I think the risk of your research being abused is very unlikely. You could do a lot of good in robotics.
Hi Commnder Zander, thanks so much for the reply. Could you elaborate on why you believe that my research is unlikely to be abused? It just seems to me that there's a high chance that architectures developed for continual learning in robots could strongly transfer to general AGI systems (even if the AGI systems are non-embodied).
^ This is a question I've tried researching online, but haven't found much discussion of it: was hoping to start some here.
For context, I am a final-year math + CS undergraduate considering pursuing a career in theoretical robotics, particularly in continual learning and the development of robots that can learn from and adapt to / navigate their environments in a human-like manner. One concern I have, however, is that such research might advance AGI timelines. Specficially, it seems possible that architectures developed for continual learning in robots could transfer to general AGI systems (even if the AGI systems are non-embodied, since capabilities such as continual adaptation and long-term objective pursuit may generalize beyond physical tasks.)
Is this a valid concern, and is it a common view within the AI safety community? I.e. would mainstream AI safety researchers view either of these directions as meaningfully contributing to AGI capabilities? Or are there strong reasons to believe that work on continual learning in robotics would not significantly accelerate AGI timelines? Would appreciate honest perspectives.
TLDR: Is it very likely that robotics capabilities research meaningfully accelerates AGI timelines? If so, why?
(For reference, I have read quite a bit of the AI safety literature, but don’t find alignment research particularly enjoyable. Hence, my [perhaps futile] hope that robotics does not meaningfully advance AGI. If people think such work significantly accelerates AGI capabilities, though, I’m happy to steer clear…)