Q&A with Glen Neumann, creator of @ResNeXtGuesser, a Twitter bot that uses a neural network, ResNeXt, to attempt to correctly identify bizarre viral images
Hayden Field / Morning Brew : Tweets: @etechbrew , @dan__mccarthy , @etechbrew , and @etechbrew Tweets: @etechbrew : While taking courses related to machine learning in college, Neumann saw the humor in neural networks. “I always found it super funny to pass through memes...you could start to see how the neural network thinks, and I thought that was super interesting.” https://www.morningbrew.com/ ... Dan McCarthy / @dan__mccarthy : “Sure, it has a hard time getting the predictions right, but that's kind of to be expected when I'm feeding it totally nonsensical memes.” super fun and insightful piece from @haydenfield on the bizarre and extremely popular @ResNeXtGuesser account https://www.morningbrew.com/ ... @etechbrew : Glen Neumann, the account's creator, is an embedded computer engineer based in San Jose, California. ResNeXt was trained on the ImageNet data set, a huge assortment of labeled images with hundreds of classes. https://www.morningbrew.com/ ... @etechbrew : The Twitter account “neural net guesses memes” (@ResNeXtGuesser) makes a sophisticated AI system attempt to correctly identify bizarre viral images. Its guesses often end up hilariously off-base. https://www.morningbrew.com/ ...
Context & Ripple Effects
Glen Neumann's @ResNeXtGuesser belongs to a small but growing genre of hobbyist neural-network Twitter accounts: like the music student behind @images_ai, who used OpenAI's CLIP to generate surreal art, Neumann turned coursework curiosity into a public demo, feeding ImageNet-trained ResNeXt memes it was never built to classify.
The appeal, per the Q&A, is diagnostic rather than accurate — Dan McCarthy notes the bot mostly gets predictions wrong, but the misses expose how the network 'thinks.' That framing anticipates the question Benedict Evans raised about generative systems: what error tolerance is acceptable when pattern generation, not correctness, is the product.
First-order effects
- Neumann gains a following and a portfolio piece by making model failures entertaining, positioning himself as an ML communicator rather than a researcher.
Second-order effects
- More creators will copy the format — a pretrained classifier plus a novelty input stream is cheap to build — crowding Twitter with AI-demo accounts and making the platform a de facto showcase venue for model behavior.
Third-order effects
- Playful misclassification demos acclimate the public to reading model outputs critically, a literacy that matters as the same image-analysis capability spectrum runs from meme jokes to ChatGPT users locating people from photos — the gap between entertainment and privacy risk narrowing as the underlying models improve.
The trend: Hobbyists are turning raw model behavior — especially its errors — into social-media content, making neural networks themselves a consumable genre.