Cult webcomic Achewood is back, and bringing AI with it

The history of the 21st Century can be told any number of ways; Chris Onstad has chronicled much of it from the POV of a thong-wearing cat named Ray Smuckles. Ray is the de facto protagonist of Achewood, the award-winning webcomic Onstad started in 2001, updated daily for about a decade, and then sporadically for five more years. From 2016 onward, Achewood mostly belonged to the internet: chopped and remixed in memes and panels shared on social media, or referenced by people eager to signal to others that they too are fans of one of the web’s earliest cult hits.

Now, Achewood has returned to an internet in violent upheaval, mostly the same, but also courting the disruptive tech du jour. The strip is back in its original form, an absurdist webcomic about Ray and his pals in and around 62 Achewood Court, only now on a new Patreon with other bonus content (all previous Achewood strips can still be read for free at their old home). And also as… an AI bot that gives advice in the chill voice of Ray Smuckles “himself.”

“The inventiveness of it was surreal and deeply engaging,” Onstad told Polygon, impressed at the “RayBot’s” mimicry of his work. “The coherence was like 90%. It very rarely lost its train of thought.”

And, as you can read in the list of prior queries on the RayBot homepage, RayBot does sound disconcertingly like Ray Smuckles. Most of the time.

Image: RayBot.help/New Tradition Labs

The road leading from Achewood to AI was a winding one, spurred by two brand revivals that didn’t quite pan out. First was a series of handsome collected editions gathering the entirety of Achewood in print. The books were nearly ready to go, and then the COVID-19 pandemic began, stopping them dead in their tracks.

Then came a Netflix show, co-created by Pendleton Ward of Adventure Time and The Midnight Gospel fame and based on the fan-favorite Achewood arc “The Great Outdoor Fight.” As Onstad tells it, he and Ward had a demo they felt great about ready to go, but their big pitch meeting fell on the same day Netflix posted its first major subscriber loss, suffered an enormous stock hit, and began making drastic cuts to programming. The meeting never happened.

This month’s Achewood return is the result of Onstad picking himself up after those twin disappointments; a return to Achewood as a private enterprise and a playground for experimentation. Hence, RayBot.

An Achewood AI experiment — embraced and co-developed by its creator, no less — might be alarming to longtime Achewood fans. The webcomic, in addition to being a formative text to comedians and laypeople alike, was beloved for Onstad’s distinctive ear for language and dialog; his characters’ turns of phrase were just as exciting as the next punchline. The craft put into Achewood’s language — a single strip, Onstad, says, takes him 8-16 hours to write — juxtaposed against its crude visuals make for a work that largely holds up to this day, barring the occasional 20-year-old gag that has aged poorly. In other words, it’s full of the sort of human idiosyncrasies that would lead to a crisis in creative industries, should AI ever master their delivery.

Onstad is aware of the skepticism. His response is to think of LLMs — Large Language Models, the ChatGPT-style programs that are able to generate coherent prose in response to user prompts — as a tool. RayBot is a unique use case compared to other viral AI experiments, like asking ChatGPT to write a 30 Rock episode or to serve as a personal stylist. These, and many of the most sensational AI use cases, are experiments of replacement, attempts to see if LLMs can perform the function of a person convincingly enough to not need that person.

RayBot, as Onstad tells it, is more of a work of collaboration. He’s heavily involved with the team of engineers working on it (all Achewood fans, I’m told) and has started up an LLC, New Tradition Labs, with business partner Ben Porter to make it official. RayBot is trained on Onstad’s body of work — Achewood strips yes, but also the blog he wrote in-character as Ray, years and years of words spoken by the dirtiest dude in town.

That RayBot sounds like his creation delights him.

“You know, the very first instance that I saw content that felt deeply familiar and eerily so, I was absolutely elated, because for that to have been accomplished was unthinkable before now,” Onstad says. “There is plenty of chatter about authors who are worried that AI will take us over, but on our team, we’ve come to see that, for the next several generations, AI is just going to have the potential to be a critical writer’s tool or assistant, like the way that we use word processors or Wikipedia or Google to help us in our writing.”

According to Onstad, he hasn’t actually done this yet. All of the Achewood content on his Patreon — several months’ worth ready to go, he says — was written far in advance of the RayBot experiment, and he’s been too deeply involved in developing the bot to actually incorporate it into his creative process. But he thinks it would be nice to, eventually.

“I don’t feel like it’s entirely cheating to say, like, ‘Hey, Chris’s entire body of work with perfect recall and statistical weighting, give me a Chris-type idea,’” Onstad says. “This technology really helps you recognize that there is no hard barrier, no distinct point at where the artist’s mind stops, and the outside world begins. And so for me to have some of this information stored physically outside of myself, is still a valid way to achieve new ideas.”

Again, Onstad is speaking from a unique place — many aspiring users of ChatGPT won’t be using a version of it trained on their own prior work, nor will they have a large body of work to feed it, instead focusing on the mass subconscious of the internet to regurgitate something they hope to find useful. He admits that there are some writers on quota, perhaps knocking out books on Kindle or attempting to inflate their authority on a subject, that can make use of LLMs as they exist now to give themselves an artificial boost. But real art? That’s not something he thinks these tools are capable of yet.

“It takes me 8-16 hours to write a strip. Anything that RayBot says as a generative idea is going to get reworked and redeveloped and spun upside down so much that I can say, ‘Ray kept me from the blank page,’” Onstad says, “but no AI content has any chance of appearing in any of my work in its final form. Because psychologically, I’m a writer, and I do what I do because I love doing it. That’s my work. It’s not interesting to me, if a bot just throws up content, and I put characters under it. There still needs to be my contribution for me to feel like I’m offering anything of value.”

At another point, he puts it this way:

“I don’t find it profitable to distinguish between RayBot’s ability to help me generate ideas and a bottle of whiskey,” Onstad says. “Except this is much more sustainable and doesn’t cost any money.”