[scubakid]

Makes me wonder: how's the HN community feeling these days about the actual plausibility / timeline of humans developing true AGI? Personally the more I learn about the current state of AI, and in comparison the way the human brain works, the more skeptical (and slightly disappointed) I tend to get.

[kromem]

I think that many people throwing their hat in the ring commenting on the unlikeliness of AGI are missing the impact of compounding effects.

Yes, on a linear basis it's not going to happen anytime soon.

But the trends in the space are developing around self-interacting discrete models to great effect (see OpenAI's Dall-E).

The better and broader that systems manage to self-interact, the faster we're going to see impressive results.

As with most compounding effects, it's slower growth today than the growth tomorrow. But a faster growth today than it was yesterday.

The human brain technically took 13.7 billion years to develop from purely chaotic driven processes, and even then it was pretty worthless up until we finally developed both language and writing so we could ourselves have lasting compounding effects from scaling up parallel self-interactions.

And from 200,000 years of marginal progress we suddenly went in less than 7,000 years from no writing and thinking the ground below our feet the largest thing in existence to measuring how long it takes the fastest thing in our universe (light) to cross the smallest stable object in our universe (a hydrogen atom).

Let's give the computers some breathing room before declaring the impossibility of their taking the torch from us, and in the process, let's not underestimate the effects of exponential self-interactions and the compounding effects thereof.

[coldtea]

>are missing the impact of compounding effects.

On the other hand those saying "it will sure happen" are missing the impact of diminishing returns.

[scubakid]

True, at a high level I think the central issue is what kind of curve we're on.

[KronisLV]

> And from 200,000 years of marginal progress we suddenly went in less than 7,000 years from no writing and thinking the ground below our feet the largest thing in existence to measuring how long it takes the fastest thing in our universe (light) to cross the smallest stable object in our universe (a hydrogen atom).

Personally, i don't doubt that AGI is possible, even though it becoming a reality might take any number of centuries or millennia, if humanity even sticks around for that long and AGI is still a goal that they pursue.

The problem lies in everyone thinking on a more human timescale: "Will we see AGI during my lifetime?" The answer to that is almost certainly no, no matter how much the industry tries to sell state machines as AI or fledgling efforts as revolutionary advances.

Being overly optimistic in regards to time scales only hurts oneself, like expecting that we'd all have flying cars or even that we'll be able to get rid of ICE vehicles or make significant improvements to slowing the pace of climate change.

[dan-robertson]

The opposite of compounding effects are compounding difficulties. Computer science is full of problems where a multiplicative increase in effort results in an additive increase in output. We call these “exponentially hard” and they crop up annoyingly frequently. So one argument is that compounding improvements will result in linear increases in output because of exponential difficulty. The counter-argument is that many of these hard problems have good but not perfect solutions which may be found more efficiently.

[toxik]

“AI research” is largely concerned with automation, not sentience or AGI. This is clearly abuse of terminology, even “machine learning” is somewhat misleading in my opinion. It’s mostly just pattern recognition of increasing elaboration, and the applications thus far are exactly that: pattern recognition.

It’s so difficult to talk about AGI, sentience, consciousness in general because there are no clear definitions apart from “I’ll know it when I see it.”

[Causality1]

Personally I think we're going to need a revolution in the fundamental physics of computation. The example I like to use is that a dragonfly brain uses just sixteen neurons to take input from thousands of ommatidia and track prey in 3D space, plot intercept vectors, and send that data to the motor centers of the brain. Calculate how many transistors and watts of power you'd need to replicate that functionality. Now multiply that number by how many neurons you think it takes the human brain to generate sapience.

It doesn't really matter what your guesses are, none of the results are good news.

[theaeolist]

I wonder if there isn't some fundamental misunderstanding here. What if it's not "just the neurons". If you found a Regency TR-1 radio you could wonder "how can this 4-transistor device produce a continuous stream of music, much like Spotify, which requires billions of transistor to run?". Of course, the radio also has an antenna, which is a completely different device than a transistor.

The device running Spotify may also have an antenna, but I hope you get the analogy. My analogy is not meant to be taken faithfully, so that we need to start looking for antennas now instead of neurons. I am just saying that maybe the neuron-counting game is not the only thing. Maybe there is something else -- not magical, not divine, but physical and as-of-yet unknown. Humanity didn't always know everything, and maybe still doesn't.

[Causality1]

Exactly my point. If all you want to do is replicate the TR-1 with four transistors, that's easy, just like making a human mind by creating a baby is easy. But making AGI with silicon, while demanding functionality completely alien to a human brain, is like making a TR-1 that can save your playlists and pause/resume the audio while still only using four transistors.

[Ginden]

> The example I like to use is that a dragonfly brain uses just sixteen neurons to take input from thousands of ommatidia and track prey in 3D space, plot intercept vectors, and send that data to the motor centers of the brain.

Human optic nerve can't send more than ~10Mbit/s. Yet, somehow, 60fps at 640x480 screen isn't best possible movie watching setup for one-eyed people, even though it delivers uncompressed 9Mbit/s.

Lots of calculations (like aggregating data to lower-quality image; eg. input of human rod cells is aggregated through interneurons) happen around of body. 16 neurons that you are referring to are likely fed with carefully processed input, not raw input.

[scubakid]

I tend to think in similar terms. There's so much going on under the surface with even the simplest creatures in the natural world that the physics and computational fundamentals seem really intimidating here. That's not to say that we could never get there -- certainly, many hold out hope for our abilities continuing to compound over time. But it's kind of a bummer to think about the glimmer of true AGI only materializing much further along an exponential growth curve that, to me, doesn't seem guaranteed to continue indefinitely.

[323]

We don't need the first AGI to be human efficient. Nobody would mind if it would require 10 data centers and a nuclear power plant to run.

ENIAC also started big and slow. Now it fits in a microSD card.

[mlyle]

> Calculate how many transistors and watts of power you'd need to replicate that functionality.

I'm curious as to your answer. Because if one's building a purpose-built analog computer for the task, my estimate is a few hundred transistors, a few thousand passives, and ... an absolutely trivial amount of power on modern process.

[Causality1]

I'm curious how we're even going to manage 420,000 pixels' worth (60,000 ommatidia, approximately 7 pixels each) of input with only a few hundred transistors, let alone do vector analysis on it.

But let's say we can. Let's say we need 320 transistors, which would be 20 transistors per pixel. That's pretending 99.7% of the seven thousand synapses each neuron has are useless for our purpose, but we'll do it. A chimp brain runs all the autonomous physical processes of a humanoid body while only having 22 billion neurons. We'll also pretend, wrongly, that chimps have no mind or emotions at all and that we only need the extra human neurons to make a sapient mind.

Humans have 86 billion neurons. Subtracting 22 gives us 64 billion, times 20 transistors per neuron gives us 1.28 trillion transistors.

1.28 trillion transistors, even with a bunch of handwaving to make it easier, and even pretending we exactly understood how sapience worked in the first place.

[mlyle]

> I'm curious how we're even going to manage 420,000 pixels' worth (60,000 ommatidia, approximately 7 pixels each) of input with only a few hundred transistors, let alone do vector analysis on it.

If you define the problem as importing 420,000 pixels, and target recognitions, and vector analysis, then you need a whole lot more computation than the organism uses. But presumably you're going to also get better results. We both know that's not exactly what's happening, I think.

That is, we know we can solve similar tracking problems with a whole lot less state.

> That's pretending 99.7% of the seven thousand synapses each neuron has are useless for our purpose

Not really... I think we can imagine a whole lot of passives / linear operations involved, along with the big nonlinear processes we need transistors for.

We're also assuming there's no net benefit to cognition that can happen using transistors, I'll note-- e.g. they have a ton of bandwidth compared to neurons, can be multiplexed more readily, etc....

> Humans have 86 billion neurons. Subtracting 22 gives us 64 billion, times 20 transistors per neuron gives us 1.28 trillion transistors.

So about half the number packed onto Cerebras WSE-2 today.

> even pretending we exactly understood how sapience worked in the first place.

This is the big problem.

[Ginden]

> 1.28 trillion transistors

So, basically, 45 x RTX 3080?

[jcims]

Feels like we're handicapping ourselves, at least in this specific domain, with digital computing.

[simonh]

We’re currently in the very early phase of our understanding of what intelligence is. The more we learn about it, the more we appreciate the staggering scale and complexity of the problem. So at the moment yes, it seems like the objective is receding into the distance faster than our progress towards it can keep up.

1960s - Herbert Simmons predicts "Machines will be capable, within 20 years, of doing any work a man can do."

1993 - Vernor Vinge predicts super-intelligent AIs 'within 30 years'.

2011 - Ray Kurzweil predicts the singularity (enabled by super-intelligent AIs) will occur by 2045, 34 years after the prediction was made.

So the distance into the future before we achieve strong AI and hence the singularity has been, according to it's most optimistic proponents, receding by more than 1 year per year.

Eventually I believe we will get a good enough understanding of the subject that we can map out a route to implementing AGI, and then our progress will accelerate towards a known and understood goal.

[joe_the_user]

The thing about these arguments for the impossibility of AI/AGI is that they inherently rest on the idea that they know what "human intelligence". So they have the same weaknesses as arguments project a set timeline for AGI.

We won't build a duplicate of the human brain - unless we have AGI first to tell us how. But we really don't know what portions of the human brain are needed for useful AGI.

You can look at GPT-3. On the one hand, never being reliable puts a crimp on practical applications. One the other hand, it does a lot of amazing things that seem human. I'd say that since we don't know where we're going in a profound way, we don't know how far we have to go.

[gameswithgo]

Nobody expected anything like supremacy in Go any time soon and then all of a sudden it happened. Maybe AI stagnates for a long time bow, maybe forever, maybe a big breakthrough happens tomorrow. Nobody knows, anyone confidently asserting anything is being foolish.

[rbanffy]

My guess is as good as any other layperson's, but I don't see much work being done for it, and no real good definition of what it is so we could plan how to create it.

OTOH, we see specialized intelligences do all soft of superhuman feats, all the time, and more impressive abilities join these all the time. These, however, are not human-like intelligences. They aren't even bee-like. They are so alien we don't see "general intelligence" in them.

So, my guess is that we'll have some extremely complex and capable systems that are extremely alien in nature well before we can have a conversation with a human-like intelligent system. They'll be useful and treated like oracles - we won't be able to understand their reasoning, but they'll be right most of the time.

It is, however, a matter of time and desire. There is nothing inherently magical in our mammalian brains and our organic bodies that can't be simulated by a sufficiently capable machine and technology for that will, eventually, become possible, then available, then practical, and then ubiquitous.

[scubakid]

To me, the term "alien" connotes a level of capability much more interesting than what I've actually seen from most modern systems. But point taken.

And I'd like to believe that you're right about it only being a matter of time and desire, but I do also worry about the possibility that we're actually on a different kind of exponential curve and will instead reach a point where we see diminishing returns.

[rbanffy]

I have no doubts there will be diminishing returns at some point, specially with narrow AIs, where increases in complexity and cost of training models will not be able to improve on what’s already good enough.

AI is a tool like many others, useful for some things and not for others.

[benlivengood]

We have superhuman performance at most narrow skills; the exceptions seem to be object manipulation with limbs/digits and semantic/logical thinking and planning. Given the advances by Boston Dynamics and others with limb-based mobility I'm guessing that's not too far off. With recent models proving a significant subset of the Metamath theorems, that doesn't look too far away either. Google/DeepMind are playing around with sparse model combinations of many useful superhuman domain models with additional layers to determine which domain to use for particular inputs.

The last and most difficult step in safe AGI is moral/value alignment. That is unfortunately probably last on the timeline of likely achievements because it requires general solutions to both planning and reasoning, and also an accurate world model and understand of physical actions and their consequences.

[hooande]

AGI is currently as likely as teleportation, time travel or warp drives. You can write a computer program to do just about anything. Artificial "General" intelligence is simply not a thing. We're not even making progress toward it.

[ethanbond]

We have natural “general” intelligence which appears to be generated by boring old chemical/thermal/electrical interactions. Why wouldn’t we be able to recreate that at some (IMO very far) point?

[nine_k]

More than that: we have literally billions of examples of human-level intelligence right here on Earth. We have not a single example of teleportation, time travel, FTL, and other staples of not-very-science fiction.

Guess what is more likely to be implemented.

[hooande]

Think about how difficult it would be to make a fly from scratch. Not editing the genes of an existing organism, but combining the raw chemical components into a form that's identical to a fly.

There are trillions of examples of insects on earth, but they do us no good when it comes to building one without using an evolved framework.

We've created a great number of things that had no natural analog. The internet, space travel, etc. I'd say our odds of doing something we haven't seen before are about even with artificially recreating a lot of things we see every day

[TheOtherHobbes]

We don't have very good general intelligence.

What we have is a fairly loose mix of categorisers and recognisers, biochemical motivators and goal systems, some abstraction, and a lot of externally persistent cultural and social programming. (The extent and importance of which is wildly underestimated.)

The result is that virtually all humans can handle emotional recognition and display with speech and body language including facial manipulation/recognition. But this doesn't get you very far, except as a baseline for mutual recognition.

After that you get two narrowing pyramids of talent and trained ability. One starts with basic physical manipulation of concrete objects and peaks in the extreme abstraction of physics and math research. The other starts from social and emotional game playing, with a side order of resource control and acquisition. And peaks in the extreme game playing of political and economic systems.

So what's called AI is a very partial and limited attempt to start climbing one of those peaks. The other is being explored in covert collective form on social media. And it's far more dangerous than a hypothetical paperclip monster, because it can affect what we think, feel, and believe, not just what we can do.

The point is that it's a default assumption that the point of AI is to create something that is somehow recognisable as a human individual, no matter how remotely.

But it's far more likely to be a kind of collective presence which doesn't just lack a face, it won't be perceived as a presence or influence at all.

[theaeolist]

Can you recreate all phenomena computationally? Could you replace the antenna of your radio or mobile phone with a special CPU? Could you bomb a country with CPUs? I don't think so.

[hooande]

A warp drive is theoretically possible, and also driven by boring chemical/thermal/electrical interactions. humans may create one of those at some very far point in the future, too

[mlyle]

> A warp drive is theoretically possible,

Dubious

> and also driven by boring chemical/thermal/electrical interactions.

Implausible exotic matter, negative energy, etc, are usually prerequisites.

Just like the existence of flying birds were a hint that flying machines might be possible, the existence of thinking creatures is a hint that thinking machines might be possible.

[simonh]

We do not observe teleportation, time travel or warp drive in nature. We also don’t have any practical theory for achieving them as depicted in science fiction. It seems unlikely we will achieve such technologies.

Do we observe general intelligence in nature though, here on Earth implemented with the materials available in our environment? If so, it’s a bold claim to make that it will always be impossible to achieve it artificially.

[hooande]

How far are we away from gene editing that will allow humans to be born with working gills or wings? Animals have these things, so we know it's possible. But having the technology to do that is very far off, if ever.

The same is true of AGI. Of course it's possible. but right now no one has any clear idea how to do it without extreme brute force.

Personally, I think it's more likely that we'll have a working Alcubierre drive before anything approaching general intelligence

[simonh]

We can extract oxygen from water right now, so we can already do this. Requiring a specific implementation technology is unreasonably stacking the deck.

You’re making the same mistake as those who critiqued the concept of heavier than air flying machines, starting from the assumption they must work by flapping their wings. As it happens now we have wing flapping drones anyway though.

“Ever” is a very, very, very long time.

[Ginden]

> How far are we away from gene editing that will allow humans to be born with working gills or wings?

Impossible due to physics limits. Human lungs have 57 square meters for extracting oxygen from fluid with 21% volume oxygen. 30°C air-saturated water have 0.5% of oxygen, so working gills for human would need surface area of 2394 square meters.

[rembicilious]

Gills work by continually “filtering” water as is flows through them. Lungs are filled and then emptied according to some pattern of breath. The difference in volume of fluid processed must be significant. Also, can’t gills have a higher surface area to volume ratio than lungs?

[more_corn]

I don't think AGI is likely, I think it is inevitable. We can make specialized neural networks that can do specific tasks quite well. There's nothing stopping us from chaining those together. We have the pieces to make neural networks that can train on new data, thus creating new layers atop previous networks. We can even train those layers based on the data generated by the action of the network itself. The pieces seem to be present, the tooling around putting them together seems to be lacking for the time being. I expect to see AGI in my lifetime, artificial super intelligence shortly thereafter and then the event horizon of the singularity.

[abetusk]

The human brain is estimated at 2.5 Pb of storage [0]. Assuming a "Moore's Law" like behavior of storage price, so that price halves every 2-3 years and assuming we use storage as a proxy for the space, access speed and computational power, the time it will take to have a $1000 computer that has the storage capacity of the brain will be in the 10-16 year time horizon.

This puts the timeline to about 2029-2035.

[0] https://www.scientificamerican.com/article/what-is-the-memor...

[nine_k]

It is not hard to create a RAID array with 2.5 Pb capacity.

The trick of the human brain is that the "processing power" is enmeshed into the "memory", so the brain must have a colossal computational bandwidth, even with pretty slow neurons. I suppose that bandwidth is larger than that of most modern GPU / TPU clusters, which also don't feature anything comparable to 2.5Pb or RAM in their disposal.

The revolution should be mostly in the architecture, much like the deep learning evolution was enabled by GPUs.

[Something1234]

In the past 10 years, I think we had a 8x increase in easily available storage going from 1TB drives being a $100 to a 16TB drive being roughly $250. So I would have to say that your time scale is way too optimistic at best.

[abetusk]

I won't check your numbers and take them at face value.

Even with your numbers, that's a 6.4x decrease in price per TB (($100/1Tb) / ($250/16TB)), which is around 2.7 halvings over the course of 10 years, which is very close to my "2-3 years per halving" statement.

Even if it's slower than a halving in price per 2-3 years, 4-5 years say, this only delays my prediction by a decade or so.

[theaeolist]

We DO have PB-grade storage facilities and they lack full AI. Moving the memory inside a single box instead of having interconnected devices is not going to bring AI just like that.

[abetusk]

I think you missed the critical point: for $1000.

Feasibility is great but economic access is the aspect that I'm focusing on.

[jjoonathan]

Have you seen the "interviews" with GPT3?

[qualudeheart]

What about them?

[jjoonathan]

You'd fail to Turing test them in casual conversation.