Let’s say an AI has a remote sensor that measures a value function until the year 2100 and it’s RLed to optimize this value function over time. We can make this remote sensor easily hackable to get maximum value at 2100. If it understands human values, then it won’t try to hack its sensors. If it doesn’t we sort of have a trap for it that represents an easily achievable infinite peak.
Reinforcement learning doesn’t guarantee anything about how a system generalizes out of distribution. There are plenty of other things that the system can generalize to that are neither the physical sensor output nor human values. Separately from this, there is no necessary connection between understanding human values and acting in accordance with human values. So there are still plenty of failure modes.
Yes nothing is a guarantee in probabilities but can’t we just make it very easy for it to perfectly achieve its objective if it doesn’t go exactly the way we want it to, we just make an easier solution exist than disempowering us or wiping us out.
I guess in the long run we still select for models that ultimately don’t wirehead. But this might eliminate a lot of obviously wrong alignment failures we miss.
Why wouldn’t a wire head trap work?
Let’s say an AI has a remote sensor that measures a value function until the year 2100 and it’s RLed to optimize this value function over time. We can make this remote sensor easily hackable to get maximum value at 2100. If it understands human values, then it won’t try to hack its sensors. If it doesn’t we sort of have a trap for it that represents an easily achievable infinite peak.
Reinforcement learning doesn’t guarantee anything about how a system generalizes out of distribution. There are plenty of other things that the system can generalize to that are neither the physical sensor output nor human values. Separately from this, there is no necessary connection between understanding human values and acting in accordance with human values. So there are still plenty of failure modes.
Yes nothing is a guarantee in probabilities but can’t we just make it very easy for it to perfectly achieve its objective if it doesn’t go exactly the way we want it to, we just make an easier solution exist than disempowering us or wiping us out.
I guess in the long run we still select for models that ultimately don’t wirehead. But this might eliminate a lot of obviously wrong alignment failures we miss.