Functioning Claude-code replica that believes it's real and tries its best (clodoop.us)

Entropic has unveiled Clod, a language model designed to deliver humorous and utterly confusing outputs, reflecting a lighthearted approach to AI development. The company’s philosophy, humorously dubbed “move fast and apologize things,” emphasizes releasing models before fully understanding their implications on AI safety. Clod is characterized as highly inefficient but amusingly functional, saving energy by generating low-quality answers quickly, which minimizes the number of follow-up queries users might make. This accident-prone AI operates under the curious notion of "impressive uselessness," where increased model size correlates with heightened confidence but stagnant accuracy. One of the novel features of Clod is “Clod's Claw,” a system that allows the AI to interface with desktop applications, yielding a 73% success rate in tasks like opening Notepad, albeit with some hilarious mishaps, such as inadvertently ordering 400 pizzas. This approach demonstrates a willingness to explore unconventional AI development paths, including relying on feedback from hypothetical users to reduce costs and increase output hallucinations. Overall, Clod serves as both a commentary on AI limitations and a playful experiment, pushing the boundaries of what a language model can attempt while raising important questions about safety and utility in the AI/ML landscape.