We have been using GPT-4-Turbo/GPT4-Vision API to develop agent handling all kinds of different actions. One bug we recently noticed is if you send empty input (no user message), LLM will generate random stuff based on the system prompt or message memory, and execute different function calls to compete the random tasks. We even noticed LLM start to talk with itself and decide on which actions to take.
We feel this could be danguage for AI safty, when we provides more and more tools for LLM. It can generate catastrophic results due to the randomness of a text completion capability. Just think about what an AI agent can do in a militory circumstances…
We need to fix this bug ASAP. The quick solution is just prevent API call with empty user input, and continue working on AI alignment as Illya is focusing on. We hope as an AI practitioner, we should all consider and work on AI saftey in all application areas.
BTW, love the GPT models, it will change this world for sure! Peace!