Peter, Nick and Fraser discuss the film Sunspring, written by an artificial intelligence. Is it a novelty or a sign of things to come?

For more information on Aleph Insights visit our website https://alephinsights.com or to get in touch about our podcast email [email protected]

Hello and welcome to the Cognitive Engineering Podcast produced by Tell Me Studios for Aleph Insights. In this series of podcasts we take a look at interesting topics and discuss what we think they tell us about analysis and decision making. I'm Fraser McGruer and I'm here with Nick Hare and Peter Coghill of Aleph Insights and this week we're discussing the short film Sunspring which was entirely scripted using artificial intelligence. So Peter, in your new guise as film critic, can you give me your review of this film Sunspring please?

So I thought it was an interesting experiment in motion picture art form. The script left much to be desired I think. It felt somewhat sort of chaotic and juvenile in places. It left much to the viewer to interpret what was going on but it was a success because of its uniqueness I think mainly as a piece of art. I think that's quite charitable.

I thought it was unwatchable and in fact and so therefore literally you know I did not watch it. I watched bits of it and I couldn't. It's only nine minutes long Fraser, that is pathetic.

No but it is, it's just awful. Well I think let's just clarify that. I mean and also we should just to just actually to compound your assessment note that humans helped the script helped the AI tidy the script up as well so in other words there was a bit of human post-processing but it does it includes lines like it's a damn scared thing to say nothing is going to be a thing but I was the one that got on this rock with the child and then I left the other two and it's those lines like that are basically entirely what it consists of. Sorry Peter,

carry on. Well it reminded me almost like a really bad Terence Malick film. Is it Terence Malick? Is that right? Is that the guy who did The Thin Red Line? Who did The Tree of Life? Is it Terence Malick? Yeah hang on I think The Thin Red Line is quite a good film. Yeah yeah well that's what they are good films. Right. But it's like that it was almost. I'd say you could you could use the adjective pinterest about some of it. Well anyway so we all have our sort of views on sort of about it so beyond sort of me saying it's unwatchable and okay look as an exercise as you say it was successful it was interesting but let's sort of go a little bit further. What can we draw on this exercise in terms of looking at artificial intelligence and you know

what it means for us? Well the the state of artificial intelligence today I think we we can draw lessons of that because so much is left to the human for interpretation and there was a great deal of filtering on behalf of post script writing editing and then at the sort of direction and producing stage there would be lots of well maybe they're talking about spaceship here maybe they're in maybe they're in an office there's a lot of there's a lot of sort of addition by humans throughout that we we haven't got AIs powerful enough strong enough clever enough to to to to replace humans just yet. Well I do want to I think we have to say how it works

I mean we haven't covered how the artificial I just think we ought to just cover that right just to explain exactly how this AI. So you're going to do that you're going to do that because at the

moment it just it actually reminds me of like a drama school exercise where you get a load of words cut them out throw them up into the air and then make a play out of it. Well that isn't that isn't far off. Right so tell it tell me how the process works in this particular case.

Benjamin the the AI is a neural network right and a neural network is a type of AI architecture which is designed to learn stuff and in particular to learn patterns so learn patterns between certain inputs and real world outputs so you set it to work to understand I don't know you could use it to try and understand you know how a car worked for example and it would have data the inputs would be all the buttons that were being pressed and the pedals that were being pressed and so on and the outputs would be descriptive data about what the car was doing and the neural net might be able to use that to learn how cars work and neural nets underpin some of the most impressive bits of AI so AlphaGo which we talked about a few weeks ago is built on a neural network which is several layers deep. It's hard to explain intuitively how they work but essentially they look for patterns is one way to think about it and the AI here Benjamin was fed a load of scripts from sci-fi that were available and it learned basically with after one particular sentence you're going to get another sentence or even word by word so if you have the word if you have the words um you know I don't uh that the next word is likely to be no and you know and so it basically it is pretty much as you describe simply finding patterns in the way sentences are arranged in uh in sci-fi scripts and then regurgitating a particular you know plausible so he's going to say okay after this word what's the most likely next word here it comes and now I've got those two words what's the most likely third word and just doing that process to generate the entire script okay now um so I think we it's being rather over egged right that because that is really quite tenuous to call that an artificial intelligence um it really is more of a sort of statistical um you know word sentence generator based on based on a training set of uh of real life scripts it isn't creating something right it and it's not what it doesn't it doesn't in particular even know that what it's what it's doing is producing words I use no in inverted commas there obviously doesn't know anything anyway but it doesn't have a model of grammar so it's not it doesn't hasn't been told look this is what sentences look like now go and find um go and build some sentences which is why some of these sentences come out completely ungrammatical so it doesn't have a even have a model of grammar and it and it what it definitely doesn't have is a model of the world so um you know not quite apart from the fact it doesn't it hasn't really been told to create sentences it it also isn't in any way thinking about plot there's no attempt to make characters or plots or anything it just is a word generator um so presumably that was by choice those options were possible well it's well yes choice and also technological constrained because uh we don't really know I mean at the moment we're nowhere near being able to um come up with with uh something which could plausibly generate you know the kinds of characters and plots that we would expect to see in even a short film so then this begs the question

then um if it I mean let I know you said for a moment there well this isn't really artificial intelligence but let's just say that it is for a moment or what what does this tell us about the limitations of artificial artificial intelligence or what does it tell us about what artificial

intelligence is good at and what's it it's not so good at Peter so the so Benjamin um you could think of Benjamin a bit like an extension of the typewriter that the scriptwriter would use to to create his create his art so it in a similar way that the typewriter is set up when you put a spool of ink in it and put some paper in it and then operate keys the IT the Benjamin was a program that was written and it was set up and it was given a certain input of data so it's a sort of similar mechanical process and I think this is typical of of all AI that has some practical function today it requires a great deal of manual intervention and really it's just a sort of next iteration of tools um and this is this has interesting philosophical but dimension when we talk about art and computer generated art it's like well you know if if arguably you could say something's art if somebody says it is so I think like this I get something from that and yeah it doesn't have any practical purpose so therefore it's sort of it's artistic um so it's the debate is whether or not we uh whether or not computers can create art they can they do and this is an example of it um is whether or not we will allow them to create artists whether or not that we demand art has got to have some sort of personal sacrifice some sort of emotional connection with the artist no isn't it more than that though it's about whether the

thing creating that what is called art knows that it's creating art even if an even if a third party can look at something and go well that's art that is that a rigorous definition enough isn't isn't what what isn't what doesn't sit well with people partly that what is creating this doesn't know

what it's creating I don't think so I mean you you can you children create art and they don't necessarily have a very form a firm definition of what art is um you you have cave paintings which are considered artistic but may have had some practical purpose and not intended to be

art so I don't think well yeah but I think more pertinently um you know who created it now did the person who built and built the neural net create it I think you could you know that's the if you've got uh so if you take something like you know Jackson Pollock he didn't intend there was no intention that every little speck of paint was exactly where it was there was a stochastic element there was still an intention to create something yeah I know I know but no but the point is that here they the people who'd built this uh built Benjamin intended to create pretty much probably something similar to what actually happened and the way that they were doing it and you could have done it with pen and paper you know you could have done it as you suggest with by by cutting up lots of bits of paper and then ran and then sort of rolling dice uh but they

chose to do it in a slightly more technologically advanced way sure but this goes back to what you're saying if you could think of this to use your analogy is that Benjamin is the typewriter okay just in so that if you if we want to call this particular thing art for a moment then the artist is I don't know the programmer or even the guy who thought this up to do this but anyway I think we're drifting off the point which is what are the limitations of AI

I think the limitation is really that they are only at the moment capable of quite simple well-defined tasks and it's going to take a lot more a lot more clever algorithms and a lot more novel ways of developing of developing heuristics and self-learning and things to become become close to develop anything like empathy that humans have for for each other or these are the emotional mechanisms that we use to to create things well I think that's that's going

a bit further than we need to look I think when we're talking about the constraints at the moment which is just creating language creating plausible sounding language and you know plausible language that's all that's in any way vaguely artistic we've got to remember that in general that in artificial intelligence things that we find easy are hard for computers to do for all sorts of reasons but particularly that we ourselves often if it's something that we find easy it's because we've got a lot of software in our brain which sort of is specialized towards that language is definitely one of those things so we have a lot of language software so learning language is very very easy for us and and you know as a side effect we find it very hard actually to explain what we do when we come up with sentences but but it's also you know the the fact that with particularly with language there is what the language means so what it actually you know what what the denotation is in other words so you know if I say uh tiger you know that you could define what that means in fairly precise terms but tiger also has with it a load of uh connotations things that aren't actually part of the definition of tiger but which um when you're writing the good prose will be drawn on so if you take an example uh so from from Shakespeare which is uh that light thickens and the crow makes wing to the rocky wood which is a sentence the literal light thickens is sort of nonsense really I mean that doesn't it's not something light can do but we all understand that light thickens means that there is this deep gloom setting in that there is a kind of the the that the the the the hour of fate is almost upon us and that's exactly the sense in which Macbeth uses that sentence um an AI could come out with that but it would be it would only be accidentally as meaningful as that because we all have a huge number of these connotations we all understand things and a lot of that is shared between humans and it's totally inaccessible to us you know we've got no means really of getting that out onto a piece of paper so a machine can use it and if we haven't got the data then then we can't expect machines to to to learn stuff

well I mean unless you could develop an AI that develops a very strong heuristic for what gloominess means um what what gloominess feels like and what what concepts are appropriate

then it might come up with yeah but does it mean to have contact with the world to do that this is I mean the the question here is you know is is is it something you can learn by just looking at what words have been written I I would suspect not I think you know you it doesn't matter you could give it every piece of literature ever written um but it wouldn't would it know what

gloom actually was no no I think I don't think yeah I think you're right but I think the the words the written word is just a convenient accessible data set the better data set would be to plonk it out in the real world and let it sort of learn let it experience for itself as

we do okay so um I'm slightly sort of lost as to where we are at the moment um so beyond what we've said um you know what what isn't what is a good question for me to ask here what what should we now be talking about um with this we've been talking about uh this experiment with sunspring we've been talking about AI limitations of AI um we've talked about why it can work well sometimes

why are I guess I guess the obvious question is the future what next no please touch on this earlier we can only build tools when we understand the task so we can only make things uh to help us do things when when we know how we are supposed to do those things in the first place um you know as a general rule and an automation happens when a task is sufficiently well understood uh for us to be able to to start to um identify the processes involved do we know what makes good fiction I'd argue that we don't we there's no clear theory of fiction which will enable us to automate it at the moment now the the the sort of neural net approach the benjamin approach is to say well we you know maybe we don't we don't need to know we can just get um this thing to learn it'll look at these scripts and learn what good fiction looks like but the the answer clearly as sunspring demonstrates is either that it's nowhere near sophisticated enough or that the goodness of a piece of fiction does not reside in the words and the order they're in as such but something else is it's what the words mean and and that means that we you know we can't expect to make progress on this unless we begin to understand that and have means of codifying that ourselves

okay that makes sense um peter um looking to the near future it's a fascinating area with huge amount of research something i spotted recently was there's a google research project called magenta which is attempting to make new make new progress in this area uh and it's one of its out stated outputs is going to be a toolkit that will be accessible to anyone with a moderate level of development to use to create their own music um with using ai and they're it's uh it's available now and it's quite i'm yet to have a play with it but it looks it looks very promising uh and they're using it as a as a basis for doing lots of research into what computer-based creativity actually is and means and so we'll have lots of interesting technical output but also lots of interesting ethical and philosophical output

yeah i mean on the music side of things it starts to intuitively to me at least become more uh understandable because i remember going back to my days in the early 90s as a hardcore raver type person and um one of the exciting things about what was going on back then and this must have been pretty crude compared today is that um djs if that's the right word were you know working off computers where you kind of you you have a program you kind of set it free and it will do certain things at a certain point and just work off an algorithm and just suddenly start pumping stuff up the beats out and um and sort of taking on a life of its own as guess what yeah well the

key the key thing about music the main difference between it and uh and written and prose is that it doesn't mean anything so music doesn't mean something so there's only one layer you have to look at and and of course music is also very uh i mean there's a lot of patterns in music so it's much more amenable yeah um and it already comes sort of ready coded you know you you've got uh music written music is fairly machine readable anyway so so you know it's much more plausible that we make progress on uh artificial generation of music in fact i think some of you know it's it's been it's 10 15 20 years really since um artificial intelligences start started producing things like you know plausible sounding uh bark preludes and so on because you know there's um you know the big question i think is whether or not we uh they can produce things that are totally different um from things that already exist and this is the kind of problem with ai creativity is if it's learned what it's trying to achieve entirely by what's been done already um it may not it may not pick up sort of the meta level which is that actually sometimes what people want is something different but um yeah i think that's a probably a topic for another podcast

sure there's a rise of a star trek episode with data um star trek the next generation where he was trying to you know who's a superb musician but i seem to remember yeah he wasn't able to create anything new or when he did it just sounded crap basically um okay i think um any final words we're gonna finish there any final words peter shaking your head nick anything from you uh no okay that's unusual um and i'm trying to think of something pithy to finish on um

well i think i think we you know to quote benjamin it's a damn scared thing to say

well said so we'll wrap up there my name's fraser mcgrewer um i've been here with nick hare and peter coghill of aleph insights you you've been listening to the cognitive engineering podcast until next time thank you very much bye