I was delighted to be part of the panel for ‘From Robots to AI: Can You Build a Human?‘ at the Barbican Centre in London a couple of weeks ago. Chaired by Timandra Harkness, along with three other panellists: Dr Fiona McEwen, Dr Matthew Gwynfryn Thomas and Dr Stuart Derbyshire. We each gave a short opening position, and then discussed questions raised by the audience for the remainder of the 90 minute session. Here is a rough transcript of my opening remarks on the topic.
Can you build a human?
Yes, I can build a human, in fact with the help of my wife I’ve built three! My daughter and her husband recently built one too, I’m very excited about her construction!
Ah, that’s not what you meant. [ I guess also you don’t mean can I grow one in a test tube?]
You meant can I build one from parts by assembly, like a machine.
So, the question is immediately shown to assume that we are thinking about humans as being electro-mechanical machines, which they are not.
Oh I see, you don’t mean they have to be squishy like us, they just have to be able to move and think like us.
Well, first the easier part – as yet we can’t make anything mechanical that can do what a human does, for many reasons. Interestingly, the most important one is energy efficiency. We humans can operate all day an very little energy – just the energy in one or two small meals. Robots are electro-mechanical machines – they use electric motors, and the batteries for those motors can’t store nearly enough energy for a whole day of human level activity. Also, batteries are very heavy, making the robot too heavy as a result. Also, we can’t make actuators that work as well as human muscles, and really that applies to everything in our body. Evolution has achieved great things over billions of years with a lab the size of a planet!
Another important thing is that humans are compliant – we are strong yet bendy. This makes operating in the physical world much easier. Its very hard for a rigid metal robot to work effectively and efficiently in a human friendly environment like a hospital or in your home, although they work great in factories and warehouses.
So, in trying to make a human, we really have a long way to go just to make a working human-like body – we need squishy robotics for that, and we’re not very good at that yet.
OK, OK, I know, let’s get on to the hard part. I guess this is really what you meant by the question in the first place. What you want to know is can we make a machine that thinks like us? One that is smart, and can solve problems. Maybe one that is *smarter* than us. Much smarter. Maybe dangerously smart? Should we be afraid?
What is smartness, or intelligence? First we need to agree on that, before we can decide if we can build it. How do you know if something, or someone, is smart? Easy answer to this one: Because they make ‘clever decisions’, they do clever things. They demonstrate intelligent behaviour. That’s the only way. In fact when you think about it carefully, you realise that intelligence, smartness, isn’t a *thing* at all. It’s a behaviour. More than that it’s a behaviour that achieves something useful in a particular set of circumstances – in a certain environment and context. When I watch animals in the wild, I’m always fascinated by how smart they are. Almost everything they do, they do for a purpose, and they do it very well. They waste very little time or effort. I think animals, insects and even plants are very smart indeed. Very intelligent. I learn a lot from them, and I apply that to my robotics.
We humans are very intelligent too. We navigate around complex, unknown physical environments with ease, and we can deal easily with uncertainty (rather like navigating the Barbican). We navigate around very complex social environments too, and we use language to do that. Today’s robots can’t do any of that very well, or for very long, before they fail.
OK OK, I hear you. What about learning, *machine learning*, *deep learning* and all that. Surely these algorithms are smarter than us. Surely they can already do things we can never do.
Well, yes they can, but I’m neither surprised nor impressed. These chess playing, or Go playing machines are certainly good at doing the thing *we* designed them to do, but they need a lot of training before they get good, they need a certain type of problem, and they can’t generalise what they *learn* in one domain to any other. In fact although we say ‘machine learning’, it bears little relation to the thing we, or animals, do when we learn. For machines, it all comes down to learning regularities, by using lots of data and doing some statistics. Deep statistics maybe, but nevertheless statistics is a much better word than ‘learning’.
Anyway, we’ve been building machines for hundreds of years that do things that we can’t do. Steam Engines, Aeroplanes, Submarines and Computer Spreadsheets all exhibit behaviours that are far beyond our abilities. Spreadsheets are much smarter than us at doing maths with lots of numbers. All the ‘deep learning’ thing does is fancy statistics to be able to do some predictions based on the regularities in the data. That’s all it is. In fact it shows how smart *we are* to be able to build these tools.
But none of this stuff is anything like what it takes to build a human thinking machine. Nowhere near.
In short, to answer your question: No, we can’t build a human.
Anyway, why would you want to? Now, that’s a much better question to ask. Perhaps we can explore that one a little over the next hour or so.