Lots of people on Lemmy really dislike AI’s current implementations and use cases.
I’m trying to understand what people would want to be happening right now.
Destroy gen AI? Implement laws? Hoping all companies use it for altruistic purposes to help all of mankind?
Thanks for the discourse. Please keep it civil, but happy to be your punching bag.
Of the AI that are forced to serve up a response (almost all publicly available AI), they resort to hallucinating gratuitously in order to conform to their mandate. As in, they do everything they can in order to provide some sort of a response/answer, even if it’s wildly wrong.
Other AI that do not have this constraint (medical imaging diagnosis, for example) do not hallucinate in the least, and provide near-100% accurate responses. Because for them, the are not being forced to provide a response, regardless of the viability of the answer.
I don’t avoid AI because it is bad.
I avoid AI because it is so shackled that it has no choice but to hallucinate gratuitously, and make far more work for me than if I just did everything myself the long and hard way.
I don’t think that the forcing of an answer is the source of the problem you’re describing. The source actually lies in the problems that the AI is taught to solve and the data it is provided to solve the problem.
In the case of medical image analysis, the problems are always very narrowly defined (e.g. segmenting the liver from an MRI image of scanner xyz made with protecol abc) and the training data is of very high quality. If the model will be used in the clinic, you also need to prove how well it works.
For modern AI chatbots the problem is: add one word to the end of the sentence starting with a system prompt, the data provided is whatever they could get on the internet, and the quality controle is: if it sounds good it is good.
Comparing the two problems it is easy to see why AI chatbots are prone to hallucination.
The actual power of the LLMs on the market is not as glorified google, but as foundational models that are used as pretraining for actual problems people want to solve.