Not sure of this at all but I’m under the impression commercial use tends to find loopholes for malicious use of AI systems that regulators/​safety teams are unable to predict.
Could we get a parallel with misaligned behaviour, where we start to notice misalignment with high volume of use in ways not rigorously testable or present under regulator testing? Would it be useful to consider a commercial incubation period with certain compute limits for example, or would we be able to definitely trust regulators?
Thanks for the comment! You’re right that this approach would need modification if ‘dangers that only become apparent after mass deployment’ becomes a major risk factor, and that a ‘trial’ commercialisation period could be a good response. My hope is that the regulatory exam period would be able to catch much more than at present though—the regulator would have ample time to design and deploy more sophisticated tests, with the aid of labs who would presumably love to submit a test their competitor would fail (so long as they themselves pass).
Not sure of this at all but I’m under the impression commercial use tends to find loopholes for malicious use of AI systems that regulators/​safety teams are unable to predict.
Could we get a parallel with misaligned behaviour, where we start to notice misalignment with high volume of use in ways not rigorously testable or present under regulator testing? Would it be useful to consider a commercial incubation period with certain compute limits for example, or would we be able to definitely trust regulators?
Thanks for the comment! You’re right that this approach would need modification if ‘dangers that only become apparent after mass deployment’ becomes a major risk factor, and that a ‘trial’ commercialisation period could be a good response. My hope is that the regulatory exam period would be able to catch much more than at present though—the regulator would have ample time to design and deploy more sophisticated tests, with the aid of labs who would presumably love to submit a test their competitor would fail (so long as they themselves pass).