Raw LLM Responses

They could have implemented an AI monitoring system for cheaper, and probably more reliably with a clustering algorithm. But the point is not to just come up with a solution, its to come up with an LLM based solution. Every tech company is focusing their efforts on making LLM’s viable as a product. Ive recently worked on a POC for an embedded LLM-based game (think classic school children’s game), involving receiving user image data as input, prompting the model to analyze the image and respond with some definitive output about the image. It fails more than 50% of the time. It would probably be better to just implement a coin-flip instruction instead of prompting the model. I could also just use openCV to do the image recognition, slap a “powered by AI” label on the app, and the customer wouldnt know the difference. Nobody would play the game either way. But thats not the point. We are just trying to figure out what kind of products we can build on top of an LLM, and so far, its a goose-egg.