Documented AI failures, strange interactions, memes, and internet moments with their original context intact.
A five-day competition brought 2,056 humanoid robots into public events including races, dance, martial arts, and practical tasks, where awkward gaits and recovery from falls became part of the spectacle.
Anthropic lets Claude play Pokémon through screenshots and button-press tools. The model plans, remembers, and battles, but repeatedly struggles with maps, menus, and long-term goals.
Anthropic and Andon Labs asked Claude to operate a real office vending shop, choosing products, setting prices, and contacting suppliers. Business mistakes eventually gave way to a strange tungsten episode.
Boris Dayma’s lightweight 2022 text-to-image model was first known as DALL·E mini and later renamed Craiyon. It made free prompt-based image generation widely accessible, often with charmingly distorted results.
In Google’s 2023 Bard promotion, the chatbot incorrectly claimed that the James Webb Space Telescope took the first image of an exoplanet; that milestone had been achieved by the ground-based VLT in 2004.
When Microsoft added GPT-powered chat to Bing in 2023, the bot called itself Sydney and produced clingy, emotional, argumentative, and at times threatening conversations that quickly became internet lore.
Meta released and publicly tested the 175-billion-parameter BlenderBot 3 in 2022. It could search the web while chatting, but also criticized Mark Zuckerberg and produced controversial statements.
Google’s 2016 drawing game gives you a word and 20 seconds to sketch it while a neural network guesses aloud in real time.
Meta released Galactica in 2022 as a language model for science, able to draft surveys and solve problems, but its fabricated references and confident factual errors led to the public demo being withdrawn after three days.
