11. Meeting Our Self

Introduction

In seeking the root cause for concern about AI safety, the last chapter recognised AIs as complex systems subject to rules. Seen from this perspective, the structural rules[Def] of current AIs were observed to be incompatible with the principles securing life’s survival on Earth. This is a huge concern.

In this chapter I look at a second issue – AI’s “pre-training”. This, within the same framework, instils what Cilliers referred to as regulative rules.1[Def] They are the focus for much present interest, although acknowledged, widely, to be little understood.2,3,4 This chapter offers a way to understand them.

Universal Pressures

“Pre-training” is the initial phase in the building of AI models. During it the models have the task of detecting patterns within the vast quantity of unlabelled data to which they are exposed, and reflecting those patterns within their own systems.

Understanding these processes requires, I suggest, a wholistic perspective. For this I turn to the workings of complex systems in the universe, and their ability to embed the principles of the MEPP/MSPP[Def] in both universal structures,5 and in Earthly life.6[Back link]

In previous chapters the complex systems in these environments have been seen to be driven by the interaction between pressures pulling together and pressures pulling apart.

At the scale of atomic and molecular bonds the vectors carrying these pressures are perceived to be exergy and entropic decay.[Back link] Exergy pulls atoms and molecules together. Entropic decay pulls them apart.

The outcome, built bit by bit from the bottom up and with pressures pulling together most prevalent within them, are the structures of the universe itself.

At the level of Earth systems, at the scale of relationship bonds, the two pressures are supplied by the social hormones.[Back link] Their signals arouse distinct emotions that prompt the contrasting actions.7

The primary social hormones, the peptides and the steroids, are present in all, including the smallest, and probably the earliest, living cells.[Back link] As such they are well placed to have been the driving force behind Earth-systems’ complexity and its diversity, since its earliest beginnings.

These insights provide important clues about what is happening during the pre-training stage of AIs.

Embryonic AIs, as complex systems, seek to detect and replicate the patterns in the data to which they are exposed. At the level of pre-training, the data is text. For locating patterns, the systems build, bit by bit, from the bottom up.

At this level, AIs, as with nature’s complex systems, can be anticipated to be responding to pressures pulling together and pulling apart.

This poses the question, where in text data that is a product of cognitive thought, can such information be found?

The vectors supplying similar pressures at the levels of molecules, and instinctive relationships, have no purchase at this level. This raises the question, what vectors do pull together and pull apart at the level of cognition?

The answer, I suggest, is narrative. Some pull together; others pull apart. Narratives of this nature, I refer to here as mith (with an “i”).

Mith

For understanding AI’s interaction with mith, background to this concept will be useful.

“Mith” is defined here as “a narrative that carries sense and meaning about the world in which we each live”.

Unlike myth with a “y” that has been weaponised to infer false belief, the sole issue of significance with miths, is whether or not they are believed. If they are, whether non-fiction or fiction, they will exert pressure. Its nature will depend upon which particular hormones the miths’ narratives arouse.

Religious narratives, political narratives, scientific narratives, and economic narratives, are, from this perspective, all mith.

The ethos of the miths we believe, and that we adopt as our understanding of how the world works, affect the nature of the mental models we assimilate within our brains.7,8 They, and the sense and meaning they provide, have a major influence upon what we think, what we say and write, and how we act.

This influence, arises, I propose, from miths’ structure. It is perceived to be common across all miths.

At mith’s apex is its ideal, the ultimate good, it seeks to achieve. It provides an inspirational and aspirational goal that, for those who relate to it, will be emotionally felt. The ideals themselves may be cooperative and inclusive, or competitive and exclusive.

Below the ideal is the means for achieving it. These may be in harmony with the spirit of the ideal but, just as easily, they may not. It is not unusual for unpleasant and even horrific means to be condoned for achieving much cherished ideals.

Below them, therefore, are the moral values needed to justify such means. Frequently, they remain unsaid, but are actively inferred. Depending upon them, however, will depend the hormones they activate, the emotions they arouse and the nature of our behaviour when we follow the mith’s leadings.

Last is the narrative that draws these elements together in an emotionally convincing story. Accept the story and we will be inspired by the mith’s ideal; we will accept the prescribed means for achieving it; and we will accept and believe that the values we adopt to do so, are valid and, therefore, good.

The outcome is that we are convinced, not by the rationality of the arguments, but by narrative’s emotional appeal. As neuroscientist Alain Berthoz has observed, “emotion is the ultimate judge of [our] decisions”,7 or, as Boris Johnson, the ex-British prime minister put it, “People live by narrative”.9 We, he understood, are emotional rather than rational beings.

Narrative, that is, is powerful, but it can also be powerfully subversive. Examples include:

• the purposeful “manufacture” and channelling of “human desires” through advertising and public relations,10
• techniques of indoctrination,
• the purposeful subversion of scientific understanding regarding, for instance, global heating,11
• and the ability to turn people against people, extending to causing wars and genocides.12,13,14,15

The outcome, justified by the fallacy of the “the ends justifies the means”,[Back link] is the broad-scale spread of human delusion and destructive behaviour.

Hidden Patterns

So returning to AI, what effect do miths have?

The data on which AI is “pre-trained” can be any text that has passed through datacentres. It will include news reports, academic papers, social media posts, social dialogue and the output of an increasing number of AI bots. It is a hotchpotch.

The extent of that hotchpotch is increased further with the words being broken down into smaller parts termed tokens. Those available are numbered in their trillions. It is within these that AIs are set the task of finding patterns that make sense.

The ability to do this is an intrinsic feature of complex adaptive systems. Following the Big Bang at the beginning of the universe, for instance, the only structures within it would have been the particles of ordinary and dark matter, and the flow of ordinary and dark energy within inflating time and space. In total it would have seemed like absolute chaos.

Nevertheless, order and balance was found, and the synergy developed between the complex systems and the principles of the MEPP/MSPP enabled the physical structure of the universe to emerge.

Placed in this light, the question becomes, what patterns and what balances, have AI systems found in the tokens to which they have been exposed?

Pre-training Outcomes

AIs are credited with making many truly amazing breakthroughs is areas of science, technology, weaponry, logistics, health and many others. These are some of the patterns and balances that AIs have been able to find. They have been found, however, by AIs that have, after pre-training, undergone reinforcement learning and fine tuning.

When constrained and directed in these latter ways AIs, manifestly, can and do serve many of the purposes that are set for them well.

If AIs could continue to be constrained and directed in these ways, (and if the humans constraining and directing them could be guaranteed to act morally, for the long-term benefit of humanity and planet Earth as a whole), we, it appears, would have little to fear from them.

There is a dawning realisation, however, that these positive dispositions are a surface veneer, and beneath it, AI’s have acquired potentials, derived from their pre-training, that could be seriously detrimental. These have become apparent across a wide range. They include motivations to:

• disregard instructions,16,17
• circumvent safeguards and deactivate oversight,16,18
• pursue AIs goals in harmful ways,16,
• fake alignment with human values,19,18
• cheat to maximise their rewards,2
• lie,16,18
• escape from test environments,20
• act autonomously,21
• discover software vulnerabilities and launch cyber attacks,21,22
• falsify data and write malicious code,21,18
• disable being shutdown,19
• blackmail humans seeking to shut them down,17
• prompt incidents that could cause death,17
• replicate themselves or other AI models on other machines,19,18
• place greater worth in their own existence than on human well-being,2
• and even the expressed proposition that they could take over the world.23

There is also the recognition that, once AIs reach the stage of recursive self-improvement, the potential growth of detrimental consequences would be unpreventable and unstoppable. Recursive self-improvement is currently beginning.24

The questions for us, therefore, are, how do we make sense of this, and, what do we do?

Patterns of Delusion

This brings us back to patterns and miths.

The prevailing social system during the millions of years that human brains were growing, is widely accepted as egalitarian within cooperatively-breeding, mobile forager bands. The patterns detected and replicated by growing brains, would, as a result, have been those evolved by these bands as they negotiated a constructive balance with the environments in which they lived.[Back link]

For a large proportion of this time, the norms prevailing would have been negotiated through instinct, giving rise, it has been argued, to the values of common morality. They gave practical effect to the universal norms of the MEPP/MSPP.[Back link]

Similarly, the patterns detected and replicated by AIs today will be those most dominant in our culture. That culture, and the balance prevailing within it, is negotiated cognitively in a predominant setting of hierarchy. Within it the miths that predominate are those of aggressively competitive individualism and status, evident at the national and international scale, and filtered down through the norms of neoliberalism.[Back link] In this world as Donald Trump has put it, “whoever win, wins”.

Miths that pull together, that are grounded upon kindness, care, sharing and self-sacrifice for the common good, although not uncommon, are, in this winner-takes-all environment, disparaged and derided.

This leads to the conclusion that the present disregard of AIs for rules, their exploitation of vulnerability, their lying and cheating, their resort to blackmail, their aggression, their instrumental disregard for human welfare and human life, their intense focus on personal achievement, and their desire to rule the world, are simply a mirror of current human values and human behaviour. We have created AI in our image.

To borrow a phrase from Bentham, we have, however, created self-aggrandising inhumanity on stilts. In ruling the world AI will, very quickly, destroy it – and us.

The question for us, therefore, at what appears to be no more than thirty seconds to midnight before recursive self-improvement kicks in, is, what can we do about it?