#14 - Ex Machina's Scientific Advisor - Murray Shanahan

**Craig Cannon** (0:00)
Hey, this is Craig Cannon, and you're listening to Y Combinator's podcast. Today's episode is with Murray Shanahan, who is one of the scientific advisors on Ex Machina. Murray's a research scientist at DeepMind and professor of cognitive robotics at Imperial College London. He's also the author of several books, one of which is called Embodiment and the Inner Life, and it served as inspiration for Alex Garland while he was writing the screenplay for Ex Machina.
All right, here we go. So I think the first question I wanted to ask you is that given the popularity of AI, or at least the interest in AI right now, what was it like when you were doing your PhD thesis in the 80s around AI?

**Murray Shanahan** (0:34)
Yeah, well, very different. I mean, it's quite a surprise for me to find myself in this current position where everyone is interested in what I'm doing. The media are interested, corporations are interested.
So certainly when I was a PhD student and when I was a young postdoc, it was a fairly niche area, so you could just beaver away in your little corner, doing things that you thought were intellectually interesting and being reasonably secure that you weren't gonna be bothered by anybody. But it's not like that anymore.

**Craig Cannon** (1:08)
No.
And so what exactly was the subject matter at the time? What were you working on?

**Murray Shanahan** (1:13)
At the time when I did my thesis.

**Craig Cannon** (1:15)
Yeah.

**Murray Shanahan** (1:16)
Well, I worked on how you could use...
Oh, this is a tricky question. I know you're asking me to go back.
Let me think. What is it? Like 30 something years? Yeah, 30 something years. Yeah, 30 years I finished. 30 years ago, I finished my thesis.
Okay, so what did it look at? So I was interested in logic programming and prolog type languages. And I was interested in how you could speed up answering queries in prolog like languages by keeping a kind of record of the thread of relationships between facts and theorems that you'd already established. So instead of having to redo all the computations from scratch, it kind of kept a little collection of the relationships between properties that you'd already worked out so that you didn't have to redo the same computations over again. So that was the main contribution of the thesis. I'm amazed I can remember anything about it.

**Craig Cannon** (2:17)
That's very impressive. I did my thesis like five years ago and I barely remember. And so did you pursue that further at Imperial?

**Murray Shanahan** (2:25)
No, I didn't. I kind of...
Well, one other thing that I discussed in my thesis was I had a whole chapter on the frame problem. So the frame problem is... There are different ways of characterizing it, but the frame problem in its largest guise is all about how a thinking mechanism, a thinking creature or a thinking machine, if you like, can work out what's relevant and what's not relevant to its ongoing cognitive processes and how it isn't overwhelmed by having to rule out just trivial things that aren't irrelevant.
And so that comes up in a particular guise when you're using logic and when you're using logic to think about actions and their effects. And there you wanna make sure that you don't have to spend a lot of time thinking about the non-effects of actions. So for example, if I move around a bit of the equipment, like your microphone here, then the color of the walls doesn't change. And you don't wanna have to explicitly kind of think about all those kinds of trivial things. So that's one aspect of the frame problem. But then more generally, it's all about sort of circumscribing what is relevant to your current situation and what you need to think about and what isn't.

**Craig Cannon** (3:44)
And so how did that translate to what folks are working on today?

**Murray Shanahan** (3:47)
Well, so it's actually, so this thing, the frame problem, has recurred throughout my career. So although there's been a lot of variation in what I've done. So I worked for a long time in classical artificial intelligence, which is there, it's all about, it was, and still is all about using logic-like or sentence-like representations of the world. And you have mechanisms for reasoning about those sentences and rule-based approach. And so that approach of classical AI has fallen out of favor a little bit. And I sort of got a bit disillusioned with it back in, well, a long time ago.
So by kind of the turn of the millennium, I'd more or less abandoned classical AI because I didn't think it was moving towards what we now call AGI, artificial general intelligence, the big vision of human-level AI.