4 ms·
It's still not good enough to manage a vending machine: https://www.anthropic.com/research/project-vend-1 https://www.anthropic.com/research/project-vend-1
by muzani 11mo ago
It's still not good enough to manage a vending machine: https://www.anthropic.com/research/project-vend-1 https://www.anthropic.com/research/project-vend-1
- sharts 11mo agoProbably by design? A simple prompt and some tools? Bro I have to continuously prompt and plan even to get a reasonably sane piece of code written w/ Claude. I’m not surprised this failed. Its like they wanted it to.
- muzani 11mo agoI doubt it was by design. It's still an autocomplete machine in the end. Claude tries to do agents better than anyone and part of this was to see where it would fail. There's plenty of blind spots in a real world situation. It was trained to answer questions and score well in exams, not to negotiate and evaluate the value of a tungsten cube, and so on. There weren't enough scams in the training set. And seeing how rare data is from a CEO's perspective, I expect it would perform much worse. It's not just vending machines, it does poorly in games with full wikis too.