AI, learning and general purpose intelligence

What problems can AI solve for us and will its 'intelligence' ever become 'general'?

Yesterday I wrote about chaos and systems collapse. Today I feel compelled to follow up with a post about AI and Artificial ‘General’ Intelligence after the United Kingdom greenlit an AI overhaul of the British state.

In yesterday’s post, I appropriated a puzzle from the reality TV show, Traitors, to explain how species navigate to higher levels of fitness thanks to the evolutionary process. I am going to stick with this metaphor today to talk about AI, intelligence and general intelligence.

To reiterate the metaphor - imagine a puzzle in which you must navigate rooms to reach a prize. Each room has 4 doors - 3 lead to elimination and 1 leads to the following room. The game is to find a pathway through the doors to reach the prize. As each team identifies the correct path, this information is shared with the next team so that they pick the correct door.

OK, so much so good. I used this metaphor to explain that evolution is all about pushing on doors and learning helpful pathways. When pathways are discovered, they are encoded as DNA, heuristics, cultural norms etc and passed on to other group members; This metaphor works for many adaptive processes. If you are setting up a business, you need to push on a lot of doors to find a pathway through. Most of the doors will be slammed in your face and if you run out of resources (money / energy) you will be eliminated from the game. Today I want to apply this metaphor to AI.

François Chollet, a brainy AI specialist who looks young enough to date my daughter, posts frequently about Artificial General Intelligence (AGI). Tech bros use impressive phrases like: ‘the singularity’ to express the same idea. The theory is that machines will one day possess the ‘general’ intelligence of humans. But the exact meaning of general intelligence is poorly specified and does not submit itself to falsification. So Chollet, a latter day Turing, has designed a test he calls ARC-AGI (the Abstract and Reasoning Corpus for Artificial General Intelligence) to ‘measure the efficiency of AI skill acquisition on unknown tasks’.

I would like to pick up some of Chollet’s ideas and squeeze them through my Traitors metaphor to offer a better understanding of what he is getting at. In my example, an organism (species / economy / company / team) pushes at doors to find a pathway to a higher level of fitness. Once a successful pathway has been found, this information is encoded as ‘knowledge’ and shared with other group members. Knowledge can be found in many places. Genetic knowledge is stored in our DNA. Cultural knowledge is stored in rules, conventions and norms. Business knowledge is stored in processes, methods and technologies. In all these examples, the purpose of the knowledge is to shorten the path to the prize. Instead of 4^4 choices every time you want to win a prize, you narrow it down to 1 pathway.

You can see how knowledge, rules, codes, algorithms and specialisation enhance productivity by shortening the means by which a goal is reached. But nature tells us that following the same path leads to diminishing returns. If we keep planting the same crop in the same soil, the quality of the soil will degrade. If we keep recombining a narrow gene pool, the variety in the gene pool will degrade leading to less strong offspring.

Humans, therefore, invented many rules about recombination of genes and crops to ensure that variation (the potential for novelty) was maintained. Salle Farm in Norfolk, England, still practises a traditional 7 year crop rotation to build up the quality of its soil. It also maintains permanent patches of grass and woodland. These enhance feedback loops between species in the overall ecosystem, flooding the fallow fields with variety and novelty.

The 7 year crop rotation cycle has been embedded in human history for thousands of years. Only since the discovery of huge energy stores during the Industrial Revolution, have we ignored the wisdom of allowing land to regenerate every 7th year.

I use this example to explain that any evolutionary process (biological or material) will require novelty and variation to stave off degradation / diminishing returns. Even if we have stored knowledge of a functional pathway, the potential of that pathway will diminish over time. Following the same, specialised pathway without any variation, traps us in degraded soil and diminishing returns.

This tells us a lot about ‘intelligence’. In our customary understanding, intelligence is a measure of individual performance in time limited tests. The IQ test is our favoured measure but school and universities offer subject proxies. Although IQ is presented as a test of general intelligence, anthropologists argue, persuasively in my opinion, that IQ tests measure embedded cultural knowledge, and not much more.

If you have followed my metaphor of doors and pathways, it should now be obvious that tests measure knowledge of pathways. And this is not nothing. If someone can perfectly remember a pathway through a long line of doors, then they perform a useful function for their tribe. But this is not the only skill needed by the tribe. It also needs people who tend to wander off the trodden path. It needs people who challenge the relevance of the chosen path. It needs people who can imagine a different pathway altogether. It needs people who can walk into a new room, with a new set of doors, and make decisions about which door to push. Finally, the tribe often needs to solve problems through collective, wisdom of crowds, reasoning.

In summary, a tribe / species / company or economy needs:

1. Learned knowledge, patterns and methods

2. Pointless meandering outside the learned knowledge, patterns and methods

3. Rigorous, awkward, annoying challenge of learned knowledge, patterns and methods

4. Deliberate experimentation to discover new knowledge, patterns and methods

5. General purpose nous to tackle novel, unpredictable problems

6. Collective reasoning to resolve novel, unpredictable problems

Only (1) is captured in an IQ test.

How much of an economy, company or tribe’s work relates to learned pathways vs. novelty and exploration depends on the lifecycle stage of each. If the tribe has moved to a new niche, it will spend most of its time tackling novel, unpredictable problems. If the tribe lives in settled, fertile pastures, with no resource constraints, it will focus much of its attention on inculcating (1) among the members of the tribe.

And that brings me to AI.

First, let me try and unpack the meaning of ‘Artificial Intelligence’. This is rather a grandiose term to describe the use of machines to perform cognitive tasks. We have, of course, been using machines to do this for a long time. Compasses, astrolabes, quadrants, sextants and clocks have all allowed humans to jump through rooms without the cognitive work of de novo, trial and error learning. Almost everything we do builds upon some prior knowledge that, where possible, has been embedded in rules, methods, tools or technology. Artificial Intelligence is simply another step along that path. The reason for the insane hype about this latest tool is, I think, the sense that the machine is ‘thinking for itself’. This is untrue. Machines apply a mixture of probabilistic reasoning (applying Bayes Theorem) and reinforcement learning (learning from feedback) to offer solutions / answers to problems. Astrolabes, calculators and personal computers do not do this. They are designed to respond to a single, non evolving problem set. They can be taught ‘if this, then this’ rules (or have it embedded in their physical construction) but they will never be greater than the sum of their parts. They are inert and single use tools.

The promise of Artificial Intelligence is that the tools are not inert. They can, it is theorised, co-evolve with their problem set. Like humans, they can, it is proclaimed, respond to novelty. For this to be true, we must assume that access to all human knowledge (taught to machines) is both necessary and sufficient to respond to all future problems, however novel.

Is this true? Does the potential to respond to novelty come from the back catalogue of knowledge (DNA) or does it require different skills? This is hard to get your head around because arguably all systems build upon prior knowledge. But novelty emerges from a greater than sum of parts input. So by definition, prior knowledge can never anticipate every novel, emergent outcome. Knowledge may be necessary but is unlikely to be sufficient in such cases. For these problems, we need a different toolkit.

It is hard to write about complexity without, at some point, being attracted to Nicholas Nassim Taleb’s work. Taleb, an awkward, difficult challenger of norms (number (3) on my intelligence list) has a name for our obsession with complex knowledge: ‘Neomania’. He is dismissive of ‘newness’ suggesting that the real test of utility is longevity. I like his point and think it applies to the question at hand. While learned pathways help us build increasingly complex constructions, each marginally more useful than a prior version, not all knowledge needs to be complex. Multi-purpose knowledge is contained in simple, rather than complex structures, that can be used on any door in any situation.

I like to think about this as the difference between the SAS and the conventional British army. The conventional army is equipped with layers of complex learning embedded in rules, processes, technology etc; But if you face a novel enemy, you drop in the SAS. Their tools are unsophisticated but multi-purpose, and with high levels of redundancy. They can open new doors much more quickly than the fancy, complex technology of the traditional army.

I think this is Taleb’s point about neomania. Complex tools are useful for known, specialised problems. But they sit alongside simple, multi-purpose tools that have been repeatedly applied to novel problems throughout human history. When children want to understand the physical properties of an object, they tend to bash it with a stone. When we try to extract a splinter or cut open a package, we reach for something sharp. When we explain directions to our house, we tell people to look out a few distinguishing marks in the landscape. Inuits use carved wooden maps to orient themselves in their land (Text and photo: www.decolonialatlas.wordpress.com)

In the face of novelty, complex technology is not that useful. Especially if it is non portable, energy intensive, inaccessible and single purpose. When the asteroid hits, you’ll most likely want a stone for smashing things, a knife for cutting things, a book to keep you sane and your dog (sorry kids). Indeed, the most useful thing at your disposal will be access to a mixed gene pool of other people. Other People are the original General Purpose Intelligence that can keep you warm, help you build things, help you catch things, build copies of itself and combine to solve any number of tough problems. And we are looking to replace this with AI??

Back to Chollet. Here is how he describes the problem of ascribing general intelligence to current forms of AI:

Measuring task specific skill is not a good proxy for intelligence. Skill is heavily influenced by prior knowledge and experience. Unlimited priors or unlimited training data allows developers to ‘buy’ levels of skill for a system. This masks a system’s own generalisation power.

To apply this to my metaphor of rooms, Chollet is saying that developers are leaping through rooms instead of building up knowledge in an evolutionary, trial and error fashion. This is giving the impression of general intelligence in the machines they construct. In fact, these programmers are creating highly specialised machines that cannot be applied to novel problems. As Chollet explains:

Intelligence lies in broad or general purpose abilities; it is marked by skill acquisition and generalisation, rather than the skill itself

It appears that something in nature’s toolkit has embedded life with a set of general purpose tools that do not rely on prior knowledge. They allow us to respond to larger than sum of parts / emergent / novel problems and they are simple and accessible, not complex and energy intensive. The toolkit lets us acquire new skills and generalise them.

This is hard to explain but I find biologist, Michael Levin’s work fascinating on this topic:

One defining features of complex life, making it distinct from our current engineered artifacts, is its multiscale nature; there is order in biology across levels of organisation, from molecules to cells, tissues, organs, whole organisations and societies / swarms. Crucially, however, this goes well beyond structured nesting: it is in fact a multiscale competency architecture

Levin is saying is that any inert, designed, sum of parts object will never have the general intelligence (competency) of an evolved organism because it cannot re-organise itself to respond to novelty.

He adds:

As evolution facilitated the increase of complexity, living things became composed of layers that cooperate and compete to solve problems in metabolic, physiological, anatomical and behavioural state spaces …

WOW …

I believe that Levin is describing the missing pieces in the intelligence conundrum. When organisms face a new door in an environmental niche not covered by earlier training data, the organism uses co-operation / competition to find a way through. The intelligence does not sit within a single part of the evolved species. It is acquired through multi-scale behaviour.

If this correct, then current models of Artificial Intelligence will never achieve ‘general’ intelligence unless the machines begin to interact with one another. For AI to solve novel problems in evolving spaces, it must learn to re-organise its parts - and recombine these with external parts - to come up with solutions. Even if all this were possible, it is still not clear that interacting machines, trained with prior knowledge, will ever know more than the sum of their combined parts.

Which brings me to the UK’s decision to greenlight AI. As a way to accelerate learning and let us leap through rooms to higher levels of complexity, great. But will the gains in productivity warrant the immense use of energy needed? And if humans are no longer forced to go through doors to acquire knowledge, how will the machines be updated? Will they feed on their prior knowledge without the introduction of new, emergent ideas? If so, the soil (variation, experimentation, recombination) will degrade and we will find ourselves stranded at the far door with no prize.

Instead of shoving everything through a specialised algorithm, we must keep increasing the potential for variation and exploration beyond learned pathways. In short, any problem where AI is the proposed answer, must be counter-balanced with an investment in other forms of intelligence (2-6). Otherwise we risk degraded, over-specialised systems that have lost their general purpose toolkit:

1. Learned knowledge, patterns and methods

2. Pointless meandering outside the learned knowledge, patterns and methods

3. Rigorous, awkward, annoying challenge of learned knowledge, patterns and methods

4. Deliberate experimentation to discover new knowledge, patterns and methods

5. General purpose nous to tackle novel, unpredictable problems

6. Collective reasoning to resolve novel, unpredictable problems