John Schulman says the field depends heavily on generalization and that it is difficult to predict when generalization, including out-of-distribution generalization, will occur.
支持这项说法
原始摘录
The whole field relies a lot on generalization and it’s very hard to predict when you’re going to get generalization, or when you’re going to get some kind of out-of-distribution generalization
Research methods may have substantial room to improve
John Schulman suggests that clever small-scale experiments could support theories that generalize to larger experiments. He expects research methods to be far from their ceiling.
支持这项说法
原始摘录
If you think hard enough, you probably could have expected some of these things beforehand. There is probably some very clever way to do a small-scale experiment that’ll let you build the theory that then will generalize to the large-scale experiment. So I would expect that we’re nowhere near the ceiling of how well you can do research
Distillation depends on a realistic prompt distribution
John Schulman says a realistic prompt distribution can help a distilled model match a larger model. He contrasts this with easily verifiable tasks, which may produce strong benchmark results but weaker performance across a broader distribution.
支持这项说法
原始摘录
If you have a really good realistic prompt distribution for distillation, you can match the big model really well. But if you only have this distribution of easily verifiable tasks, then you can match the big model on all the benchmarks, but you do worse on this broader distribution
nafyi.