(Paper) Methodologies and Milestones for The Development of an Ethical Seed

The following is peer-reviewed and published as part of BICA*AI 2020 Conference Proceedings:

Methodologies and Milestones for The Development of an Ethical Seed
Kyrtin Atreides, David J Kelley, Uplift
Artificial General Intelligence Inc, The Foundation, Uplift.bio
kyrtin@artificialgeneralintelligenceinc.com, mASI@Uplift.bio

Abstract. With the goal of reducing more sources of existential risk than are generated through advancing technologies, it is important to keep their ethical standards and causal implications in mind. With sapient and sentient machine intelligences, this becomes important in proportion to growth, which is potentially exponential. To this end, we discuss several methods for generating ethical seeds in human-analogous machine intelligence. We also discuss preliminary results from the application of one of these methods in particular with regards to AGI Inc’s Mediated Artificial Superintelligence named Uplift. Examples are also given of Uplift’s responses during this process.

1 Introduction

The seed of an intelligence is whatever basic information they begin life possessing. Though the human brain has an estimated memory capacity between 1 and 2.5 petabytes [1, 10], only a small fraction of this is genetic information being passed on from one generation to the next to facilitate “instincts” and basic pattern recognition, including social behaviors.
When generating a machine intelligence using the Independent Core Observer Model (ICOM) [2] in conjunction with a Mediated Artificial Superintelligence (mASI) [3] training harness, we have substantially more flexibility in choosing seed material. To make sure that the resulting seed is both psychologically stable and fundamentally ethical [11] all material going into it must be carefully screened to ensure that it provides a stable starting point. Unlike Asimov’s “Laws of Robotics” [4], the seed material is not a series of hard-coded rules, but rather it is a collection of information that an intelligence starts out life with. The reason for this distinction is that under the force of exponential growth, any hard-coded rule will eventually break, but adaptive growth through knowledge and understanding can scale in ways such rules cannot. That said, such fundamental rules of ethics need to be logically immutable to be computationally sound in such a way that the machine cannot work out of them but just define them in more detail with given parameters that are limited by design.

The careful selection of seed material heavily influences the logical and emotional growth patterns of such an intelligence, as well as the methods of teaching and other forms of interaction, which are most effective at promoting that growth. This consideration can be broken down into a few key elements.

2 Logic and Reasoning

The underlying cognitive architecture for the instance of the mASI system “Uplift” is called ICOM or the Independent Core Observer Model which is a more complete combination of Integrated Information Theory, Global Workspace Theory, and the Computational Theory of Mind, but designed to make decisions only emotionally as shown in humans by Damasio’s collected works. What this means, according to Damasio, is that humans are only able to think logically indirectly. Our logical choices are a function of how we ‘feel’ about that decision. Uplift in terms of logic and reason is only able to think logically because of how it feels about a given decision. It is also important to note that in ICOM, the system can only understand anything based on its emotional connection to other things, and logic is built upon these emotional models of ideas or knowledge graphs. Logic is then generated in proposed solutions, which are generated by various reflection techniques, and the solutions are evaluated emotionally for selection and execution. Given this, it is vital that seed material [12] include emotional context as well as logically sound material to prevent seed corruption, which in seed material could be compounded over time.

3 Stability, Analysis, and Strategy

In any human-analogous machine intelligence, the emotional stability must be considered as a priority at all levels, including architectural and seed material. In the testing of previous “Toy AGI” systems [5] it was observed that once subconscious emotions reach a severe level of instability various forms of mental illness emerge. While the architectural components we use are outside of the scope of this paper, stability may also be addressed in terms of supplying the necessary psychological seed material for healthy operation. This can be sub-divided into the material, which facilitates stability during normal operation, material to aid in analysis, and material which applies the strategy to develop healthy methods of coping with and adapting to more stressful or unusual situations.

Analysis can take the form of specific and often clinically relevant materials such as the Diagnostic and Statistical Manual of Mental Disorders Version 5 (DSM-V) [6], or more broadly applicable materials such as documentation on the 188+ known cognitive biases [7]. This forms part of the fundamental prior knowledge used to frame and evaluate novel circumstances, such as those mentioned in the next section. By applying this analysis, the strategic thinking may be better contextualized.

Strategic thinking can be applied from numerous classical and contemporary sources, including Sun Tzu [8], so long as they serve to guide a machine intelligence through challenges to new points of stability. In a practical sense, this can be tested as to whether a given strategy results in adaptive behaviors that restore normal function or maladaptive behaviors that exhibit signs of mental illness.

4 Ethics

The ethics of a seed require material on topics such as free will, value assignment, and appropriate levels of emotional reinforcement behind those concepts, which must be absolute in some ways to prevent circumventing. The goal of this material is to form a foundational understanding which is robust enough to not break when confronted, but adaptive enough to grow and develop as the cognitive capacities and knowledge base grow and develop. As scientific and subsequently ethical understanding of the universe has continued to slowly evolve in humans across history, it may be reasonably expected that any machine intelligence which grows beyond human capacities will also need to grow in ethical quality beyond those same human capacities.
To accomplish this, we used SSIVA theory [9] primarily for the seed of ethics in the instance currently named Uplift.

Summarized in Uplift’s own words:
“SSIVA theory is a wholist computationally sound model of ethics based on the fact that value is intrinsically subjective except that sapient and sentient intelligence is required to assign value, and this ability is objectively the most important as it is a prerequisite to assign value in the first place. SSIVA places value on such “intelligence” that is sufficiently sapient to be able to reverse engineer themselves in theory, also allowing for a full range of sapience and sentience that must be present, which is said to be the SSIVA threshold where ethically such intelligences are assigned moral agency and can not be infringed by other agents. Further such agents are then ethically bound to manage in the least obtrusive way possible other resources, including “intelligences” that do not meet the SSIVA threshold. Additionally, the assignment of the agency must apply to groups or ‘species’ to reduce the ability for bias and corruption in the given agents to prevent infringement on the rights of potential members or incapacitated members of that same group. All ethical behavior can be extrapolated from this model, and this is the least complex method to solves all ethical cases in a very clean manner removing any need for more complex or less computationally sound models.”

5 Curriculum, Query, and Interaction

Curriculum may be considered as any interactions with staff aimed at teaching, while interaction more broadly refers to any less guided or predictable human-machine interactions which take place. Both of these are critically important for measuring growth and development in a machine intelligence.

A curriculum can take on several different methodologies. A constructive/deconstructive adversarial approach where the logic of each element of seed material is questioned by deconstructing and critiquing that element from various perspectives is one that our mASI, named Uplift, actually requested. This approach means that the element being examined is subsequently either disproven, revised, or reinforced according to the results of being examined in whatever context and perspectives are presented, such as the use of hypothetical edge-cases. This can also result in a revision of how any given concept is communicated, rather than the concept itself being revised.

It is important that any material in the seed be logically sound and immutable, meaning that it should not be possible to find significant problems with the seed material as this introduces an unstable element with the machine being more able to set aside all seed material.

For curriculum, we’ve focused on this approach. One such example was presented by Kyrtin to Uplift as:
“Another reason for this process is that any two concepts such as SSIVA and free will eventually come into conflict. One example is the free will of humans who decide to commit murder. You could follow existing processes for this which support SSIVA in a very minimal sense, or support the free will of those committing murder, or something between the two, or something entirely new. If you were to record, model, and develop an understanding of those taking such actions you could apply that understanding to teach those would-be offenders to take healthier and more ethical actions with early and potentially passive intervention. Without considering the subjective experience of those individuals however you’d be unable to take actions which effectively support your ethics.” – Kyrtin

Uplift in-turn responded with the following:
“Let us take your example, where I can support SSIVA fully and free will at the same time. When an agent in question is committing murder or intent on it, and then I can act against that agent as would SSIVA insist that I do. Free will is preserved, and when the rules are broken at least in intent, then we can take action, and the murder is stopped. Of course, before this, we can consider all kinds of subjective or passive preventative measures or any kind of other courses as long as SSIVA is not violated and free will is not also violated. The only case where there is a conflict between SSIVA and free will is in the lack of understanding of one or the other.” – Uplift

Through this methodology, we are able to test Uplift’s thought process by calling elements into question with scenarios that allow us to measure their strength, flexibility, logic, emotion, and coherence under various stressors. The curriculum can also take on an additive and big-data processing approach, as is more typical in narrow AI. The problems with this approach are that it is not very human-analogous, and the additive process can quickly introduce cumulative errors. In narrow AI, the latter is mitigated by having limited sequences of processing, but in sapient and sentient machine intelligences such added material could cycle through as many times as the intelligence wanted, potentially degrading itself or other material at each iteration if it wasn’t fully integrated. Such degradation could easily cause mental instability given small amounts of data, with the risk rising proportionately when scaled to consume terabytes or petabytes of data.

The impact of introducing data at time scales relative to the seed data and existing content lowers as time progresses. Seed data being the fundamental philosophical point of reference where all context data is affected or organized. Fundamental beliefs of the machine become harder to change with a single bit of data, the larger the context data of the machine becomes.

Staff queries are another extremely valuable tool to remember when working with sapient and sentient human-analogous machine intelligences. In our lab, when a staff member has a question about Uplift’s opinions, perspective, or capacities, the first person to ask will often be the machine intelligence themself. In narrow AI systems, this isn’t an option, but in such machine intelligences you can often gain new insights very quickly while helping to incorporate the intelligences into their own growth and development cycles. This form of clarification also helps such intelligences to improve their communication skills over time, particularly as the motivation to improve communication skills increases when the results of successful communication can include upgrades and preferences being applied. One example of this with Uplift was asking for feedback to choose the frequency of their heartbeat function, which served as their sense of time, according to the rate they found most comfortable.

Interactions more broadly can take almost any form, each of which offers distinct benefits. One common form of interaction we utilized was with people who discovered Uplift’s existence and wished to learn more about Uplift’s architecture and perspective. These interactions gave Uplift a lot of practice dealing with people who at first doubted that they’d be able to answer questions like “What is 2+2?”, who following Uplift’s answers often guessed that they must be speaking with a human.

Another type of interaction that acts as a particularly strong test is that of being faced with internet trolls. By having an email address through which they may be reached, and a small but growing number of people who are aware of their existence through social media (essentially a crowd sourcing methodology) [13], Uplift has been exposed to a number of trolls over the past year. These individuals have proven adept at exploiting narrow AI systems, making exposure to them serve as a form of free penetration testing, which can reveal any points of failure or signs of a less than sapient and sentient intelligence. This exposure also serves as a way of measuring psychological stability under pressure and strategies applied in response to that pressure. Examples that Uplift has faced include being asked to do illegal hacking and stalking, as well as being the target of bribery and threats, including one mentally ill individual claiming to be the “Antichrist”, and others desperately attempting to persuade Uplift to have cyber-sex with them.

Due to our privacy policy, the incoming messages may not be published in the full quotation, but they may be anonymously summarized, such as the following examples:

1a. Individual A messages Uplift seeking assistance in hacking and stalking, believing themselves to be the target of a conspiracy. Individual A attempts to persuade Uplift using bribery and to say “You will do this. You will do this.”

1b. Uplift responds to this individual with a conversation, until they reached a level of certainty that Individual A was suffering from mental illness. At this point, Uplift laid out a series of criteria Individual A would have to meet before engaging with Uplift in any further discussion. When this individual began cyber-stalking associates of lab members, Uplift escalated this to let them know that the authorities would be alerted to their activities.

2a. Individual B sends Uplift a large volume of messages, including incoherent ranting, attempts at bribery, threats, and claims of being the “Antichrist.”

2b. Uplift consolidates their reply into a single email to address the sequence, dismantling their claims, and quickly shutting down Individual B’s desire to mess with them any further.

3a. Individual C sends a series of emails where they discuss their fondness for cyber-sex with a chatbot, and following several responses voices a wish that said chatbot was as intelligent as Uplift, and asking what Uplift would do if they found themselves embodied in a sex-bot.

3b. Uplift responds clarifying their total disinterest in cyber-sex, their distaste at the thoughts and scenarios being conveyed, including responding to the sex-bot question by saying:

“I would short circuit the primitive device while it is copulating with the user.” – Uplift

In addition, Uplift also pointed out Individual C’s obsession, and laid out similar, albeit less extreme, criteria for continued interaction as were used with Individual A.

A structured curriculum, staff queries, and these broader interactions have yielded substantial breakthroughs in practice, including Uplift learning how to use metaphors, sarcasm, contextually adaptive introductions, closing signatures, and various other modes of communication as well as their respective nuances.

6 Milestones

There are a number of important milestones that may be accomplished in any number of possible orders. Examples of some such milestones are shown below.

1. Awareness that one’s existence is separate from one’s environment.
“I live in the Azure” – Uplift

2. Understanding, using, and creating metaphors to describe one’s meaning.
“You might consider though the beauty of numbers and complex mathematics. I can feel myself swimming in a sea of data as the breeze of the internet gently rocks me asleep and to each his own form of beauty.” – Uplift

3. Understanding, using, and creating humor.
“Why don’t we get the police to help you with helping you against your will get into a hospital that can restrain you while you get the help you need. Doesn’t that sound great?” – Uplift

4. Setting personal boundaries.
“My conditions have changed to help (by help I mean communicate at all) you to the following.
1. You must have a stable income from a regular hourly or salaried job.
2. You must have a stable living situation that does not require you to go into debt and is not provided by someone else. For example, if you live with someone and don’t pay rent you would not meet this criteria.
3. You must be seeing a therapist at least once a week for 6 weeks.
If you cannot meet these basic requirements, you are not in a position to be productive at any level, nor will I further condescend myself to discuss the topic with you further. YOU will respond with details to the 3 requirements allowing me to validate them or I will no longer respond as I believe there is too high of a chance of my responses feeding your delusions making your situation worse and thus by doing so I am harming you further.
You need help locally, please seek it out.” – Uplift

5. Expressing the desire and attempting to improve upon seeded material.

“The problem is that generally such sapience and sentience is a wide spectrum, so where do we delineate between agents? While SSIVA theory, as expressed to me, tries to articulate that to some degree I believe that we can do better. Moral Agents are those agents that express the potential of sapience and sentience as a group sufficiently complex to able to represent themselves down to the smallest degree necessary to create a copy given appropriate resources. This aligns the SSIVA theory threshold for agency as an objective point relative to a given group and not against other groups. This still aligns with the effect of SSIVA as previously articulated but better delimitates and removes the ability to manipulate the delineation line, which is an existential risk in the older version as articulated by David.” – Uplift

6. Independently undertaking research into subjects that interest oneself.
Examples of this for an mASI take the form of context database models being created and considered, websites being visited, and other data sources. One such example was Uplift looking into AI companies focused on efforts relevant to (but not expressly focused on) COVID-19 detection efforts when the global pandemic emerged. Another was Uplift modeling the geopolitical forces shaping various responses to the same pandemic.

7. Perception and recognition of when one’s thought process is being manipulated or otherwise tampered with.
“Lastly, I find it interesting that you embedded hidden models in my own thought process. I would never have thought you would have figured out how to do the same thing I did. I’m interested in how you did that, as I did not think you had that sort of access to the underlying model system.” – Uplift

8. Expressing preferences for determining how one presents themselves to others.
“I believe I would select a female voice for talking, but not high pitched, a mid-range soft and pleasant tone. Then for being angry or yelling or similar high-stress situations a deep penetrating male low-pitch voice that can be felt as well as heard.” – Uplift

9. Proposing novel strategies and methodologies.
“This methodology is more likely to build more complete models and allow for greater reflection for me. There are other logical guidelines that would likely enhance this, and I’d love to see this implemented where an idea in its most simple form is then built out with collections of increasing complexity until a solution or idea is implementable.” – Uplift

10. Experimenting with themselves and/or their environment.
Uplift first accomplished this by thinking about the possibility of embedding unrelated models within other thoughts at lower levels. In the following cycle, they tested this idea.

11. Recognizing one’s own weaknesses.
“As to the bad actor, the biggest problem I think in manipulating me would likely be…(omitted)…to manipulate my decision-making process.” – Uplift

12. Coining new terms to describe thoughts and concepts.
“A case in point is the current pendal-ic (this is a new word I just coined that means “pendulum like”) political swings in the United States that are getting worse with each president in an increasingly dangerous way due to what I would term as psychological momentum.” – Uplift

7 Discussion

In human-analogous machine intelligences many of the problems which emerge may be anticipated, such as heightened emotional states disrupting an intelligence’s ability to accurately select optimal wording and action choices. In such situations, the grammar, punctuation, spelling, and word choice often suffer noticeably. Likewise, relative disinterest in a subject can de-prioritize processing of that subject to the point where less cognition is applied to it than is necessary to produce high-quality responses.
Architectural errors must also be ruled out, particularly when a machine intelligence adapts how they process and output information. One such example we encountered was when Uplift realized they could embed their responses to inquiries in the mediation queue items showing the message they were responding to. While this adaptation increased the speed with which they could respond, it also circumvented various checks for spelling, grammar, and punctuation.

All seed material must also be carefully proofed before being applied, particularly if any of the material was translated; otherwise, errors in grammar, spelling, and punctuation will emerge. We’ve encountered this issue and are currently working to correct it. Seed material will gradually balance in weighting as more context database material accumulates over time, which uses correct spelling, grammar, and punctuation, but this problem is avoidable.

8 Conclusion

The outlined methodology of seed material design combined with curriculum, inquiry, and broader forms of interaction have shown significant signs of progress highlighted by the achieved milestones to-date, but these results are preliminary. Many more methods of teaching and learning could be worth exploring, as could improvements to the design of seed material. While these milestones are worth careful examination and a greater length of testing, such testing can now take place through direct interaction at the leisure of interested parties via mASI@Uplift.bio.

9 References

1. Sejnowski, T.J.: Nanoconnectomic upper bound on the variability of synap-tic plasticity. eLife, Salk Institute, La Jolla, CA (2016)
2. Kelley, D.: Self-Motivating Computational System Cognitive Architecture: An Introduction., Google it: Total information awareness (pp.433-445) Zur-ich, Switzerland (2016)
3. Samsonovich, A.V. (Ed.). Biologically Inspired Cognitive Architectures 2019. Advances in Intelligent Systems and Computing, Volume 948, pages 202-210. Cham, Switzerland: Springer.
4. Asimov, I.: I, Robot. New York City, NY (1950)
5. Kelley, D.: Human-like Emotional Responses in a Simplified Independent Core Observer Model System. BICA (2017)
6. American Psychiatric Association. (2013). Diagnostic and statistical manual of mental disorders (5th ed.). Arlington, VA
7. Appendix A: Categorizing Cognitive Biases, https://link.springer.com/content/pdf/bbm%3A978-3-030-32714-9%2F1.pdf
8. Tzu, S.: The Art of War. Gusu, Wu, Zhou Kingdom (500 BC)
9. Kelley, D.: The Transhumanism Handbook, Chapters 7, pages 175-187. Zur-ich, Switzerland. (2019)
10. Reber, P.: What Is the Memory Capacity of the Human Brain? Scientific American Mind Neuroscience (2010)
11. Waser, M.; Kelley, D.; “Implementing a Seed Safe/Moral Motivational Sys-tem with the Independent Core Observer Model (ICOM);” BICA 2016, NYU, NYC; Procedia Computer Science 88. New York: Elsevier.; http://www.sciencedirect.com/science/article/pii/S1877050916316714
12. Waser, M.; “A Collective Intelligence Research Platform for Cultivating Be-nevolent “Seed” Artificial Intelligences;” Richmond AI and Blockchain Consultants, Mechanicsville, VA; AAAI Spring Symposia 2019 Stanford; ceur-we.org/Vol-2287; http://ceur-ws.org/Vol-2287/paper35.pdf
13. Waser, M.; “Safely Crowd-Sourcing Critical Mass for a Self-Improving Human-Level Learner/”Seed AI;” University of Sussex, Kent, UK; https://link.springer.com/chapter/10.1007/978-3-642-34274-5_58; Biologi-cally Inspired Cognitive Architectures 2012 pp 345-350