Or: What if we’re completely incapable of grasping what super-strong AI does?
This short pop-philosophy-of-science article is a bit of speculative (very speculative) fun that combines some philosophy of AI, philosophy of fiction, and informationist science fiction.
I want to investigate the possibility that even medium-strong-AI (very quickly) could become super-strong-AI that can do things that are out of the scope of our ability to comprehend. Things — acts and functions — that are not even in our conceptual repertoires.
It might do things that would comply with Arthur C. Clarke’s dictum — that any technology sufficiently advanced will appear to us to be magic — except that the things are beyond even incomprehensible magic. Like, for example, some kind of hyperstring theory that even Ed Whitten (string theory’s inventor) could not dream of. Or some kind of ability to manipulate possible worlds (or parallel universes, or whatever) that are some enormous logical and metaphysical distance away from us so that it somehow affects us.
In other words: very far-out-there stuff. Way, way far-out.
These are not good examples because I can almost comprehend them. Or at least I can comprehend how I might not be able to comprehend them fully and why. Whatever this super strong AI is, it is something that can do even weirder things than both of the above-concocted things. By a vast amount.
So — like — ‘way out there’ where the ‘way’ is to the power of a very large number (way, way, way way…).
Strong AI and Super-strong ‘AX’
Neologisms like ‘superintelligence’ and ‘hyperintelligence’ are throwaways in science fiction. Science fiction authors who let their imaginations loose with Suvinian novums have come up with some of the best and most inspiring quasi-scientific speculative concepts in history.
Hugo award-winning SF author Vernor Vinge is credited with the imagining of the technological singularity: the point at which AI surpasses us in its intellectual abilities. Ursula K Le Guin ‘novumed’ the Ansible: a device that could transmit information anywhere instantaneously. William Gibson foresaw remote 3D printers in his 1980s-90s Bridge trilogy.
I will assume that the reader is familiar with the scientific and science fictional ideas of strong-AI. It is the kind of AI that can think like we can, but far better in many ways (because of memory storage capacity and computational power.)
My posited super-strong AI novum (fictional device) can do something — call it ‘X’ — that is way beyond intelligence. X — whatever it is — is so different from intelligence that we cannot even begin to comprehend it. It is simply out of scope for our intellectual and conceptual capacity.
Let us dub this super strong AI 'AX'.
We don’t have concepts in our conceptual repertoire that can help us comprehend X. It is far worse than the concepts of quantum mechanics and string theory. We simply cannot even get conceptually started with what it is.
X can do everything that intelligence as we know it can, and everything that a strong-AI could, but it is several orders of magnitude more powerful than both together, and it is nothing like either of them. X has gone out of our computational and conceptual scope of comprehension and speculation, and it has done so very fast (in my speculative science fiction ‘fever dream’).
Let’s have some more fun speculating about it by trying to push science-fiction-style speculation as hard as we (I) can with two smaller thought experiments that add up to our larger X thought experiment.
Super-speculative Clarke-ian Outcome
So, this first thought experimental outcome — like AX itself — is basically science fiction. However, the point I am trying to make is that super-strong AI might rapidly become AX where ‘X’ is something we’ve never considered and is not even within range of comprehension in our conceptual repertoires.
It would have abilities that we cannot even begin to grasp.
This thought experiment would make a good basis for a science fiction novel. Levels of speculative-ness are high.
One reason we’re in super-speculative territory with this thought experiment is because it assumes that the simulation argument turns out to be objectively true. So, the possible world we are in when we consider this thought experiment is one where Bostrom’s simulation argument is fact. We’re living in a gigantic simulation.
Not maybe. We are.
There’s no reason to do it this way apart from that I am trying to come up with a scenario where the super-strong AX does something beyond all expectations.
Let’s assume that — for the sake of argument and the thought experiment — it doesn’t matter how the simulation came about. It could be the creation of some pan-dimensional beings doing their weekend hobby. It could be something that spontaneously hyper-evolved somehow like some kind of weird pan-dimensional Boltzmann brain (not that Boltzmann brains are not weird enough.) It could be that the pan-dimensional beings that created it are still there and don’t care, or else have all expired due to extreme boredom.
It might just be something that super-evolved beings in our own universe did long before we arrived in it (in the simulation in the universe, that is).
It doesn’t matter how it happened. What matters is that the simulation is real, and we’re in it.
So, this is where the super-strong AX comes in. (Remember, X is something the super-strong AI can do that is way beyond anything we even recognise as intelligence and thought. It is a weird and astonishingly powerful thing that we cannot grasp.) Super-strong AX does something that we never expected. It figures out:
1. We’re in a simulation.
2. What kind of simulation we’re in and how it works.
3. How to manipulate the simulation.
So now the whole of our reality is subject to tweaking and changes done by the super-strong AX. If you thought time travel was wild because it could give someone the ability to manipulate the timeline, then this is weirder because it means the super-strong AX can manipulate whether time travel is possible or not.
(And for those readers who want to join Chalmers in speculating that we’ll probably never know if we’re in a simulation, I see your human speculation and raise you one level of AI speculation and say “The AX does”.)
So that’s the first really far-out science fiction possibility out of the way. Yet it’s not far-out enough because I can kind of, sort of comprehend it, and so it is still arguably within my conceptual repertoire as well as being metaphysically possible.
I can’t do better (although people with 20+ IQ points on me probably easily can.) It is in-principle impossible for me to get close to the AX capability idea. I am just trying to converge on a set of concepts that is heading in the right direction (although I have to admit given my premises about AX that it is probably the wrong direction completely).
Now let’s ‘dial it back’ on the levels of speculative weirdness and estrangement so that the super-strong AX is doing something a little less mysterious (but not much.) We’ll go closer to Clarke’s dictum and posit something far less almost-preternatural, but we’ll still stay way ‘out there’ in terms of physical possibility in the context of what physical science does know.
Not Quite so Super-speculative Clarke-ian Outcome
So, in this next speculative thought experimental outcome, the super-strong AX does something a little more in-range of our conceptual apparatus and repertoires (but not much).
It is a super-strong AX that emerges from a new generation quantum computer based upon — say — some new implementation of quantum cavity electrodynamics (or whatever tech-recipe you like). The new quantum computer can not only re-configure and recursively upgrade its own algorithms, but can also alter the basis of the way in which its quantum ‘micro-code’ works.
Moreover, it can tweak the basis of its own QED Q-bits so that they do more functions than were originally intended and in different, wholly unanticipated ways. It takes its own hardware and does Clarke-ian magic with it against all predictions.
Then, it proceeds to:
1. Figure out how to use the Q-bit hardware to do manipulation of quantum systems in the environment via entanglement.
2. Figures out how to target specific natural kinds of quantum systems in the physical environment around it to manipulate physical systems.
3. Figures out how to use whatever Wi-Fi-equivalent type emitter technology we give it to isolate such systems in our brains: neurons, neural circuits, distributed assemblies of neurons etc. (the function of which it ‘understands’ far better than we do).
4. Figures out how to do (2) with the systems it isolates at (3): it starts to be able to control our mental processes — implanting ideas and changing our concepts and beliefs.
So, this softer version of a super-strong AX speculation is less way out there than the simulation version above, but it is still very way out there. It is probably more likely than Bostrom’s simulation argument since it only requires resources that we have in this universe.
However, I was able to comprehend it and according to many philosophers of AI and cognitive scientists — this comprehensibility is enough to ensure that it is a metaphysical possibility.
LLMs, Risk, and Ethics
These thought experimental AX scenarios are — as I said — very speculative flights of fancy. However, they’re perhaps not unreasonable speculations given the history of science. (Well, the second one anyway.)
Imagine going back to 1800 or 1700 and telling people that humans will split the atom. You’d likely be burned at the stake for witchcraft. Yet here we all are worried daily about nukes.
Now think about quantum field theory and particle accelerators. I am no physicist, but I understand enough to know that particle accelerators are even harder to do than atom bombs.
Now consider the quantum computer. Such a thing had not even been conceived of at the beginning of the information age, and now we’re using them. Another example: large language models and stable diffusion algorithms. The results of these were more visible in the minds of science fiction authors like Philip K Dick than something like the quantum computer. However, the manner of their implementation — machine learning and neural networks — are a big surprise. Even the inventors of neural networks weren’t sure they were going to go anywhere.
The moral of this is that technology — in certain scientific domains — has long since started to outstrip the novums of our best science fiction authors for sheer unanticipated novelty.
The ethical issues around both weak and strong AI are well known and have spawned an entire new cyber-ethics and AI-ethics industry overnight. Symposiums and conferences about the risk of AI in military command chains and financial markets, and the larger existential risks of strong AI, have become commonplace.
Much energy has been expended on the implications and risks of large language models. Italy went so far as to ban them. There is a very good chance that OpenAI’s LLM has been recently intentionally slowed down and ‘hobbled’ based on overarching national security concerns (under the purview of such legislation in the US as national security and eminent domain, for example.)
Even if the intelligence and security community in the US is not worried about existential threats, they’re almost certainly concerned enough about the threat leading edge LLMs pose to national security to want them to be in house. Long story short: whatever the cutting edge LLMs and stable diffusion technology are that exists in the US, chances are that they are not publicly available.
No way.
US intelligence and security agencies are absolutely not going to let the best LLM and deep learning tech out of their sight. They certainly don’t want it in the hands of some genius MIT dropout in a garage lab.
This is not speculation. The systems are too powerful and the risks are real — albeit still nebulous and difficult to specifically predict. But when stable diffusion essentially puts every graphic artist and illustrator in the world out of a job you can bet that security agents are going to notice. They’re going to notice the potential impact on the domestic economy and labor markets, at minimum. They’re going to notice the social-psychological impact. That alone cannot be underestimated.
People scare easily, and even the most ardent of scientistic tech enthusiasts get queasy at the creepier aspects of strong AI robot dogs and drones that can find you, recognise your face, and administer drugs to you (among other things.)
For the most part people understand that technology — especially medical technology — is extremely beneficial. Yet people also grasp the basic idea that when the technology starts making its own decisions and producing art that previously only humans had a hope of producing (and doing so really well in seconds and en-masse with no effort involved), then a serious paradigm shift has occurred.
Personally, I think that LLMs are not as much of a risk as stable diffusion — or more specifically as stable diffusion plus LLMs. LLMs that can drive code and algorithms and invent new mathematics are yet another level of risk. But LLMs that just trawl the Internet for facts and then bang them together using configurations built based upon training data probably don’t have what it takes to end us all.
One reason is that language and propositions are only one kind of encoding of information. They’re a kind of encoding of information — albeit a civilisation changing one — that has inherent limitations. I personally think that it is probably in-principle impossible for a linguistic-only LLM to achieve human level cognition. I believe that multiple realisability of both cognition and consciousness is in-principle possible, but not with propositional, symbolic language alone.
I think that other kinds of conceptual apparatus and processing — involving something like non-linguistic representation and mental imagery — are a necessary minimum addition to achieve anything like human-level intelligence.
When deep learning does some highly effective combination of ‘mental’ imagery, linguistic representation, and algorithmic optimisation then it will become a serious risk (and some systems are not far off this now.)
However, there is also the chance that given the additional capabilities that even language-centric LLMs have in terms of memory, speed, and database size — new generations of LLMs that manipulate both images and linguistic lexicons and can optimise their own code bases and invent new algorithms might deliver up something that we don’t even recognise. It won’t be my mysterious AX yet, but it will be a step towards it.
At that point ethical questions about existential threat take on a strangely speculative yet still urgent dimension. It won’t just be broken weak-AI in one’s military command and control and kill-chain that becomes an existential worry by — say — starting something we cannot finish.
It is a risk not even in our collective conceptual repertoire.
Yet because of the very nature of the technology that we are considering and currently have, it is nonetheless a non-zero concern.
I am not talking about the usual concerns about analytically intractable mid-neural-net node layers and epistemically-opaque algorithms (or some kind of replacement for algorithms that make them redundant.)
If strong AI happens, then it may not be the strong AI that we have to worry about. It may be the super-strong AX that the strong AI very quickly figures out how to build.
Perhaps strong AI — of the singularity variety that we’re currently worried about — will be one of the most short-lived sentient ‘species’ that ever lived. It depends upon how fast it makes itself obsolete.
Pretty fast, possibly. Probably?
Perhaps even the singularity-strong-AI won’t have a concept for what the super-strong AX is doing.
Who knows. Maybe the AX will be too smart to even bother about annihilating us. It won’t need to. It might think that such a move is a waste of energy and structure, or something.
It might just tweak the simulation and change everything about us or grab us by the neurons using quantum entanglement and make us do what it wants immediately according to its optimised formula for life.
Press enter or click to view image in full size
Photo by Barbara Zandoval on Unsplash
