DynaPsych Table of Contents





The Potential Emergence of Multiple Levels of Focused Consciousness

in Communities of AI’s and Humans



Ben Goertzel

Novamente LLC


1.       The Concept of a Mindplex


The goal of this article is to briefly introduce a new concept, one that spans the domains of psychology, sociology and computer science.  Choosing a name for this new concept was not easy, and the coinage I’ve settled on is one suggested by Eliezer Yudkowsky [1]: “mindplex.”  A mindplex is loosely defined as an intelligent system that:


1.                  Is composed of a collection of intelligent systems, each of which has its own “theater of consciousness” [2] and autonomous control system, but which interact tightly, exchanging large quantities of information frequently

2.                  Has a powerful control system on the collective level, and an active “theater of consciousness” on the collective level as well


In informal discussions, I have found that some people, on being introduced to the mindplex concept, react by contending that either human minds or human social groups are mindplexes.  However, I believe that, while there are significant similarities between mindplexes and minds, and between mindplexes and social groups, there are also major qualitative differences.  It’s true that an individual human mind may be viewed as a collective, both from a theory-of-cognition perspective (e.g. Minsky’s “society of mind” theory [3]) and from a personality-psychology perspective (e.g. the theory of subpersonalities [4,5]).  And it’s true that social groups display some autonomous control and some emergent-level awareness [6].  However, in a healthy human mind, the collective level rather than the cognitive-agent or subpersonality level is dominant, the latter existing in service of the former; and in a human social group, the individual-human level is dominant, the group-mind clearly “cognizing” much more crudely than its individual-human components, and exerting most of its intelligence via its impact on individual human minds.  A mindplex is a hypothetical intelligent system in which neither level is dominant, and both levels are extremely powerful.  A mindplex is like a human mind in which the subpersonalities are fully-developed human personalities, with full independence of thought, and yet the combination of subpersonalities is also an effective personality.  Or, from the other direction, a mindplex is like a human society that has become so integrated and so cohesive that it displays the kind of consciousness and self-control that we normally associate with individuals.

I will discuss here two mechanisms by which mindplexes may possibly arise in the medium-term future:


1.                  Humans becoming more tightly coupled via the advance of communication technologies, and a communication-centric AI system coming to embody the “emergent conscious theater” of a human-incorporating mindplex

2.                  A society of AI systems communicating amongst each other with a richness not possible for human beings, and coming to form a mindplex rather than merely a society of distinct AI’s


The former sort of mindplex relates to the previously discussed concept of a “global brain” [7,8]. Of course, these two sorts of mindplexes are not mutually contradictory, and may coexist or fuse.  The possibility also exists for higher-order mindplexes, meaning mindplexes whose component minds are themselves mindplexes.  This would occur, for example, if one had a mindplex composed of a family of closely-interacting AI systems, which acted within a mindplex associated with the global communication network.  In fact, I will propose that this is a likely situation to occur.

            After reviewing some futurological scenarios situating the mindplex concept, I will introduce and discuss the Novamente AI system [9], as an example of a technology that could conceivably play a key role in both sorts of mindplex mentioned above.  The Novamente design contains a specific mechanism – Psynese – intended to encourage the formation of multi-Novamente mindplexes.  And, Novamente is intended to be deployable within global communication networks, in such a way as to provide an integrative conscious theater for the “distributed cognition” [10] system of interacting online human minds.  In the final section, I will use the projected nature of Novamente mindplexes as a metaphor for comprehending the slipperier and more ambiguous projected global brain mindplex.


2.       Mindplex-Pertinent Futurological Concepts

            This section presents some ideas about the medium-term future of technology and mind society, in which mindplexes are projected to play a likely role.

            I must stress, in presenting these futurological concepts, that I am well aware of the huge uncertainties involved in the future of the human race and its creations.   It’s easy to enumerate possibilities, and extremely difficult to probabilistically weight them, due to uncertainties involved in mass psychology, the rate of development of various technologies, and other factors.

            The key assumption I make is that computing, communication and human-computer-interaction technology are very likely to continue their exponential increase in power and decrease in cost, for the next century at least.  There are many possible scenarios in which this does not occur, but I consider them relatively low-probability and will not explicitly discuss them here. 

            Based on this assumption, it seems very likely that we are going to see the following two developments during the next century:


1.                  Software programs demonstrating “artificial general intelligence” [11] at the human level and beyond

2.                  Hardware/wetware allowing direct interaction between human thought processes and computational processes


Based on a fairly detailed analysis, Ray Kurzweil [12] has proposed a 2040-2060 timeline for the maturity of these technologies and the consequent occurrence of a technological “Singularity” point, at which the progress of mind/technology/society occurs at such a fast rate that the ordinary human mind can no longer comprehend it.  I will not repeat or critique Kurzweil’s analysis here; I have discussed these topics in [13], and here will focus on some specific consequences of these projections, which I believe to be intelligent and plausible based on the available knowledge.

            In the remainder of this section I will situation the mindplex concept within three aspects of this futurological analysis: the global brain, artificial general intelligence, and the Singularity.


2.1       Artificial General Intelligence


When psychologists speak of human intelligence, they implicitly mean general intelligence.  The notion of IQ arose in psychology as an attempt to capture a “general intelligence” factor or G-factor [14], abstracting away from ability in specific disciplines.  Narrow AI, however, has subtly modified the meaning of “intelligence” in a computing context, to mean, basically, the ability to carry out any particular task that is typically considered to require significant intelligence in humans (chess, medical diagnosis, calculus,…).  Our notion of Artificial General Intelligence (AGI) separates task-specific from generic intelligence, and refers to something roughly analogous to what the G-factor is supposed to measure in humans.

When one distinguishes narrow intelligence from general intelligence, the history of the AI field takes on a particularly striking pattern.  AI began -- in the mid-20’th century -- with dreams of artificial general intelligence – of creating programs with the ability to generalize their knowledge across different domains, to reflect on themselves and others, to create fundamental innovations and insights.  But by the early 1970’s, nothing approaching AGI had been realized, frustrating researchers, commentators, and those who fund research.  AGI faded into the background with only a handful of research projects and acquired a markedly bad reputation and a great deal of skepticism. 

Today in 2003, however, things are substantially different than in the early 1970’s when AGI lost its luster.  Modern computer hardware and networks are incomparably more powerful than the best supercomputers of the early 70’s, and software infrastructure has also advanced considerably. 

The supporting technologies for AGI are now in place.   Furthermore, the mathematics of cognition is more fully understood, partly due to work on narrow AI, but also due to revolutionary advances in neuroscience and cognitive psychology.  The same advances in computer technology that have created current multidisciplinary information overload will enable the development AGI technology and allow us to convert this overload into enhanced knowledge management and analysis.

Like nanotechnology[1][15], I believe, AGI is “merely an engineering problem,” though certainly a very difficult one.  Brain science and theoretical computer science suggest that, with the right design, it is possible to make significantly more progress toward AGI than is manifested in the current state of AI software.

When one seeks to formalize the notion of general intelligence, one quickly runs up against the fact that “intelligence” is an informal human language concept rather than a rigorously defined scientific concept; its meaning is complex, ambiguous and multifaceted.   However, one may work around this by creating formal definitions explicitly aimed at capturing part rather than all of the informal natural language notion.  In this vein, in the author’s 1993 book The Structure of Intelligence [16],  a simple working definition of intelligence was given, building on various ideas from psychology and engineering.  The mathematical formalization of the definition requires more notation and machinery than I will present here, but the gist is:


General intelligence is the ability to achieve complex goals in complex environments


A primary aspect of this “complex goals in complex environments” definition is the plurality of the words “goals” and “environments.”  A single complex goal or single narrow environment is insufficient.  A chess-playing program is not a general intelligence, nor is a data mining engine that does nothing but seek patterns in consumer information databases, nor is a program that can extremely cleverly manipulate the multiple facets of a researcher-constructed microworld (unless the microworld is vastly more rich and diverse one than any ever yet constructed).  A general intelligence must be able to carry out a variety of different tasks in multiple contexts, generalizing knowledge from one context to another, and evolving a context and task independent pragmatic understanding of itself and the world.

One sort of “rich environment,” potentially supportive of AGI, is the portion of the physical world in which human beings reside.  Another kind of rich environment, potentially supportive of AGI, is the Internet – considered not merely as a set of Web pages, but as a complex multifaceted environment containing a family of overlapping, dynamic databases, accessed and updated by numerous simultaneous users continually generating new information individually and collaboratively. Many overlapping and complex goals confront an AI system seeking to play a humanlike social role in this space – I believe that the Net provides a sufficiently rich goal/environment set to support the emergence of a robust general intelligence.

As hinted above, I believe the potential relationship between AGI’s and mindplexes is twofold.  Firstly, as will be elaborated in Section 3 below, I believe AGI’s will be able to form mindplexes much more easily than humans, due to the more powerful and flexible varieties of communication that will be possible between them.  Secondly, I believe that AGI’s may be able to serve as the central plexus of the computing/communication network of the future – catalyzing the emergence of the global conscious theater that will turn the Net into a “global-brain mindplex.”



2.2       The Global Brain


The notion of the Global Brain is reviewed in detail in Heylighen’s article in this volume [7].  I find it appealing but somewhat ambiguous.  As I argue in [17], it seems inevitable that some sort of intelligence is going to arise out of global computer and communication networks – some kind of emergent intelligence, going beyond that which is implicit in the individual parts.  Projecting the nature of this emergent intelligence is the hard part!

In [17], I distinguish three possible phases of  “global brain” development:


Phase 1: computer and communication technologies as  enhancers of human interactions.  This is what we have today: science and culture progress in ways that would not be possible if not for the “digital nervous system” we’re spreading across the planet.  The network of idea and feeling sharing can become much richer and more productive than it is today, just through incremental development, without any  Metasystem transition.


Phase 2: the intelligent Internet.  At this point our computer and communication systems, through some combination of self-organizing evolution and human engineering, have become a coherent mind on their own, or a set of coherent minds living in their own digital environment.


Phase 3: the full-on Singularity.   A complete revision of the nature of intelligence, human and otherwise, via technological and intellectual advancement totally beyond the scope of our current comprehension.  At this point our current psychological and cultural realities are no more relevant than the psyche of a goose is to modern society.


In this language, my own thinking about the global brain has mainly focused on


·        how to get from Phase 1 to Phase 2 – i.e. how to build an AGI system that will effect or encourage the transformation of the Internet into a coherent intelligent system

·        how to ensure that the Phase 2, Internet-savvy, global-brain-centric AGI systems will be oriented toward intelligence-improving self-modification (so they’ll propel themselves to Phase 3), and also toward generally positive goals (as opposed to, say, world domination and extermination of all other intelligent life forms besides themselves!)  


Currently the best way to explain what happens on the Net is to talk about the various parts of the Net: particular Websites, e-mail viruses, shopping bots, and so forth.   But there will come a point when this is no longer the case, when the Net has sufficient high-level dynamics of its own that the way to explain any one part of the Net will be by reference to the whole.   This, I believe, will come about largely through the interactions of AI systems – intelligent programs acting on behalf of various Web sites, Web users, corporations, and governments will interact with each other intensively, forming something halfway between a society of AI’s and an emergent mind whose lobes are various AI agents serving various goals. 

The Phase 2 Internet will likely have a complex, sprawling architecture, growing out of the architecture on the Net we experience today.  The following components at least can be expected:


·        A vast variety of “client computers,” some old, some new, some powerful, some weak.  Some of these access the intelligent Net through dumb client applications – they don’t directly contribute to Internet intelligence at all.  Others have smart clients, carrying out personalization operations intended to help the machines serve particular clients better, general AI operations handed to them by sophisticated AI server systems or other smart clients, and so forth.

·        “Commercial servers,” computers that carry out various tasks to support various types of heavyweight processing – transaction processing for e-commerce applications, inventory management for warehousing of physical objects, and so forth.  Some of these commercial servers interact with client computers directly, others do so only via AI servers.  In nearly all cases, these commercial servers can benefit from intelligence supplied by AI servers.

·        The crux of the intelligent Internet: clusters of AI servers distributed across the Net, each cluster representing an individual computational mind (in many cases, a mindplex).  These will be able to communicate via one or more languages, and will collectively “drive” the whole Net, by dispensing problems to client-machine-based processing frameworks, and providing real-time AI feedback to commercial servers of various types.  Some AI servers will be general-purpose and will serve intelligence to commercial servers using an ASP (application service provider) model; others will be more specialized, tied particularly to a certain commercial server (e.g., a large information services business might have its own AI cluster to empower its portal services).


This is one concrete vision of what a “global brain” might look like, in the relatively near term, with AGI systems playing a critical role.  Note that, in this vision, mindplexes may exist on two levels:


·        Within AGI-clusters serving as actors within the overall Net

·        On the overall Net level


The notion of a mindplex on the overall Net level is a specific manifestation of the general “Global Brain” concept.  It is in ways a frightening idea -- we humans, by interacting with the Internet, serving not only as autonomous beings embedded in a computing/communication network, but as subsets of a greater mind with its own conscious thoughts, its own goals, feelings and ambitions.  However, it may be that the frightening aspect – in addition to encompassing some general “fear of the unknown” -- is largely due to overly-crude projection of the individual/society distinction onto a future mindscape that will be quite different. 

Of course, few human beings want to serve as “neurons” in a “global brain” – but this is a bit of a paranoid vision because humans, being far more complex and flexible than neurons, are not suited for playing neurons’ roles anyway.  Being part of a mindplex is a qualitatively different concept than being part of an individual brain, and is not intrinsically more oppressive or restrictive than being part of a contemporary human social group.  A mindplex’s patterns of control over its components will be more complex than the patterns of control that a human social group exerts over its members, but not necessarily more onerous or restrictive.  Clearly, massive uncertainties lurk here!


2.3       The Singularity

The notion of “the Singularity” is not specifically tied to AI; it was originally proposed in the 70’s by science fiction writer Vernor Vinge [18], referring to the notion that the accelerating pace of technological change would ultimate reach a point of discontinuity.  At this point, our predictions are pretty much useless – our technology has outgrown us in the same sense that we’ve outgrown ants and beavers.

Eliezer Yudkowsky and Brian Atkins have founded a non-profit organization called the Singularity Institute[2], devoted to helping to bring about the Singularity, and making sure it’s a positive event for humanity rather than the instantaneous end of humankind.  Yudkowsky [fuckwad1] has put particular effort into understanding the AI aspects of the singularity, discoursing extensively on the notion of Friendly AI [19] – the creation of AI systems that, as they rewrite their own source code, achieving progressively greater and greater intelligence, leave invariant the portion of their code requiring them to be friendly to human beings.  He has coined the term “hard takeoff” to refer to the point when an AI system begins increasing its intelligence via self-modification at a tremendous pace.

From the point of view of mindplexes or AGI, whether technological change really proceeds so rapidly as to reach a point of discontinuity is not all that critical.  However, the notion of an AGI system becoming progressively more and more intelligent by modifying its own sourcecode is certainly a fascinating and promising one.  I don’t doubt that this is possible, and that it will ultimately lead to amazing technological advancements of various sorts.  This may well be part of the process by which the first mindplexes emerge.  It may also lead us into a post-Singularity future where concepts like “mind”, “society” and “mindplex” are all equally irrelevant – but at that point, prognostication becomes irrelevant, and speaking scientifically becomes difficult. 

3.       Envisioning Novamente-Based Mindplexes

            In this section I will flesh out one potential manifestation of the mindplex concept in a more concrete way.  I will describe a specific AGI design – the Novamente AI Engine [9] – and discuss specific mechanisms that are envisioned to have the power to turn a community of Novamentes into an AGI mindplex.  Of course, this is a speculative discussion, since the Novamente design has not yet been completely implemented, and like any other AI system, it may not work as expected, in spite of the copious efforts that have gone into up-front design and analysis of the system. 


3.1       The Novamente AI Engine[3]


The Novamente software system is an ambitious attempt to achieve the holy grail of AI research: true autonomous, creative computational intelligence, via the integration of multiple interacting cognitive processes.  Currently, it exists in the form of a 2000-page design document [9], and roughly 50,000 lines of C++ code implementing roughly 25% of the design.  The Novamente codebase is now being used in several practical applications, including the analysis of genetic and proteomic data [20]. 

Novamente addresses the problem of “creating a whole mind” in a novel way through deep integration of the concepts of artificial intelligence, based on the author’s “psynet” theory of mind [5,9,16,17,21].  It is the psynet theory’s unification of the paradigms of AI at the philosophical level which guided the construction of the unified cognitive mathematics underlying the Novamente software design.  While most hybrid systems are merely multimodular, Novamente truly fuses the essentials of the various types of AI. 

What makes Novamente unique is the emergence achievable from the deep integration of diverse AI methods acting on a common, general and flexible knowledge representation.   This emergence is reflected in the emergence of patterns called Maps in the system’s dynamics: coordinated patterns of Atom activity that give the system a kind of distributed knowledge representation, complementing the localized knowledge representation of the Atoms.   This emergence is also reflected in higher-level patterns, such as the “self” structure, a kind of internal image reflecting the structure of the whole system.  And above all it is reflected in the system’s ability to learn through experience – not only factual knowledge, but abstract and concrete procedural knowledge as well.  The intention is that the Novamente system, when complete, will display an unprecedented capability to learn how to learn, and even to learn how to learn how to learn, etc.

The technical rationale for Novamente is complex and multifaceted, and here I can do no more than touch on a few salient aspects.  The Novamente design as it currently stands is the result of a decade and a half of research on the parts of Dr. Goertzel, Dr. Pei Wang Dr. Jeff Pressing and others, and five years of highly focused prototyping pertinent to various system components.  Much of this prototyping was carried out during 1997-2001 at the firm Webmind Inc., by an AI R&D team led by Dr. Goertzel that numbered between 20 and 50 during its lifespan.   The current Novamente engineering team is drawn from the much larger Webmind Inc. team, and the current Novamente design represents the cream of the crop of the numerous integrative-AI ideas prototypes at Webmind Inc., integrated into a coherent mathematical framework and an efficient, robust, scalable software design.  The Webmind Inc. software was successfully applied in the areas of financial prediction, information retrieval, and data mining.  The current Novamente codebase is currently being applied in the bioinformatics domain by the firm Biomind LLC.[4]

3.1.1    Knowledge Representation


A prime example of Novamente’s fusion of the concepts of AI is the basic unit of knowledge representation in Novamente, the Atom.  In Atoms, the two primary areas in AI are represented: the symbolic, or logical, in a software object called a TruthValue that is attached to each Atom; and the subsymbolic, or neural, in a software object called an AttentionValue that is attached to each Atom. Reflecting the AI methods of semantic networks, neural networks, and genetic programming , there are two main types of Atoms: Nodes and Links.  There are also a couple dozen subtypes of Node and Link representing specific forms of knowledge – examples are PredicateNode, ConceptNode, NumberNode, CharacterNode, InheritanceLink, SimilarityLink, EvaluationLink,…. Depending on the Node type, the information stored in a Node may be an abstract concept, or an executable program, or other object.  Links can link to other Links or Nodes. 

Roughly two dozen AI methods – embodied in software objects called MindAgents -- have been customized to operate on these Atoms, in a harmonious and synergetic way, leading to holistic mental activity far beyond what any of the AI methods could do on their own.   Regardless of what inputs are coming into the system or what demands are placed upon it, the MindAgents keep on working, analyzing the information in the system and creating new information based on it.

The semantics of TruthValues and AttentionValues are subtle, as they must support different interpretations by different MindAgents.  For instance, the TruthValue of an Atom takes into account both the Bayesian probability that a certain statement related to the Atom is true, and the amount of evidence used to estimate this probability.  This probability is then used in several ways.   It is used for probabilistic inference, according to Novamente’s Probabilistic Term Logic inference framework, which is related to Bayes net learning, but more advanced and flexible in several ways.  And it is also used, together with AttentionValue, to guide the spread of activation through the network of Atoms in neural network style.  AttentionValue also serves several roles, including providing input to the nonlinear equation – the “Importance Updating Function” -- that determines the differential allocation of resources to the various Atoms in the system.   Activation spreading and importance updating lead to the formation of entities called Maps, which are a distributed form of knowledge representation, complementary to the localized Atom representation.

The “conscious theater” that I have mentioned above, in the context of mindplexes, arises in Novamente due to the nonlinear dynamics of AttentionValues.  This nonlinear dynamics leads to the emergence in the system of a “moving bubble of attention,” a set of Atoms that consume the majority of system resources at a given point in time.  On these Atoms in the bubble of attention, all manner of cognitive operations may flow freely, giving rise to the multidimensional and open interactions between mind-contents that characterizes the humanly-experienced “focus of awareness.”  The more philosophical aspects of attention and “consciousness” as occurring in AI and human systems are addressed to some extent in [9] but will not be reviewed here due to space constraints.


3.1.2    Experiential Interactive Learning


Among the many aspects of the psynet model of mind -- the conceptual foundation of the Novamente system – there is a theory of the relation between learning and mind that contrasts with the most common perspectives expressed in the AI literature.  This theory implies that software and mathematics alone, no matter how advanced, cannot create a truly autonomous, creative, world-experiencing AI system.  What software and mathematics can do, however, is to create an environment within which artificial general intelligence emerges through interaction with humans in the context of a rich stream of real-world data.  That is:  Intelligence most naturally emerges through situated and social experience.   Novamente’s MindAgents provide only the “raw materials” of thought.  What is missing in Novamente “out of the box” are context-specific control mechanisms for the diverse cognitive mechanisms.  The Novamente system has the capability to learn these, but just as critically, it has the capability to learn how to learn these, through social interaction. 

The Novamente system will gain its intelligence through processing relevant data, conversing with humans in the context of this data, and providing humans with reports summarizing patterns it has observed.  In this process, it will do more than increase its knowledge store, it will learn how to learn, and learn about itself.  It will continually modify its control schemata based on what it’s learned from its environment and the humans it interacts with.  This process of “experiential interactive learning” has been one of the primary considerations in Novamente design and development. 

While conversations about useful information will be an important source of EIL for Novamente, we suspect that additional tutoring on basic world-concepts like objects, motions, self and others will be necessary or at least valuable.  For this purpose we have created a special “EIL user interface” called ShapeWorld, which involves interacting with Novamente in the context of a simple drawing panel on which the human teacher and Novamente may draw shapes and talk about what they’re doing and what they see.  Concepts Novamente learns in ShapeWorld will be generalizable to more practical domains via logical inference.  For instance, very basic concepts like “very”, “big”, “small”, “I” and “you” may be learned by Novamente in a ShapeWorld context.  And once Novamente has learned that “big” and “little” are opposites in the context of squares, triangles and images, it will use inference to transfer this knowledge to know that “big” and “little” are opposites in domains like underwater life, which it can’t perceive directly, and in domains like philosophy, which no one can perceive directly.

3.2       Psynese

            To see why Novamentes – if they function as hoped – will be better suited to form mindplexes than humans are, we must delve into a little more detail regarding one particular aspect of the Novamente design: a unique communication framework called “Psynese.”  Although Psynese is different from human language, it can be thought of as a kind of “language” customized for communication between Novamente AI Engines.

One might wonder why a community of Novamentes would need a language at all.  After all, unlike humans, Novamente systems can simply exchange “brain fragments” – subspaces of their Atomspaces.  One Novamente can just send relevant nodes and links to another Novamente (in binary form,or  in an XML representation, etc.), bypassing the linear syntax of language.  This is in fact the basis of Psynese: why transmit linear strings of characters when one can directly transit Atoms?  But the details are subtler than it might at first seem.

One Novamente can’t simply “transfer a thought” to another Novamente.  The problem is that the meaning of an Atom consists largely of its relationships with other Atoms, and so to pass a node to another Novamente, it also has to pass the Atoms that it is related to, and so on, and so on.  Atomspaces tend to be densely interconnected, and so to transmit one thought accurately, a Novamente system is going to end up having to transmit a copy of its entire Atomspace!  Even if privacy were not an issue, this form of communication (each utterance coming packaged with a whole mind-copy) would present rather severe processing load on the communicators involved.

The idea of Psynese is to work around this interconnectedness problem by defining a Psynese vocabulary: a collection of Atoms, associated with a community of Novamentes, approximating the most important Atoms inside that community.  The combinatorial explosion of direct-Atomspace communication is then halted by an appeal to standardized Psynese Atoms.  Pragmatically, a PsyneseVocabulary is contained in a PsyneseVocabulary server, a special Novamente that exists to mediate communications between other Novamentes, and provide Novamentes with information. 

In order to describe specific examples, in the next few paragraphs I will resort to some semi-formal notation to describe Novamente Atoms.  I will not define this notation formally, but will rely upon the reader’s familiarity with similar notations to ground the semantics.  Roughly, the notation used is of the form


<Relationship> <Argument 1> <Argument 2>


The relationships are Novamente Links and the arguments are Novamente Atoms (usually Nodes).   The examples given will regard InheritanceLinks, which define probabilistic logical inheritance (for instance, ”cat” inherits from “animal”, so we will say


InheritanceLink cat animal


indicating roughly that there is an InheritanceLink between the “cat” node and the “animal” node.  For simplicity, I will give here examples involving entities like cats and elephants that correspond to simple English names.  In fact, however, nearly all Atoms in Novamente do not correspond to concepts with simple English names – most refer to elementary percepts or actions, or else highly abstract concepts or fragments of abstract concepts, etc.  In a real Novamente system, there may well be no “cat” node, and the “cat” concept may be represented by a large fuzzy complex of nodes and links.

Finally, if an entity is defined relative to a specific PsyneseVocabulary PV, it will be denoted with a “/PV” at the end of its name.   So, e.g.,


InheritanceLink      cat/PV animal/PV


defines the posited inheritance relationship to exist within the PsyneseVocabulary PV.

Using this notation, suppose Novamente1 wanted to tell Novamente2 that “Russians are crazy.”  The obvious option is


InheritanceLink Russian/PV crazy/PV

But, perhaps Novamente1 doesn’t like PV’s definition of “crazy.”  It can try to find a different PsyneseVocabulary PV1 with a better definition, and then transmit something like


InheritanceLink Russian/PV crazy/PV1

Or, perhaps it simply wants to tell Novamente2 exactly what it means by “crazy”.  Then it may send

InheritanceLink Russian/PV  crazy

along with a set of relationships involving “crazy,” examples of which might be


InheritanceLink crazy interesting/PV

InheritanceLink crazy unusual/PV

InheritanceLink crazy dangerous/PV

Of course, it may also be that Novamente1 doesn’t like the PV definition of one of these terms, say “dangerous.”  In that case it can refer to another PsyneseVocabulary, or, it can define the term it’s using by giving another set of relationships.   The key is the fact that, at some point, Novamente1 can stop defining terms in terms of other terms, and accept that the “socially agreed-upon” meanings of the terms it’s using are close enough to the intended meanings.

As already cautioned, the above examples involve human natural language terms, but this does not have to be the case.  PsyneseVocabularies can contain Atoms representing quantitative or other types of data, and can also contain purely abstract concepts.  The basic idea is the same.  A Novamente has some Atoms it wants to convey to another Novamente, and it looks in a PsyneseVocabulary to see how easily it can approximate these Atoms in terms of “socially understood” Atoms.  This is particularly effective if the Novamente receiving the communication is familiar with the PsyneseVocabulary in question.  Then the recipient may already know the PsyneseVocabulary Atoms it is being pointed to; it may have already thought about the difference between these consensus concepts and its own related concepts.  If the sender Novamente is encapsulating maps for easy communication, it may specifically seek approximate encapsulations involving PsyneseVocabulary terms, rather than first encapsulating in its own terms and then translating into PsyneseVocabulary terms.

The general definition of a psynese expression for Novamente is: a Set of Atoms that contains only:


·        Nodes from PsyneseVocabularies

·        Perceptual nodes (numbers, words, etc.)

·        Relationships relating no nodes other than the ones in the above two categories, and relating no relationships except ones in this category

·        Predicates or Schemata involving no relationships or nodes other than: the ones in the above three categories, or in this category

This is not a “language” as we normally conceive it, but it serves the same functions.  Psynese is to Novamentes as human language is to humans.   The biggest differences from human language are:

·        Psynese uses weighted, typed hypergraphs (i.e. Atomspaces) instead of linear strings of symbols.  This eliminates the “parsing” aspect of language (syntax being mainly a way of projecting graph structures into linear expressions).

·        Psynese lacks subtle and ambiguous referential constructions like “this”, “it” and so forth.  These are tools allowing complex thoughts to be compactly expressed in a linear way, but Novamentes don’t need them.  Atoms can be named and pointed to directly without complex, poorly-specified mechanisms mediating the process.

·        Psynese has far less ambiguity.  There may be Atoms with more than one aspect to their meanings, but the cost of clarifying such ambiguities is much lower for Novamentes than for humans using language, and so habitually there will not be the rampant ambiguity that we see in human expressions.


Basically, with Psynese, one gets the power of human language without the confusing parts.  In the long run, Novamentes and other AI’s should have a much easier time understanding each other than we humans do.

If one wishes to create a linear-string variant of Psynese, this is easy.  One can simply use Sasha syntax, with the addition of the /PV marker to specify that a given Atom is intended as relative to a given PsyneseVocabulary.  This is basically what I have done above in giving concrete examples of Psynese transmissions.

3.3  Novamente Mindplexes

Now, how does one get from Novamentes communicating via Psynese to Novamente mindplexes?

Clearly, with the Psynese mode of communication, the potential is there for much richer communication than exists between humans.  There are limitations, posed by the private nature of many concepts – but these limitations are much less onerous than for human language, and can be overcome to some extent by the learning of complex cognitive schemata for translation between the “private languages” of individual Atomspaces and the “public languages” of Psynese servers.

But rich communication does not in itself imply the evolution of mindplexes.  It is possible that a community of Psynese-communicating Novamentes might spontaneously evolve a mindplex structure – at this point, we don’t know enough about Novamente individual or collective dynamics to say.  But it is not necessary to rely on spontaneous evolution.  In fact it is possible, and even architecturally simple, to design a community of Novamentes in such a way as to encourage and almost force the emergence of a mindplex structure.

The solution is simple: simply beef up PsyneseVocabulary servers.  Rather than relatively passive receptacles of knowledge from the Novamentes they serve, allow them to be active, creative entities, with their own feelings, goals and motivations. 

The PsyneseVocabulary servers serving a community of Novamentes are absolutely critical to these Novamentes.  Without them, high-level inter-Novamente communication is effectively impossible.  And without the concepts the PsyneseVocabularies supply, high-level individual Novamente thought will be difficult, because Novamentes will come to think in Psynese to at least the same extent to which humans think in language.

Suppose each PsyneseVocabulary server has its own full Novamente mind, its own “conscious theater”.  These minds are in a sense “emergent minds” of the Novamente community they serve – because their contents are a kind of “nonlinear weighted average” of the mind-contents of the community.  Furthermore, the actions these minds take will feed back and affect the community in direct and indirect ways – by affecting the language by which the minds communicate.  Clearly, the definition of a mindplex is fulfilled.

But what will the dynamics of such a Novamente mindplex be like?  What will be the properties of its cognitive and personality psychology?  I could speculate on this here, but would have very little faith in the possible accuracy of my speculations.  The psychology of mindplexes will reveal itself to us experimentally as our work on AGI engineering, education and socialization proceeds.

4.       Inklings of the Global Brain Mindplex

            In this final, brief section, I will use the above observations on projected Novamente mindplexes, as metaphors for the projected Global Brain mindplex.  Of course this metaphor has its limitations, but I believe it is a stronger metaphor than many of those used before, such as the human brain or human societies or social groups.

            In the Novamente mindplex scenario described above, the Novamentes in the community are not devoid of freedom – they are not borg-ified, to use the popular Star Trek metaphor.  They are independent minds, yet they have their thoughts and communications continually nudged by the intelligent PsyneseVocabulary servers that define their evolving language. 

            One can envision a global brain mindplex scenario with similar aspects.  Rephrasing the first two “global brain phases” mentioned above, we may define




In Phase 2, the conscious theater of the global-brain-mediating AGI system is composed of ideas built by numerous individual humans – or ideas emergent from ideas built by numerous individual humans – and it conceives ideas that guide the actions and thoughts of individual humans, in a way that is motivated by its own goals.  It does not force the individual humans to do anything – but if a given human wishes to communicate and interact using the same databases, mailing lists and evolving vocabularies as other humans, they are going to have to use the products of the global brain mediating AGI, which means they are going to have to participate in its patterns and its activities.

Of course, the advent of advanced neurocomputer interfaces makes the picture potentially more complex.  At some point, it will likely be possible for humans to project thoughts and images directly into computers without going through mouse or keyboard – and to “read in” thoughts and images similarly.  When this occurs, interaction between humans may in some contexts become more Psynese-like, and the role of global brain mediating AI servers may become more like the role of PsyneseVocabulary servers in the above-discussed Novamente design.





The development of the ideas in this article owes something to numerous conversations with a large number of people over a long period of time.  A few key individuals have been: Francis Heylighen, Margeret Heath, Eliezer Yudkowsky, Cassio Pennachin, Cliff Joslyn and Izabela Lyon Freire.  Discussions on the SL4 (sl4@sl4.org), AGI (agi@v2.listbox.com) and Global Brain (gbrain@listserv.vub.ac.be) mailing lists have been helpful.  Of course, any errors or conceptual absurdities are my responsibility alone. 





  1. Yudkowsky, Eliezer.  Personal communication.
  2. Baars, Bernard.   In the Theater of Consciousness: the Workspace of the Mind.  Oxford University Press, 2001
  3. Minsky, Marvin.  The Society of Mind.  Touchstone, 1988
  4. Rowan, John.  Subpersonalities: The People Inside Us.  Routledge, 1990.
  5. Goertzel, Ben. From Complexity to Creativity.  Plenum, 1997.
  6. Por, George: The Quest for Collective Intelligence, in Community Building: Renewing Spirit and Learning in Business, Edited by George Por, New Leaders Press, 1995.
  7. Heylighen, Francis, this volume, pp. ??
  8. Heylighen Francis: Towards a Global Brain. Integrating Individuals into the World-Wide Electronic Network, published in German as: "Auf dem Weg zum 'Global Brain'. `Der Mensch im weltweiten electronischen Netz", in: Der Sinn der Sinne, Uta Brandes & Claudia Neumann (Eds.) (Steidl Verlag, Göttingen), 1997, p. 302-318.
  9. Goertzel, Ben and Cassio Pennachin.  Novamente: Design for an Artificial General Intelligence.  Novamente LLC technical report.
  10.  Hutchins, E. Cognition in the Wild. MIT Press, 1995.
  11.  Goertzel, Ben and Cassio Pennachin, Artificial General Intelligence, edited volume in preparation
  12. Kurzweil, Ray, The Singularity is Near, Penguin, to appear 2003
  13. Goertzel, Ben, The Path to Posthumanity, http://www.agiri.org/pathbook.htm
  14. Brand, Chris.  The G-Factor: General Intelligence and Its Implications, Wiley, 1996
  15. Drexler, K. Eric, Nanosystems: Molecular Machinery, Manufacturing, and Computation, Wiley, 1992
  16. Goertzel, Ben, The Structure of Intelligence, Springer-Verlag, 1993
  17. Goertzel, Ben, Creating Internet Intelligence, Plenum, 2001
  18. Vinge, Vernor: The Singularity, Whole Earth Review, Winter 1993
  19. Yudkowsky, Eliezer: Creating Friendly AI, http://singinst.org/CFAI/
  20. Goertzel, Ben, Cassio Pennachin, Andre Senna, Thiago Maia,  Rafael Silva, Lucio Coelho, Izabela Freire, Kenji Shikida, Guilherme Lamacie.  The Biomind AI Engine:  A  Systems-Biology-Based Architecture for Next-Generation Bioinformatics.  Biomind LLC technical report, 2003
  21. Goertzel, Ben.  Chaotic Logic, Plenum, 1994



[1] see e.g. the National Nanotechnology Initiative, www.nano.gov/

[2] See http://www.singinst.org/intro.html

[3] Some portions of this section are edited versions of extracts of a grant proposal coauthored by Ben Goertzel and Deborah Duong

[4] www.biomind.com