[godelski]

  > But our models make it clear that such an [intelligence] explosion may not follow if there are diminishing returns (“ideas become harder to find”) or if feedback loops become bottlenecked.

How is this not obvious to everyone? As we advance it becomes more difficult to advance. You obviously make most advancements around the things that are easiest to improve. Then all the easy things are done. So you go onto the next easiest things. They're "the easy things" from that standpoint but that doesn't mean they aren't harder than "the easy things" when you started. Complexity increases as precision increases.

[setgree]

The goal of a modeling exercise like this, which you don’t have to buy, is to generate a simple set of initial conditions that can explain things we already know. Then, we can manipulate some initial parameter value to make predictions about things we don’t see, but might.

Likewise, it is obvious that gravity exists, but a simple model that explains where it comes from (in quantum terms) would be a big breakthrough iff it came with plausibly testable implications that could be tested via experiment.

[bryanrasmussen]

>How is this not obvious to everyone?

because people have other ideas https://en.wikipedia.org/wiki/Technological_singularity

[Majromax]

> because people have other ideas https://en.wikipedia.org/wiki/Technological_singularity

In a weak sense, singularities are common and should be expected every so often. In a mathematical sense, a singularity is where a model 'blows up' and neglected terms become part of the dominant balance.

Industrialization hit a point of diminishing returns, but the industrial revolution was nonetheless a 'singularity,' where life afterwards was qualitatively unpredictable to people who lived before. Likewise, agriculture was such a technological singularity to hunter-gatherer ancestors.

I could even make a decent argument that writing and literacy were such a singularity, making inconceivable social organizations routine.

In that weak sense I expect AI to be a singularity, recursive self improvement or no. Life in 2050 may be completely unpredictable to someone who was taken out of time in the year 2000.

The remaining questions are speed and intensity, and both of these questions are related to RSI. If RSI works, then the 'fast takeoff' visions become more plausible where society transforms over months to a few years – at least locally where the enabling technologies have diffused. If not, it might take a couple of decades.

[bryanrasmussen]

> Life in 2050 may be completely unpredictable to someone who was taken out of time in the year 2000.

Well the end of AMOC would also be a singularity by that measure, maybe about the same time.

[WJW]

People have other dreams, and have written about them in science fiction. But science fiction can just handwave away resource problems that exist in real life so that the plot can happen.

Just as the popular "grey goo" nanomachine disaster can easily be shown to be impossible due to resource imbalances and energy shortages, AI recursive self improvements rapidly slows down due to problems with complexity, training data and compute availability (whether due to actual processors or just due to energy demands).

[ben_w]

> Just as the popular "grey goo" nanomachine disaster can easily be shown to be impossible due to resource imbalances and energy shortages,

Depends on specifics and how dramatic they're being, given that "mold infestation" can come in grey.

[bryanrasmussen]

whether or not the singularity would be best thought of as science fiction, it was not presented as such, and has not been taken as such by a great number of people, and has formed the basis of a lot of opinions as to how things will pan out.

It is these opinions that make the conclusion the parent poster supposed should be obvious to everyone not obvious to so many.

[godelski]

The problem is in cases of the singularity it's being treated like reality rather than fiction

[vidarh]

Because they depend on whether the rate of improvement of self-improvement outpaces the rate of increase in difficulty or not, and at some points they clearly do - e.g. a lot of skills makes the relative rate of subsequent improvement easier for a while.

It may seem obvious that it can't last, but showing the conditions where it can't still matters.

[tossandthrow]

> As we advance it becomes more difficult to advance

I don't think this follows.

We have advanced tremendously over the past 200 years, and we are likely going into a time with rapid advancement again.

With advancement, we also develop tools (eg. Llms) that assist advancing.

[Windchaser]

To put it another way, the cost per advancement increases. That’s “as we advance it becomes more difficult to advance”. However, because of the prior advances, you also have more resources to throw at future advancements.

So, then, the question is whether the “profits” on the last advance are enough to pay for the next one. We can define a new term, “affordability”, as what % “profit” you can expect from each advance relative to its cost, telling us whether it becomes relatively easier or harder to continue to advance.

[tossandthrow]

You could argue that there is a capital built up - but even that is difficult as a lot of the knowledge for, eg., building LLMs can not demand rent. Everybody can build their own transformer networks.

> To put it another way, the cost per advancement increases.

This is not true when you don't use capital that demands rent. On the contrary, the cost per advancement is actually decreasing as we develop more knowledge.

LLMs are a cognitive technology (as contrasted to a communication technology). And it will help us tremendously manage knowledge such that you can utilize it better decreasing the cost of advancement.

[RandomLensman]

Whether or not it decreases the costs of advancedment would depend on what is causing advancement to have costs. Some things might be addressable by cognitive technologies, others might not.

If capital doesn't create a return evenutally, why would it invest? (Or did you mean "rent" in the narrow economics definition? But then not sure how that applies here).

[breuleux]

> We have advanced tremendously over the past 200 years

Would most of that have happened if we hadn't found oil, though?

It's difficult to precisely identify how much of progress is owed to intelligence, and how much is simply energy availability. Energy is intrinsically transformative -- the dumbest of organisms can take over the world in a blitz if they can process available energy better than the others, whereas the smartest of entities is going to be ineffective without fuel. Intelligence can, of course, unlock the capability to exploit new energy sources, but there's still an intrinsic physical distribution of it that's outside of anyone's control.

[tossandthrow]

Reducing progress to increasing entropy makes it even more clear that making progress will likely become easier, not harder.

Yes, oil was important back then. Now we can augment with wind, solar, nuclear, etc.

And more technologies will come.

[breuleux]

Maybe, maybe not. The point is that you can't simply will energy into existence. Its availability, including the ease of exploitation, is a physical given. Which is true of resources in general. Arbitrarily advanced technology can't simply make exploitation arbitrarily easy, so for example there can always be situations where the very best of all possible technologies will be less productive than setting fire to petrol would have been, if there had been any. It's complicated. It's not just what we do, it's not just what we have, it's both.

Probably the best test of how the future might trend long term is to see how well we can extract value from all the stuff and waste we are currently producing, because these will inevitably become the primary resources from which we have to build everything new. Unfortunately that shift is likely to happen extremely quickly and catastrophically, because typically you only start processing bad materials once you're done with the good.

[windexh8er]

Then use batteries as an example. Going on 200+ years now most of the major changes have been an easy swap: chemistry. Lead acid --> nickel cadium --> nickel metal hydride --> lithium ion. Improvements in batteries are hard and the surrounding ecosystem has gotten better. Better BMS and better motors have also made gains to increase the efficiency of the battery. Compounding gains within the ecosystem doesn't truly improve the battery directly, but improves them for implementation and use cases.

All the remaining problems with batteries are a hard solve because it's tradeoff after tradeoff. Lithium sulfur, for example, could be amazing but die after a couple hundred cycles at most [0]. So... If LLMs are so novel, as many continue to claim, why hasn't this been solved? Probably because LLMs are an interpretation of our current understanding of everything. An LLM currently only does what any human can do, albeit in a compressed timeframe, mostly.

I think it follows. LLMs have made some considerable improvements, but use "reasoning" (really just a mechanical, autoregressive, loop: generate tokens --> append tokens to context --> feed the expanded context back into the model --> generate the next token --> repeat) as an example. You can do this by hand as well in a traditional chat volley, but it's slower. So we improved LLMs by putting them in an automatic loop (oversimplification - but at the end of the day holds water).

Until LLMs showcase true external breakthroughs that aren't driven by human guidance (not happening anytime soon) we are in this loop of hacks being used to improve the 80% (the LLM itself). Notable jumps? "Reasoning" and with Mythos-like models we now have "Advanced Reasoning" (by leveraging a more capable looping framework such as an agentic harness).

[0] https://advanced.onlinelibrary.wiley.com/doi/10.1002/adfm.20...

[skinfaxi]

It seems the rate of advancement has increased over time though. Like from ancient times to 1800s the rate of progress was slow, then basically a relative explosion of progress compressed in 200 years.

[godelski]

  > It seems the rate of advancement has increased over time though.

Effort and expenditures are not fixed. You cannot hand wave away these variables.

Here's an example let's compare two investment strategies:

  - Alice puts $1 into VOO every day. 
  - Bob puts in $10000 into VOO every day.

You don't get to say "wow, Bob is a much better investor than Alice. He's so smart, look how fast his money grows!" You get to say "Bob is investing more money than Alice."

Now pretend Bob is investing into a slower growing account than Alice. His account will still grow faster as long as his account is not losing >99.9% per day. He could have a worse strategy and still grow faster because he's throwing more money at it

[bobson381]

Because of access to a giant store of energy. Our advancement has also created a future need to consume more and more energy to fuel increased complexity.

[godelski]

Which has created all sorts of other problems even if we exclude climate change.

I'm not arguing we shouldn't live in a modern world, but rather that complexity increases

[red75prime]

> How is this not obvious to everyone?

That RSI can be bottlenecked? I guess this is obvious to many people. Whether RSI will be bottlenecked (at some not very interesting stage) is another question.

[RandomLensman]

Why would it not be bottlenecked? Is intelligence so valuable as such?

[ben_w]

Our intelligence was enough to turn some balding apes into atom-bomb wielders and astronauts over a few thousand generations.

Consider a hypothetical: it is possible to make a semiconductor clone of a human brain, that runs at semiconductor speeds, has semiconductor size-scale, and has a power requirement of the Landauer limit.

This copy would be smaller than a pea, and you'd get to pick on a sliding scale between "same power draw as human but thinks faster than us to the same multiplier that we jog faster than continental drift" or "thinks as fast as we do while using around a few µW of power".

We do not know how to do this. We don't know how far we are from figuring out how to do this. We do know* evolution made our brains, and we do use simulated evolution as a standard technique in machine learning.

But it may well be that just as no human knows how to design a human mind, we find the best AI we can make don't know how to make a better AI, at which point they're stumbling blind in the dark: while evolution is a neat method, it is limited, blind to what the best next step is at any moment.

* well, those of us who are not creationists, at least.

[defmacr0]

Well at least it seems pretty implausible to me that a machine learning model trained to reproduce human text can in principle generate something that is significantly above human text production ability. If the "pea-sized semiconductor brain" is not a surprisingly shallow problem that you can just solve by interpolating existing research, I don't really see the LLM-approach to AI being the thing that makes something like it happen.

>we do use simulated evolution as a standard technique in machine learning.

Well, not really. For large-scale AI models it's almost exclusively some form of gradient-based non-linear optimization. Genetic algorithms (which is just hill-climbing optimization with extra steps) and genetic programming (which is really cool and not well-understood) do not perform all that well in practice and I'm not aware of any notable applications.

[ben_w]

> Well at least it seems pretty implausible to me that a machine learning model trained to reproduce human text can in principle generate something that is significantly above human text production ability.

These are not the only currently-in-use AI models. However, recent news has even this particular category of AI solving multiple previously unsolved Erdős problems.

(Even if they also do stupid things on a frequent basis).

> Well, not really. For large-scale AI models it's almost exclusively some form of gradient-based non-linear optimization. Genetic algorithms (which is just hill-climbing optimization with extra steps) and genetic programming (which is really cool and not well-understood) do not perform all that well in practice and I'm not aware of any notable applications.

I said "standard technique" rather than "best" for a reason ;)

This is proof-of-possibility: natural selection did it, we know how to mimic that, but we don't know enough to be sure we're doing it with a reward function that will actually give us minds like ours on an interesting timescale with a probability high enough to care about.

[godelski]

  > solving multiple previously unsolved Erdős problems.

So what?

Machines have solved tons of unsolved problems in mathematics. That's not a proof of intelligence.

If you brute force a solution, we congratulate you on your effort.

If you stumble into a solution, we congratulate you for being lucky (if we can distinguish)

If you find a unique solution that no one else imagined, we congratulate you on your intelligence.

These are categorically different things and the difference matters. It's the whole distinction. Though in the real world success usually requires all three (and more), complicating evaluation.

When we're talking about intelligence you can't distill it to "getting the answer". If you do then I'll direct you at an abacus, a calculator, a watch, or google search if you want to look at super intelligence.

  > natural selection did it

Did what? Gradient descent? If that's the argument, you need to read more

[red75prime]

Autoregressive pretraining (text/images/video prediction) produces a foundational model. You can look at it as a highly compressed conditional probability distribution of the human brain output. The information-theoretically optimal compression of the data is a program that reproduces functionality of the process that generated the data.

So, it stands to reason (and observations) that such a model captures not only surface statistics of the data, but a part of functionality of the system that generated the data (the human brain for text, physics for video).

[RandomLensman]

Is it possible to clone without destroying the original brain (and is a semiconductor the right substrate for it)? Thinking of the no cloning theorem here...

[ben_w]

The no cloning theorem applies to quantum states; when it comes to information processed by it, it's unlikely that a big wet messy hot thing like a cell is usefully treated as "a quantum state", let alone an entire brain.

However, this is irrelevant to the thought experiment, for which it is sufficient to merely be as competent as a human, rather than a completely quantum-perfect replica. Clone in the sense you would use the photoshop brush, or perhaps in the sense you might be sued for trademark infringement.

[RandomLensman]

What is the thought experiment then to show other than a hypothetical piece of compute hardware? What's the experiment and its implication beyond the existence and its properties?

[jamesrcole]

> As we advance it becomes more difficult to advance. You obviously make most advancements around the things that are easiest to improve. Then all the easy things are done.

This isn't some foregone conclusion. It completely depends on the rate at which the intelligence and abilities of the AI increases. If that rate was high enough, then the harder and harder problems would become easier and easier for it.

[dinfinity]

Actually, evolution seems to show the opposite: The rate of advancement has only sped up, with billions of years between significant changes going to millions, to thousands, to tens and arguably to mere years now.

Having said that, we're probably looking at an S-curve with the physical limits of reality getting in the way in the end.

[podocarp]

The more you evolve I guess there's more surface area to evolve on. Similarly cavemen could only do so much. But today you can take a degree in physics, art, history, etc. and it's probably impossible for one person to know all that ever was and is. But it's possible for a caveman shaman to know everything the tribe knew up to that point. Probably. So while it's harder to make progress there's also a lot more places you can progress.

[godelski]

You're ignoring a lot of important factors. I'll give you most obvious ones. We often do genetics testing on fruit flies because they reproduce quickly and the initial population is large.

Your claim ignores this fact. The base population and number of interactions has increased.

[dinfinity]

I am not ignoring anything, I'm looking at the broader picture, which includes non-biological evolution. Simple rebuttal to your specific point: The population of self-improving AIs will also go from 0 to many more.

In a broader sense evolution moved from very static simple domains to dynamic malleable complex domains. Biological evolution speed is glacial compared to cultural evolution speed. Even then, cultural evolution is fairly slow compared to technological evolution.

[mpalmer]

Just like fully extending a ladder brings the moon closer.

[FrustratedMonky]

"How is this not obvious to everyone? As we advance it becomes more difficult to advance. "

It's a type of recent change blindness. Because for the last few years there are seemingly impossible breakthroughs every few months. Literally, things everyone said would take 30 or 100 years, happen within months. It is really easy to think this will continue. The new normal. Really, we haven't seen it slow down at all, so why would we expect the miracles to stop all of a sudden?

It seems like accelerating, and decelerating, are both fair game as future directions at this point.

So it is nice that the paper put some weight behind the argument that this could all grind to a halt for a lot of different reasons beyond the hype.

[breuleux]

> Literally, things everyone said would take 30 or 100 years, happen within months.

"Everyone"? Plenty of people in the 50s said we'd have sentient humanoid robots before the year 2000. And flying cars. Sure, many people underestimate the rate of change. But techno-optimists have consistently overestimated it, and they still do.

[FrustratedMonky]

I was sloppy. I didn't mean to include science fiction authors, or extreme futurists.

"everyone", being, a typical average of realistic predictions.

In the 90's. I remember people talking about voice recognition will take 50 years. Let alone understanding. Now this is an easy base line feature.

Let alone protein folding.

[breuleux]

> I didn't mean to include science fiction authors, or extreme futurists.

Science fiction authors are quite influential, and I don't think they were the only ones either. I expect that plenty of actual scientists had similar views. It wasn't extreme.

> In the 90's. I remember people talking about voice recognition will take 50 years.

Sure. Who? How many people said that? How many people said the opposite? If we're doing anecdotes, I remember that circa 2010 it was a fairly normal belief that fully self-driving cars would be widespread by 2020. I mean, I thought they would.

Point is, many people say many different things, there's never been any widespread agreement about what the future holds. They underestimate, they overestimate, but mostly reality moves sideways in directions nobody thought of and which are somehow simultaneously way more impressive and way less impressive than our expectations.

[FrustratedMonky]

Technically correct. But kind of backs every argument into a corner. I'm assuming also, anything like a paper with data, can also be questioned in the same way. Don't object, I've seen that done. So basically, don't talk about anything, because anything you say is contradicted by something somewhere. Everything is True, Everything is False, everything is subjective. There is no reality. Every person has there own point of reference that is completely subjective and invalid.

[breuleux]

> I'm assuming also, anything like a paper with data, can also be questioned in the same way. Don't object, I've seen that done.

Unless you've seen me do it, I don't see why I can't object. What a strange thing to say.

Look. You've supported your argument with vibes and anecdotes, so I'm offering contradicting vibes and anecdotes of my own. All that means is that it's not as obvious as you thought, and it is worth digging deeper (or not, it's not that important). A paper with data would certainly make a stronger argument. And of course it can be questioned, but it cannot be questioned in the same way. That you've "seen it done" is irrelevant.

[godelski]

Fair but there's two things I think need to be added

1) predictions are usually made assuming either constant effort or continued growth. They will be extremely conservative when the environment changes and you start dumping trillions of dollars into solving those problems.

2) while there's been real progress that doesn't mean it isn't being over sold. We do need to differentiate the signal from the noise

[exe34]

> How is this not obvious to everyone?

Probably because it's not true. We had shitty neural networks for decades before the recent explosion. That particular branch may be a dead end, but there could be others lurking and waiting for their time.

[soulofmischief]

Progress is often lumpy, modulated by new discoveries and their applications.

[EarlKing]

> How is this not obvious to everyone?

Because everyone's thinking around intelligence is incredibly muddled by a variety of factors, and no one is particularly motivated to actually correct anyone's mistaken notions on the matter.

[godelski]

I'm conflicted. On one hand I think we should more openly call people idiots and push back. On the other hand there's Descartes argument for idiots in good company.

I just wish all the people that claimed to care only about truth would actually care about truth. Feels like society is more that trope where someone says a joke to a crowd and no one hears it except one charismatic person who repeats it and gets all the laughs. In reality it feels like the repeated version of the joke doesn't even make sense, it is just vibes.