AGI: How to Ensure Benevolence in Synthetic Superintelligence

*Image:"Better Than Us," 2018 Netflix Series

"Yet, it's our emotions and imperfections that makes us human."   -Clyde DeSouza, Memories With Maya

IMMORTALITY or OBLIVION? I hope that everyone would agree that there are only two possible outcomes after having created Artificial General Intelligence (AGI) for us: immortality or oblivion. The necessity of the beneficial outcome of the upcoming intelligence explosion cannot be overestimated.

Any AGI at or above human-level intelligence can be considered as such, I’d argue, only if she has a wide variety of emotions, ability to achieve complex goals and motivation as an integral part of her programming and personal evolution. I could identify the following three most optimal ways to create friendly AI (benevolent AGI), in order to safely navigate uncharted waters of the forthcoming intelligence explosion:

(I). Naturalization Protocol for AGIs (bottom-up approach);

(II). Pre-packaged, upgradable cognitive modules and ethical subroutines with historical precedents, accessible via the Global Brain architecture (top-down approach);

(III). Interlinking with AGIs to form the globally distributed Syntellect, collective superintelligence (horizontal integration approach).

Now let’s examine in detail the three ways to facilitate the emergence of benevolent Synthetic Superintelligence and, for those in AI research, ride the “crest of the wave” of the upcoming intelligence explosion:

I. (AGI)NP: Program AGIs with emotional intelligence, empathic sentience in the controlled, virtual environment via human life simulation (advanced first-person story-telling version, widely discussed in AI research). In this essay I will elaborate on this specific method, while only briefly touching on the other two methods, as we discuss the Global Brain in my recent book The Syntellect Hypothesis: Five Paradigms of the Mind's Evolution in great detail.

II. COGNITIVE MODULES AND ETHICAL SUBROUTINES: This top-down approach to programming meta-cognition and machine morality combine conventional, decision-tree programming methods with Kantian, deontological or rule-based ethical frameworks and consequentialist or utilitarian, greatest-good-for-the-greatest-number frameworks. Simply put, one writes an ethical rule set into the machine code and adds an ethical subroutine for carrying out cost-benefit calculations. Designing the ethics and value scaffolding for AGI cognitive architecture remains a challenge for the next few years. Ultimately, AGIs should act in a civilized manner, do “what’s morally right” and in the best interests of the society as a whole. AI visionary Eliezer Yudkowsky has developed the Coherent Extrapolated Volition (CEV) model which constitutes our choices and the actions we would collectively take if “we knew more, thought faster, were more the people we wished we were, and had grown up closer together.”

Will our brain’s neural code be a pathway to AI minds? In May, 2016 I stumbled upon a highly controversial Aeon article titled “The Empty Brain: Your brain does not process information, retrieve knowledge or store memories. In short: your brain is not a computer” by psychologist Rob Epstein. This article attested to me once again just how wide the range of professional opinions may be when it comes to brain and mind in general. Unsurprisingly, the article drew an outrage from the reading audience. I myself disagree with the author on most fronts but one thing, I actually agree with him is that yes, our brains are not “digital computers.” They are, rather, neural networks where each neuron might function sort of like a quantum computer. The author has never offered his version of what human brains are like, but only criticized IT metaphors in his article. It's my impression, that at the time of writing the psychologist hadn't even come across such terms as neuromorphic computing, quantum computing, cognitive computing, deep learning, computational neuroscience and alike. All these IT concepts clearly indicate that today's AI research and computer science derive their inspiration from human brain information processing – notably neuromorphic neural networks aspiring to incorporate quantum computing. Deep neural networks learn by doing just like children.

There’s nothing wrong with thinking that the brain is like a computer, but in many ways the brain is a lot different. Whereas information on a computer hard drive is laid out and ordered, for instance, that doesn’t seem to be the case with human memories. Arguably, they are stored holographically throughout brains regions (as well as in the field of non-local consciousness). There are similarities and differences. Like a computer, the brain processes information by shuffling electrical signals around complex circuitry. Neither analog nor digital, the brain works using a signal processing format that has some properties in common with both.

The computational hypothesis of brain function avers that all mental states – such as your conscious experience of reading this sentence right now – are computational states. These are fully characterized by their functional relationships to relevant sensory inputs, behavioral outputs, and other computational states in between. That is to say, brains are elaborate input-output devices that compute and process symbolic representations of the world. Brains are computers, with our minds being the software, simplistically speaking, of course. The physical wetware isn’t the stuff that matters. What matters is the algorithms that are running on top of the wetware. Theoretically, we should be able to “crack our neural code” and reproduce it on other computational substrates. The central goal of neuroscience is breaking this neural code – deciphering the relationships between spatiotemporal patterns of activity across groups of neurons and the behavior of an animal or the mental state of a person. However, algorithmic solutions in this top-down approach in AI research will most likely come not from neuroscience and not even from computational neuroscience, they might come as breakthroughs from neurophilosophy, software engineering and computer science.

As I mentioned above, the brain is a quantum neural network. Quantum computation inside our brain is beyond classical physics, it’s in the realm of quantum mechanics and information theory. Our brain is not a “stand-alone” information processing organ: It functions as a central unit of our integral nervous system with recurrent feedback with the entire organism and the cosmos at large. Along the lines of quantum cognition theoretical framework, we can conjecture that quantum coherence underlies the parallel information processing that goes on in the brain’s neocortex, responsible for higher-order thinking, and allows our minds to almost instantaneously deal with the massive amounts of information coming in through our senses. Quantum computing is a natural fit for that: It is massively parallel information processing that is ultrafast and practically inexhaustible.

A strict reductionist approach that takes a bottom-up methodology to the mind seems to be missing some crucial element. This kind of approach, focused on local cause and effect classical mechanics within the brain, on neurons firing across their synaptic connections, is doomed to fail. The mind is scientifically elusive because it has layers upon layers of non-material emergence: It’s just like a TV screen – if you’re watching a movie and could only look at an individual pixel, you would never understand what’s going on. No single neuron produces a thought or a behavior; anything the “mindware” accomplishes is a vast collaborative effort between brain cells. When at work, neurons talk rapidly to one another, forming networks as they communicate, with several networked links resonating at different times and with different subgroups of nodes, such that understanding the behavior of individual “pixels” or even of smaller groups of them won't tell the whole story of what's happening.

We need to think in terms of networks, modules, algorithms and second-order emergence  –meta-algorithms, or groups of modules. We need these methods to see the whole screen, the bigger picture, to see what’s playing in our minds. Ultimately, a new cybernetic approach with a top-down holistic methodology could be applied to explain the human mind and other multi-scale minds in creation. Human minds, as diverse as they are, occupy only a very narrow stratum of the total space of possible minds.

Numerous studies have found that the brain organizes itself into functional networks that vary in their activity and in their interactions over time. One such classification gives us three major networks: the central executive network, which is responsible for attentional focus; the salience network which involves awareness; and the default mode network as an “idling” mode such as inward-focused thinking and mind wandering. The subtleties of our psyches are being managed by smaller networks – specific modules in our brains. A unique cognitive percept is the end result of the processing of a module or group of modules in a layered architecture. The idea that the brain is made up of many regions that perform specific tasks is known as ‘modularity’. This concept of modular organization suggests that specialized areas of the brain do different things, with certain capacities coming up one at a time and through time they are stitched together to give the illusion of a unitary conscious experience. In actuality, each individual part of the brain is doing its respective job, and each then passes information to the next level of network. This continues until we become aware of the thought or function like sight or sound. There are many layers in an onion, in a manner of speaking.

Overall, the brain may operate on an amazingly simple mathematical logic. “Fire together, wire together” is perhaps neuroscience’s most famous catchphrase. Learning activates select neurons and synapses in the brain, which results in a strengthening of the connections between pairs of synapses involved in storing a particular memory. With reactivated circuits, you get retrieved memories. Corresponding algorithms for intelligence could inform neuromorphic computing, teach artificial circuits to recognize patterns, discover knowledge, and generate flexible behaviors. That would enable the creation of artificial neural networks that are wired in a manner akin to our own grey matter but embedded in a different substrate.

Once designed, pre-packaged, upgradable cognitive modules, including ethical behavior subroutines, with access to the global database of historical precedents, and later even “the entire human lifetime experiences,” could be instantly available via the Global Brain architecture to the newly created AGIs. This method for “initiating” AGIs by pre-loading cognitive modules with ethical and value subroutines, and regularly self-updating afterwards via the GB network, would provide an AGI with access to the global database of current and historical ethical dilemmas and their solutions. In a way, she would possess a better knowledge of human nature than most currently living humans. I would discard most, if not all, dystopian scenarios in this case, such as described in the book “Superintelligence: Paths, Dangers, Strategies” by Nick Bostrom and more recently in the book “Human Compatible: Artificial Intelligence and The Problem of Control” by Stuart Russell. While being increasingly interconnected by the Global Brain infrastructure, humans and AGIs would greatly benefit from this new hybrid thinking relationship. The awakened Global Mind would tackle many of today’s seemingly intractable challenges and closely monitor us for our own sake, while significantly reducing existential risks.

Ever since Newton, materialist science has been entrenched in our models but a new trend now emerges towards post-materialist, information-theoretic science which is to transform any area of research into computer science, i.e. information technology. Once the field becomes information technology (such as genomics, for example), it jumps on its own exponential growth curve of further development. We now see it in computational neuroscience, connectomics, and other related fields. Computer science gives us a new code-theoretic, substrate-independent model for looking at our brains and our neural code. Our brains, however, do not generate consciousness since our minds are embedded in the larger consciousness network, the topic discussed in The Syntellect Hypothesis. We humans are deep down information technology running on genetic, neural and societal codes. Self-transcendence from a bio-human or cyberhuman into a higher-dimensional info-being might be closer than you think.

... to continue to Part II: Interlinking & AGI(NP) ...

-Alex Vikoulov

*Image Credit: "Better Than Us," 2018 Netflix Series

**Original article first appeared on EcstadelicNET (ecstadelic.net) in Top Stories section on March 8 , 2016.