Extreme AI: Closing the Loop and Opening the Spiral

This informal paper proposes a simple model for artificial intelligence (AI) systems that can learn and teach, as well as create mutations of themselves, so that subsequent generations of AI systems can be increasingly more intelligent than their predecessors, all without the need for human intervention. A single generation constitutes the loop. Subsequent generations constitute the virtuous spiral.

Beyond the discussion here in this paper, I have a much more extensive paper on AI, Untangling the Definitions of Artificial Intelligence, Machine Intelligence, and Machine Learning.

Levels of Artificial Intelligence

I see artificial intelligence (AI) as having five levels:

  1. Weak or light AI — individual functions or tasks, in isolation.
  2. Moderate AI — integration of multiple functions and tasks, as in a robot or driverless vehicle.
  3. Strong AI — incorporates roughly or nearly human-level reasoning and some significant degree of learning.
  4. Extreme AI — systems that learn and can produce even more capable systems that can learn even more capably, in a virtuous spiral.
  5. Ultimate AI — essentially Ray Kurzweil’s Singularity or some equivalent of superhuman intelligence.

Whether my conceptualization of extreme AI is comparable to the intelligence of Ray Kurzweil’s Singularity is left for a separate debate. I am content to suggest that it is a big step towards the Singularity, maybe.

The central thesis of this informal paper is that an Extreme AI is capable of near-human or even super-human level of learning as well as the ability to produce other systems and teach them how to learn and teach as well. This chain of learning and teaching is referred to as closing the loop. The capacity to produce new systems that can learn and teach on their own with increasing levels of capability is referred to as the spiral.

This paper makes no distinction between robots and non-robotic software systems. Robots simply have mobility and the ability to physically manipulate objects in the real world, while a purely software system cannot physically move or act physically per se, but can communicate with other systems which are not physically near by, or even transmit a copy of itself or a created software system electronically to another system. In some sense a software system can move, but not in the robotic, physical sense.

To be clear, the Extreme AI model proposed in this informal paper is an unsolved research problem. It is more of a thought experiment than an implementation-ready plan.

Traditional AI

Traditional AI takes several forms:

  • Heuristics — clever algorithms that give some of the effects of intelligence without true intelligence per se.
  • Smart tasks — individual functions or tasks, in isolation.
  • Smart systems — limited coordination of individual smart tasks.
  • Autonomous systems — such as space probes or driverless vehicles.
  • Machine learning — discovering patterns and rules in data.
  • Robots — mobile, quasi-intelligent systems.
  • Software agents — quasi-intelligent software modules that communicate with other agents or traditional systems over communications networks, and may even be able to move across networks or create instances of new agents across networks.

Limited learning

The main limitation of traditional AI is that any intelligence must be mostly pre-programmed, with only limited ability to learn from the environment.

Generally, traditional AI systems have a training phase. In essence the training data set is really just another form of pre-programming — the AI system doesn’t have the opportunity to pick its own training data.

Even so-called Deep Learning or neural networks really only discover relatively limited patterns and rules of behavior, not learning about the environment in any real, human sense. The developer is responsible for setting up the training data — the system is not free to choose what it wants to learn about.

Learning and teaching

The heart of Extreme AI is general learning and general teaching.

To be clear, these are unsolved research problems.

For our purposes here, we define learning as:

  • Observing the world around us, gathering all available data.
  • Discovering or discerning all possible patterns, rules, principles, logics, limits, and constraints discernible from that data. This constitutes knowledge, including reasoning.
  • Continually monitoring for new data, comparing and contrasting it with previously discerned knowledge, updating the knowledge base accordingly.
  • Speculating and creating hypothetical models and thought experiments, validating them against all available data and previously discerned knowledge, updating the knowledge base accordingly.
  • Accepting knowledge from other entities, computational or human, discerning how and if it can be blended with the existing knowledge base.

How to do all of that is an unsolved research problem.


We define teaching as simply conveying all of a distilled subset of the accumulated knowledge base of an entity to another entity, computational or human. How to do that is an unsolved research problem.

Armed with only the ability to learn and to teach, an AI entity could become quite intelligent, but would still not be a complete Extreme AI entity. As initially noted, learning and teaching are the core, but they are not sufficient by themselves to produce a full-blown Extreme AI system.

Learning how to learn

The primary constraining factor for producing AI systems at the present time is the need for intensive pre-programming of the system, even if it can be somewhat automated using an automated training process with training data. We can produce systems with a limited capacity to learn, but we haven’t managed to produce systems that can learn how to learn — we have to pre-program them how to learn, and even then for a limited degree of learning.

Pre-programming an AI system to learn how to learn is a necessary requirement for an Extreme AI system, but it remains an unsolved research problem.

Learning how to reason

The conception of learning proposed by this paper includes not only isolated facts, patterns, and connections, but principles, logics, and reasoning as well. Yes, that’s a very tall order and definitely an unsolved research problem, but essential for Extreme AI. It won’t be sufficient to have only pre-programmed reasoning.

Learning how to teach

We haven’t yet been able to design AI systems that can teach other AI systems complex, general knowledge, let alone achieve the required step of developing an AI system that can itself learn how to teach. In essence, rather than directly solve the problem of pre-programming teaching per se, we should focus on how to enable an AI system to itself learn how to teach.

That is indeed a very tall order, most certainly an unsolved research problem, but is nonetheless a necessary step in the development of Extreme AI.

Teaching how to learn

If learning how to teach is an extremely hard research problem, we can only imagine how difficult it will be to discover how to develop AI systems which can themselves teach other AI systems how to learn. Nonetheless, this is a necessary step in the development of Extreme AI.

Teaching how to teach

And if learning how to teach is a difficult research problem, imagine how difficult it will be for an AI system to teach another AI system how to teach.

Actually, there is a shortcut — if we have Extreme AI to learn how to teach and Extreme AI to teach how to learn, we can use simple composition to cobble together Extreme AI processes so that an Extreme AI system can learn how to teach how to learn how to teach how to teach. Cool.

Closing the loop

To summarize, there are three basic capabilities required for Extreme AI:

  1. Learning how to learn.
  2. Learning how to teach.
  3. Teaching how to learn.

Combined, they can provide us with the fourth required capability of teaching how to teach.

With these four capabilities we can come full circle, producing an AI system that can itself produce a functionally complete AI system that can both learn and teach, in a general, human-like sense.

I call this closing the loop.

In summary, the Extreme AI loop consists of:

  1. Learning how to learn.
  2. Learning how to teach.
  3. Learning and accumulating knowledge about the environment, the universe.
  4. Reasoning, speculating, creating, and experimenting, accumulating additional knowledge.
  5. Sharing knowledge with other Extreme AI systems.
  6. Instantiating a new AI system.
  7. Teaching it how to learn.
  8. Teaching it how to teach.
  9. Rinse and repeat.

The virtuous spiral

AI systems running in a pure circle, recreating themselves endlessly, like a dog chasing its tail is not exactly useful. The key revelation is that each cycle of the loop has the potential for evolving the AI system. I call such a continually evolving AI system a virtuous spiral — the loop is still there, but at the end of each cycle the system is now at a whole new level rather than returning back to its starting point.

This is what the title of this paper means when it says closing the loop and opening the spiral.

Monitoring and coaching the student

In order to prevent endlessly looping without progress, we need to monitor and coach the student. Sure the student could learn completely on their own, but it would probably take too long to learn everything from scratch on each iteration in the spiral.

Monitoring and coaching permit the spiral to be stretched.

Coaching must be moderated so that the teacher is not limiting the student, but simply guiding them.

The end goal is that the student’s knowledge and capabilities exceed those of the teacher.

Suggestions, hints, and alerts

Actually, monitoring and coaching of students by teachers would simply be specialized instances of more generalized monitoring and hinting mechanisms for sharing knowledge. The intention is not for the teacher to dictate what the student should believe, but to offer suggestions and hints for both directions worth considering and directions for which less attention is warranted or which are likely to be less fruitful or even fruitless.

Monitoring would be a specialization of an alerting mechanism, to allow one Extreme AI system to be notified about changes in selected areas of the knowledge base of another Extreme AI system.

Student vs. teacher vs. peer

Ultimately, once a new Extreme AI system gets up and running, any sense that it was a student of another Extreme AI system would gradually or quickly vanish as it learns more knowledge, so that at some point it may well become a peer of its teacher.

Standing on the shoulders of giants

Extreme AI systems are constantly monitoring their environments for new data, including the knowledge bases of other Extreme AI systems, enabling each new Extreme AI system to start from a higher base of knowledge than those who came before it, enabling them to stand on the shoulders of giants.

This is another key aspect of transforming the simple closed loop into a virtuous spiral.


Monitoring and coaching are also designed to facilitate evolution. Otherwise, the student would rarely be more intelligent than the teacher.

The evolutionary mechanism includes:

  • Random mutations
  • Fitness function

With AI and modern computing capabilities we have the opportunity of creating large numbers of students or spirals, each with subtly different initial knowledge bases and guidance, so that each loop and spiral can achieve its own potential.

Evolution can have the effect of allowing a student to quickly become more capable than its teacher, although in some cases the mutation of the student will not be as fit and conceivably even fail.

Incremental improvement vs. quantum leaps

One nuance of evolution is whether the mutation should be relatively minor, an attempted optimization, or a major change to enable a quantum leap. Both are useful.

Disruption vs. threat

Quantum leaps can be great, but they frequently tend to be disruptive. It is typically best to maintain a large degree of order with only a modest level of disruption at each stage, comparable to changing one or only a few variables at each stage so that the system can learn something of significance from the experiment, so that the application of the fitness function will permit explicit learning about which parameters were more fit and which were less fit.

That said, some degree of occasional major changes can be quite beneficial.

The goal is to enable enough disruptive change to give the environment a net creative sense of progress without threatening the entire ecosystem. An existential threat to the ecosystem is not a good outcome.

Enabling the Singularity?

It seems an open question whether the model of Extreme AI proposed in this paper would be necessary to enable Ray Kurzweil’s Singularity. I think it would be a step in that direction, but I am unable to argue that there might not be some other more optimal path.

Is it sufficient for the Singularity?

Would the model of Extreme AI proposed here be sufficient for the AI of Kurzweil’s Singularity, or would an even more powerful form of AI be needed? I simply could not say.

My current feeling is that it would be at least close to sufficient, but that remains an open, presently unverifiable assertion.

As noted, virtually all of the components of my model of Extreme AI are unsolved research problems.

Self awareness

This paper does not address the thorny issue of self-awareness of AI systems. Presumably Kurzweil’s Singularity would accomplish self-awareness. Maybe Extreme AI should include at least some minimal level of self-awareness, but I’ll leave that question itself as an unsolved research question.

The main reason that this paper sidesteps the self-awareness issue is that the primary focus here is on learning, teaching, and evolution — closing the loop, and initiating and promoting the spiral.

Beyond Turing machines

Yet another thorny, unsolved research problem is whether classic Turing machines, with traditional, algorithmic computation are sufficient for the level of learning, reasoning, and teaching required for Extreme AI.


This paper also sidesteps all issues related to reasoning itself, instead focusing on learning, teaching, and evolution.

Speculation and creation

This paper also sidesteps all issues related to creativity and creation, including speculation, creation, and experiments.

In addition to creating new knowledge, a practical AI system may also be able to create objects in the real world, whether with 3-D printing, robotics, or simply creating plans for real people or traditional non-AI fabrication equipment.

And, of course, an AI system can certainly create digital media artifacts, including text, documents, images, slideshows, presentations, audio, video, VR, and social media.

How long until Extreme AI is commonly available?

I am unable to even guess when it might be likely for Extreme AI systems to first appear, let alone be common. Five years? Ten years? I would hope so. Twenty years? Geez, Kurzweil’s Singularity is supposed to happen by 2045 (28 years), and just a few days ago he predicted that computers will have human-level intelligence by 2029 (12 years.)

Will Extreme AI have human-level intelligence?

I would say that Extreme AI will have near-human level intelligence. In some areas it may fall far short, but in other areas it could well exceed human-level intelligence. In any case, I am not predicting or asserting true, human-level intelligence.

Technically, Strong AI should be human-equivalent level of intelligence, but I would personally call that more of an aspirational goal rather than a minimum bar for success of Strong AI.

The goal of Extreme AI (and even Strong AI) is not to fully replace human intelligence, but simply to allow us to construct software systems that exhibit reasonably intelligent, near-human level behavior with a minimum of human effort.

What’s next?

Progress on evaluating the proposed model for Extreme AI depends on:

  • Deciding whether this model seems plausible.
  • Decomposing the knowledge, learning, and teaching components into more realistic research areas that realistic projects can be focused on.
  • Attracting the requisite level of technical research talent.
  • The small matter of funding.
  • Extreme patience, extreme focus, and extreme passion.

I don’t personally have either plans or intentions to personally engage in any of this work, although I will write on occasion about related topics. The point of this informal paper is simply to record and promote a model of AI capable of closing the loop and enabling the virtuous spiral of learning, teaching, and evolution.

For further reading, as mentioned in the introduction, I have a much more extensive paper on AI, Untangling the Definitions of Artificial Intelligence, Machine Intelligence, and Machine Learning.

For more of my writings on artificial intelligence, see List of My Artificial Intelligence (AI) Papers.

Freelance Consultant