On why A.I. scientists haven't already far exceeded the capacity of brains:
"Well, maybe it's the way in which the brain is not perfect that accounts for some of our abilities. So that, for instance, our memories are not perfect - we don't remember everything, and that explains some of the things that we are able to do. And when you produce a perfect system that does whatever you tell it to, and stores every piece of information you've ever got, you can't produce the same kind of behaviour".

On the difference between a conventional computer and a neural network:
"With a neural network you've got no central set of instructions, and they're particularly useful where you have something where you don't have a rule to describe the relationship between the inputs that it gets so - say, the sensory information that it gets and the behaviour that you want. What you've got is a set of input units, and a set of output units and you've then got a weight between them which is analogous to the idea of the synapse between neurones in the brain. You then train the network, as in you've presented it with a set of inputs, you've told it what output you would like it to produce, so it adjusts the weights so that it does produce the right outputs. You can then give it a set of inputs that are similar, but not the same as the ones it saw. It will then produce an output that is like the output. So it's flexible in the way it operates."

On whether machines can have the same kind of capabilities as humans:
"I think that you could model what you think it is to be intelligent of conscious, and you'll know more about it as a result, and you can check out your explanation and make sure it works, because you've built a model. But you're still only modelling it and I don't think you've done anything like actually creating something that's really having experience of the world."

On the possibility that A.I. will have a sort of consciousness which is even better than ours:
"I think because of something about how conscious evolved, you're not going to be able to create anything like it. Because you can use evolutionary mechanisms to create something that you could say was similar, but you're using them towards a goal, you're indirectly programming the thing to do it. So you're rather back to the original symbolic A.I.

If you look at what we have achieved, at what systems we have that seem to be intelligent, we don't have systems then really. You could model an aspect of intelligence, but we don't have anything that is a whole intelligent system. And my hunch is that it is in principle impossible to go further. If you create an artificial system its not integrated with the environment and actually motivated by itself. You've still got an external person who is making it do that."

On the importance of evolution in A.I.:
"You can use evolutionary techniques. It's interesting to see what kind of behaviour you can evolve with a simple reactive mechanism, but then the hardware isn't evolving - it's abstraction of the real thing."

On the future:
"I think that we'll certainly end up with robots doing simple tasks for us where you don't need to worry about whether they're intelligent or not, for example, vacuuming the floor. Or similarly using neural networks, you could perhaps have a neural network that made your car go at the appropriate speed or not bump into the car in front. So I think we will certainly see artificial intelligence used in our daily environment in that kind of way."