Hacker News Reader

Comment by nradov | Hacker News Reader

It would be foolish to claim we've made any meaningful progress toward a true AGI that can pass the Turing Test until someone demonstrates a computer as smart as a mouse across the full spectrum of activities.

View on HN

Once we get computers as smart as a mouse, we'll be at most 3-5 years or so computers as smart as a human. We will have solved at the major challenges in AGI and it will simply be a scale problem then.

Saying we don't have mouse level AGI is simply saying human AGI is greater than 5 years away, which isn't a remotely contentious statement.

The difference in intelligence between an amoeba and a mouse is enormous compared to a mouse and us. People greatly under appreciate how intelligent and close to human a mouse/bird/pig are in the grand scheme of things. Emotions, behaviors, motivations, goal setting, memory it's all there already. A flat worm, an ant, a fly - those are the large stepping stone accomplishments.

Think about rate of very long distance communication in humans. It took us tens of thousands of years to get to 2.4 kps dial up modems, and only few decades to get to common 300Mbps. The important signal is seeing a 100bps modem, not a 100Mbps connection.

So the real question is how long until we can replicate a worm's intelligence?

> worm's intelligence

Done.

http://openworm.org/

Not done. OpenWorm is a project to create a worm simulation. An important distinction.

More like “overdone”, it’s a cellular simulation including the brain as a first step.

The brain in question is 302 neurons, which is probably fewer than most of us have lost whilst participating in this discussion! But they also need to simulate a nervous system, mobility, food needs, interaction etc to have a remotely persuasive argument that a 302 node neural network actually resembles a worm brain. It can't be as smart as even a really stupid worm until it can perform analogous tasks to the worm, which means having some sort of artificial body to wiggle.

(parallel arguments but for human/mouse level complexity of bodies and stimuli responded to would suggest that whole brain emulation is going to be an incredibly painful way to attempt to achieve AGI)

> But they also need to simulate a nervous system, mobility, food needs, interaction etc to have a remotely persuasive argument that a 302 node neural network actually resembles a worm brain. It can't be as smart as even a really stupid worm until it can perform analogous tasks to the worm, which means having some sort of artificial body to wiggle.

It has those things[1]. There’s a video of its simulated body wiggling around on the project’s github repository.

[1] except possibly food, I was skimming the page.

Sure, I've even seen the wiggles; I was making the point that's why they needed it to yield anything that could be compared with the real thing (though I'm not sure it has food and reproduction simulated yet, and I'd imagine a large portion of the worm's limited capability for cognition is ultimately directed to pursuing those ends)

It needs to be able to survive and reproduce in an environment like a real worm does to prove that it's just as intelligent, and not just wiggle around.

"To get a quick idea of what this looks like, check out the latest movie. In this movie you can see a simulated 3D C. elegans being activated in an environment. Its muscles are located around the outside of its body, and as they contract, they exert forces on the surrounding fluid, propelling the body forward via undulutory thrust. In this model, the neural system is not considered and patterns of muscle contraction are explicitly defined"

http://docs.openworm.org/en/0.9/projects/

> In this model, the neural system is not considered and patterns of muscle contraction are explicitly defined.

In other words, the simulated worm brain is not yet even capable of causing the wiggling seen in the video. So the question remains, what can the simulated neurons do, f anything?

Alas, it's not even close to "done." It's a work in progress and a surprisingly difficult one.

For context, the worm (c. elegans, at least) has a very stereotyped nervous system with 302 neurons. The anatomy, down to the cellular level, is known incredibly well. Their behavioral repertoire is not huge and they're fairly easy to study. Nevertheless, we can't even simulate a worm very accurately. (There was a good twitter thread about why yesterday: https://twitter.com/OdedRechavi/status/1086992699528544256)

The human eyeball has about 120M rods, 6.5M cones, and projects to a brain containing ~86B neurons, which is about 8-9 orders of magnitude more cells. The number of possible interactions scales even faster. In summary, we're not close, not at all....

When biology comes into play, you can see how weak our understanding and abilities are. We can barely simulate small proteins of a few thousand atoms. Accurately simulating an entire cell is (at the moment) in the realm of science fiction. Only inaccurate abstractions can be used to model it.

However I have to object in a way about the brain. To me, there's an unanswered question: Is the rest of the human brain as simple and "generic" as the convolutional neural networks we made inspired by the vision system? Or is each networks' architecture and "algorithms" developed specifically for a task? In the latter case we might still be a very long way from anything resembling AGI.

However my personal estimation is that most of the things we do can be modeled using existing tools when scaled and modified appropriately (ie RNNs). There's also the ugly job of stitching those systems together, but it's not that different from what happens in nature.

> Once we get computers as smart as a mouse, we'll be at most 3-5 years

But how long did it take Nature to get from a mammal with mouse-level intelligence to a human-level brain. I think 200-ish million years [0].

You might be right that a mouse is a good indicator of high-level intelligence and that you don't need human-level intelligence to make a good AI, but there might still be some considerable way to go until we have an AI that can significantly outperform us.

[Edit - I agree that natural selection wasn't aiming or directed, and thus wasn't forced to be as fast as we could be. But a human's higher brain functions might not be simple incremental improvements over a mouse's, and there could still be a long way to go]

[0] https://en.wikipedia.org/wiki/Mammal

How long did it take Nature to get from nothing to a mouse? About 20 times as long.

So I think the estimate of 3-5 years might be realistic, but the artificial mouse is a long way away, IMO.

Nature wasn’t really aiming. So it’s not a valid basis of time estimates.

How long did it take nature to go from T-Rex to chickens?

There’s no reason to believe human level intelligence to be an inevitable result of evolution. It just happend.

Keep in mind that nature uses a somewhat directional random process. How much of those 200-ish million years were spent waiting on selection pressures to promote bigger brains?

> But how long did it take Nature to get from a mammal with mouse-level intelligence to a human-level brain

Not that I agree with the sentiment in the GP, but it took a relatively short time from the first development of multicellular life until nervous systems developed and an even shorter amount of time to go from small mammal intelligence to human intelligence. However, evolution isn't about "progress" as we understand it. The most we can say with regards to intelligence and evolution is that human intelligence satisfied a niche that existed at a certain place and time.

I was thinking about human intelligence today and thought about how anything at the tail end of a normal distribution usually produces quite a perverse result. Then I realized we are at the extreme tail end of the intelligence distribution in the animal kingdom. No wonder we manifest all manner of odd results.

I'm not sure creating an intelligence that even supercedes our own will lead to anything good. If anything I'd expect things to get even more perverse.

Maybe this isn't what you meant, but I'm curious why you would put a Fly ahead of an Ant in the stepping stones.

Can you make a Turing Ant without a Turing Ant Colony?

> Once we get computers as smart as a mouse, we'll be at most 3-5 years or so computers as smart as a human. We will have solved at the major challenges in AGI and it will simply be a scale problem then.

Why? What challenges?

You have no rational basis for that 3 - 5 year estimate. That's just picking numbers out of the air.

The basis is the complexity growth trend that we have already observed in computing (and most other human endeavors)

There is zero evidence that progress toward AGI follows Moore's Law. And we observe much slower complexity growth in most other human endeavors.

But there's a lot of evidence that computer power follows Moore's law. All that matters is that the growth is exponential.

Evolution took a billion years to evolve multi-cellular life, but the jump from apes to humans took far less than a million years.

Computer processing power is not growing at an exponential rate. The physical limitations of Moore's law are well known and imminent.

Moore's law is essentially tracking the transistor density on silicon. We may be pushing up against physics in that area, but that is not the same thing as processing power. Our systems grow ever more complex. Each generation enables new tools that enable the creation of the next. When Moore's law finally crashes and burns, we will compensate using other technologies. Multicore and multiprocessing, shifting work to the cloud, advanced materials, etc. Hell, even quantum could take off in the next decade or two. I see no reason whatsoever to believe that we just give up and rest once we've reached the limits of silicon transistors.

> We may be pushing up against physics in that area, but that is not the same thing as processing power

I didn't say we're pushing up against the limits of processing power, I said that processing power is not growing exponentially, which is true, despite the gains that other advancements and innovation have provided.

Moore's law is reasonably dead.

We're moving faster than Moore's Law, see "Hyper Moore’s Law":

https://www.extremetech.com/computing/256558-nvidias-ceo-dec...

That's a total non-sequitur. So far there is zero evidence that increases in computing power are getting us any close to true AGI. It's entirely possible that we've been moving sideways, or even backwards, relative to that goal.

And we can't reliably extrapolate growth in computing power more than a few years into the future. It's possible that the curve isn't really exponential, but rather an S-curve which will eventually flatten out.

Now you're going out of context. We very well may be on the wrong track for making an AGI, but the OP's premise was that once a mouse-level AGI is achieved, human-level AGI won't be far behind.

I'm more comfortable predicting that computing power will continue to grow than to predict that it will peter out and everyone will simply sit back and be happy with what we've got.

Moores Law has not been working anymore like what, last 5 years?

mouse, crow, worm, human, all running on neurons. You simulate one you can simulate a berzillion.

I think you're right. I haven't read the article, but I think Kurzweil is perpetually too optimistic about what we can achieve, at least in how fast it will happen. There are cool things happening, and lots of advances being made, but we are nowhere near anything that could really be called AI.

Plus, I think all the marketing use "AI" is giving a very distorted and inflated view to the average person of what software is actually doing, and what it's capable of.

It's a buzzword, full stop.

Kurzweil really, really, really does not want to die. So all of his predictions are always timed just so that the technology needed to live forever will be within his projected natural lifespan.

That doesn't make him wrong, but that's the personal bias he's operating under.

Yet die he will.

> Kurzweil really, really, really does not want to die

What a shit life must that be. And I say this in a very sympathetic way. However, feeling you are almost in reach of eternal life, but not being sure you'll make it in time, being constantly afraid of an accident, or illness, taking that away from you... It's a recipe for anguish and panic.

Dying is not that terrible when you know everybody else will too, sooner or later; but try accepting the idea of being among the last to die..

> Kurzweil really, really, really does not want to die.

So strange, considering that non-existence is the one thing that every conscious being is guaranteed to never experience. Why run from something that can never catch you?

That is a logical argument, but presumes existence is logical. - Another thing to cast into doubt when staring as deep as you do seem to into the abyss.

The near term "reptilian" fear of death seems axiomatic. The longer term one I can only charcterize as FOMO.

He values continuing to experience, like most people.

Because creatures who felt compelled to run from it had a better chance of passing on their genes than creatures who didn’t, leaving us all with an instinctual desire to avoid death.

I'm really happy that the world is shifting from billionaires that just want more money to billionaires that understand that it doesn't mean anything if they don't finance solving the problem of dieing.

I view all the polititians who have the power of advancing healthcare research but not doing it stupid.

> it doesn't mean anything if they don't finance solving the problem of dieing.

There's no solution to death. You can only put it off, but something will assuredly kill you in time. If it's not aging, then it will be cancer, heart disease, an accident, etc. Ultimately, entropy will get you one way or another.

Personally I didn't do a lot of fun stuff (like riding motorcycle, extreme sports) to decrease probability of an accident, even though I'd love to (I'd tried them, loved doing them, but not doing them regurarly).

As for cancer and heart disease, both are linked to aging or genetically inherited mutatations. Heart disease is a natural result of damages in the human body not being reversed.

https://www.cell.com/fulltext/S0092-8674(00)80567-X

I can see a world with the means to escape natural death be one of social, cultural and technological stagnation.

Just living longer doesn't mean that humans become any wiser on average. There will maybe be some benefits of longer-lasting first-hand experience of historical events (pushing the 'historical horizon' to more than 100 years) but to me it's like switching from a simulated annealing method (or stochastic gradient descent) to a simple local gradient descent in terms of getting society/culture/technology to adapt and find anything better than the status quo.

Worst case, such a technology serves to create an almost eternal ruling class. Best case, it results in societies with either two classes of people (those who may extend their lives longer and those who may not) or societies that tightly regulate who may have children.

Getting rid of suffering and cancer is one thing, getting rid of natural death carries a rat-tail of consequences.

> Worst case, such a technology serves to create an almost eternal ruling class. Best case, it results in societies with either two classes of people (those who may extend their lives longer and those who may not) or societies that tightly regulate who may have children.

You should read (or watch) Altered Carbon.

If we get rid of cancer, are you going to object to getting rid of heart attacks? If we get rid of heart attacks, are you going to object to getting rid of telomere shrinking, an equally pernicious disease? There won't be an immortality pill, it'll just be lots of preventative treatments that eventually add up.

> That doesn't make him wrong, but that's the personal bias he's operating under.

And in the meantime he can sell his vitamins and supplements to "make people live longer" despite zero evidence. Good business both ways.

Isn’t his argument basically that humans suck at estimating exponential growth, tending to be biased towards a linear expectation?

Wait but why had a nice article series digging a bit deeper into that https://waitbutwhy.com/2015/01/artificial-intelligence-revol...

It's kind of funny since I would say the singularity is a result of being bad at estimating exponential growth, since all exponential growth eventually hits some limiting factor and slows down, like a sigmoid.

What does it mean to be as smart as a mouse? If you specify a handful of tasks that demonstrate it, someone will be able to purpose-build an "AI" to do those things well.

Navigate a forest floor looking for food and avoid predators.

Not even that, mice are very social creatures, they make friends with other animals. They have so many other micro traits

(most humans can't do this task and survive)

Huh? Of course most humans could do this. Obviously humans who have lived their entire lives in modern human society will have serious difficulty, but this is true of literally any animal taken out of a wild habitat.

Pretty sure a typical human would die in the first 72 hours in cold climate.

So would my cat, yet feral cats survive the winter.

So basically we just need to show a computer can beat PacMan?

Basically it means having "agency." What most people are looking for when they think of "intelligence" is not the ability to master specific tasks, but to choose which tasks to perform using one's own "free will," ultimately leading to behavior that humans find novel and feel they can connect with.

But what does it mean in the context of a mouse? The mouse isn't using it's free will to decide whether to become a computer programmer or a doctor, it's responding to stimulus and environment. If an AI is trained to mimic the responses of a mouse is that intelligent?

Agency in the context of a machine seems purposefully impossible to reach - its decisions are always somehow tied back to how it was programmed to react.

The mouse reacts to stimuli and environment in a qualitatively different way than our programs do. It does continuous and essentially free-form learning of the environment around it, and engages in what looks to us as dynamic formulation and achievement of goals. In "AI" we have today, the learning is very shallow (despite the "deep learning" buzzword), it's usually neither free-form nor continuous, and goals are set in stone.

You are applying what you know about the learning process in current AI. But if you simply observed behavior between a real mouse and an AI mouse, especially if the latter was trained to mimic the behavior of the former, can you tell that they react in a completely different way?

If your AI mouse would behave like the real mouse for couple hours of observation, I'd conclude you've done a good job.

I'm not trying to make a Chinese room argument (which I don't buy), implying there's some hidden "spark" needed. I'm just saying that currently existing "AI" programs are pretty far from mouse brain, both in individual capabilities and the way they're deployed together (i.e. they're not). For instance, deep learning is to mice brains what a sensor/DSP stack is to a processor. We seem to be making progress in higher-level processing of inputs, but what's lacking is the "meat" that would turn it into a set of behaviors giving rise to a thinking entity.

The goals for a mouse are also set in stone and simple: maximize brain dopamine. Almost everything a mouse does can be described in terms of maximizing that 'reward', and that can lead to a host of other emergent behavior.

I don't see much difference between that and Open AI's engine: https://openai.com/five/. Watch some of those games and you definitely see the same dynamic formulation and complex decision-making, none of which was directly programmed.

What’s the difference compared to a human, whose decisions are always tied back to how its atoms are arranged?

I put agency in quotes because it's really a convincing illusion of agency that we're going for. In the end, I agree with those making the point that even we don't have free will.

Ultimately it just has to be able to convince humans that "wow, there's an actual thinking and learning 'being' in there."

What makes you think a mouse (71M neurons) has agency\free will? Does a cockroach (1M neurons), fruit fly (250K), jellyfish (5k) have agency? I don't think we're gonna get far by relying on a phenomenon that we can't clearly define or even (externally) observe.

Indeed. Human beings have many, many examples to suggest that we lack agency, as well. Why do addiction, obesity, crimes of passion, etc exist?

Without the baggage of the limbic system and dopamine-seeking behaviors, it's quite easy to argue that an artificial intelligence is potentially capable of even greater degrees of agency than humans.

Many people overcome these addictions, though.

That doesn't mean we lack agency, it just means agency is complicated by other factors. It's not an either-or thing.

"Basically it means having 'agency.'"

Well, in these definitions of intelligence, what one often ends up with is some combination of "deal robustly with it's environment" and a bunch of categories defined in terms of each other. That's not to say categories/qualities/term like "agency", "free will", "feel they can connect with", "find novel" and such are unimportant. It's just saying people using the terms mostly couldn't give mathematically/computationally exact definitions of them. And that matters for any complete modeling of these things.

To use machine learning parlance, such a solution would (likely) be overfitting the problem, and not generalize well. If one instead changed the setup to be: 1) Specify a handful of tasks for the AI system to complete 2) Test the performance on a _separate_ (un)related set of tasks

The test set has to be unknown to the system developers.

If the system can realize the unknown tasks without further input from researchers, in the same way that a mouse can, then we have some level of generalizable intelligence.

What is the baseline? How well does a mouse perform when placed in an "unrelated" task for the first time? The mouse also gets an explicit reward function (food, pain, etc) - does the "unrelated" task use the same reward function as what the AI was optimized for?

Also, is it ever really the "first time" for a mouse when behavior has been ingrained and tuned over millions of years of evolution? Is this different than training an algorithm?

My point is just that it's really hard to define these tasks and how to evaluate performance for a machine and a mouse.

The baseline performance would be whatever a set of mice would do in the same situation. What constitutes an "unrelated" task is quite a difficult question, we would probably need to iterate a lot on that. If we are to have "hidden" tasks available we need to come up with a lot of new task formulations/variations anyway.

I think that replicating mouse-level adaptability in an intelligent agent while allowing 'inherited' behavioral traits will already be an achievement. And probably take us quite a while.

Reply on news.ycombinator.com