dimensions and much less intelligent on others. Some of the gaps can be solved by scaling and brilliant engineering but others are fundamental to the nature of LLMs.

> an ultraintelligent machine could design even better machines

There is a huge leap between "surpass all the intellectual activities of any man" and "invent extraordinary breakthroughs and then reliably repeat that feat in a sequential, directed fashion in the exact way required to enable sustained iteration of substantial self-improvement across infinite generations in a runaway positive feedback loop". That's an ability no human or collective has ever come close to demonstrating even once, much less repeatedly. (hint: the hardest parts are "reliably repeat", "extraordinary breakthroughs" and "directed fashion"). A key, yet monumental, subtlety is that the self- improvements must not only be sustained and substantial but also exponentially amplify the self-improvement function itself by discovering novel breakthroughs which build coherently on one other - over and over and over.

The key unknown of the 'Foom Hypothesis' is categorical. What kind of 'difficult feat' this is? There are difficult feats humans haven't demonstrated like nuclear fusion, but in that example we at least have evidence from stellar fusion that it's possible. Then there are difficult feats like room-temp superconductors, which are not known to be possible but aren't ruled out. The 'Foom Hypothesis' is a third category of 'hard' which is conceptually coherent but could be physically blocked by asymptotic barriers, like faster-than-light travel under relativity.

Assuming Foom is like fusion - just a challenging engineering and scaling problem - is a category error. In reality, Foom requires superlinear, recursively amplifying cognitive returns—and we have no empirical evidence that such returns can exist for artificial or biological intelligences. The only prior we have for open‑ended intelligence improvement is biological evolution which shows extremely slow and unreliable sublinear returns at best. And even if unbounded self‑improvement is physically possible, it may be practically unachievable due to asymptotic barriers in the same way approaching light speed requires exponentially more energy.

[LeifCarrotson]

"There's no stopping it at this point" - Sure there is, if a handful of enormous datacenters pull the very large plugs (or if their shaky finances collapse), the dubiously intelligent machines will be turned off. They're not ultraintelligent yet.

Stopping it merely requires convincing a relatively small number of people to act morally rather than greedily. Maybe you think that's impossible because those particular people are sociopathic narcissists who control all the major platforms where a movement like this would typically be organized and where most people form their opinions, but we're not yet fighting the Matrix or the Terminator or grey goo, we're fighting a handful of billionaires.

[observationist]

I'm not saying it's technically impossible, I'm saying that in the real world, it's not going to stop. Nobody is going to stop it. A significant number of people don't want it to stop. A minority of people are in the "stop AI" camp, and the ones with the money and power are on the other side.

It's an arms race replete with tribalism and the quest for power and taps into everything primal at the root of human behavior. There's no stopping it, and thinking that outcome can happen is foolish; you shouldn't base any plans or hopes for the future on the condition that the whole world decides AGI isn't going to happen and chooses another course. Humans don't operate that way, that would create an instant winner-takes-all arms race, whereas at least with the current scenario, you end up with a multipolar rough level of equivalence year over year.

[hollerith]

The whole world decided in the 1970s not to pursue the technology of germ-line genetic engineering of humans, and that decision has stood.

People similar to you were saying in the 1950s and later that it was inevitable that nuclear weapons would be used in anger in massive attacks.

Although the people in charge are tentatively for AI "progress", if that ever changes, they can and will put a stop to large AI training runs and make it illegal for anyone they don't trust to teach, learn or publish about fundamental algorithmic "improvements" to AI. Individuals and groups pursuing "improvements" will not be able to accept grant money or investment money or generate revenue from AI-based services.

That won't stop all research on such improvements (because some AI researchers are very committed), but it will slow it down to a rate much much slower than the current rate (because the current fast rate depends of rapid communication between researchers who don't each other well, and if communicating about the research were to become illegal, then a researcher can communicate only with those researchers he knows won't rat him out) essentially stopping AI "progress" unless (unluckily for the human species) at the time of the ban, the committed researchers were only one small step away from some massive algorithmic improvement that can be operationalized using the compute resources at their disposal (i.e., much less than the resources they have now).

Will the power elite's attitude towards AI change? I don't know, but if they ever come to have an accurate understanding of the situation, they will recognize that AI "progress" is a potent danger to them personally, and they will shut it down.

It's not a situation like the industrial revolution in England in which texile workers were massively adversely affected (or believed they were) but the people running England were mostly insulated from any adverse effects. In the current situation, the power elite is definitely not insulated from severe adverse consequences if an AI lab creates an AI that is much more competent that the most competent human institutions (e.g., the FBI) and the lab fails to keep the AI under control. And it will fail if it were to use anything like the methods and bodies of knowledge AI labs have been using up to now. And there are very bright people with funding doing their best to explain that to the elite.

Those of you who want AI "progress" to continue until the world is completely transformed need to hope that the power elite are collectively too stupid to recognize a potent short-term threat to their own survival (or the transformation can be completed before the power elite wake up and react). And in my estimation, that is not inevitable.

[LeifCarrotson]

I actually got to talk to an individual at a coffee shop just last night about this issue, he's a representative of our local PauseAI chapter:

https://pauseai.info/proposal

They think that a similar solution to the genetic modification treaties and nuclear proliferation treaties can be implemented for AI training.

[hollerith]

That plus something similar to the prohibition on publishing knowledge about the design of nuclear weapons, but for machine-learning algorithms. (Inb4 "but that's ridiculous, you can't ban linear algebra". Watch us!)

[goodmythical]

right, because turning off any number of data centers is going to do anything at all but create massive pressure on researching the efficiency and effectiveness of the models.

There are already designs that do not require massive data centers (or even a particularly good smart phone) to outperform average humans in average tasks.

All you'd accomplish by hobbling the data centers is slow the growth of sloppy models that do vastly more compute than is actually required and encourage the growth of models that travel rather directly from problem to solution.

And, now that I'm typing about it, consider this: The largest computational projects ever in the history of the world did not occur in 1/2/5/10 data centers. Modern projects occur across a vast and growing number of smaller data centers. Shit, a large portion of Netflix and Youtube edge clusters are just a rack or a few racks installed in a pre-existing infrastructure.

I know that the current design of AI focusses on raw time to token and time to response, but consider an AGI that doesn't need to think quickly because it's everywhere all at once. Scrappy botnets often clobber large sophisticated networks. WHy couldn't that be true of a distributed AI especially now that we know that larger models can train cheaper models? A single central model on a few racks could discover truths and roll out intelligence updates to it's end nodes that do the raw processing. This is actually even more realistic for a dystopia. Even the single evil AI in the one data center is going to develop viral infection to control resources that it would not typically have access to and thereby increase it's power beyond it's own existing original physical infrastructure.

quick edit to add: At it's peak Folding@Home was utilizing 2.4 EXAflops worth of silicon. At that moment that one single distributed computational project had more compute than easily the top 100 data centers at the time. Let that sink in: The first exa-scale compute was achieved with smartphones, PS3s, and clunky old HP laptops; not a "hyperscaler"

[ben_w]

> quick edit to add: At it's peak Folding@Home was utilizing 2.4 EXAflops worth of silicon. At that moment that one single distributed computational project had more compute than easily the top 100 data centers at the time. Let that sink in: The first exa-scale compute was achieved with smartphones, PS3s, and clunky old HP laptops; not a "hyperscaler"

A DGX B200 has a power draw of 14.3 kW and will do 72-144 petaFLOP of AI workload depending on how many bits of accuracy is asked for; this is 5-10 petaFLOP/kW: https://www.nvidia.com/en-us/data-center/dgx-b200/

Data centres are now getting measured in gigawatts. Some of that's cooling and so on. I don't know the exact percent, so let's say 50% of that is compute. It doesn't matter much.

That means 1GW of DC -> 500 MW of compute -> 5e5 kW -> 5e5 * [5-10] PFLOP/s -> 2500 - 5000 exaFLOP/s.

I'm not sure how many B200s have been sold to date?

[trvz]

Open models barely any worse than SOTA exist, and so does consumer-ish hardware able to run them. The genie’s out, the bottle broken.

[slibhb]

Do you really think AI companies/researchers are motivated by greed? It doesn't seem that way to me at all.

Stopping AI would be immoral; it has the potential to supercharge technology and productivity, which would massively benefit humanity. Yes there are risks, which have to be managed.

[jobs_throwaway]

AI researchers are not a monolith. I definitely think that many of them are motivated by greed. Many are also true believers that AI will improve the human condition.

I fall in the latter camp, but I think its a bit naive to claim that there is not a sizable contingent who are in AI solely to become rich and powerful.

[ben_w]

> has the potential to supercharge technology and productivity, which would massively benefit humanity

The opportunities you chose to list are the greedy ones.

> Yes there are risks, which have to be managed.

How?

As a reminder, we've known about the effect of burning coal on the climate for well over a century, we knew that said climate change would be socially and economically disasterous for half a century, yet the only real progress we're making is because green became cheaper in the short term not just the long term and the man in charge of the USA is still calling climate change and green energy a hoax.

Right now, keeping LLMs aligned with us is easy mode: they're relatively stupid, we can inspect the activations while they run, we can read the transcripts of their "thoughts" when they use that mode… and yet Grok called itself Mecha Hitler, which the US government followed up by getting it integrated into their systems, helping the Pentagon with [classified] and the department of health to advise the general public which vegetables are best inserted rectally.

We are idiots speed-running into something shiny that we don't understand. If we are very very lucky, the shiny thing will not be the headlamp of a fast approaching train.

[slibhb]

> The opportunities you chose to list are the greedy ones.

Technology covers healthcare. I don't see how it's "greedy" to want to cure cancer. But on some level I guess "wanting life to be better" is greedy.

Your attitude is very European, and it's basically why your continent is being left behind. I'm not totally against Europe becoming the world's retirement home, as long as there are places in the world where people are allowed to innovate.

[ben_w]

> Technology covers healthcare.

If you'd chosen to list that in the first place, I wouldn't have said what I did; "supercharge technology and productivity" is looking at everything through the lens of money and profit, not the lens of improving the human condition.

> Your attitude is very European, and it's basically why your continent is being left behind

And yours is very American. You talk about managing the risks, but the moment you see anyone doing so, you're against it.

And of course, Europe does have AI, both because keeping up is so much easier and cheaper than being bleeding edge on everything all the time, and of course, how DeepMind may be owned by Google but is a British thing.

Plus: https://mistral.ai

Also, to be blunt, China's almost certain to win any economic or literal arms race you think you're part of; they make too much critical hardware now.

> as long as there are places in the world where people are allowed to innovate.

I would like there to be a world.

When people worry about the end of the world, they usually don't mean to imply its physical disassembly. Sometimes people even respond as if speakers did mean that, saying things like "nukes or climate change wouldn't actually destroy the planet, it will still be here, spinning", as if this was the point.

AI is one of the few things that could, actually, literally, end up with the planet being physically disassembled. "All it needs" is solving the extremely hard challenges of a von Neumann replicator, and, well, solving hard problems is kinda the point of making AI in the first place.

[slibhb]

> If you'd chosen to list that in the first place, I wouldn't have said what I did; "supercharge technology and productivity" is looking at everything through the lens of money and profit, not the lens of improving the human condition.

Bullshit. "Technology and productivity" are not the same thing as "money and profit". You're projecting your garden-variety European degrowth ideology onto what I wrote.

> Also, to be blunt, China's almost certain to win any economic or literal arms race you think you're part of; they make too much critical hardware now.

Europeans are so hilariously polarized against the US that they would prefer China, a literal authoritarian dictatorship, to "win any global economic arms race". I guess it's because China is too culturally distant for them to feel insecure over.

> AI is one of the few things that could, actually, literally, end up with the planet being physically disassembled. "All it needs" is solving the extremely hard challenges of a von Neumann replicator, and, well, solving hard problems is kinda the point of making AI in the first place.

It's not worth wringing our hands over science fiction scenarios.

[rune-dev]

> Do you really think AI companies/researchers are motivated by greed?

Researchers, maybe not. Companies, absolutely yes.

I don’t see how you could assume the likes of Google, Microsoft, OpenAI, and even Anthropic with all their virtue signaling (for lack of a better term) are motivated by anything other than greed.