“there is no significant chance of Friendly AI without this project”
Has to mean over time to make sense.
People who will be running DARPA, or Google Research, or some hedge fund’s AI research group in the future (and who will know about the potential risks or be able to easily learn if they find themselves making big progress) will get the chance to take safety measures. We have substantial uncertainty about how extensive those safety measures would need to be to work, how difficult they would be to create, and the relevant timelines.
Think about resource depletion or climate change: even if the issues are neglected today relative to an ideal level, as a problem becomes more imminent, with more powerful tools and information to deal with it, you can expect to see new mitigation efforts spring up (including efforts by existing organizations such as governments and corporations).
However, acting early can sometimes have benefits that outweigh the lack of info and resources available further in the future. For example, geoengineering technology can provide insurance against very surprisingly rapid global warming, and cheap plans that pay off big in the event of surprisingly easy AI design may likewise have high expected value. Or, if AI timescales are long, there may be slowly compounding investments, like lines of research or building background knowledge in elites, which benefit from time to grow. And to the extent these things are at least somewhat promising, there is substantial value of information to be had by investigating now (similar to increasing study of the climate to avoid nasty surprises).
People who will be running DARPA, or Google Research, or some hedge fund’s AI research group in the future (and who will know about the potential risks or be able to easily learn if they find themselves making big progress) will get the chance to take safety measures. We have substantial uncertainty about how extensive those safety measures would need to be to work, how difficult they would be to create, and the relevant timelines.
Think about resource depletion or climate change: even if the issues are neglected today relative to an ideal level, as a problem becomes more imminent, with more powerful tools and information to deal with it, you can expect to see new mitigation efforts spring up (including efforts by existing organizations such as governments and corporations).
However, acting early can sometimes have benefits that outweigh the lack of info and resources available further in the future. For example, geoengineering technology can provide insurance against very surprisingly rapid global warming, and cheap plans that pay off big in the event of surprisingly easy AI design may likewise have high expected value. Or, if AI timescales are long, there may be slowly compounding investments, like lines of research or building background knowledge in elites, which benefit from time to grow. And to the extent these things are at least somewhat promising, there is substantial value of information to be had by investigating now (similar to increasing study of the climate to avoid nasty surprises).