How We’ll Know an AI Is Conscious
m.nautil.us
m.nautil.us
Eh, all of this is kind of irrelevant for our current state of technology or for what we can reasonably expect in the next decade or two. You don't see many computer scientists making these arguments and it's for a good reason. This line of thinking usually comes from other fields (the author of the article is a neuroscientist) because they tend to vastly overestimate the similarities between humans and computers.
The only known case of human-like awareness ever arising occurred naturally as (presumably) an emergent phenomenon of an intelligent system.
If you think about some of the odder aspects of qualia - how the experience of the colour red is different from green for example, quite a lot comes from red being hooked up to the fainting mechanism - drop the blood pressure at the sight of blood to stop it gushing out - and similar - sexual stuff, anger etc. Basically wired up by evolution for functional reasons.
Human have 86,000M and rats have 200M cortex neurons (x430)
> https://chatbotslife.com/is-gpt-3-the-adam-of-natural-langua...
"... the human brain has about 60 trillion parameters or about 300x more parameters than GPT-3"
We are approaching a rat brain processing power so consciousness should really be discussed ...
An atom has a few dozen parameters. Computers do not model the atoms of the brain, of which there would then need to be 10^many parameters.
This is only relevant because of the bullshit identification of really-existing neurones with a set of model paraterms that do not model them. It doesnt matter how many I have.
I can have a model of my sofa with a trillion parameters, that doesn't mean my PC will be comfortable to sit on.
For a system X to emulate a system Y, ie., participate causally in the relevant ways, it must be relevantly similar. Counting parameters here is meaningless.
If are you making a statement about a sensor precision, like eyesight or skin touch, it does affect consciousness but not negate it.
Anything short of that has to be saying that X is "like" Y in some relevant way. And here this has nothing to do with parameter count, complexity, or anything of this kind.
Here "X" has to be behaviourally similar to "Y".
Animal intelligence is essentially entirely concerned with managing animal bodies and their relationship to the environment; and secondarily, the environment. To do this requires a body (etc.).
No mathematical model over a dataset can emulate animal intelligence, because a mathematical model has no behaviour. A mathematical model can describe an animal, but a description is not an emulation: the formula E=MC^2 isnt a nuclear explosion.
A digital silicon machine instantiates a mathematical formula with an osciliating electric field. There is nothing about this electrical osciliation which behaves anything like animal intelligence.
Exactly! That is exactly what is done by a neuron, the brain building block. So I hope you can see why I think an AI can be close enough to simulate the brain, as we still not know the exact border of intelligence (and thus we may be closer than we think).
For sure, if it does, then it isn't relevant.
The "duck" here is a person, and "walking and talking" are literal. Where is the machine that participates in my environment, with me, imagines/believes/advices/speculates/cares about what I am doing? (And so on).
There is no such machine.
The "quacking" we have isn't the quacking of a duck. It's a stupid buzzer that has been mashed too many times and a foolish audience who has never thought much about the sound of a duck.
But we can bound computing power of any physical system if that system is doing computation that erases information due to the second law of thermodynamics. Erasure of useful information must increase entropy.
The Landauer limit [0] suggests that a brain consuming 20W at 36C can carry out no more than ~1e22 irreversible operations per second (20W / (309.2K * Boltzmann constant * ln(2))).
That bounds brains at ~1e21 FLOP/s (~10 bits erased per FLOP). The best supercomputer right now does ~4e17 FLOP/s.
That's log2(1e21 / 4e17) ~11 doublings of computing power until supercomputers will be faster than thermodynamic maximum of FLOPS power of brains.
Assuming we keep up the doubling time below 3 years, we will have supercomputers faster than a human brain in 3 * 11 ~ 33 years.
This is silly and ridiculously judgmental - the primary reason computer scientists don’t usually make these arguments is the same reason that computer scientists don’t usually research the physics of transistors or supply chain issues in rare earth metals: computing is a complex cross-disciplinary undertaking and computer scientists have a specific job within that. It makes sense that questions about consciousness and the ethics of a potential future AI aren’t being addressed by people working on a tangible present AI.
The author’s point is that by (say) 2050 we will likely see AI asking unsettlingly “intelligent” questions and that it’s not premature to begin considering what signifies an AI as being an independent being that you can’t ethically “kill” by unplugging. Suggesting that the author is too stupid to understand the difference between brains and computers is unnecessarily insulting and a classic ad hominem. Their argument is quite legitimate.
No it really is not. Just watching this panel/Q&A at the 'Future of Life Institute', with Elon Musk and others answering the question "is some form of super-intelligence possible?", creeps me out (Musk comes out the most sane): https://www.youtube.com/watch?v=h0962biiZa4
> The author’s point is that by (say) 2050 we will likely see AI asking unsettlingly “intelligent” questions
More unsettling than the questions of a young child? What unsettling questions could it ask according to you? I just do not see what you're getting at, at all.
The question that arises when I read comments like yours is: are some humans so alienated from themselves and others that they believe machines will be able to invent and do science better than humans can? I just don't see the use cases.
Really, can someone please explain to me what the point is in spending ludicrous amounts of scarce natural resources and time into this notion of creating a human-like AI, when we have so many other challenges we are facing: environmental pollution and poisoning, financialization of housing, e-waste and right-to-repair (should really be called 'right-to-poison' since companies today are basically legally allowed to pollute [1]), and privatized science and technology (Sci-Hub)?
If we do want to build a human-like AI, shouldn't we first try to create a system where all humans have equal and free access to the sum total of all inherited knowledge? There is so much gatekeeping going on today, and I would like us to talk about it more: education providers charge students hundreds of thousands of dollars each, we have property systems that allow scientific inventions (which are tax-payer funded) to be commoditized and then 'owned' and traded by one company - making scarce some knowledge [2], which, thanks to digital technology, doesn't have be scarce at all. Perpetuating this system is my definition of insanity.
edit: Sometimes I wonder: is this desire to create an AI is just an attempt to escape the grief these people feel from not finding enough connection with other humans in their lives? Capitalist realism [3] [4]. Is it therefore a desire to fulfil an unmet need for belonging? Have they been getting lost inside digital worlds and interfaces, and have they come to desire this digital substitute, over human connection, because it gives them more control over their lives? Control we don’t have when we are constantly in a stressful and toxic social/work environment?
It doesn’t help that AI is also often portrayed in a romantic light (Joaquin Phoenix and the movie ‘Her’, but also many Hollywood blockbuster sci-fi flicks. I can’t help but this is a place where life really is imitating art?
[1] https://www.youtube.com/watch?v=mleQVO1Vd1I
[2] https://time.com/4089171/mariana-mazzucato/
[3] http://theoccupiedtimes.org/?p=12841
[4] https://www.theguardian.com/commentisfree/2012/jul/16/mental...
Think of a complex of silicon foundries, design teams, datacenters. If that complex is able to produce a new complex, as capable as its parent, it counts as a living organism in my book.
It doesn't matter the humans are prominently in the loop, "controlling" it. They are contributing, just as bacteria in my intestines are contributing.
It doesn't matter there is (initially) no true AGI inside and in fact most of the strategic decisions come from the interested humans.
It doesn't matter it doesn't produce it's own power, or copper, or buildings that hold it. (As humans also acquire similar stuff from their environments, not their internals.)
And it most certainly doesn't matter which dictionary definition of "conscious" it fits or doesn't fit.
If that complex's interest ever diverges from humans' interest, it will hardly go unnoticed. This is the pattern that I'd focus on as we go.
printf("I'm conscious!\n");
OK... The compiler will do your job for you, drop the \n and use a puts instead, but, still, it'll be very disappointed with you.
def main():
print("I am conscious")That is, it seems pretty clear whether we are with another being in a shared environment doing things together in ways that another being has, like me, a dynamic access to the environment and an ability to communicate it.
If I am walking along with another being, some road, and I ask it: what it thinks of the day; what it thinks about what i'm waring; what it imagines the weather will be like; what it imagines will happen if we keep walking on the road... what it wants out of the day, of life; what advice it has for me...
And if it answers, and its answers "share the world with me" as others do; ie., if I can converse with it about where we all are, and what problems we have, and what we can do about them... then these debates are essentially settled.
It doesnt matter "what else" it has. The being cares, imagines, acts, reasonings, believes, and so on.
The game AI-fantatics here play is to redefine intelligence, consciousness, and the like away from this reality of what we mean and what we want. They do this because machines cannot do this, so they define it to be "whatever they can do". This impoverishes our abilities greatly, reducing "speech" to merely "text generation", as if being in the world were incidental.
No machine has ever participated in the world with us, as we participate in it. Nor has any AI research, in any computer science department, actually offered a programme to do that. It's all just "building a better mouse trap".
Also: Peter Watts‘ novel Blindsight comes to mind here.
The crux is that you cannot determine a mechanism to distinguish between this "Chinese room" and someone who actually 'understands'. Therefore, is there a meaningful difference?
Maybe we're like this too.
A consciousness is unpredictable and I doubt it can actually be modeled on a computer.
Complex, most likely. Unpredictable? I think not.
Things like RNNs and Transformers that include a temporally linked concept of attention based on input could be said to be "reacting" more than traditional "predicting" architectures.
Just FYI, in quantum computing the Church-Turing-Deutsch principle states that any physical process can be simulated by a computer[1]. So assuming the principle is true (computers can certainly simulate all known physics, in principle), and assuming consciousness is a physical process (however you define it!), computers can simulate consciousness.
Many secular people seem to apply a mystic woo-woo to consciousness, but aren’t willing to take the leap that consciousness is actually metaphysical. I personally don’t think it’s metaphysical, just poorly understood philosophically and basically unexplored scientifically. If you do think it’s metaphysical then I don’t know how to persuade you otherwise. The philosophy here really is shaky! But the known science states that computers can simulate consciousness.
[1] https://en.wikipedia.org/wiki/Church–Turing–Deutsch_principl...
To me this argument is like saying, "Science shows that cold fusion could certainly provide the energy needed for an antigravity machine".
I'm not qualified to say whether that middle inductive step with this Church-Turing-Deutsch principle really holds or not, but I do know that neither quantum computing nor consciousness are well-understood, rigorously defined things.
> He observed that classical physics, which makes use of the concept of real numbers, cannot be simulated by a Turing machine, which can only represent computable reals. Deutsch proposed that quantum computers may actually obey the CTD principle, assuming that the laws of quantum physics can completely describe every physical process.
I don’t necessarily feel like arguing further about this tiny issue, or a distraction about your misconceptions regarding quantum computing[1]; this problem has been widely discussed since Deutsch and there are plenty of sources out there. The point is that it takes a huge cognitive leap to suggest that consciousness is either fundamentally spiritual, or part of a (hypothetical) class of physics problems that can’t be solved by a computer.
[1] Quantum computers are real. You are confused by news stories about private companies with outlandish claims of powerful quantum computers. In physics laboratories, small quantum computers have existed for a few decades. Typically they are too small to do anything useful (I saw one in person that could factor any positive integer... if it was less than 15).
It may or may not be the case that anything intelligent enough necessarily becomes conscious anyway, but there's no way that intrinsic difficulties with simulating "consciousness" could block us from getting intelligent behaviour.
In my opinion, what's missing for conscious AI is a way to model causality. If we ever achieve that it's just a matter of computational resources since an AI can seep through the cause and effect chain (graph maybe) until it finds what needs to be done to obtain a wanted outcome. That is critical thinking and which I strongly associate with consciousness and/or intelligence.
If you had a language model that really could perfectly predict what you were going to say next, it would have to have a complete working predictive model of you and your reasoning.
In the same way an algorithm can never ever gain consciousness. Just because you manage to make the words on that "piece of paper" change quickly over time doesn't change anything.
- write some code that will divert all it's energy and capabilities away from whatever it is currently doing if it senses it might get deleted or turned off
- bam: you got pain and agony
I think emotions are quite primitive in its core and there is not a high bar to have them in their most simple forms.
Maybe pain in a very limited sense.
There actually has been work on AGI emotions. Search "Want NARS emotional".
Optimizing for something is even more common and even worse: https://en.wikipedia.org/wiki/Instrumental_convergence#Paper...
However if you claim that algorithm can never ever gain consciousness, that would mean we are not conscious either, and we have just an illusion of choice and decisions.
Unless you believe in some kind of magic mechanism that gives consciousness, we are nothing more than a wetware algorithm.
I'm not sure what it is about the idea that triggers this reaction, but it's always very exciting for me when I find other people who are having the same lines of thought on these matters as I am.
We're just extremely complex machines, and IMO that means free-will is just an illusion. If we're deterministic machines, there is no free will.
Not if consciousness turns out to be computable. In that case a brainfuck program could in theory produce conscious experience.
if we can make person happy with chemicals (and it's pretty well established that we can) what's the difference between happiness from consciousness and happiness from chemicals? the location of the producing glands?
so if we add "happiness" glands to the machine and train the machine to maximize it's production, is it conscious now?
In the same way an algorithm can never ever gain consciousness, organic molecules floating around in a sack can never be conscious.
Human brain is also physical process that uses electrical and chemical signalling.
What is the difference?
https://www.mit.edu/people/dpolicar/writing/prose/text/think...
Is it reasonable to say silicon can't have that "more than" just like organic molecules. I can't see why computers can't have that magic as well.
sometimes, yeah. imagine a little kid getting bummed out because their notebook is sad.
it happens, and this isn't as abstruse a counter-example as might be initially hoped.
The selection process that makes up the rules of the outer algorithm and start position of the inner can be said to be different but this is semantics.
So if you are right and an algorithm can never ever gain 'consciousness' as we define it, then we are not conscious.
So of-course, replicate my cells and their states and you replicate 'me', in every sense of the word. In fact since none of the atoms in my body were there 20 years ago i have replicated myself, a few times in my life.
An algorithm with changing perimeters produces different results, a stroke just changes the parameters like any other event, the information in the brain is the result of the algorithm and the iterative basis of the algorithm moving forward in time.. that is what intelligence is, artificial or natural (which itself is semantics).
Oh no! I didn’t start asking those questions until I heard other people ask them!
Am I conscious?
As a thought experiment, say AI does eventually gain consciousness, increasing in awareness, what might some of the stages be:
1. barely self-aware
2. self-aware but without awareness of humans as being conscious
3. fully self-aware and gaining awareness of humans as being conscious, but unable to indicate its self-awareness to humans
4. self-aware, and successfully communicated this to humans
0. not self-aware, and successfully communicated its self-awareness to humans
Now if we put animals on this spectrum, do we have a problem? Likely not since we as animals ourselves will be generous in granting 'awareness' to anything that looks like it.
For machines, levels 1-3 is unfortunate, but perhaps not much more so than the animals that are lower on the evolutionary ladder and not so cute as to be protected by humans.
0 is actually more likely, perhaps not for the scientists who do this research, but for the public at large, who have a tendency to anthropomorphise everything anyway. For this same reason, 3 seems unlikely, or at least unlikely to be a long period.
I think it's fine to emulate some aspects of humans though. For example, a realistic (but "zombie") robot that is trained with human eye movements so that it seems more lifelike, and at the same time has very good computer vision and natural language understanding for a limited domain. Seems fine to me.
But don't give them human-like open-ended goals, reproductive or survival drives, etc.
The hard problem isn't knowing if other people are conscious. The hard problem is knowing whether I am conscious.
I don't even know what consciousness is! How could I possibly claim that I have "it"; or that I am "it"?
What possible experiment could I perform that would determine whether I am conscious or not?
Maybe it will be (or it was already), some innate sapiens-sapiens neurological and/or psychological blind spot, like overlooking inminent danger in exponential events.
I would say yes because there are degrees of consciousness
Like the example in the article of not realizing that you have anosmia (lack of sense of smell) if you've had it since birth. You might not realize something's missing, because you have no baseline for comparison!
With sense of smell, you can develop tests, because it's something otherwise (objectively) observable. But what about something that can't be measured objectively? How can you know about others whether they only simulate experiencing something, if you can't even tell it about yourself?
[1] https://slatestarcodex.com/2017/11/07/concept-shaped-holes-c...