Um, I’m working on a message to someone who is smart but knows nothing about computer science. It includes a brief description of AI in regards to safety concerns, but I’m not an AI researcher, so I might not have done a good job. Can someone competent enough to judge this tell me whether this is misleading in an important way? (It might not be technically true to suggest that an AI doesn’t just react to inputs, but I’m fine with technical inaccuracies, so long as they are minor.)
The real problem isn’t easy to understand, but I’ll try to summarize it as briefly as possible. Basically, an AI can be thought of as an agent that
– Learns and models things about its environment
– Has a utility function that specifies which goals are desirable or undesirable (anything else is neutral)
– Based on the above, runs a search for policy options (i.e. possible actions) that will maximize its utility function
The principal difference to a pocket calculator is that there a pocket calculator has explicitly programmed responses to every input. This doesn’t work for an AI because there are too many possible scenarios. We can’t address each one individually, so instead we only tell it what the goals are and let it figure out what to do to meet them.
Um, I’m working on a message to someone who is smart but knows nothing about computer science. It includes a brief description of AI in regards to safety concerns, but I’m not an AI researcher, so I might not have done a good job. Can someone competent enough to judge this tell me whether this is misleading in an important way? (It might not be technically true to suggest that an AI doesn’t just react to inputs, but I’m fine with technical inaccuracies, so long as they are minor.)
The real problem isn’t easy to understand, but I’ll try to summarize it as briefly as possible. Basically, an AI can be thought of as an agent that
– Learns and models things about its environment
– Has a utility function that specifies which goals are desirable or undesirable (anything else is neutral)
– Based on the above, runs a search for policy options (i.e. possible actions) that will maximize its utility function
The principal difference to a pocket calculator is that there a pocket calculator has explicitly programmed responses to every input. This doesn’t work for an AI because there are too many possible scenarios. We can’t address each one individually, so instead we only tell it what the goals are and let it figure out what to do to meet them.