Elon Musk is scared of the wrong AI

Nobody is likely to confuse me with Elon Musk, but we share something apart from the burden of an unusual first name and a penchant for pink-trimmed trackie tops: we might be the only two techies on the planet who aren’t thrilled by the thought of Artificial General – i.e. human-level or better – Intelligence.

But I think Mr Musk is missing a trick: the real threat to humanity isn’t clever tech’s potential to become a tool of evil dictators. It’s something far more likely, far more insidious.

Want to know more? Watch Mr Musk making his case for existential terror while rocking the aforementioned athleisurewear. Check out my budget fashion alternative, and read on to find out why I’m even more panicked than someone who runs three high-risk companies, can’t get a good night’s sleep, and has incurred the social media-fuelled wrath of Azealea Banks.

What Mr Musk is scared of

Mr Musk – for those of you too impatient to sit through the video or too distracted by his outfit or general fabulousness – fears that AGI - Artificial General Intelligence - will be so much smarter than humans that its superior reasoning powers will be used by crazy bad guys to effect nefarious triumph and nix human civilisation.

That’s a pretty scary thought, but how likely is it to come true? How many shopping days left before annihilation?

AGI today

Look around the world pretty baby and you will see robots beating human experts at board games, filing typo-free legal claims, answering idiot questions when we’re out of reach of a keyboard and saviing us 40% on air conditioning.

None of this is what we’re talking about when we talk about AGI.

All the clever bots above are examples of Artificial Narrow Intelligence – robot systems trained to carry out specific tasks. They are powered by various flavours of machine learning, from simple strings of algorithms (decision-trees of instructions) to artificial neural networks, structures where different algorithms can hang together, provoke the best from each other and somehow (we can’t track exactly the processes) come up with solutions to messy, imprecise problems.

All those Artificial Narrow Intelligence robots look scarily impressive until we ask them to apply their intelligence in new ways. Challenge AlphaGo to a game of Old Maid, and the system will freeze. Refer Siri or Alexa back to something you said earlier and they will blank recollection like a teenager nagged about tidying. DoNotPay can help you file legal papers but has no clue about insurance claims. Artificial neural networks trained to guide autonomous cars have failed to tell a crashed schoolbus from a snowplough. If life or even work was easily broken down into discrete, highly-structured and precisely repeatable tasks, Narrow AI would be all we needed.

Sadly, life’s a mess.

General Intelligence – evolution’s Marie Kondo

Intelligent creatures – higher mammals and cephalopods – create order from life’s mess using General Intelligence – the ability to learn from trial and error, to extrapolate from examples, to remember, connect, adapt and come up with new ways of doing stuff and ways of doing new stuff. Mammals do this using their brains, cephalopods via a distributed nervous system.

Fascinating as squid and octopus can be, let’s leave them to one side-dish for the moment and focus on ever-fascinating me (okay, and the rest of you boneheads who make up the human race).

Human general intelligence is a funny bunch of stuff. As career teachers stress, real-world smarts are not just about good test results (though that’s basically what psychometricians’ g-factor measures). Psychologist Howard Gardner identified nine different types of intelligence, including visual-spatial (not getting lost), intrapersonal (not losing it), and interpersonal (being sought after by others). Research into the drivers of job success points to a range of different intelligences, or competencies, some of them counter-intuitive: Google’s mega-research into high-flyer tech competencies famously found that test-measurable coding skills came bottom of the factors which correlated with successful performance for tech leaders. Feelings matter, too: neuroscientific research has pointed to the role of emotions in intelligence-rich activities as unexpected as decision-making. And there are some process factors that make a difference: a musician stopping her practice every time she hits a wrong note sounds irritatingly pointless but actually turbo-charges learning.  

None of the robots around today are anywhere near exhibiting that kind of range or interaction of intelligences. Each Artificial Narrow Intelligent solution starts pretty much from scratch and is a pain – a long, expensive and difficult pain – to set up. Algorithms need to say exactly what they mean (when was the last time any of us managed that?), or misunderstandings and biases get built in to the process. Pattern-recognition machine learning (most Artificial Narrow Intelligence) needs each training data-point to be accurately and comprehensively labelled – which of course it isn’t. Even so-called unsupervised artificial neural networks need reviewing, calibrating and adjusting.

Bunching together narrow solutions to box and cox has failed to achieve much. There are a couple of interesting research papers out there on blended artificial neural networks and symbolic software combos, and a radical first-principles approach based around prediction accuracy maximisation promulgated by a guy who likes talking to random people in public parks, but no practical application so far has shown the general intelligence of a puppy, let alone a fully-functioning human.

So let’s go back – as I’m sure Mr Musk would – to first principles. If we want human-level intelligence, let’s look at human brains, then see if we can replicate that biological machinery using silicon and electrical circuits.

Copying brains – don’t try this at home

Replicating the brain sounds super-cool. The only problem is, we don’t really know how brains (or distributed nervous systems, for all you squid readers out there) produce intelligence.

True, we have the foundations of a theory. Sort of. Neurons seem to be at the heart of the process – okay, the neurons in our hearts are not actually intelligent neurons; the ones we’re interested in live in the brain in vertebrates and all over the nervous system in invertebrates – but so long as we’ve fixed that fMRI software bug, we can watch them connecting to each other electrochemically when intelligent stuff happens.

Problem is, neurons are frustratingly unpredictable. Two similar intelligent creatures doing the same activity in what looks like exactly the same way do not necessarily fire up the same neurons along the same pathways of connections. In general, more neurons and neural connections suggest more intelligence, but higher-intelligence humans have fewer neural connections than their thicker brothers and sisters; perhaps cleverclogs get more bang for every neural buck?

Cut to the chase: we have a way to go in understanding natural intelligence (one researcher I know came up with the compelling but I suspect spurious figure of 15% to represent our current knowledge), so if we are to create artificial general intelligence anytime soon, either we get a lot better at understanding ourselves, and/or we look to different models.

Phew! For a minute back there I got scared

Not so fast, sunshine.

Much as experience has taught me that everything takes twice as long, costs three times as much and is a tenth as successful as predicted, I still think it is not a question of if we create Artificial General Intelligence, but when.

We are getting better at seeing what is going on inside brains. fMRI combined with quantum analysis looks promising, as does quantum imaging. Once we can see in detail how neurons connect when people carry out specific tasks (such as routine jobs) in specific contexts (outside laboratory conditions), there is no reason we cannot reproduce that mapping using silicon and electrical circuits. 

We do not need to know how or why brains do strange things like assign a dedicated Jennifer Aniston neuron, in order for us to reproduce that neuron, and the neural pathways that connect to it, and the intelligence that results from the firing up of those pathways.

My guess is that AGI will be very context-specific, because that is what natural intelligent systems are – an elephant’s neural pathways fire up in millions of ways when it moves its trunk; human pathways focus more on social connection (hence Ms Aniston’s importance to a worrying number of middle-aged men and women). We might first develop a Generally Intelligent Customer Service Agent (don’t laugh); then a Generally Intelligent Doctor; then a Generally Intelligent Lawyer (you can laugh now).

Looking at non-biological models, we can work backwards from the job – create a system that brings together different types of intelligence and other capabilities and focuses them on a role, a context, an environment, a field of play. Think of it as Base-level Artificial General Intelligence: BAGI for short.

Is BAGI going to kill me?

It’s not impossible to envision a BAGI that’s focused on global destruction. But there are a number of reasons why that vision is unlikely ever to become reality.

First, we don’t have a Destroyer-of-Worlds job description. It’s one thing to break down the activities, tasks, responsibilities and interactions of a truck driver, or a lawyer, and determine the different types and combinations of intelligence needed to carry out the role. It’s quite another to hypothesise a set of intelligences for a hypothetical role. One of the first things I learned about interviewing was that asking “What would you do if…?” questions produced answers that told me a lot about the storytelling creativity of the interviewee but nothing about their other, more relevant competencies. In contrast, asking “What did you do when…?” produced accurate, predictive data on what they did well, where they faltered and what their jobs actually involved.

Even if we did map the Destroyer-of-Worlds role, I still think we are safe. Humans may be utterly useless at protecting the planet from the side-effects of fossil fuel bingeing, but we have staved off nuclear destruction for close on three quarters of a century by ploughing enough smarts, money and effort into defence spending to keep innovation rocking, military secrets more or less secret, international cooperation and coercion active and results-focused, and by ensuring defence systems keep pace with weapons technology so the world remains safe for democracy and a select range of alternative growth-focused governance structures.

Also, let’s remember that, unlike the Maxim gun, AGI is being developed all-over. It’s unlikely – very unlikely – that any one player will dominate any time soon. For now, we can breathe easy.

The real threat of AGI

Only we shouldn’t. There is a clear and present threat from AGI – and it speaks to our greatest weakness as a species.

Our greatest weakness is, of course, the flipside of our greatest strength: our ability to adapt. Humans are creative, we find good-enough solutions, we make the best of a bad deal and muddle through our frustrating, imprecise, unfair and imperfect world. With our big problem-processing brains, our uniquely broad ability to work with others, our physical shortcomings and our successful improvisation in the face of a whole range of fuel, environments and category challenges, we are creation’s greatest adaptive technology.

All this is brilliant stuff, but it comes with a downside. Humans are too easily, too rapidly pleased. We are impatient and superficial. We are lousy at deferring gratification and assessing long-term risk. We default to the simple rather than to the important issues. We jump to conclusions, we take a partial and biased view of the evidence. We shoot first and ask questions later.

Base-level AGI plays to exactly these weaknesses. Just as we have let open price competition drive down the experience of air travel – we unerringly pick the cheapest flights, no matter how much we moan in surveys that we want more legroom, better refreshment options, shorter queues at checkin – so we will choose the 25-40% cheaper RoboTruck over the human trucker, the infuriating ChatBot over a Customer Service Representative.

So what? Isn’t good-enough what most jobs require?

I’m not so sure. We are increasingly finding that human richness matters. Solitary confinement sends people mad. The root of all disease may well be inflammation – irritation at a cellular level. There is a reason humans have built cathedrals, composed the Songlines, dressed our graves and painted our faces and paid over the odds for football tickets and done a host of other things that have nothing to do with basic utility and everything to do with pleasure and value for ourselves.

I’m sure Mr Musk would agree. For all his engineering brilliance, what makes him special is his intuition for human delight: The red sportscar in space. The twin rockets landing base-down in perfect synchronisation. The electric car that plays games and dances.

None of us wants to live in a grey, utilitarian world. And unless we get serious about Better-than-Base-Level AI, that’s what we are facing.

So, my inclination is to meet the threat of AGI not by retreating, but by doubling down. Let’s create AGI that can delight us, thrill us, comfort us and inspire us just as well as humans can – and let’s not accept half-baked substitutes.

Or maybe, if humans prove too slippery or inadequate a model for delight-ful intelligence, we should take another look at those squid…