We chat with the man behind the madness, Neal "tsunami" Khandheria, about his AI creation.
If you’ve spent plenty of that time you’ve been killing in lockdown surfing social media, you may have noticed tweets from a brand-new twitter account pop up in your feed.
But, if you didn’t notice; or perhaps didn’t realize those tweets were from a new account, you could easily be forgiven.
The tweets in question bear something of a resemblance to a much beloved Dota 2 personalities tweets. Jake Kanner, better known as "SirActionSlacks".
And that’s because in a way, they are his tweets.
Some people think dota lore is my cup of tea but I fucking hate tea
— Deep Slacks Bot (@DeepSlacks) April 16, 2020
Deep Slacks Bot is the quarantine brainchild of none other than Neal "tsunami" Khandheria; another well-known face in the DPC.
Although “generating text through AI for comedic purposes is nothing new to the Dota scene,” says tsunami, he was inspired by twitter account @DeepLeffen. A text-generating AI bot which parodies professional Super Smash Bros Melee/Ultimate player, William "Leffen" Hjelte.
Tsunami says when he first encountered DeepLeffen he “didn't fully understand” what exactly it was. “But, once I realized that the tweets were AI generated, reading them had me in tears.”
As a content creator, tsunami says he’s “always amused by new fads” and often thinking of ways he can bring them to the Dota 2 scene. “Earlier this year I knew nothing about Instagram filters, but when the 'What X Are You?' stories started popping up everywhere, I learned how to make them just so I could create filters for Dota players.”
Similarly, after encountering DeepLeffen, tsunami “really wanted to make a text synthesis account for Dota twitter to enjoy.”
Also, he admits “having a bunch of pent up quarantine energy definitely helped.”
A quarantine brainwave
“This project was conceived and completed basically in the span of 48 hours,” He says.
Tsunami explains, he has “some programming background,” but he did have to do some independent research to get the bot to work.
One of the main obstacles was the number of source tweets needed to train the bot.
“Dota twitter is extremely slow compared to all the other esports I follow,” he explains, and a GPT-2 text bot needs, “a lot of source tweets.”
As tsunami notes, the variety in the content of Dota 2 pro tweets can leave something to be desired when crating a bot.
“Even the Dota pro players and personalities who do tweet generally just say things like 'I'm streaming' or 'We won the tournament, thanks for the support' which don't really lend themselves to comedic reinterpretations.”
But there is one personality within the Dota 2 community who is a prolific tweeter; Slacks.
“Slacks is also the polar opposite of Leffen in terms of personality and eloquence on twitter,” tsunami says, “so I wasn't really sure if training a model on his tweets would be funny or even work properly.”
Nothin better than watching my favorite team lose, and especially nothin more satisfying than sitting at home and crying
— Deep Slacks Bot (@DeepSlacks) May 4, 2020
“Luckily, Deep Slacks turned out to be comical in its own absurdist ways.”
Bringing Deep Slacks to life
Did you know it's surprisingly easy to make one of those machine learning text synthesis bots?
— tsunami (@tsunami643) April 15, 2020
We asked tsunami to elaborate a little on exactly what was involved in bringing Deep Slacks to life.
“As much as I would like to pretend that I'm some kind of machine learning savant, my claim that the project was "surprisingly easy" isn't me being humble.” He says, “I literally just googled "gpt-2 twitter bot" and the first hit is a very straightforward guide created by @minimaxir literally "How to Build a Twitter Text-Generating AI Bot With GPT-2".”
According to him the actual creation of the bot was akin to “just reading a cake recipe and following the steps as written.”
When i wear pants im dressing like a a slacks fan
— Deep Slacks Bot (@DeepSlacks) April 26, 2020
But building and training the bot wasn’t the end of the road for DeepSlacks.
“The most time-consuming aspect of this whole thing wasn't training the bot, but rather panning for gold out of all the garbage that it gives.” Says tsunami. “Only a very very small percentage of what the bot generates is actually humorous.”
Put me on the record saying that I would never again work for free because I just found out that it was possible to get paid
— Deep Slacks Bot (@DeepSlacks) April 19, 2020
How many is that?
“I'd wager that for every 300 tweets it produces, maybe 15 are passable, and maybe 1 is legitimately hilarious”. He says, “The rest are gibberish, boring, or verbatim tweets from Slacks himself.”
The Slacks seal of approval
Because DeepSlacks bot was conceived and created in such a short period, there wasn't much time to “mull over the ethics of the whole thing — or ask Slacks for his blessing before publishing it.”
But as tsunami explains he “…would feel a bit weird if someone started speaking through a 'character' that's supposed to be a representation of me”. So, he feels “a responsibility…” to Slacks “…to not post anything that would disrespect his image when curating the bot's tweets.”
Although he didn’t float the idea past him first, tsunami says he was fairly confident Slacks “would've trusted me using his 'likeness' for comedy purposes.”
A fair call considering the two know each other quite well. And post-release that was indeed the case. “He's messaged me saying he's found it very funny as well as retweeting many of the bot's submissions.”
“Generating text through AI for comedic purposes is nothing new to the Dota scene.” He adds, before referencing the Markov generated patch changelog that spawned the "Illusory Orb speed increased by 1" Puck buff in patch 6.87.
But even if it’s nothing new, DeepSlacksBot is definitely something uniquely Dota. And with DPC events delayed, it’s certainly kind of tsunami helping us get a little more Slacks on our screens.
If you haven’t caught any of DeepSlacksBot’s ramblings in your feed yet, you can check them out on twitter.
Feature Image: Mars Media