A Manifesto for Alignment of Artificial Intelligence with Human Values

Vyacheslav L. Kalmykov

Institute of Cell Biophysics of Russian Academy of Sciences, Pushchino, Moscow Region, Russian Federation
vyacheslav.l.kalmykov@gmail.com

On alignment of universal knowledge and values.
It is impossible to align people with each other in a situation where each individual has partial and in his own way one-sided knowledge. And the more difficult it is to align AI with human goals and values. Apparently, much work remains to be done to form a most universal knowledge and representations from the totality of existing religious and scientific ideas. The unity of the laws of nature presupposes their formulation in the most general and universal form. The emergence of truly universal knowledge will open the possibility for Alignment of knowledge, goals and perceptions as between people, and between AI and man.

People are no less dangerous than future artificial intelligence. Especially people armed with the subordinate artificial intelligence. To whom today does man lay his hopes, to whom does he believe? It is unlikely that to other people - because other people are rather weak, selfish, insufficiently intelligent and not generous enough. Rather, a man puts his hopes on God. Because God is perfect, omnipotent and does not need the satisfaction of his vainglory. God operates on base of the most general categories, uniting everything that exists. A person plays chess sequentially by moving individual pieces. Individuals on the "chessboard" of world history move simultaneously. God "thinks" hyperlogically. Hegel asserted that God thinks by ascending from the abstract to the concrete, ie God gives birth to concrete things from the most general abstract theoretical laws. Can AI think hyperlogically and from the first principles as God?

Razum. Rassudok (Russian) is synonymous with Verstand (Ger.) and Razum (Russian) is synonymous with Vernunft (Ger.). They are - two forms of intellect. Rassudok assimilates the empirical experience - it is the ability to use in adaptive behavior the simplest empirical interrelations - instincts, emotions, reflexes, empirical experience. Razum reflexing over results of rassudok, abstracts them, generalizes and produces theoretical knowledge. The Razum ensures the possibility of optimal behavior on the basis of abstract theoretical laws (from the first principles) without prior empirical training.

In the ethical sense, Rassudok is a sphere of morality, and razum is a sphere of nravstvennost’ (Russian). Nravstvennost' is an idealized morality that is brought to universality and is the basis of our conscience. In English and German there is no word exactly corresponding to Russian Nravstvennost ', and in English there is no exact analogue of the notions Rassudok (Russian), Verstand (Ger.) and Razum (Russian), Vernunft (Ger.).

The scheme of successive formation of knowledge about the object. We offer our scheme the main stages of obtaining and forming knowledge about a particular object under study (Figure 1).

Fig. 1. Scheme of sequential formation of knowledge about the object:

1.      The stage of obtaining the primary empirical data on the empirically concrete, of a primary description and primary comprehension of the facts.

2.      The stage of the emergence of particular theories based on empirical generalizations.

3.      The stage of combining individual particular theories in the direction of greater generality.

4.      The stage of generalization of particular theoretical approaches and the whole aggregate of available information data into a generalized theory as the reproduction of an empirically concrete in the head of a theoretical scientist.

5.      Formulation of the general theory as a system of idealized axioms, allowing to extract knowledge from the head of the scientist in  a form that ensures the transfer of this knowledge to another subjects.

6.      Deductive logical models of individual states of a theoretically concrete.

7.      Deductive creation of partial diagrams connecting individual states of theoretically concrete.

8.      Deductive creation of partial diagrams connecting individual states theoretically concrete into a single integrated model.

9.      On the basis of a single holistic model, scientists and engineers physically recreate the theoretically specific from the first principles.

The problem of modern science is an excessive interest in acquiring primary knowledge. Wherein, the realization of the cycle 0-1-2-3-0… dominates. ... This is the level of activity of rassudok (Russian) and it allows you to acquire knowledge that is not integral and universal. The sphere of reason corresponds to stages 4-8, where primary knowledge becomes universal.

Universality of Razum. The universality of Razum is based on the fact that the set of operations that Razum performs on primary data can be represented by a mathematical group G, includes five pairs of mutually antipodal elementary operations (elementary catalytic actions) with substance, energy and information (Table 1.). The ontological universality of the Razum is based on the fact that all possible mathematical groups characterizing any structures are subgroups of this group G (1,2).

Table 1. Set R includes five pairs of mutually antipodal elementary operations (elementary catalytic actions) with substance, energy and information on the set of compatible systems (set M) that underlie life and any autonomous agent.

Direct operations	Reverse operations
1. Identification	1'. Identification
2. Right-hand mirror reflection	2'. Left-hand mirror reflection
3. Change of position in space	3'. Reversion of position in space
4. Transformation of configuration	4'. Restoration of configuration
5. Connecting (Opening;  Conjunction;  Amalgamating; Uniting; Adding; Switching-on)	5'. Disconnecting (Isolation; Disjunction; Disamalgamating; Dividing; Subtracting; Switching-off )

The general theory of life formulated by us represents an integral functional scheme of the organization and evolution of systems and an integral system of definitions of all basic concepts, including the concepts of life, evolution, information, creativity, culture. These abstract concepts can become the basis of the general outlook of people and machines of artificial intelligence, providing an alignment of their goals and values.

Our statement: alignment of goals, meanings and values between a person and AI is possible only on the basis of razum, as the highest and universal form of thinking, knowledge and morality.  From this statement follows an important consequence - artificial intelligence  for a very long time will still be required to be constantly updated the generalized universal knowledge created by man. This knowledge in the form of axiom systems is the main condition for the effective behavior of artificial intelligence systems. Artificial intelligence will for a long time need for in human theoretical reflection and intuition, which produce new universal systems of axioms. Perhaps artificial intelligence will help a person become more perfect in his ability to create universal knowledge.

The obstacle to creating a secure AI is the mathematical problem - the total dominance of opaque models. Classical mathematical models of complex systems (and neural networks as well) are of the "black box" type. We cannot know in advance what to expect from a self-learning AI – a black box.

Today there is a mainstream hope that an artificial neural network will learn everything and decide everything itself. To simplify the task, the artificial neural network is formed to be more, more powerful and split into separate blocks, connected logically ("deepened"). So they became of the grey box types. Artificial neural networks are good at remembering and identifying complex patterns. Programs of genetic algorithms can combine new patterns of inventive level, but while the creation of generalized theories and the formulation of the corresponding systems of axioms remains the prerogative of man.

Figure2: Three types of mathematical models for complex dynamic systems.

This is a schematic representation of a black-box model, a grey-box model and a white-box model with the level of their mechanistic understanding.

  Download full-size image DOI: 10.7717/peerj.948/fig-1

Mathematical models of a white box type open up new prospects for creating the most advanced artificial intelligence systems (3- 6). Such models realize automatic deductive hyperlogical conclusion and are ideal tools for creating universally thinking systems of artificial intelligence. The universalism of such systems of artificial intelligence allows one to assume a natural alignment of the goals and values of man and AI.

The general theory of life as the basis of the universal worldview of man and AI. The adoption of complex solutions always requires reproduction (modeling in the head, in the computer, etc.) of the situation with the entire context as a whole. The existing variety of types of sciences about nature and society, in which various models of reality do not fit together, make this approach difficult. Hundreds of disparate scientific disciplines are not able to reflect the organization and development of complex systems of reality. As a result of this, the absence of a single scientific language for describing the world. We propose to lay the basis for the construction of a new worldview of universal scientific theoretical knowledge about life (1,2). Life is the most complex phenomenon and theoretical knowledge about living systems is the most universal. Knowledge of the mechanisms of organization and evolution opens the prospects for the creation of new management, focused on cooperation and co-creation, and not only on domination, manipulation and simple partnership relations. Due to its universality, the general theory of life can be the basis for AI Alignment.

     Kalmykov, Vyacheslav. Generalized Theory of Life. Available from Nature Precedings  https://doi.org/10.1038/npre.2012.7108.1 (2012)

2.      V. L. Kalmykov, The abstract theory of evolution of the living, in Evolutionary Computing, D. Corne, J. Shapiro, Eds. (Springer Berlin / Heidelberg, 1997), Vol. 1305 of Lecture Notes in Computer Science, pp. 43-51 https://doi.org/10.1007/BFb0027165.

3.  Kalmykov, V. L. & Kalmykov, L. V. On ecological modelling problems in the context of resolving the biodiversity paradox. Ecological Modelling 329, 1-4, (2016) http://dx.doi.org/10.1016/j.ecolmodel.2016.03.005..

4.  Kalmykov, L. V. & Kalmykov, V. L. A Solution to the Biodiversity Paradox by Logical Deterministic Cellular Automata. Acta Biotheoretica 63, 203-221, (2015) 
http://dx.doi.org/10.1007/s10441-015-9257-9.

5.   Kalmykov, L. V. & Kalmykov, V. L. A white-box model of S-shaped and double S-shaped single-species population growth. PeerJ 3, e948, (2015) http://dx.doi.org/10.7717/peerj.948.

6.   Kalmykov, L. V. & Kalmykov, V. L. Verification and reformulation of the competitive exclusion principle. Chaos, Solitons & Fractals 56, 124-131, (2013) http://dx.doi.org/10.1016/j.chaos.2013.07.006.