SERI 2022: Timelines for Transformative AI and Language Model Alignment | Ajeya Cotra
EA Talks
00:00
Using Ai Systems to Prevent Superintelligent Style Scenarios
In the future, we're going to have systems that are smarter than humans. That brings particular challenges in controlling them and making sure they don't do something human doesn't want. The challenge of trying to prevent this future outcome is that systems now aren't powerful enough to take really scary, open ended actions for their goals. Ai systems might know more physics than a certain group of humans, or something. It's sort of not obvious how to get those humans to give use feedback to these models without insentivizing the models to just deceive the humans. I onyu say, useful? You mean like a insems, like usefulness for lineand research? Or usefulness for for
Transcript
Play full episode
