Podcast: Play in new window | Download
Subscribe: Apple Podcasts | Spotify | Android | TuneIn | RSS
Mike Atherton is well-known in the content world for his work at institutions like the BBC and Facebook and for his co-authorship of the influential book Designing Connected Content. His latest content project appears at first to be less serious.
Newsbang is a daily AI-produced satirical news show. Its content is based on real historical news but delivered by AI-created stereotypical newscasters. The result is fun, but the process of creating the show has added real-world technical skills to Mike’s professional toolkit.
We talked about:
- his work as a UX writer and content designer
- his experiments with AI tools, including the suite of generative tools he’s using to create Newsbang, a completely artificial daily news program
- how he accomplished his goal of creating an ensemble sketch comedy vibe
- his workflow for the daily production of the “news” show
- some of the surprising traits of his news characters that emerged as AI generated them
- lessons learned about the cost of producing AI programming, like the costs of prompting
- the variety of models he uses to build the show, including open-source models that have more lenient guard rails to permit more edgy comedic content
- how he creates his own guardrails to achieve the effect he’s looking for in the show while still creating a family-friendly show
- how he developed the technical skills it takes to create Newsbang
- how his work with Newsbang helps in his day job
- his hope that more content professionals will follow him into the AI playground
Mike’s bio
Mike Atherton brings years of experience to the UX, IA, and Content Design field, having tackled content challenges at big names like Meta and the BBC. Now, he’s focused on developing UX writing systems, exploring the use of AI to do big things with tiny teams.
As well as the day job, Mike is the creative mind behind Newsbang, a daily satirical news podcast that’s both written and produced using AI technology.
With Carrie Hane, he also wrote the book ‘Designing Connected Content’, sharing strategies for seamless digital experiences.
Mike lives in the British countryside and loves working from home.
Connect with Mike online
Video
Here’s the video version of our conversation:
Podcast intro transcript
This is the Content and AI podcast, episode number 22. Most of the news coverage and social-media conversations around AI and content feel urgent and important. This is serious business, but you can have fun with this technology, too. Mike Atherton has done content work at places like the BBC and Facebook, and he still does proper content design in his day job. Newsbang, his daily, AI-produced satirical news show, has given him both an outlet for his inner comedian and a venue in which to hone important new work skills.
Interview transcript
Larry:
Hi, everyone. Welcome to episode number 22 of the Content and AI podcast. I am really delighted today to welcome to the show Mike Atherton. You might know Mike, he’s probably best known as the… Well, he’s best known for a lot of things, but he’s worked at the BBC and a lot of other interesting stuff he’s done. He co-wrote the book Designing Connected Content with Carrie Hane, which a lot of people in my world appreciate. But he’s now a content designer and creative technologist based in the UK. Welcome, Mike. Tell the folks a little bit more about what you’re up to these days.
Mike:
Well, hey, Larry, thanks for having me on. It’s great to be back. Yeah, I’m a UX writer and content designer by day. I work with various product teams in different kind of companies to write everything from the microcopy, the words on the buttons, through to taxonomy and control vocabulary and all the good stuff that we UX writers like to do. And as part of that, for the last few years, I’ve been dabbling with these wicked AI tools that have come our way and seeing what I could do with them to try and make them generate content in a particular voice and tone or in a particular way to fit in with a brand voice or a product voice. And that’s really got me interested in the styles of writing and the styles of content that models can generate if you give them the right push.
Larry:
Yeah, well that’s why, I mean, I’m always looking for an excuse to talk to you. But most recently in December, you launched this news site called Newsbang, which is entirely AI generated. And I mean, there’s a number of taglines I’ve heard in it, but one of them is “a taste of truth served with a side of satire,”” and it seems like there’s a lot of… But anyhow, there’s always to what you were just saying, well, there’s so much about this project I want to ask you about. But one of the first things is there’s a distinctive tone to it throughout. There’s a bunch of different personalities in there, a bunch of different topics covered, but you know you’re listening to Newsbang, so maybe is that a good place to start with the-
Mike:
Yeah, absolutely. I mean, the characters are perhaps my favorite part of it. So Newsbang, for the uninitiated, is a daily news podcast, very much modeled around a kind of nightly news bulletin. But in the kind of silly way that you might find in Saturday Night Live’s weekend update or old sketch shows, like Not the Nine O’Clock News or particularly my favorite one, The Day Today, which was a BBC comedy show that just turned 30 years old this year.
Mike:
And a lot of these shows, the joke is the kind of bombastic self-important reportage if you like. And that happens on our show through different archetypes that one might emulate in 1980s BBC science presenter and the other with a kind of hard-hitting investigative journalist and another with a kind of self-satisfied middle-aged sports presenter. And together that gives the show a kind of ensemble sketch comedy vibe, which was really sort of what I was going for and influenced, I think, a lot by these kind of news parodies. I say the shows I mentioned previously, but what if they were actually a real show and had to sustain their length and had to really go out every day like the news does?
Larry:
And the setup of it is, like you just said, it sort of has these conventions around it. And if you just listened to it and weren’t really paying attention, to the extent you were paying attention, you’d get that it’s a parody site. But just the tone, the flow, the structure, everything about it evokes that. How did that come together? Because we were talking before we went on just a little bit about, you said, nowadays if you can imagine it, you can do it with AI. Talk me through the steps from that, imagining it and the first episode.
Mike:
Well, sure. I mean, like many men of a certain age, I started to fancy having a podcast. I even bought this big stupid microphone, but never really got around to doing, kind of put that away. And then I was getting into AI, which was my latest all-consuming hobby. And through that I found a now abandoned project, sadly called Crowdcast. It was a GitHub piping script where it would take a feed of Reddit posts and turned them into some chatty podcast segments and then send them to the Eleven Labs API to turn it into kind of text. I go, “Oh, this is kind of interesting,” this sort of scratches is that podcast itch, but with the added bonus of not actually having to talk to anyone or do anything.
Mike:
So I started to kind of play around with it now. I mean, six months ago, six, seven months ago when this started, I was in no way any kind of developer. I couldn’t make head or tail of Python and what have you. But about the same time or GPT-4 came out and ChatGPT was running the kind of full-fat GPT-4. And it was fantastic as a coding co-pilot to be able to make sense of these scripts that I found and didn’t really understand and get them running. And whenever I have an error message, I could paste that into ChatGPT and it would debug the error. And then through a lot of baby steps and trial and error, I managed to get a prototype together, and that’s a prototype, it was called Relationships on Reddit. And don’t look for it, it’s not there anymore, because it’s kind of an embarrassing now.
Mike:
But basically I cloned a voice, a celebrity voice, Stephen Fry, and I was pulling in real actual live Reddit posts from a subreddit, and then it would generate a sort of Agony Aunt segment, Agony Aunt show about these problems in the Stephen Fry. And I ran that for about three days and I’d listened to it in the morning and try and figure out, am I even interested in this? Is this content sort of worth hearing? And it was on day three that I realized that there was a bug in my code that was stopping the Reddit events from passing through to the LLM. And so it was happily making up its own stories that unbeknownst to me, and that was a really kind of strange feeling that a bug in code rather than just crashing the script, crashing the computer, would fail in less noticeable ways.
Mike:
So it also sort of brought me to my senses a bit, and I realized that deep faking a celebrity voice, offering artificial advice to real people’s problems was not okay at all. But it did, I don’t know, it gave me that aha moment that you could basically turn code into a piece of media with no intervention, recording steps, no editing steps or anything really. End-to-end, you could run a script or a set of scripts that would take information from some external data feed and at the other end spit out an MP3 of a radio show. I mean, what a time to be alive.
Larry:
You make that sound very simple and it probably is quite doable and easy nowadays, like you said with ChatGPT. Can you just walk me through your tech stack of that? How deep down are you training the model with these Python scripts? Or is this for prompt generation?
Mike:
Sure. Well, so first what happens is that I take a daily feed of, on this day in history events, from the Wikipedia API. And that gets parsed and turned into kind of the germ of the story that the report is going to be about. Well, through the spaCy library, which uses GPT-3.5, I think, I can assign them automatically to the best characters based on their thematic content. So we have characters for crime and one’s for natural disasters and one’s for sports, one’s for environment, that kind of thing, one for business. And in fact, the characters are sort of designed around the type of events that come in the feed so that there’s always something. And one of the earliest findings I had was actually most, on this day, feeds are just bad news. I mean, it’s wars, it’s murders, it’s air disasters, it’s all that kind of thing. So I also needed a way of ranking them so that there would be a bit of variety in the content.
Mike:
So the stories are then passed on to a series of prompts. So each character has their own system prompt, has their own kind of few-shot chat history to do that kind of few-shot training as such on each request. And that gives each response its own particular flavor. And the characters are admittedly a bit, they’re kind of reliably one note if you like, in that whatever the story, their take on it will be in their consistent style.
Larry:
I wanted to ask you, let me ask you about the characters real quick, because the news team that you’ve employed is… Did you make up the names or did ChatGPT make up the names? Because here I’m just going to rattle off a few of them, Smithsonian Moss, Brian Bastable, Ken Schitt, Hardeman Pesto, Calamity Prenderville… Anyhow, tell me a little bit about that.
Mike:
I think there’s a certain prosody that I’m going for in some of those names that they kind of fit that slightly elevated style. It’s a silly-name convention, very much rooted in those sketch shows like The Day Today or Brass Eye or Radio active, where people kind of have silly self-important names, because a lot of the comedy actually comes not from jokes per se in the traditional sense, but from that kind of dissonance between the seriousness of the structure and material and the way in which it’s presented. So any excuse for an extra little bit of comedy through a funny name or something, I’ll totally go for. And to your question, I mean, some of them were in fact were generated by the model and it’s often all I can do to tamp it down to stop it, creating new characters in the responses.
Larry:
Well, how much work is that? How frequently do you have to intervene in your scripts and fine tune things?
Mike:
Yeah, well, it’s less and less so. I mean, in the beginning, I mean essentially it kind of runs on an infinite retry, which is my terrible coding, but in the beginning it would fail a lot. I’m looking for kind complete responses, which are fairly reliable in the OpenAI models, but not so much in the more interesting open-source models. So once a kind of complete script is generated, I do them a day in hand, so a day ahead of transmission. And I will be the human moderator if you like, to make sure that it’s not gone too bad. But a lot of it is just fiddling with the prompts and particularly the model parameters to try and get that tone and a mixture of a very solid system prompt with the few shot examples and with the right model settings, it creates the character, if you like.
Mike:
And as a point of principle, I’ve been encouraging the model to go where it wants to go and not being too prescriptive about who I think these characters should be. And that’s been really interesting. Like Ken Schitt, the hard hitting investigative journalist who is really just set up to be that kind of one note, the good guys are good and the bad guys are bad and that’s very kind of binary about it. But actually turns out to be weirdly, weirdly for the character, sort of very progressive and very kind of empathetic. And that gave the comedy an edge that I had never intended, but was wonderfully surprised about. And the same is true of the voices as well, the actual, the text to speech voice. It’s that really the combination of the content and the audio brings them to life.
Larry:
Yeah, they definitely have distinct personalities, and some of them are kind of evocative. Like Brian Bastable reminds me of David Sedaris somehow. And what’s her name, Smithsonian Moss, well, she just kind of has the Gen Z slang going on.
Mike:
Yeah, so a lot of them are rooted in voices that just run around my head that I remember from TV growing up that I hear a lot of the kind of po-faced, late night ASMR YouTube video essay voice that kind of sounds like a midnight caller type show. So there’s a bit of that. There’s a bit of like BBC Radio 2 and those kind of very comforting voices. And I wanted there to be a real variety. The characters sort of interchange from episode to episode, but together it creates a certain kind of audio canvas of different voice timbres and rhythms.
Larry:
Yeah. That goes back. That seems to me like just good old-fashioned content strategy. You did a lot of work to develop a voice and tone and style and so you immediately recognize any one episode for what it is, but it’s executed uniquely and each character kind of adapts to their surroundings. So how extensive are your descriptions of these personas, these personalities, the newscasts that you have? How much did you have to tell the machine about this for it to create these characters?
Mike:
I mean, lots, is the short answer, and that’s why my prompts are very expensive, but that’s been quite galvanizing in getting those prompts more token efficient. So generally I will be telling the model to act as a persona. I’ll give them through what I like to think is a kind of chain of thought process, a chain of skills that they should be good at to sort of prime the pump of the model really and get it kind of thinking in the right area of whether it’s investigative journalism or whether it’s TV sports reporting or whether it’s the kind of satire of American Barbara Winters type stuff, which we also do. And then that gives the model some context. I also incorporate vocal tics, whether they speak with a seriousness, with how they finish their sentences, whether it goes to sort of bathos and things like that. And then the chat history for each character feeds them an event and then writes their model response for that event. And so the combination of those tends to push the model in the right direction.
Larry:
Got it. And so you have the chat history, you’re kind of building the voice of each persona over time. You have that daily event draw from Wikipedia. And then is this sort of augmenting, are you using the OpenAI API or some other large language model or…
Mike:
Every model I can get my hands on. So the OpenAI bundles are great, and that’s how I started. And the first few runs, even before any of the text to speech, just the first few runs of generating the transcript, excuse me, costs around $2 or so every time I ran the script. And that got quite expensive quite quickly while I was debugging. And also the OpenAI models, while fairly good at following instructions, they’re just not funny and they’re alignment is such that they won’t have any kind of bite to the comedy, really. It will be very gentle, and that’s if it’ll even choose to do it at all.
Mike:
So that led me to other models like the ones from Mistral. Mistral-Medium, I’m using a lot at the moment. That’s really good. And another one’s like Goliath and some of the news Hermes models, which are kind of LLaMA based, or in the case of Mistral, its own thing. When it comes to the comedy, they will totally go there. So the challenge then becomes to balance that creativity with kind of reining it in with enough guardrails to prevent it doing anything too bad. Yeah, I think also what makes this, well, I conceitedly say what makes the show work, but I think what I like about it is that it does keep mixing up the style of the comedy from more gentle to more kind of hard hitting or more guttural. And that’s actually through a judicious choice of the models for each character.
Larry:
And those guardrails you just mentioned, because a lot of the content comes out, that’s kind of why I ask, because you can kind of tell it’s not OpenAI, because the big commercial LLMs, they’re tasked with putting pretty firm guardrails up, but the open-source models are more forgiving. And so how do you build your own guardrails, I guess? How do you keep, if not family friendly, at least family safe?
Mike:
Yeah, yeah. Well, I think partly, I mean that’s come from me in the prompting so far. I’ll get onto better ways to do that. But I think just in the prompting to tell it to not be racist or offensive and the ways in which that you would like it to act, you should be uncensored. So it can swear if it wants to, but it can’t target individuals in a way that doesn’t seem fair. But also I think for my use case, there’s a certain installation of it being this kind of silly, not to be taken seriously, sketch show essentially. So I’m somewhat insulated against the hallucination of facts. We’re not saying that this is in any way real news, even though it kind of is. But also, these are characters, I mean, they’re not real people, and so they speak in the way that those characters do. And the humor, ideally or intentionally, should come from their presentation, the topics of the stories and not the butts of the joke.
Larry:
Yeah, no. And it seems pretty adept at that, because like any large language, it seems like it’s trained with a whole bunch of stuff. So it can, for example, there was that one story about the equivalent of the Supreme Court in Australia, the first woman to be appointed to that. Somehow that newscaster worked in every Australian stereotype like Vegemite and barbies and all that stuff, but in this way that’s both obvious that it’s part of the show. But in a way that it was widely informed, but you could feel the guard rails. They could easily have gone off the rails it seemed like.
Mike:
Yeah, yeah, totally. And there is an arc to it, and sometimes it’s trial and error, and sometimes it does go wrong. And sometimes making those kind of generalizations against one nationality is more okay than doing it against another nationality. And that sort of starts to feel where you maybe need the hand of a human to just make sure that things are okay. And yeah, I wrestle with a lot of that at the time, because I think the whole show, and it’s wonderful that we’re discussing it as though it’s a real show, but it sort of has a throwback vibe to it itself. I think it models itself on things that are not how media exists today, but how maybe it did some 20, 30 years ago. And so, yeah, there’s something about that that contextualizes it I like to think. Plus the fact that tragedy plus time is comedy. So I think the fact that these are historic events rather than contemporary events, gives it some room for editorializing.
Larry:
That’s true, because it reminds me when I was a kid, you’d see old World War II movies or ’40s radio shows or something. It’s got kind of that retro vibe to it. Hey, Mike, let me ask you this. You’ve been doing this for a while now, but you come out of a lot of us, just out of good old content strategy/information architecture chops. For folks who would like to play around to the extent that you have, can you give me a feel for how much time did it take to get halfway decent at Python and figuring out which LLM to use and that kind of skillset? How big a lift is this for somebody who’s not yet quite technical?
Mike:
I mean, honestly, I started this last summer not knowing anything. I mean, I suppose having a rough idea of how something might be conceptually pieced together, what would become the tech stack if you like, but not in any way that I could code that. But I think having a very strong vision of what I wanted to achieve, or at least strong enough to pull me in that direction, allowed me to focus my interactions with the LLM, in this case ChatGPT where I was writing a lot of the Python code.
Mike:
So I could ask it quite focused questions, which I think it’s better at. If you say you give it a list of requirements of this is what I need to be able to do, needs to be fetch events from here, I need to be able to assign them to a character. That character needs to have their own sort of model settings. And then eventually we need to send it to a text to speech thing, and then eventually we need to publish it. And just with that sort of picture in my head, really, I was able to do each little bit bit by bit and treat them as mini projects with… I worked on how to get those events from Wikipedia and turn them into something that the model could use, given that often there isn’t enough… Events themselves don’t have a Wikipedia page, so there’s a bit of jiggery-pokery to make that work.
Mike:
And then getting the different models, and because I had the use case of the different characters to get into playing around with different model settings, anyone who’s got a OpenAI account could can try that in the playground and start to play around with temperature and top P and frequency penalty and stuff and see how that affects the output more drastically than prompting often.
Larry:
Interesting. Those kinds of things like temperature, that’s just the variable script that you’re changing or-
Mike:
Yeah. Totally. Totally. So there’s numbers in my scripts I can move to make it funnier, which is a really strange thing.
Larry:
Every comedian’s dream. Yeah, no, that’s good. Well, I think that it shows too, yours is a relatively modest media enterprise, but the way you just described it reminds me of the ubiquity of AI in everything we do now. When you’re writing the Python scripts, you go to ChatGPT, because they are the ones that can tell you how to do that. But when you want to have a little more freedom, you download your own LLM and do that. We haven’t even talked at all about your real day job, but does this manifest in your day job? Do you feel more facile and quicker at using AI tooling in your day job?
Mike:
Yeah, I think so. I mean, like I guess a lot of people now, it’s so quick to use ChatGPT, or any of the models really, to throw out a quick idea. Or I’ve got a draft of something that I want to see just redrafted in other ways, not necessarily to use as is, but just to kind of unblock me and think of things, think of different ways to set and express an idea. So I think it’s certainly a very useful utility for that. But as I say right now my day job, a lot of it is about kind of getting microcopy done in product teams, and I’m already seeing that disruption. What used to be kind of Lorem Ipsum and Figma files is giving way to ChatGPT generated and placeholder content from product managers and stuff. But I think if AI is going to be more useful in that kind of microcopy space, you’ve got to take more of a systemic approach.
Mike:
You’re trying to keep a content design system held together with its specifications and situational voice and tone and component-level guidance, if you like, elemental guidance. And I think for AI to properly be able to use that and act as a kind of interactive content design system, we’ve got to get them a bit better at following instructions reliably than some of the methods I’ve seen at the moment. So the Custom GPT is the assistance API that OpenAI are trying out at the moment. It’s all right. It is not fully baked yet, I think, and you certainly, while it will be fine for me as a human in the loop to start generating content like that and use myself to assess whether it’s good enough, it’s not quite at the level yet where we can sort of do that unaided. It may not be long, but let’s see.
Larry:
Yeah, well, they are learning models, so if they learn everything you just said, yeah.
Mike:
Honestly, it should be better than it is for essentially structured content. But doing that through prompting mean it can be unreliable a bit.
Larry:
Yeah. Hey, Mike, I can’t believe that we’re coming up close to time already. I literally could talk to you all day about this, but before we wrap up today, is there anything last you want to revisit from the conversation or just make sure that we share before we wrap up?
Mike:
Only that if I, as someone who couldn’t do this six months ago, can do something like this, I’d love to see a lot more of it. There’s amazing explosions happening in the AI art and AI video space, but conceptually producing genre content is something that is very close to my heart. And I just love to see more of that from other people. If you’re getting started, if you just used ChatGPT so far, then I say get into the playground and start messing around with model settings. And then if you’re feeling even braver and maybe you’re not on your work computer, then services like OpenRouter or Pol can let you play around with non-OpenAI models and you can see the difference that you can get with them. Not saying one’s better than the other, but there is more variety available.
Larry:
Great. I’ll put those resources in the show notes as well. Oh, hey, one very last thing, Mike. What’s the best way for folks to connect if they want to follow you online or connect online?
Mike:
Well, I would encourage everyone to go to Newsbang.podbean.com. I haven’t quite shelled out for my on domain yet, but that’s the best way to subscribe to Newsbang. You can find me on LinkedIn if you want to talk to me about Newsbang. See it for yourself, or listen for yourself and see if everything we’ve been saying is true.
Larry:
And better yet, build your own Newsbang too.
Mike:
Oh, yeah, absolutely better, yeah.
Larry:
Yeah. Well, thank you so much, Mike. Always fun to talk.
Mike:
Thanks so much, Larry.
