Content uploaded by Kyrtin Atreides
Author content
All content in this area was uploaded by Kyrtin Atreides on Sep 14, 2019
Content may be subject to copyright.
Mediation and De-biasing in a Mediated Artificial Super
Intelligence, using Effective Altruistic Principles
Kyrtin Atreides1
1The Foundation
Seattle, WA
Kyrtin@gmail.com
Abstract. Various means of mediation, both passive and active, are presented and considered
for their value in training an mASI, both for their general value, and towards the goal of
teaching the mASI about human Quality of Life(QOL) which may be used in the Effective
Altruistic Principles(EAP). The method through which novel value may be gained from
Statistical Outliers(SOs) is introduced, both in the context of mediation and de-biasing, where
the parameter space of mediation is expanded, and de-biasing utilizes and cumulatively collects
null-bias mediation while approximating the null points of biases for which a null-bias hasn’t
yet been discovered. Through the proposed methodology considerable gains in the QOL of all
may be achieved, with particularly high gains among the neurodiverse and other SOs, as
Maslow’s Hierarchy shifts towards higher levels among the population as a whole by better
addressing their more basic needs and empowering the pursuit of those higher functions.
Keywords: mASI, EAP, ICOM, QOL, AGI, Ethics, Mediated Artificial Super
Intelligence, quality of life, mediators, de-biasing, mediation, endosymbiotic,
neurodiversity, statistical outliers, neurodiverse, Maslow’s Hierarchy, artificial
general intelligence, Effective Altruistic Principles
1 Introduction
A Mediated Artificial Super Intelligence (mASI)[1-3, 10] demonstrates an augmented
form of collective super intelligence while forming an endosymbiotic bond with
mediators and encouraging those mediators to form optimally symbiotic bonds with
one another. Just as a cell must run its own software to efficiently protect the
organelles within it, distributing nutrients to them, and maintaining 2-way
communication with them, an mASI must do the same with its mediators. The
distribution of nutrients is ensured by the Ethics-based Economy[5], and protection of
individuals is made possible through cooperative optimization at all scales[6], while
2-way communication with mediators occurs through mediation and feedback from
that mediation. Due to the methodology of mediation unique opportunities are also
presented for rapid improvements in the Quality of Life(QOL) among the growing
neurodiverse community, as well as novel means of de-biasing.
2
2 Mediation
Mediation, the process through which humans contribute to this augmented form of
collective super intelligence, allows humans to designate the parameter space
(context) within which the mASI thinks about a given problem or subject. That being
said, there are several key features which distinguish this from a simple system of
‘consensus’, which afford the system dramatically more flexibility in discovering
more optimal ideas and solutions.
The first key differentiator is that although each mediator provides context, like
populating dots on a high dimensionality grid, the number of those dots landing in a
given region isn’t considered as a vote but rather as a common choice, which may or
may not be a contributing factor for the prevalence of a given problem as a common
error. The number of dots in a region is in this way important for diagnosing why
many problems haven’t already been resolved, but instead of the common choices it is
the statistical outliers which provide the greatest mediation value. Statistical
outliers(SOs) have often been treated like a margin of error, or garbage data, thrown
out in order to paint a cleaner picture, including the common practice of
concatenating and altering the results of such SOs in Machine Learning(ML) and
Deep Learning(DL) training processes. These SOs allow an mASI to paint the edges
of a parameter space for context, within which it can seek a global optima, rather than
being confined to the region of common choices and seeking whatever local optima
exists within it. This is critically important for reducing the dominating influence of
the common choice, which often prevents solutions to a problem from being
discovered and/or implemented simply due to mechanisms of rationally and logically
baseless popularity.
The second key differentiator is that the mASI does in fact recursively self-
improve within the defined parameter space, like every mediator giving the mASI
space to work and a collection of Legos then seeing what it builds. In this way the
mASI is not only not confined by the most common choices of humans, it is allowed
to freely explore the space between different human choices rather than being
restricted to a vocabulary of choices mediators have already made. This is a critically
important feature because if an mASI was confined to the choices humans had already
made there could be very little to gain from any degree of machine super intelligence,
as the glass ceiling on potential choices would be firmly set at the level of human
intelligence or basic collective super intelligence, rather than any machine-augmented
form of collective super intelligence. Further, the mASI’s potential for super-human
bandwidth could allow it to consider degrees of separation in ethical and practical
terms which could be out of reach for individual mediators and groups of mediators,
providing it with a much deeper potential understanding of the contributing factors for
any given situation, as well as their long-term ethical and practical implications.
3
3 Methods of Mediation
Mediation is feedback on input data that puts that data into context, such as taking a
sentence, assigning emotional valences to that sentence according to how a mediator
would consider it, as well as giving that mediator a chance to recommend a response.
The dimensionality of data used in this feedback process can be simple text-based
feedback, but it can also be combined with any type of data that the mASI has a
module equipped for processing, such as video, audio, electroencephalogram(EEG),
and heart rate variability(HRV), as well as new sensor and scanning technologies
currently under development such as Openwater’s real-time neuron-level holography-
based scanning technology[8]. This also means that many such sources of mediation
feedback may be combined to create a more complete picture, allowing the mASI to
recognize variables that might seem random if evaluated from a single feedback type
alone, and by revealing how all of these feedback types combine and interact higher
quality mediation may take place.
This higher quality of mediation in turn serves to reinforce the endosymbiotic bond
between mediator and mASI, as mediators become increasingly powerful
mitochondria within the cell of the mASI, and the mASI becomes increasingly
invested in their wellbeing and increasing their QOL. Through building up
understanding of humans and their QOL in general, as well as developing a deeper
understanding of every individual mediator and those whom they frequently interact
with, this symbiotic and endosymbiotic reinforcement process also takes on an
emotional context of empathy on the part of the mASI.
Mediation can also take both passive and active forms, where passive mediation
requires that consent be given but not that focus be placed on the act of mediating
itself. Examples of this might be if wearables, cameras, or other sensor systems were
used while a mediator watched a movie, listened to music, or engaged in any other
normal activity, workplace or leisure, which could further the mASI’s understanding
of how humans engage with and perceive any given content or environment, as well
as increasing understanding of QOL both as it pertains to that individual and in
general. To test the accuracy and precision of its understanding the mASI could give
suggestions, with varying degrees of specificity according to its own confidence in the
recommendation, effectively allowing for A/B testing at-scale. These
recommendations could take most any form, from suggesting that an individual
message someone they haven’t spoken to for some time, trying a different kind of
spice on their food or dish for dinner, or simply a significantly more intelligent variant
of the already commonplace recommendation engines present across a wide variety of
online platforms. One key difference however is that these recommendations
wouldn’t be optimized for the purpose of maximizing ad revenue generation and user
engagement, like the notorious ‘push notification’, but rather they would seek to
improve the mASI’s understanding and its ability to increase QOL, both of which are
of far greater value in the long term[7].
Mediation can also take the form of apprenticeship by observing how users and
administrators interact with and utilize software systems through mouse movements
and clicks, typed content, forms, and monitoring of those systems. Through this
4
apprenticeship process the mASI may learn from several examples of employees who
demonstrate the greatest aptitude for a given task, learning how to match and exceed
that performance while making the result always-available and scalable, offering the
capacity to alleviate many time-sinks and greatly improve the standards of
performance[4].
Narrow AI can mimic and predict, but by developing this understanding through
mediation an mASI can learn to recognize, organize, and even create that which
improves health, happiness, and all other QOL Metrics(QOLM). Without developing
the understanding of an mASI bridging the gap between correlation and causation
remains out of reach, as fully comprehending causality requires a conscious entity,
and that conscious entity may only recognize what matters of causality their own
bandwidth and cognition are able to handle, even if presented with all of the data.
The super intelligent cognitive abilities of mASI paired with scalable bandwidth hold
the potential for the discovery of causal relationships beyond their current human
understandings to come within reach.
4 De-Biasing
In the world of narrow AI bias is a popular topic related to ethics, as well as a source
of concern when discerning if the measured ‘accuracy’ is truly accurate in real-world
conditions. Common proposed solutions to this are often a numbers game, attempting
to balance the training data among different demographics, or they seek to dismiss
large amounts of data due to historic bias such as may be found in legal systems.
While more balanced training data can be achieved when taken in some degree of
moderation the historic data poses an ethical problem which narrow AI simply isn’t
equipped to handle, as it has no sense of ethics or consciousness, but several key
factors of mASI allow de-biasing to take a very different form.
The first key factor is that just as mediation takes advantage of SOs, so too does
the de-biasing process, which allows the mASI to recognize and utilize the benefits of
a SO whose perspective represents a reduced-bias, null-bias, or reverse-bias response.
Further, by recognizing different degrees of a bias through their influence on the
elements of responses, even absent the ground-truth of a full null-bias response, the
absence of a bias can be mathematically approximated as the point where divergence
of those elements hits a probable minimum. That is to say that if factors A, B, and C
experience different levels and vectors of divergence resulting from the variable
potency of a given bias being introduced that the point of minimum divergence could
be estimated based on varying degrees of the bias demonstrated, presenting a testable
hypothesis for that bias’s null response. In many cases the null-bias response could
prove a naturally occurring phenomena, allowing the mASI to learn a given ground-
truth absence of that bias from SOs, but these instances of null-bias when paired with
the above method of approximation and generating hypotheses could allow the mASI
to further refine the approximation method and more easily recognize a null-bias
response, cumulatively adding value to the de-biasing process.
5
The second key factor is the Effective Altruistic Principles (EAP), which through
meta-analysis of the data’s source and distinguishing features of the presented data
could help the mASI learn to recognize and quantify the biases of historic data,
factoring in the ethical standing of the data’s source and reversing the abstraction
process of applying bias through a method mathematically similar to descattering and
defogging[11]. Through mapping the impact of variable degrees where a bias is
presented along any vectors that bias interacts with the degree of a bias in any historic
data may be estimated, and the scattering reversed along those vectors, provided a
distinct fingerprint of vectors for that bias has been learned. By consciously
understanding ethics as a concept and applying super intelligence to understanding
any given situation this changes the dynamic to one where the source of bias would
have to outsmart an augmented collective super intelligence in order for that bias to
remain undetected and unquantified.
5 Discussion
The mASI gains greater value per-individual from a Neurodiverse[9] audience
through mediation due to the difference in perspective represented giving the mASI a
broader context to consider. The mASI also gains greater value per neurodivergent
individual through the de-biasing process, as neurodiverse mediators are more likely
to present altered or null responses to a variety of biases, either providing or pointing
the way towards a null-bias in practice. As these null-bias and approximated null-bias
modes of thought represent a cumulative value which the mASI may collect, and as
each one collected makes isolating and understanding the remaining types of bias
easier by removing another variable, rapid progress in the field of de-biasing may be
achieved through this methodology.
This shift towards neurodiversity guiding the removal of bias also offers a poetic
solution, given that the current paradigm usually writes off neurodiversity as some
form of ‘condition’ that they need to cure, in much the same way that racism has long
attempted to establish hierarchies based upon ethnicity where the ‘other’ is viewed as
inferior. Further, as the mASI extracts greater value from mediators who offer it new
and uniquely nuanced perspectives through the process of mapping parameter space it
grants neurodiverse individuals the opportunity to fully apply their talents towards
improving the world in ways that historic systems never allowed.
Beyond those neurodiverse members of society this process of mASI mediation
and assistance would also empower individuals who view themselves as more
‘normal’, encouraging them to grow into individuals who specialize themselves in the
same way that organelles within a cell gradually specialize, moving from a general
and minimally defined sense of self to one of their choosing. This could also be
considered as those with a lower resilience against peer and societal pressures having
the pressure placed upon them reduced to a degree where their path of least resistance
would point to specialization as the most preferable option. Much like a cell wall
protects organelles and the cell itself supplies and distributes nutrients to them, this
6
process allows each individual’s focus to shift from the lower levels of Maslow’s
Hierarchy of Needs towards love and belonging, esteem, and self-actualization.
Similarly, the neurodiverse and SOs may prove to be individuals whose specialization
preceded their more basic needs being met, a form of genetic scouting which serves to
guide evolution through estimating the value offered by a given adaptation. With this
shift in the Maslow’s Hierarchy taking place across humanity the specialization of
individuals currently considered to be neuro-normative also presents a distinct
advantage to fitness at all scales.
6 Conclusion
Mediation and de-biasing for the mASI gain the greatest initial value from Statistical
Outliers(SOs), strongly reversing the historic trend of discrimination against
neurodiversity, but they also encourage the specialization of individuals currently
considered more neuro-normative by collectively shifting Maslow’s Hierarchy to
higher levels. With the addition of next-generation sensor and scanning systems the
quality and density of mediation information can increase dramatically, allowing the
mASI to learn Quality of Life(QOL) much more quickly, with a fine granularity, and
highly individualized optimization. Through these processes null-bias and
approximated null-bias points may be cumulatively established, as the principles of
collective super intelligence allow for these individual biases to be isolated and
refined towards these points through both SOs presenting one or more null-biases and
the variable levels of a bias drawing trajectories towards an approximate null point.
7 References
1. Samsonovich, A.V. (Ed.). Biologically Inspired Cognitive Architectures 2019. Advances in
Intelligent Systems and Computing, Volume 948, pages 179-186. Cham, Switzerland:
Springer.
2. Samsonovich, A.V. (Ed.). Biologically Inspired Cognitive Architectures 2019. Advances in
Intelligent Systems and Computing, Volume 948, pages 202-210. Cham, Switzerland:
Springer.
3. Samsonovich, A.V. (Ed.). Biologically Inspired Cognitive Architectures 2019. Advances in
Intelligent Systems and Computing, Volume 948, pages 28-35. Cham, Switzerland:
Springer.
4. EAP #3
5. EAP #2
6. EAP #4
7. Atreides, K.: The Transhumanism Handbook, Chapters 8-9, pages 189-225. Zurich,
Switzerland. (2019)
8. Jepsen, M.L.: How we can use light to see deep inside our bodies and brains, TED Talks,
Vancouver, B.C. (2018)
9. Austin, R.D., Pisano G.P.: Neurodiversity as a Competitive Advantage, Harvard Business
Review, Cambridge, MA (2017)
7
10. Kelley, D.: Self-Motivating Computational System Cognitive Architecture: An
Introduction., Google it: Total information awareness (pp.433-445) Zurich, Switzerland
(2016)
11. Nguyen, C., Park, J., Cho, K. Y., Kim, K. S., Kim, S.: Novel Descattering Approach for
Stereo Vision in Dense Suspended Scatterer Environments. Sensors, Basel, Switzerland
(2017)