Three Viable Paths to True AGI
According to my current best guesses...
The deep neural nets and other ML algorithms that are absorbing most of the AI world’s attention today are, in my own view, fundamentally unsuited for the creation of human-level AGI. For a primer on why, see my fireside chat with Gary Marcus at the AGI-22 conference, at around 4hrs 40 min of
As I noted in my last blog post, the absolute upper bound for which these deep nets or any vaguely similar methods could be sensibly hoped to achieve would be what I’d call “closed-ended quasi-AGI” systems which could imitate a lot of human behaviors — but which, due to the fundamental lack of ability to innovate, abstract or generalize, would be incapable to address difficult unsolved science and engineering problems, or to perform the self-modification and self-improvement needed to serve as seed AIs and launch a Singularity.
OK but if deep neural nets and such aren’t the path then what is?
I’ll briefly run through here a few approaches I think could probably work: a cognition-level approach, a brain-level approach and a chemistry-level approach.
None of these are radically new approaches. And of course, neither are deep neural nets remotely new — they were first proposed and experimented with in the 1960s. But even when I taught about them in university in the 1990s back when I was an academic, the computing hardware available was not adequate to enable them to show their potential very well in a practical sense.
Just as the advent of the Internet, larger RAM units and GPUs enabled deep neural nets to shift from sluggish research into scalable practice — so may the ongoing advanced in various sorts of compute hardware enable other historical AI paradigms to come to the fore in the near future, showing their potential in a more practical sense and just maybe leading us toward AGI.
COGNITION-LEVEL APPROACH — Hybrid Neural/Symbolic/Evolutionary Metagraph-Based AGI Inspired by Cognitive Science and Optimized via Funky Math
The option I’m most psyched about at the moment is, not surprisingly, the one I’m working on — OpenCog Hyperon, the new version of the OpenCog system, currently under active development by myself and colleagues at SingularityNET and OpenCog Foundation. A lot was said about the underlying theory and current progress on Hyperon in the workshop on Day 1 of AGI-22 last week:
Hyperon is a neural-symbolic approach but it’s not just that — deep nets, attractor neural nets, evolutionary program learning, probabilistic logical inference, probabilistic programming and other methods are integrated into a common framework allowing them to co-update a large knowledge metagraph. Advanced math formulations are used to enable these diverse algorithms to be implemented using a small common set of mathematical operations action on this “Atomspace” metagraph.
Right now in the Hyperon project we’re focusing on creating an efficient large-scale distributed Atomspace, on improving the interpreter for the new MeTTa programming language we’ve created for implementing Hyperon’s AI algorithms, on experimenting with early implementations of these algorithms in MeTTa — and together with Simuli, on creating a custom AGI board with custom chips for accelerating Hyperon operations.
While there are some formal neural nets around in the Hyperon design, there’s not a lot of brain modeling in any serious sense. What there is, however, is quite a lot of cognitive modeling… the types of nodes and links in the Atomspace, and the particular assemblage of learning and reasoning and memory and attention algorithms used in Hyperon, are drawn on decades of careful study of human cognitive psychology. Very loosely, part of the idea here is to figure out what key functions are carried out by different networks in the human brain, and how these functions interoperate, and then emulate these networks and functions and their interactions using computer-science algorithms that operate efficiently on current hardware. And then optimize as much as possible by implementing these algorithms on a common metagraph infrastructure in a way that makes their underlying calculations look as similar as possible (so they can benefit from the same mathematical and software optimizations).
BRAIN-LEVEL APPROACH: Large-Scale Nonlinear-Dynamical Brain Simulation
Current neural net models are “neural” only in a very loose sense. Computational neuroscience remains a fascinating discipline (in which I did some work years ago on a multi-year IARPA-funded research project), and often involves running computer simulations of various regions of the brain using models of neurons and their interactions that are radically different from the simple “formal neurons” used in today’s deep neural models.
I remain a huge fan of the brain modeling approach pursued in Eugene Izhikevich’s book Dynamical Systems in Neuroscience … but nobody has so far made a serious effort to take this sort of more-biologically-realistic, chaos-theory model of neural nets and use it to actually model the brain.
This would be a mammoth undertaking, given the hundreds of importantly distinct brain regions and multi-regional brain networks, along with the various different sorts of neurons and neuronal columns and glia and astrocytes and so forth. Henri Markram’s Human Brain Project was aiming in this general direction, but famously failed to live up to its promise. I think this failure had a few key causes .. some were organization and project management related, resultant from the generally suboptimal way the EU manages Big Science projects … some were rooted in Markram’s own human peculiarities … and then there was the absence of a core conceptual grounding in nonlinear dynamics and the science of complex systems.
Thing is, even the best possible stab we could take at brain simulation given our current knowledge of the brain wouldn’t be SUCH an accurate stab really — there is just too much we don’t know about how the brain works, and our computer hardware is just too different from neural wetware. So the right way to think about a non-BS Human Brain Simulation Project would be as: Building a complex, self-organizing nonlinear dynamical network that incorporates as much knowledge about the human brain as we have, and then exposing it to environmental stimuli and feedback in a manner calculated to induce the self-organization of complex structures with as much human-like intelligent behavior as possible.
If Izhikevich and Edelman had been put in charge of the HBP, and it had been run either as a unified Manhattan Project or a self-organizing decentralized network of freewheeling scientists, then it might well have succeeded in meeting its goals. But it’s not too late to try again. The world routinely blows tens of billions of dollars on far less worthy pursuits.
A variation on this approach would be to try to simulate the brains of simpler organisms first. We don’t even have good simulations of the brain of a cockroach or a bee, at this point — let alone a mouse or a monkey. A disadvantage of this animals-first direction is that we have less knowledge about how these animals’s brains work. OTOH there are a lot fewer neurons and other cells to worry about.
An interesting unknown in this whole brain-simulation direction is how much additional mileage would be obtained by putting one’s simulations to work/school in humanoid robot bodies, versus humanoid avatars in highly tuned simulators. I’d suspect a little of both would be valuable. Critical would be to do whole-system testing of models of all brain regions working together controlling real embodied systems, starting at a fairly early stage in the project.
A downside of the brain-emulation approach to AGI is that when you’re done all you get is a… a very expensive, awkwardly-implemented human brain analogue. We have a lot of smaller, cheaper, lower-power human brains already. But I’m willing to give the approach the benefit of the doubt here — once we really understood the human brain well enough to make a full-on emulation in software, we’d almost surely be able to vary and improve on this system and in some not-huge number of years move toward superhuman AGIs of some form or another.
CHEMISTRY-LEVEL APPROACH — Massively Distributed, AI-Optimized Artificial Chemistry Simulation
The human brain is the existing highly-generally-intelligent system we know best. Part of what we know about it, however, is that it has a lot of special and peculiar characteristics that are not at all necessary for human-level general intelligence. We also know that it operation relies heavily on peculiarities of biological cells and molecules that are difficult and expensive to simulate in today’s digital computers.
What if instead of simulating the brain, or going a level higher up and simulating human cognitive science in a non-biological way like Hyperon tries to, one shifted a level further DOWN and created a lower-level complex self-organizing system, and then coaxed it to produce loosely brain-like structures and dynamics at the emergent level?
We lack either the knowledge or the computing power to try to simulate a human body at the molecular level, let alone to simulate the whole emergence of life from prebiotic soup up to humanity at the molecular level. These would be awesome things to try, but we’ll need to wait till after the Singularity, or lacking a Singularity, at very least we’ll need to wait several decades for improvements in biochemistry knowledge and computing infrastructure.
However, the discipline of complex systems has taught us that, sometimes, the critical aspects of a real-world complex dynamical system can be captured in quite different complex dynamical systems, which are composed of different sorts of elements that may be easier to efficiently manipulate at scale on current computer systems. Wolfram’s New Kind of Science argued extensively for this perspective, but of course the idea long predated Wolfram’s entry into the complex systems arena. The Santa Fe Institute probably did more than any other organization for promoting this “complex systems science” perspective, which is also closely allied with nonlinear dynamics aka “chaos theory.”
Could we come up with some sort of cellular automaton model that loosely brainlike structures could tractably pop out from? How about a CA living on a graph or metagraph rather than a square array? How about a soup of computer programs that rewrite each other to produce new computer programs — as in Walter Fontana’s good old Algorithmic Chemistry work, or Kristinn Thorisson’s more recent Replicode based systems?
This potential approach to AGI is very blue-sky-researchy and high-risk, yet also philosophically very attractive — and it’s never really been tried at modern scale. Bruce Damer’s Evogrid project aims to use distributed computing to do computational-chemistry simulation of origin-of-life models at large scale; one could take a similar approach to algorithmic chemistry models. And one could put such models behind humanoid robots and their simulated avatars, just as Replicode was put behind humanoids in the HUMAN-OBS project, with interesting preliminary results. Thorisson thinks of his work as “constructivist AI” rather than artificial chemistry, but the conceptual direction is largely the same.
I love this sort of approach — and I spent some time on it in the mid-90s before getting frustrated with the inadequacy of hardware at that time to emulate large self-organizing networks — but every time I start thinking about it, I start thinking next that it would work even better if one were able to apply an advanced pattern mining and reasoning engine to the task of optimizing one’s primordial algorithmic-chemical soup.
What if instead of trying 10K randomly configured algorithmic-chemical soups on different machines, and evolving this population slowly and hopefully and expensively in the direction of interesting emergent structures and dynamics — what if instead, one used an observer AI system to notice what characteristics led an algorithmic-chemical soup to display more interesting structures/behaviors … and then seeded new algorithmic-chemical soups based on the observations of this AI systems.
I.e. what if one treated the evolution of an algorithmic-chemical soup leading to AGI less as a pure evolutionary-learning problem and more like an Estimation of Distribution Algorithm (EDA)?
But once one does this, one is pushed into the problem of making EDAs work robustly on complex problems — which leads one right back to designing and building OpenCog Hyperon, which if I’m right will be the world’s best EDA engine by far … and just the right tool for mining patterns among algorithmic-chemistry soups and figuring out how to seed new ones.
And this leads to my Cogistry proposal from a few years back — doing algorithmic chemistry in the OpenCog infrastructure, which is the perfectly place for also building and running the EDA to guide the distributed chemical evolution. One gets a pleasant potential recursion this way … as the algorithmic chemistry soups start to do useful things, one can leverage them as part of the pattern-mining/reasoning/generation engines studying the new generation of algorithmic-chemistry soups.
We thus come pleasantly full-spiral … and we basically only have two approaches in my list, Hyperon (or other similar systems) and large-scale brain simulations… with artificial chemistry appearing in large part as one promising potential way to leverage Hyperon systems.
Making It Really Happen
All this is significantly more complicated that current deep neural net systems — but then, so is the human brain and so is the human mind. It’s true that one can sometimes get astounding complexity to emerge from apparent simplicity — but that needs to happen through a sort of self-organization and autopoiesis and evolution that deep neural net architectures patently lack.
Pursing closed-ended quasi-AGI systems using end-to-end trained conglomerations of deep neural nets is not entirely stupid, though it’s not how I’d choose to spend my own time. But if we want to really get to AGI that can solve hard science and engineering problems that currently stump humans, and that can seed the Singularity, we’ll need to do better. Nobody knows for sure what “better” direction is most likely to succeed, but I’ve summarized here my own intuitions, which are based on quite a lot of experience thinking and prototyping and experimenting in AGI, narrow AI, computational neuroscience, algorithmic chemistry and related spaces.
After the 2008 financial crisis, President Obama (whom I like quite a lot compared to most politicians) allocated a few trillion USD to bailing out large banks and insurance companies. Thus decreasing the odds of increased volatility in global financial markets as well as accelerating the increase in wealth inequality. What if instead a trillion USD had gone into hybrid AGI systems, a trillion into large scale brain simulation, a trillion into algorithmic chemistry. It’s not so outlandish to think we’d have experienced the Singularity already.
But events in our world have a certain flow to them, often for reasons our human minds don’t fully comprehend. The mass mind of humanity doesn’t understand itself or the world or AGI well enough to make that sort of resource allocation. But we’re getting there anyway. At the moment the right sorts of brain simulation are not being funded or otherwise pursued at the scale needed to meet massive success there anytime soon — but we’re proceeding pretty well on the Hyperon project, and others in the hybrid-AGI space are making their own progress as well. So even though our species prefers to bail out banks and create massive deep neural nets fueling advertising engines than to work directly toward beneficial and creative full-scale AGI … the Singularity is near even so, and gives the appearance of proceeding on some approximation of the schedule Ray Kurzweil has projected.