A place for the lighter side of generative AI.
Researchers at the University of Zurich and Cologne University are creating large language models trained solely on historical documents up to specific dates. The team, led by Daniel Göttlich, Dominik Loibner, Guohui Jiang, and Hans-Joachim Voth, describes their project on GitHub. The models, called Ranke-4B, contain four billion parameters and are trained on 80 billion …
Anthropic’s Claude AI model recently took over the management of a vending machine in the Wall Street Journal newsroom. The experiment aimed to test the autonomy and business logic of AI agents in a real-world setting. Joanna Stern reports for the Wall Street Journal that the project quickly devolved into financial chaos and social engineering. …
A new iOS app called “touch grass” helps users break screen addiction by blocking their most distracting applications until they physically touch grass and verify it through the app’s computer vision technology. The app aims to help people reclaim their time from mindless scrolling in a tongue-in-cheek way while encouraging connections with nature. Basic features …
Just as unwanted junk advertising is commonly known as “spam”, unwanted AI-generated junk content could be called “slop”. At least, that’s the opinion of Simon Willison, who noticed the term and suggested it in a blog post. And I think it fits very well. He points out that not all AI-generated content is automatically slop: …
Redditor ForceTypical asked ChatGPT to come up with memes. As expected, the results are strange.
Give it a listen. It’s a gimmick, sure, but at the same time it shows the current state of a music AI like Suno.ai very nicely.
AI image generators often fail because they don’t understand (yet?) what they’re creating. Mirrors are a good example. Source: Reddit
Poem/1 is a special kind of clock: it uses ChatGPT to generate a poem at the current time and displays it on an e-paper display. You can support the project on Kickstarter.