Oh boy. I will certainly be interested in reading these. AI is my main political issue.
Some obvious topics that come to mind:
Open vs closed-source. Does it really matter if the AI race is won by liberal countries if we release the models open source, leaving it easy for Iran and other countries to download them and release them?
The goal orthogonality problem: If we are making agentic AI, how do we teach it to follow any goal we have for it? How do we know?
The compositionality problem: Right now, LLMs like image generators struggle to generate exactly what they are told. I.e., ask one to generate a pink square on top of a green triangle, resting on a purple rectangle, and it will get it wrong most of the time. Why is this?
The rule-following problem: Why can no LLM play chess?
Oh boy. I will certainly be interested in reading these. AI is my main political issue.
Some obvious topics that come to mind:
Open vs closed-source. Does it really matter if the AI race is won by liberal countries if we release the models open source, leaving it easy for Iran and other countries to download them and release them?
The goal orthogonality problem: If we are making agentic AI, how do we teach it to follow any goal we have for it? How do we know?
The compositionality problem: Right now, LLMs like image generators struggle to generate exactly what they are told. I.e., ask one to generate a pink square on top of a green triangle, resting on a purple rectangle, and it will get it wrong most of the time. Why is this?
The rule-following problem: Why can no LLM play chess?