I’m Annie Rauwerda, and I started @depthsofwikipedia on Instagram, Twitter, and TikTok to highlight my favorite trivia and internet rabbit holes. I partnered with Bullish Studio to launch this newsletter. While most editions involve Wikipedia deep dives, today I'm interviewing the creator of my favorite Twitter bot, @ResNeXtGuesser! Here's my favorite quote from our conversation:
"In my opinion, the funniest and best projects are those where you take some highly professional, advanced, or academic subject/technology and throw in a great deal of immaturity."
You're about to meet my favorite twitter bot, so here are the bots that you recommended in my Instagram story question box.
@ATOMlCATS@THATONE23902017@TiuttiChilly@DailyWerehog Your video is ready. Do you want it removed? contact @/LuisMayoV
— Ace Attorney Court Bot 🏳️⚧️ (@aceCourtBot)
Nov 3, 2021
Alcohol and Drug Abuse Lake
(Manual post)
— TMNT Wikipedia Titles (@wiki_tmnt)
May 1, 2021
we call x^2 “squared,” we call x^3 “cubed,” but people forget there’s a word for x^8 and that it’s zenzizenzizenzic
— Depths of Wikipedia (@depthsofwiki)
Oct 30, 2021
The account's creator is a 23-year-old embedded computer engineer in San Francisco. He prefers to keep his professional life separate from his online life, so I'm using Dave in place of his real name (nod to the Dave Rule).
A few days ago, I discovered the delightful @ResNeXtGuesser bot, which tries its hardest to make sense of memes. When given bizarre, cursed images, the bot labels the photo's contents — sometimes with hysterical inaccuracy. When I scanned the Github code (available at this link), I realized that the program only gives 1,000 categories of images, which explains why the bot seems so creative with its classifications.
I asked the account's creator to answer a few questions about his internet masterpiece, and in the few days since I first reached out, the account's following has EXPLODED. In the past day alone, it has gained 20k followers, and it's almost to 100k. I'm so excited to share the interview.
Image prediction: pineapple
Confidence: 99.3%— neural net guesses memes (@ResNeXtGuesser)
Nov 2, 2021
Annie: When did you start @ResNeXtGuesser? What inspired you to create it?
Dave: I launched the account in June of this year. I recently graduated college, and although AI and machine learning weren't my main area of focus, I was really interested in the subjects and had taken a handful of classes on them. I now work as an embedded computer engineer, so I never even come close to dealing with machine learning. I wanted to do a project that gave me a chance to play with neural networks (NNs) again.
In my opinion, the funniest and best projects are those where you take some highly professional, advanced, or academic subject/technology and throw in a great deal of immaturity. In college, I had various AI/ML assignments that dealt with classifying images, and it was so funny to me to see what happened when I sent through bizarre or random memes. What was particularly interesting was that even though the NN got it wrong 99% of the time, you could often see/understand what the NN was thinking or why it came up with its prediction.
I recently learned about the ImageNet competition, and how neural nets are already better than humans at predicting the contents of an image. But the problem is that those NN's are tested using boring images, such as cats, dogs, cars, pencils, etc. I made a video on my youtube account putting random memes through the ResNeXt NN to see what happened. In my opinion, the results were really funny. A couple months later, I thought that running memes through a NN would make for a really entertaining Twitter bot, so I got to work on it.
Image prediction: skunk
Confidence: 99.52%
Submission by @waffluffe— neural net guesses memes (@ResNeXtGuesser)
Nov 1, 2021
Annie: How long was the programming process? What were the main steps of the project?
Dave: The programming process was actually quite quick — I'd say it took me maybe a month of half-hearted, over-the-weekend type work. There are a lot of pre-existing Python libraries that make interfacing with the twitter API a lot easier. Besides the Twitter stuff, the Python library PyTorch provides a bunch of pre-trained, ready-to-use NN models that are super easy to get up-and-running. Ironically, for a bot about neural networks, the neural network code was the easiest part.
The real pain was automating grabbing submissions through the DMs. Twitter's documentation on the subject is quite sparse, and the libraries I was using to interface with Twitter had even worse documentation. But I got it in a usable state, albeit with a handful of bugs. But now, the account has exploded in popularity, which I am so happy to see. However, the high volume of DMs causes my code to send so many API requests that I hit the request limit and Twitter times me out, which breaks my code.
Image prediction: refrigerator
Confidence: 20.69%— neural net guesses memes (@ResNeXtGuesser)
Oct 31, 2021
Annie: You work in embedded computer systems, you create Twitter bots and Youtube videos, and you like memes. Do you have any other hobbies?
Dave: I love working on personal projects. Right now, I'm working on getting a website up and running. It's fun! I like learning new things, and it gives me a chance to make neat stuff. I love cooking and photography (possibly the only two "artistically creative" things at which I'm somewhat competent). And I'm really into stocks & finance — I'm a huge math nerd, and quantitative finance is neat to me.
Image prediction: dingo
Confidence: 69.09%— neural net guesses memes (@ResNeXtGuesser)
Oct 30, 2021
Annie: What are some things on the internet you’ve enjoyed lately and would recommend to others?
Dave:
Image prediction: pizza
Confidence: 23.88%— neural net guesses memes (@ResNeXtGuesser)
Oct 27, 2021
Annie: Do you have any favorite Wikipedia pages?
Dave: Absolutely! I'm a Wikipedia junkie which is why I love your account. Here are four in particular:
Annie: Thanks again for answering these questions, and best of luck as you keep working on computers, working on bots, and collecting cursed images!!
That's all for today! As always, feel free to forward this to a friend and send your thoughts to [email protected]. I really like your emails!
Deep dives into Wikipedia rarities, curated by @depthsofwikipedia