Making an LLM that doesn’t hallucinate is probably literally impossible, and I’d be very surprised if Apple’s AI bullshit isn’t built on one.
You don’t need to wonder, Apple has said as much that their AI is built on LLMs, just like everybody else. While hallucinations are still a major unsolved problem, that doesn’t mean they aren’t able to be reduced in frequency and severity. A ChatGPT like chatbot is going to hallucinate because you’re asking it to give extremely open ended responses to literally any query. The more data you feed it in the prompt, and the more you constrain its output, the less likely it is to hallucinate. It’ll likely be extremely rare that using the grammar check or rephrasing tools in Apple AI will be affected by hallucinations for that reason. Siri is more comparable to ChatGPT with regards to open ended questions, but it’s likely that they will integrate LLMs primarily for transforming inputs and outputs rather than the whole process. For example, the LLM could be prompted to call a function based on the user’s query. Then, that function finds a reliable result, either using existing APIs for real time information like weather, or using another LLM with a search engine. The output from this truth-finding process is then fed back into an LLM to generate the final output. The role of the LLM is heavily constrained at every step of the way, which is known to minimize hallucinations.
You arguing that this is an unsolvable problem is defeatist and not helpful to actually mitigating the real issue.
