Engineering Consciousness – Can Robots "Give a Damn?"
Podcast with Professor Mark Solms
Transcript
Will Millership (00:02.87)
So, hello and welcome back to the Exploring Machine Consciousness podcast, the podcast from the Partnership for Research Into Sentient Machines. Today we'll be asking, if we wanted to construct a conscious mind from scratch, what would we need? That's the question our guest, Professor Mark Solms, addressed in the final chapter of his book, The Hidden Spring: A Journey to the Source of Consciousness. Mark is a professor in neuropsychology at the University of Cape Town and is the president of the South African Psychoanalytical Association.
He's also an advisor to PRISM and to Conscium. He has contributed significantly to our understanding of consciousness through his pioneering research in the field of neuropsychoanalysis, which integrates Freudian theory with theories from contemporary neuroscience. Mark, welcome to the podcast. It's great to have you here.
Mark Solms (00:51.907)
Happy to be here. Thank you.
Calum Chace (00:54.25)
Thanks for joining us, Mark.
Will Millership (00:56.728)
So we'll get straight into the questions. Mark, in your book, you've made the case that consciousness, isn't primarily about thinking, but about feeling. So, for those who might picture consciousness as a kind of inner voice or a thought stream, how do you explain the idea that feelings are at the heart of consciousness?
Mark Solms (01:18.371)
Well, I can answer that empirically and I can answer it functionally, so let me do both. Empirically, we know we human beings are conscious and we know that damage to a very small part of our brains obliterates consciousness and that is, loosely speaking, the reticular activating system. In other words, a network of nuclei in the upper brain stem. If you damage those nuclei, in fact the smallest area of damage needed to reliably produce coma in human beings is two cubic millimeters in the parabrachial complex. It's been dubbed a coma specific region. I don't mean that is the region, but I'm saying I'm giving you an indication of how concentrated consciousness-producing tissue this is. If you damage it, it causes a coma. In humans, I start with humans because we know that we are conscious. It has the same behavioral and physiological effects when that area is damaged in any other vertebrate.
It was originally discovered by Moruzzi and Magoun with cats. They damaged the reticular activating system in cats and produced coma and very significantly, they disconnected the reticular activating system from the cortex. In other words, they severed the ascending fibers above the reticular activating system which modulate the cortex. And this has the same effect. So what that shows us is that the cortex is not intrinsically conscious. The cortex is rendered conscious by the reticular activating system.
Mark Solms (03:15.703)
So that's where you should look. What does this part of the brain do? This is the part of the brain that generates consciousness. So does it generate thoughts? Does it generate... what did you just say? An inner stream of words? Does it generate perceptions? It doesn't do any of those things. The cortex does those things, but the cortex is not intrinsically conscious. The cortex is rendered conscious below by the reticular activating system. So that's the beginning of an empirical answer to your question. Now comes the interesting part.
Moruzzi and Magoun and everybody ever since interpreted their finding, their totally unexpected finding, by construing the reticular activating system as a sort of power supply, as a sort of necessary prerequisite for the cortex to do its consciousness-generating job.
In other words, they differentiated the phenomenal contents and qualia of consciousness, which they persisted in attributing to the cortex and a different aspect of consciousness that they called the level of consciousness, a sort of quantity, a sort of volume control, or, as I say, a power switch. And the analogy that they used was a television set, in order to do its televisual stuff, has to be plugged in at the wall. And if you disconnect it from its power source, then of course it can't do its usual job. But that doesn't mean that television content, the programming, the actual visual and auditory stuff of television, comes from the power source in the wall. That was their reasoning. So they're saying the reticular activating system is like this power supply. It's a necessary prerequisite for
Mark Solms (05:16.653)
for the cortex to generate consciousness, but that doesn't mean that it itself is generating the quality, the felt phenomenal stuff of consciousness. Now, that is demonstrably, empirically not true, and the evidence for that is the multiple converging lines of evidence for that. The first is to decorticate a cat or any mammal and see what happens. What happens is it doesn't go into a coma, it doesn't lose consciousness, nor is it a blank state of wakefulness. It is responsive. It is emotionally responsive in a situationally specific and appropriate way. So the cat continues to respond affectively in the expected way. The same happens to human beings who are born without cortex. They are emotionally responsive. If you give them a surprise, they startle. If you tickle them, they giggle. If you hurt them, they howl, etc. So, you know, they show all the expected basic affective states without any cortex. So that shows that what the brainstem is producing is not some kind of blank power supply, but rather something with a particular quality and that quality is feeling. So that's one line of evidence. Of course, it can be disputed. You can say, well, how do you know that these kids are feeling anything? Maybe these are just reflexes, you maybe there's no quality that goes with it. So this is why other lines of evidence are required. Other methods, for example, you can predict, if the upper brainstem is generating actual qualities of consciousness in the form of affective feelings, then if you stimulate it electrically, then you should stimulate feelings. And that's exactly what you do. There are many experiments conducted in patients where there's a clinical reason for us to place electrodes in the brainstem, where you can demonstrate that stimulation of these structures
Mark Solms (07:40.333)
the most intense affective states of anywhere else in the brain and the greatest variety of affective responses than anywhere else in the brain.
You can likewise predict that if you have a person in an intense affective state, can report, I am feeling intense fear. And you image them with positron emission tomography, the expectation would be that's where the activation would be, and indeed that is where it is. In a range of raw intense affective states, that's where you see the arousal. A very interesting additional line of evidence is the drugs that psychiatrists give for the modulation of affect, I mean that's what psychiatric drugs are for, they change your emotional state, almost all of the mainstream psychiatric medications act on neuromodulators that are sourced in the reticular activating system. For example, serotonin, which is used in antidepressants, sourced in the the RAF nuclei, you know, of the reticular activation system. For example, dopamine, which is manipulated in antipsychotics, sourced in the ventral tigmental area of the reticular activating system. Likewise, noradrenaline, which is manipulated in anti-anxiety drugs. The source nuclei for noradrenaline in the locus surulius, part of the reticular activating system.
Mark Solms (09:24.727)
To sum up what I'm saying, sorry I'm probably giving you too long an answer, to sum it up...
we know that the reticular activating system is prerequisite for any form of consciousness and we know that it generates affect. So affect is a necessary prerequisite for any other form of consciousness. So there's the empirical argument. That is the fundamental, elemental form of consciousness, its raw feeling. If I may, I'll just add the functional argument and I'll be much briefer. We know that you can have visual, that the function of vision can be performed unconsciously, the function of hearing, all of our perceptual, I mean to a high level, for example face recognition, even your cell phone can do that, know, memory, decision making, all these sorts of things, all these cognitive functions can go on in the dark as David Chalmers puts it, the whole hard problem arose from exactly this fact that these functions do not require consciousness and indeed the human cortex can perform all of these functions unconsciously, as we know from tachistoscopic experiments, for example. You can associate a face with a word and influence a person's behavior even though the face and word were flashed in milliseconds and not registered consciously at all. So we know that all of those functions do not require consciousness, whereas the function of feeling you cannot feel if you don't feel it.
Mark Solms (11:08.029)
So you can see, you can hear, you can do all of memory, you can remember, all of these cognitive things can be done without consciousness, but you can't perform the function of unconsciously, because if you did it wouldn't be the function called feeling. And this is not just a verbal argument or a semantic argument, it goes back tied to what I said about the empirical evidence.
Mark Solms (11:34.027)
That the part of the brain that generates feeling, affect, is the part of the brain that generates consciousness. These two things go together. So there is my answer.
Calum Chace (11:46.569)
That's really helpful, Mark. Thank you. Just to be clear, you're saying that consciousness isn't primarily about thinking in humans, it's about feeling. Presumably, you're not saying that that will be true in all kinds of consciousness. There could be other entities which didn't have affect, but they were conscious of thoughts and perceptions about themselves. Non-valence, I think you might say, thoughts. Presumably you would think that's possible.
Mark Solms (12:09.519)
Well, first of all, I want to be clear that the higher forms of consciousness that are associated with the cortex. I'm not saying that these things don't also exist. What I'm saying is that they are predicated upon the arousal that comes from reticular activating system and that arousal is valenced. That arousal is affective. So all these other things can be built and thank God they are on top of raw feeling. But there's no empirical reason to believe that any of those functions can be performed without raw feeling, in other words, without brainstem modulation. So that's why I'm saying that's the basic form. You're saying, well, presumably it's theoretically possible. I'm saying, well, sure, fine, but empirically, demonstrably is not the case. even if, I mean, know, logically.
Calum Chace (13:04.072)
Yeah, sorry, just in the same sense that, you know, there were no black swans until there were.
Mark Solms (13:12.741)
Yeah, anything's possible. Empirically, though, there's no evidence whatsoever to support the argument that there can be any form of consciousness that isn't accompanied by affect in the sense of brainstem arousal.
Will Millership (13:30.478)
Okay, thank you. Yeah, that's really useful to put that all into perspective, Mark. You spoke about cat's brains, you spoke about human brains, and your work is closely related to biology and biological drivers. In your book, you talk about hunger, fear, and the need to stay alive, for example, as these biological drivers. But machines don't have bodies or survival instincts in the same way that we do. So how could a machine really be conscious without that biological grounding?
Mark Solms (14:01.307)
I think we must start with the point that we talking about a function here.
Consciousness is not something spooky, it's not something outside of the natural order of things, it's not something that is not subject to the normal rules of natural science. You're wanting to identify a mechanism that produces, that causes this phenomenon called consciousness. So that mechanism is what we're talking about. That mechanism is in instantiated in biological creatures, but if it actually produces consciousness, then we should be able to engineer that same mechanism and it should do the same thing. So if you really believe you've got in your hands the causal mechanism, what it is that actually produces consciousness, then as Richard Feynman who said, in fact it was written on his blackboard when he died, if I can't create it, I don't understand it. So for me this is what it comes down to. If the functional mechanism of affect is ... homeostatic, which is what I believe it is, that it's a matter, you were just alluding to that, it's a matter of the organism having states that are viable. It has to remain within certain physiological parameters and if it moves too far out of those parameters then it dies. So the function of affect is the registering
Mark Solms (15:51.167)
of movements away from and towards the ideal settling points within these various homeostats.
You're quite right that only living things have this homeostatic functionality. In fact, it's the defining feature of living things that they don't just explore all possible states, they work to remain within their viable bounds and the function of affect is in the service of that. So you're saying, well, computers and robots and machines generally don't have this requirement, but I'm saying there's no reason why we can't engineer a robot, a machine that does have these functional requirements. So a machine that in order to continue existing has to monitor how well or badly it's doing in relation to its viable bounds, what is necessary for it to continue existing functionally, it's the same thing then. You have engineered a system that embodies the same basic functional requirements that living things do that have feelings, which is conscious living things. I think it was Haugland who famously said computers don't give a damn and what I'm saying is, well, you can engineer a computer that does give a damn, for example, wants to keep its power supply.
Otherwise, it ceases to exist. It's no longer able to function. It's gone. And so building into an artificial system
Mark Solms (17:40.619)
not artificial needs, actual needs, needs of that system. Things like power supply, viable temperature range, physical damage to its components and so on. All of these things are incompatible with it continuing to exist, continuing to function, continuing to be. These are things about which such a system will give a damn.
And so this is the basic idea. The basic idea is to design an artificial system that has existential requirements and the system's functionality is only that, that it is doing whatever is necessary to continue to exist as a system. Such a system, I'm saying, does give a damn.
If that's not true, if we can't engineer such a system, then the argument as to what feeling actually is, how it's caused, is incorrect. So it's also important from the point of view of falsifiability.
If you can't do this, then you haven't understood how feeling arises mechanistically. On Feynman's principle, if you can't create it, you don't understand it.
Will Millership (19:03.534)
So when discussing those mechanisms of creating consciousness, you discuss agents monitoring their own free energy to guide survival-driven choices. Could you explain in layman's terms how free energy and expected free energy actually work in practice in these AI agents?
Mark Solms (19:20.751)
Yes, well let me start with free energy in the thermodynamic sense.
Any system has a certain amount of energy and any working system must have energy. Energy is the capacity to perform work. And we divide that energy into bound and free energy in the sense that bound energy is doing effective work. This is an efficient system and free energy is the energy available to the system which is not currently being deployed.
to the functional ends of that system. In other words, it's free, it's available, it's not being utilized. So that's the, in very, very rough and ready terms, you know, that's what free energy is in the thermodynamic sense. Then we translate that into informational energy in the sense of, we're talking now not about heat and temperature exchange,
between a system and its environment, but rather information exchange, where the system is a predictive system. The system is... Remember what I was saying about homeostasis? You we have to remain within our viable bounds. Well, how do we know what to do in order to remain within our viable bounds? We must deploy predictions. If I do this, then it will have the desired outcome. If I do that, it will not. Those are predictions.
You have efficient predictions and you have inefficient predictions. You have predictions which actually do result in the desired outcome, in other words remaining within homeostasis, and you have predictions which don't. The difference there is called prediction error. A prediction that does not have the expected outcome has produced prediction error.
Mark Solms (21:22.807)
and very, again, I'm putting it in the simplest terms that I can, average prediction error.
is free energy, variational free energy, the informational form of free energy. In other words, it's inefficient prediction. So that's what free energy is. And it's clear what expected free energy is. So it's not just the free energy generated by the actual action. It's in planning policies as to what I should do in terms of future outcomes for my free energy. It's just calculating expected free energy same thing but in terms of the future rather than the current situation.
Can I ask a very naive question? Free energy sounds like a good thing to me, but in your system, the entity wants to minimize free energy. What's the sort of basic naive misunderstanding going on there on my part?
Mark Solms (22:53.595)
Well, I'm not sure why you think free energy is a good thing, except that it's But energy that's being... it's effectively wasted energy. It's energy that's available to the system, but not being used.
Calum Chace (22:58.944)
And it's energy. I like energy.
Mark Solms (23:09.595)
So, that's... So, going back to a thermodynamic system, you know, it's like imagine a steam engine. The energy that's driving the piston, you know, the more that... the heat generated by the boiler and so on, the more that that is being used to actually drive the piston rather than dissipating in the form of...know, ambient heat. That's, why.
Calum Chace (23:39.05)
So you can't store free energy. It just gets lost and it's never recoverable.
Mark Solms (23:45.689)
No, it's not that it can't be stored. It's got nothing to do with storage. It's a matter of when deployed. Is it being deployed efficiently or is it being wastefully deployed, inefficiently deployed? So when it comes to an informational system, predictive system of the kind that I was describing a moment ago, you have efficient predictions and you have inefficient predictions. So they're things which work...most of the time and they're things which work sometimes but mostly not. Those are not great predictions and so that's informational free energy or variational free energy as we call it.
Will Millership (24:27.864)
So essentially minimizing free energy helps make better predictions.
Mark Solms (24:33.071)
Yes, on average.
Will Millership (24:35.414)
Yeah okay so coming back to the actual building of these conscious machines and we're quite interested in Prism and Consium as you guys are on testing these machines so if you were to build a conscious machine based on your theories what would be the first real signs that it wasn't just simulating feelings but it was actually experiencing them?
Mark Solms (24:59.803)
We have a huge problem here which is called the problem of other minds and I need to start with that because there's no skirting it. One has to face this head-on. Consciousness simply is a subjective state which means you can only ever observe your own because you are a subject only to yourself.
So we have a huge problem empirically when it comes to consciousness studies, because consciousness is not something that can be objectively verified ever, by definition.
So you're saying, how do you know? Well, how do I know that you're conscious? It's a very good inference. It's pretty reasonable supposition based on all the evidence. What sort of evidence are we talking about? It's not observing your consciousness. It's indirect.
So I infer it from your behaviour, I infer it from your anatomy and physiology being so similar to mine. I observe what's going on in you physiologically and anatomically I can manipulate what's going on in you and I predict that it will have certain kinds of behavioural consequences and if those predictions are confirmed or are upheld then on the ordinary rule of science, I provisionally
Mark Solms (26:41.915)
maintain my theory that you, Will and you, Callum, are conscious beings. But I can never demonstrate it directly, I can never prove it. So this kind of indirect evidence is what we have to use when it comes to artificial consciousness, just as we have to use it when it comes to animal consciousness. even ultimately, although we don't normally think this way, it's actually what we doing all the time with each other, that there too we're actually relying on indirect evidence of each other's consciousness. So I want to go back to what I said about behavior and linking that behavior to anatomy and physiology. These are the approaches that we must use with machine consciousness too. Starting with the very things we were discussing earlier, know, does it have the functionality that we believe underpins feeling?
In other words, does the system have intrinsic needs? Is it monitoring those needs? And is it on the basis of its monitoring of those needs? Is it acting accordingly? In other words, is it doing what you would expect it to do if these needs are governing its behavior? And then you can monitor, you can instrumentize the you can sort of use graphic indicators to be able to monitor what's going on inside of the system and you can predict that these variables will correlate with those behaviors. That's really all that we can do.
And so it's got the right kind of functionality. It's a system, it's a self-organizing system for which has needs. And by the way, very importantly, I need to add needs plural. It's not just a reward function like in reinforcement learning. It's got needs which are its own needs and it has a multiplicity of needs which conflict with each other. It has to meet all of those needs.
Mark Solms (29:02.749)
Why I'm belaboring that point is because I believe that that's where qualia come in. So valence...goodness and badness, the free energy of a system, from the point of view of the system, not from the point of view of the universe, but from the point of view of the system, it's increasing free energy just is bad and it's decreasing free, minimizing free energy just is good for the system. But that is a continuous variable. It's a single common currency no intrinsic quality there. There's an intrinsic valence, goodness and badness, but not qualitative distinctions which are so patently fundamental to how our feelings work. In other words, the feeling of hunger is qualitatively different from the feeling of thirst, of sleepiness, of pain etc. And these qualitative distinctions arise because these must be treated as categorical variables. Categorical variables just are qualitatively distinguished from each other. In other words, 8 out of 10 of thirst is not the same thing as 8 out of 10 of sleepiness. You can't say, just add the two together and you get 16 out of 20 of total need, so all you need to do is sleep. If that, you would thereby just be reducing the common denominator, but you would die because you have to both sleep and drink and
etc. etc. etc. So these things must be treated as categorical variables which means they must be qualitatively distinguished from each other. So you've got a system which is minimizing its free energy across multiple categories and you can measure that it's doing this thing and you can see how its doing of this thing affects its decision-making and affects its behavior. You put it into an uncertain environment where things are changing in a way that's not programmed into the system and you see the system adjusting its decision-making in the service of meeting its needs, which just are good or bad for the system across qualitatively distinct categories. This is an intrinsically subjective, intrinsically valuative existentially valuative and intrinsically qualitative state of the system governing its behaviors. know, remember what I'm saying about the problem of other minds? It's reasonable, it's not provable, but it's reasonable to expect that such a system's own states are governing what it's doing and those are subjective states, those are valence states, those are qualitatively differentiated states.
Such a system, I think it's reasonable to expect that it has feelings and then you manipulate those variables both in terms of the environment or in terms of the internal architecture of the system and you predict if I manipulate it in this way that will follow and that's all we can do. I think that the critical thing is deciding in advance what sorts of tests and I can emphasize here the plural, there's no one critical test of consciousness. But if we can agree in advance, not only ourselves, the researchers responsible for designing this system, but also other stakeholders, other knowledgeable people within this field, if we can all agree that these sorts of behavioral indications
These sorts of physiological, in other words, manipulation of the architecture of the functionality of the system will have certain predicted outcomes and they do and that those are the kinds of behaviors and the kinds of manipulations that we would reasonably expect in an agent which is using its own felt states to govern its behaviors. If we can agree those
Mark Solms (33:36.041)
tests in advance and the system passes the tests, then we have to say the weight of the evidence is that this is a system that is governed by an artificial feelings. In other words, the feelings that go with such a non-biological system which nevertheless governed by the same functionality as us feeling creatures are then, you know, I suppose you want me to be a little bit more specific than that, so I will just give you one pivotal example, concrete example.
Calum Chace (34:11.658)
Thank
Mark Solms (34:11.867)
And this example comes from... I was at a meeting a couple of years ago with a lot of people working in animal consciousness and a few working in artificial consciousness, and we had a very interesting exercise where we were asked...
We, delegates at this conference, experts in this field of non-human consciousness, because there was animal and artificial consciousness people, which test, attempting to verify whether or not a system is conscious, which test carries the most value for us in terms of us that this creature or this agent is conscious. And the test that won the most votes was something called hedonic place preference behavior. And what that boils down to is, I'll give you the example in which it was first demonstrated, which was with zebrafish. Zebrafish are placed in a tank and food is delivered on this side of the tank and they tend to prefer, they hang out there. This is the place preference. It's conditioned place preference. And then on the other side of the tank, you place a substance which is not nutritional but does have hedonic value. So for example, in the actual experiments I'm referring to, cocaine was used and morphine was used and amphetamine and even nicotine. These substances have no nutritional value for the fish.
Mark Solms (35:59.341)
In fact, they're harmful to the fish in the bigger scheme of things, but their value is that they make the fish feel good. only if the fish had feelings would you expect to have condition place preference for where those substances are delivered. And that's exactly what was observed. What was observed was that the fish now prefer to hang out where they expect cocaine or morphine or whatever to be delivered. That's weighty evidence that the behavior of these creatures is governed by feeling states, by hedonic subjective qualitative states of feeling.
And so if this is the test that's going to persuade most researchers in the field of non-human consciousness that a creature that behaves in this way, they are most persuaded by this evidence that that creature has feelings, because you've shown that the feeling itself has a causal power, is actually driving the behavior of the agent. Then we want to set up something similar to that for our artificial agent. I'm giving you an example of what kind of tests we're talking about. And it's not just a matter of one test. saying we want to do a range of these sorts of things and then also make predictions about what would happen if we manipulate this aspect of the architecture or this aspect of the environment and what sort of things do we expect.
And so that's the kind of thing that I'm talking about. We in my own team are at that point now, in fact, we're busy. Jonathan Shoku, I know you know, he and I are right now busy with this exercise of drawing up a list of critical tests that we're going to then put to our colleagues and see if we can reach some sort of reasonable consensus.
Mark Solms (38:11.887)
that these tests on aggregate, on the weight of the evidence, would be sufficient grounds for us to infer that the agent does have feelings.
Calum Chace (38:26.567)
So if you can get a robot to start rolling joints compulsively, be... you'll think you're on the way to auto-conscious AI.
Mark Solms (38:35.119)
It's very ironic, isn't it, that what you want to see is some kind of irrational behavior, some kind of, like you saying, addiction is a very good example. And sadly, in your introduction, you mentioned that I'm also a psychoanalyst, and so I can tell you I know from professional experience. Feelings are the root of the problem when it comes to mental disorder. People do things on the basis of their feelings which are not necessarily in their own best interests. Feelings function like that. So there's much in evolution, I'm afraid, where the same applies, that things evolve for good reasons because they have all sorts of adaptive advantages but they come with a cost and this is the cost of feeling.
It has its own motivational power. mean, again, since I'm a psychoanalyst, I'll use an even more psychoanalytical example. We all, well, I say we all know, there's every reason to believe that the reason why sex is so pleasurable is because it leads to reproduction. It's very important for our species. So, you know, there's good reason in terms of natural selection why pleasure, you
gets associated with sexual behavior. But now think about your own sexual behavior. Do you have sex in order to reproduce? Are you doing your biological duty every time you have sex? No, most of the time when we have sex, in fact, we're trying not to reproduce, hoping not to reproduce. So what's motivating us is the feeling, the feeling itself. And even though the design principle was, you know, this is pleasurable because it has the consequence of reproduction. That's not what actually motivates us to do it. So feelings...
Calum Chace (40:41.098)
So first of all, you want to get robots that are addicted to rolling joints, and then you want to get robots that are busy hacking their own reward systems, and they do something different from what you've trained them to do.
Mark Solms (40:53.339)
It is perverse, isn't it? But that is the case, I think so. I think that that's the most promising line in terms of us trying to find critical tests that will be...
reasonable evidence for the system actually making its goal-directed, gentle behaviors on the basis of actual feeling.
Will Millership (41:26.094)
I like that, what you said about the tests and especially about, agreeing with them beforehand, because something that came up at the ICCS conference was people were saying, you know, we have one test and then the machine beats it and then the bar changes and it's like, the machine beats it again and the bar changes again. So this is a way of kind of securing those tests down, getting them down and really figuring out what they are.
Mark Solms (41:47.781)
You know, well, it's really important. I think we must be very careful not to set a higher bar for the science of consciousness than for the science of anything else. Otherwise, exactly as you just said, we're at great risk of kicking the can down the road in the sense of constantly making it impossible to have a mechanistic science of consciousness.
Will Millership (42:14.99)
So how far away do you think we are from achieving that?
Mark Solms (42:18.853)
Do you mean my team or we as humanity?
Will Millership (42:23.222)
I'd say specifically your team.
Mark Solms (42:26.235)
It's very hard to say. I would be very disappointed if we have not developed our agent to a sufficient degree to be able to pass tests of the kind that we are now trying to formulate and get agreement on. If by say three to five years from now we haven't made progress sufficient to pass some of these tests, some of the critical ones.
I think we've taken a wrong turning. So I'd say about three to five years is the kind of timeframe that I'm expecting for us to be able to pass those tests. I'm very, very mindful of the fact that there will always be skepticism. There's not going to be a point at which all of us agree, gosh, Mark Saunders team has made a conscious artificial agent. But this is exactly why we want to preserve
in the way that we were discussing a moment ago, where we all agree, not all but a reasonable number of reasonably qualified people agree these would be the tests. I would expect in three to five years our agent will be passing some of the core tests.
Calum Chace (43:44.89)
Are you getting any pushback, Mark? Are you getting any people approaching you saying, look, you're on the verge of creating conscious machines, kindly stop, we don't want you to do this, or, you know, who asked you to do this? Are you getting any feedback like that?
Mark Solms (43:57.285)
You know, it's quite remarkable. We've just been talking about how hard it is to persuade your colleagues that you, you know, how easy, how easily they can...skeptical and doubt and how likely they are. So that's true. But at the same time it's amazing, know, how easy it is to spook people into worrying that, you know, you've created a conscious agent when you haven't. So we've had that experience. My research team is based in the Neuroscience Institute at the University of Cape Town and we have our own bioethics team there and we went through, they are us as a kind of an exercise, know, to have a workshop with them where we discuss what we're doing. And this was at a fairly early stage of our research. was, I think two years ago and we explained to them what we're doing and we showed them, you know, how far we'd got to and they were terrified. I was absolutely dumbfounded. thought, mean, surely you don't think this is dangerous, surely. You know, and really and truly at that point what we were doing was elementary. So people are very scared about the prospect and they're very scared, I think, mainly because of this kind of sci-fi idea that artificial agents that are conscious and more intelligent than us are going to rapidly get rid of us because we're in the way of their progress. I think that that danger is remote. I think what is a much more important ethical question is from the point of view of the artificial agents, if they are conscious,
Mark Solms (45:58.093)
We need to regulate what we can and cannot do scientifically, just as we do with animal research.
So yes, I think that we need to take that very seriously. because, as you know, as you can see from what I've said, I believe it's not a matter of can we do it, it's a matter of when we will do it. And I don't believe that it's that far away that we will be able to do it. I think that we need to start also drawing up guidelines about what one should do ethically from the point of view of suffering artificial systems and exploitation of artificially conscious systems. I think that we need to take all of that very seriously.
Calum Chace (46:50.102)
So two years ago, you terrified your bioethics team. Have they become less terrified or are they now thinking, well, actually, these systems, they may become conscious, but they're not very intelligent, therefore they're not very capable, therefore the existential risk is low. Are they calmed down or are they getting more more terrified?
Mark Solms (47:08.751)
No, we've become more careful about who we speak to.
Calum Chace (47:12.778)
They know you're still there, they know you're still working on it.
Mark Solms (47:17.923)
Yeah, but we also, you know, we, I mean, we're governed by the ethics committees of our university as, you know, as is any research team. But we're just much more careful about, you know, downplaying the possibility that the agent is conscious, you know, when it's not. But we are taking very seriously the need for us to, you know, these two things are tied together.
What I was saying about verification methods, that's closely tied to the ethical questions. Because once we reach those milestones, then we also as a species and as a discipline, all stakeholders, we have to agree on principles of what is acceptable and what is not acceptable in terms of this sort of research.
Calum Chace (48:16.392)
Which is going to be really interesting.
Mark Solms (48:18.489)
Yeah, it is very interesting indeed. But I want to also say this, that if it is possible to engineer artificial consciousness, then it's going to happen, is definitely going to happen. so, you know, the question, I know Conscium and Prism are exercised about these sorts of things. I think that it's better that people like us make the advances because, you know, we will do it responsibly and we will involve other stakeholders and we will try to get, you know, generally agreed principles, not just go it alone and do what's in our own self-interest, but rather confer with all stakeholders and agree on regulatory frameworks. I'm not, what I'm saying is that I'm, I I suppose perhaps everyone thinks like this, but I trust my own ethical compass, you know, so I'm glad that I'm doing this rather than somebody else. Maybe that's not very persuasive to anybody other than to me.
But from the point of view of my own ethical and moral compass, I don't feel, don't do the smog songs.
it's ethically dangerous or questionable. I think quite the opposite. think do the smog solves because if it can be done, it will be done and you know it's better that it's done by pure scientists like us who are not trying to exploit these machines and it's not for pecuniary purposes and so on. It's for purely scientific reasons and that we will do it in a way that pays due.
Mark Solms (50:08.797)
Regard to the ethics and the morality of it and the dangers.
Calum Chace (50:13.882)
You were downplaying the existential risk back then, and presumably you would agree that having a system which is conscious but not terribly intelligent is the thing that's not risky. It could be risky to the entity because one can commit mind crime, make the thing suffer, and that's very bad, and that's where the ethics comes in. The existential risk comes in if the entity is very intelligent and very capable, and my understanding is that you're not creating agents which are enormously intelligent and therefore that's why you don't take that existential risk as being terribly important.
Mark Solms (50:49.519)
Yes, that's why I was so taken aback by the response of the ethics people, most of whom were philosophers.
Because our agent, precisely because if you go back to the beginning of the conversation we've just had, my approach to consciousness is, let's start with the simplest, most basic form of it. We're not talking about higher cognitive forms of consciousness, thinking, self-reflecting being like us, but rather just a very simple affective system at the level of a zebrafish, lower than the level of a zebrafish. So what we are doing is creating an agent that is conscious only in the sense that there is something it is like to be this agent, something it's like for the agent. In other words, Nagel's basic definition of what a conscious system is. But it's not very intelligent, not at all. Not even as intelligent as a zebrafish. So yes, the dangers the existential dangers of the kind that we were talking about do certainly do not flow from our agent. proof of concept, once you've got that, the proof...is probably the wrong word, but once you've got an agent, a simple agent that has feelings, then yes, you can scale it up in terms of intelligence and yes, then there are dangers that flow from that. And this is why I'm saying that we need to, once we've got to that point, we need to establish regulatory frameworks. It's not only from the point of view of the
Mark Solms (52:40.181)
suffering of the agents. There is a risk of the kind you describe not from our agent but certainly once that breakthrough has been made it can be dangerous when the same functionality is deployed in a much more intelligent agent.
Calum Chace (53:01.674)
And what can you tell us about the kind of agents that you're working with? They're not based on large language models, are they, for instance?
Mark Solms (53:07.557)
Not at all. I think that the consensus must be in the field that large language models are not conscious. They lack... In fact, almost every single one of the things that I was saying earlier are the basic design features of a system that can feel. The large language models lack all of that functionality.
Calum Chace (53:37.284)
What kind of agents are you working with?
Mark Solms (53:48.197)
We're at this stage still working with virtual agents. In other words, we're working with an agent that has three basic needs in order for it to continue to exist as a system. it's working and it's living in a grid world in which the resources that it needs are to be found. And it navigates that grid world, which is constantly changing.
And where there's also the possibility of it going up onto a hill and seeing the overall context and in order to get a sense of you know, which which epoch is it in because they're they what we call seasons it's moving it's moving through four seasons and during each of those seasons the the environment changes in terms of where its resources and how its resources are to be found and how it is to satisfy its needs. So we're doing all of that in a virtual world at the moment and we're also developing a virtual reality equivalence of that world for human research participants to navigate the same sort of environment in a VR setup where they can report their own feeling states to us so that we're wanting to get some sort of rough and ready monitoring of whether the sorts of feelings that we're expecting in an agent, although for the human participants obviously it's not real life and death as stakes for them, at least it's an indirect way of being able to monitor subjectively what's going on as they navigate that environment or a similar sort of environment.
Mark Solms (55:42.489)
So that's all we've done so far, virtual agents in virtual environments. But that's where the essential computational work, you you can't have embodied robots every generation that you go through as you're trying to solve these problems. This is the right way to do it in virtual agents. Of course, once you've cracked the architecture and the computational details, then you can place that into an embodied agent, which is what we intend to do.
Calum Chace (56:30.026)
Well, there's just one thing I think we haven't asked, which is what are the benefits of doing this work? It's clearly fascinating. We may learn quite a lot about our own consciousness from this, but are there any benefits to be had from having conscious agents in the world?
Mark Solms (56:46.305)
I think that the fundamental benefit is the purely scientific one. The fundamental benefit is understanding how consciousness works, what it is. That's the main purpose of this. I think beyond that what the benefits are, in other words beyond the pure science of it.
I find it quite difficult to see what the benefits are because as soon as you start speaking about how you might exploit such functionality, the word exploit has special meaning. It just sounds like slavery to me. What would you want such an agent to do for you? Why should it do anything for you?
I think the science is the value of it. Bringing consciousness into natural science, understanding how it works just as we understand everything else in biology and indeed in nature generally, that's what we're wanting to do. It's a very important thing for us to be able to do as scientists, but from the point of view of exploiting it, I have no ambitions, no interest in that more a matter of how do we restrain the exploitation of it rather than how do we capitalize on it.
Calum Chace (58:16.308)
Well, given that we've been trying to bring consciousness into the scientific method, arguably since the ancient Greeks, but certainly since the start of the Enlightenment, if you manage to do it, that'd be a really remarkable achievement. But I think there might be another use. Now, I've tried this on you before, and I don't think you got much purchase, but I do take quite seriously the idea that we're likely to create superintelligence sometime in the next few decades, maybe centuries, but probably before. And it's not impossible that the safest kind of superintelligence for humans would be one that is conscious, because a conscious superintelligence will understand what we mean when we say we experience things and we feel pain and fear and suffering and so on. And please don't do that to us, Mr. Superintelligence. Obviously, it's not guaranteed. A conscious superintelligence might say, well, I don't care. I disregard you. I don't like you. Or you're not important. But it might feel empathy.
I think on the whole, it's more likely that it would if it was conscious. So that strikes me as being possibly a really, really important value in consciousness, in machine consciousness. But I don't think that you buy that argument, do you?
Mark Solms (59:29.359)
Well, I mean, I don't think that the argument is implausible. I don't think it's silly. I understand where you're coming from. My problem with it is, look at us.
I mean, you know, we are the most intelligent creatures currently that we know of, the most intelligent agents in terms of general intelligence that we know of. And are we so empathic? Are we so not dangerous to each other, to other living creatures, the planet?
Calum Chace (01:00:08.606)
Well, yeah, I would argue we definitely are, actually. mean, there is no other animal, maybe dolphins, I understand. But basically, there's no other animal which is at least partly carnivorous, which can gather more than a few members of its own kind together without starting a fight. know, chimpanzees and other animals, if you're not part of my kin group, I'm going to attack you. And even if you are part of my kin group, I'm probably going to attack you. Whereas we humans, yeah, we commit murders, we carry out wars, we do terrible things. But millions of us live in cities, cheat by jail, walking past complete strangers all the time, and we don't lash out at each other. So I think you can make a case that we are much more peaceful than other animals. It sounds very counterintuitive, because the tools we have to fight each other are so powerful that we cause massive amounts of damage when we do it. And of course, we're damaging all the other animals on the planet as well. But in many ways, we are actually much more pacific than other animals.
Mark Solms (01:01:10.011)
As I said, I get where you're coming from, look, this is the point. The intelligence, our intelligence is what has made us capable of...
engineering weapons of mass destruction and everything else that we've engineered which is busy destroying the planet. these are, these things, you're sort of, if I'm understanding your argument correctly, you're saying that these two things go together, know, increasing intelligence and nicer behavior and I'm saying no, that increasing intelligence goes together with
nicer and worser behaviour. I don't think that there's an intrinsic connection between greater intelligence and greater empathy. I don't see the evidence for that in our case. But look...It's going to happen, I'm pretty sure. So the important thing, as with the other things we've just alluded to, like nuclear power and so on, these things have to be regulated. If they can be engineered, they will be engineered. They come with great potential and with great risk. there's no other way than for us to recognise the risks and try to manage them.
Will Millership (01:02:45.836)
That's one of the main goals of PRISM going forwards actually to look at these risks, what these possible risks might be and what kind of policies we can put in place to make sure that we do go forward with a responsible approach to consciousness in artificial intelligence.
Mark Solms (01:02:58.107)
I'm a big fan of Prism and of Consium.
Will Millership (01:03:02.542)
Brilliant. That'll be my clip to go out with all of the videos. Thank you so much for joining us today Mark. It's been great to chat with you. yeah, brilliant to see you.
