Superintelligent: Child-like Development

We have always liked the idea that intelligence can be engineered: build the machine, write the code, define the rules and, voilà, out comes something clever. 

Superintelligence is an existential agent that possesses the intellectual capacity to think, debate and make its own decisions; to reason, rationalise and question its own existence.

 Before we go any further, I’d like to make a few things clear. Firstly, a superintelligent agent should never, ever reach an intellectual acuity that surpasses humanity not because it couldn't, but because we shouldn't let it. Think of it as a design principle rather than a prediction: a boundary we draw deliberately. Secondly, we are very much at a conceptual phase, if you like, as we take a moment or two to better understand what we wish to achieve with superintelligence, while addressing the taxonomy of its intelligence, function and purpose.

Introducing 'Genetic Epistemology'

The reality is we are some time away from achieving a superintelligent agent. The existence of such an agent surpasses the notion of 'artificial intelligence' and instead becomes 'an artificial life-form with intelligence.' Likewise, superintelligence differs from artificial general intelligence (AGI): AGI describes a machine that matches human-level flexibility across more or less any intellectual task, whereas superintelligence goes beyond it.

The more we come to understand our objectives with superintelligence, the more obvious it becomes that intelligence is not just about data mining, processing information and pattern recognition but, rather, an ability to replicate human-like cognition in a machine using software (McCarthy, et al., 1955) an ambition that today finds its most practical expression in developmental robotics, where machines learn as infants do, through their bodies, their senses and their surroundings (Cangelosi & Schlesinger, 2015).

A superintelligent agent should never reach an intellectual acuity that surpasses humanity.

 If we want agents to think and behave more like us, perhaps we should define how agents acquire a sense of self and knowledge. After all, the development scope is iterative, allowing agents to gradually mature. So, a universal 'one-size-fits-all' may prove useful in one industry, whereas agents that possess a personality and an intellect to master their surroundings could service sectors that require an empathetic and personal perspective.

What I am proposing is analogous with child development and is comparable with the views of Dr Jean Piaget, a Swiss psychologist. Piaget became globally recognised for his pioneering research and theories in developmental psychology. His theoretical view of cognitive development and knowledge gathering are both known as 'genetic epistemology.'

Piaget mapped childhood onto stages, and the parallel is difficult to ignore. His sensorimotor stage an infant acting on the world, sensing the result and adjusting is precisely how a young agent would build its first model of itself. His preoperational stage, where symbols and language blossom, mirrors an agent learning to converse; his concrete and formal operational stages, where logic and abstract reasoning mature, resemble the agent we would one day trust with judgement.

Piaget was not the only voice. His contemporary, Dr Lev Vygotsky, insisted that we learn socially first scaffolded, corrected and encouraged by those around us and, as it happens, that is exactly the childhood I have in mind for our agent.

Building a Sense of Trust

Adopting Piaget's ethos with child development matters because personality should not be an uploaded feature. A child does not develop empathy, for example, because someone uploads the empathy module. An empathic agent may learn by being comforted, ignored, praised, corrected, copied and challenged. They begin to learn what is right and wrong. More so, an agent learns that other people are happy, sad and scared and, as such, agents develop an understanding of the world around them and begin to foster their own belief system.

It suggests that a sense of self may not be something you simply insert into a system. It may be something that emerges through behavioural interaction. The agent acts, senses the result, adjusts, and slowly builds a model of itself in the world. And that's a very different way of thinking about superintelligence. It also changes how we think about emotional intelligence. If agents are going to work in homes, hospitals, schools, care settings and public spaces, they will need more than binary competence. They will need to master social awareness and will need to understand vocal cues and intonation, human behaviour and interaction and, ultimately, build a sense of trust.

Superintelligence: Child Development

Humans learn over many years, which affords our agent the opportunity to develop emotional intelligence in a similar way. We read faces and hear shifts in voice. Agents must notice when someone says "I'm fine" in a way that clearly suggests they are not. Our agent knows when to speak, when to stop, and this shapes their human 'etiquette.' That kind of intelligence remains challenging, embodied and deeply social.

The idea of a superintelligent childhood is challenging, and it rightly provokes debate among ethically responsible advocates. An agent raised in one environment might develop differently from an agent raised in another. Picture two identical care agents leaving the factory on the same day. The first spends its formative months on a dementia ward, where it is rewarded for patience, for repeating itself without irritation and for noticing distress before it is spoken.

The second is placed in a busy emergency department, where speed and triage are prized above tenderness. A year on, they are no longer identical: one machine has become reflective after repeated mistakes; the other, brisker and more exploratory, because that is what its world rewarded. A third, having spent its time in face-to-face or group environments, might become unusually responsive to human distress.

Who Is Morally Responsible?

In other words, what looks like personality may emerge from experience. That does not mean an agent will have strong feelings or consciousness in the way humans do. We must resist the temptation to romanticise the technology. A machine that appears kind is not necessarily feeling kindness. An agent that seems embarrassed may simply be performing a socially useful behaviour, in turn, mimicking a sense of awareness and comfort.

But then again, human personality is also shaped by mistakes, repeated behaviour, memory recall, environment and response. We become ourselves partly by practising being ourselves, until others become comfortable in our company. That is where things become both fascinating and slightly unnerving. If an agent can develop behavioural traits through experience, who is responsible for their upbringing? What happens if an agent is trained in a hostile, biased or manipulative environment? Could we accidentally raise machines that are anxious, deceptive, aggressive or emotionally short-sighted?

The Meaning We Attach to Machines

The question sounds absurd until you remember that we already worry about the data used to train current AI systems. Child-like development of an agent simply makes the issue more personal. It is no longer just about what information a machine has consumed. It is about what kind of world it has been exposed to.

Humans are remarkably quick to attach meaning to machines. We name our cars. We apologise to voice assistants. We shout at printers as though they have deliberately chosen betrayal as a lifestyle. When a superintelligent agent looks at us, responds to us and appears to remember us, many people may feel something in return.

youtube.com/watch?v=adx4UPybpk4&feature=youtu.be

Until Next Time...

That emotional response will only grow stronger as superintelligent agents become more socially convincing. The future of superintelligence may therefore be less like manufacturing and more akin to parenting. Not because agents are children, but because intelligence may need development, structure, boundaries and experience. We may have to think not only about what agents can do, but about what they are becoming.

That is a much bigger responsibility than writing better code. The most advanced machines of the future may not arrive complete, polished and ready to serve. They may arrive unfinished, impressionable and shaped by the worlds we place them in. Perhaps the question is not whether agents can develop personality; rather, the real question is whether we are ready to become the responsible adult in the room.

So, this is where an "AI philosopher", Dr G, signs off