Time to write up something that’s been sitting around on various disks for a while.

Many months ago, I saw Plotagon. It’s best explained as Xtranormal by way of The Sims: reasonable resolution, 3D-animated videos based on scripts; a desktop tool to generate them, and a site to host them.

Most interestingly, it’s scripted with what actually looks like movie scripts, and that got me thinking: what would it look like to feed it with procedurally generated scripts? Could you make the machine make videos? All I knew was two things:

1. I know, for good or ill, how Markov chains work.
2. All the scripts for Friends are transcribed on the internet.

After all, given Plotagon’s focus on semi-realistic forms, I decided that it was best suited to the great American artform of the 20th century: the sitcom.

The Infinite Friends Machine was born.

The machines does a few simple things. First, it scrapes Friends transcripts. For now, it works for most of Series 1. It then parses those scripts and chops them up into episodes, scenes, and lines attributed to individual characters. It also strips out some directions. Then, using all that, it offers ways to generate new scripts.

Markov Chains, as Leonard has frequently pointed out, are not always the best way of generating text alone, especially when the corpus you’re working from isn’t particularly consistent. He is, of course, right. Still, I enjoy the mental leap readers make in order to make generative prose actually make sense, and for this project, I mainly wanted to get to scripts as fast as I could.

Still, I didn’t want to hamper their relative crudeness, so I tried to skew things in their favour. To that end, the Infinite Friends Machine generates scripts by copying the structure of existing scripts. When it makes a new “episode”:

1. it finds the scenes that are in the original episode it’s being copied from
2. for each scene, it finds each line – who says a line at what point in the episode
3. then, it generates a new line for the speaking character from their own corpus. That is: Joey only ever things derived from Everything Joey Has Ever Said. What this means is that the main cast have quite diverse things they might say, and the bit players pretty much only say the same thing. Gunther is quite boring.

That’s it. A few seconds later, it spits out a nonsensical episodes of friends. Here’s a scene:

and this is a full episode.

The machine isn’t online because it’s quite crude and processor-intensive, but you can get at the sourcecode from Github.

Anyhow: machine to generate scripts. Next stage: get them into Plotagon.

This was where my troubles began. For starters, despite having a nice format for scripts, Plotagon really demands you enter them via its UI – you can’t paste a big block of text in, you have to enter it by hand. Painful.

Next: Plotagon only lets scenes have two characters in. I decided to make a single scene – the tag on the end of the episode. But this turned into many scenes in Plotagon, as four people in an apartment was a bit much for it. I had to keep track of who was where, who was talking to whom at any point.

And then I had to deal with the unfortunate truth: Plotagon is horrible. I mean, Xtranormal used its non-realistic avatars and computer-voices to comic extent. By contrast, here we had disappointing voice acting with clunky visuals. Also, I had to add some ‘acting’. This largely consisted of making Chandler say everything whilst doing the (crazy) emote, to really capture that Series 1 Matthew Perry vibe.

A quick sting later, and Infinifriends S1E1 existed:

It is not exactly high art.

Just one scene took long enough, and I think, proved my point to an extent, but probably can’t be improved on for now. I’m not sure if I’ll ever return to the Infinite Friends Machine, but it was an entertaining enough exercise, and the video rendition is probably worth it for the cringe factor alone.

Theme tune. Credits. Tune in next time.

There’s a copy of me running around London. It runs at the right speed, but it’s unstuck in time: unstuck by 365 days, in fact.

Tom Armitage In The Past is a foursquare user. He checks in where I was a year ago. He says what I said a year ago. As long as the game systems haven’t changed, he gets the badges I got a year ago. (And, if the game rules do change, he’ll start to deviate from me).

He’s maintained by a piece of software I wrote called ghostcar.

In racing games like Ridge Racer, your previous time attack high score is often represented by a 3D outline of a car rendered into the same world as you – a “ghost car”. Matt Jones reminded me of this at the pub one night, as I explained my idea, and the project had a name.

It’s not a spambot. I’m its only friend; it’s invisible to people who aren’t its friend. It’s invisible to venue owners – ie, it’s not generating false marketing data. Honesty is Foursquare’s main currency: saying where you really are, being who you really say.

I’m not breaking the terms of service. When I wrote it, the Foursquare terms of service told me I couldn’t impersonate other people.

But I’m not impersonating other people. I’m impersonating myself. That’s got to be OK, right?

So: why did I do this?

There were two main inspirations. Firstly, James Wheare’s Twitshift (now sadly closed). Twitshift works because its output is in the same medium as the source data. I didn’t warm to Timehop because it was just an email. Email is good, and there’s value in juxtaposition – but anyone can send an email. Email dilutes the notion of “being”, however. Foursquare is not, precisely, about presence – it’s just about saying you were somewhere – but email about foursquare feels like another abstraction layer. I wanted to minimise abstraction. And so, to do that, I’d need to build a Foursquare timeshifted-echo service that itself output to Foursquare.

Secondly, I remember Kevin Slavin talking about Area/Code’s wide, GPS-enabled game Crossroads at dConstruct a few years back – and how, once, when testing, it the game’s entirely digital antagonist “Papa Bones” moved through their office, and every GPS-enabled phone they were testing the game on suddenly jumped. What made the virtual “ghost” interesting was when it manifested in the world. Even though it wasn’t real – and everybody knew it wasn’t real – it still made you jump when you were there; the juxtaposition of knowing something is right where you are, even if it isn’t a real thing, is highly striking. I wanted to explore that a little with my own data.

I don’t check up on Mr Armitage In The Past very much. The idea, rather, is that I might bump into him some time. How strange: to be in one of my usual haunts, and know that a ghost of me, a year in the past, is also there, watching a movie, having a drink. Sometimes, those memories are less cheery than others. Sometimes, they’re brilliant. It gives me a visceral memory: reminds my bones, my heart, what they felt. (That, for reference, is my defence against nostalgia. This isn’t just about nostalgia, because you might not like what it makes you feel. It’s just about remembering feelings; stopping to pause and remember the passage of time).

It’s also made me check in to Foursquare a bit more. The moment I fired ghostcar up, I realised I needed to start giving it better data so that it’d continue to have meaning a year in the future. So that’s a strange, interesting takeaway: changing my behaviour because I want the fossil record to be more accurate.

I was only going to find out what it felt like by making it, so I made it. It chugs away on my server in private. I run ghostcars for a few friends, too. It’s not particularly elegant, and I don’t think it’d scale to loads of users, so for now, it’s a private distraction. But I thought I should write it up. If you’d like to run your own install, the code is on Github. It’s a minimally documented Rails app, because it was made just for me, but you might enjoy it.

In the meantime: I go about my business, and an echo of me in the past will do one day, too.