Raw LLM Responses

> Must self-improving General AI have access to its source code? If it does have access to its source code, can self-improving General AI really have effective safeguards and what would they be? If it has access to its source code, could it simply change any safeguards we have in place? Could it also change its goal? I think such an AI would be easier to write in a dedicated DSL (Domain Specific Language). The AI could modify all parts of its behavioural code; but is ultimately confined by the constraints of the DSL. You could in theory make an AI (let's assume C) that modified its own source and recompiled itself before transfeering execution to the new source. In this case it would be confined by the hardware the code was executed on - that is unless you assume that the AI can for example learn to pulse the voltages in a way to create a wifi signal to connect to the internet without a network card. Given an infinite amount of time; sure - that'll happen, but I don't think it is resonable to expect an AI to evolve to that stage in our lifetime (I imagine that would require another order of magnitude faster evolution).