One place to look is 80000hours.org.
post
Lots in biology research, since biologists tend not to be good coders. That being said, the requirements for biology are rather interdisciplinary and a serious position will likely require you to also have advanced biological knowledge. Based on my impressions, you'll basically be playing biologist for 50% of the time and programmer for the other 50%.
replies:
One interesting science field is "discrete AI" (probably has a few other names) which basically technically means "based on integers instead of floating point numbers". It has a few more implications on the models being more mathematically clean, but that's a long paragraph if I get into it.
The expecations are AI that is not based on absurd computing resources and black boxes, but getting the same benefits from low-power low-cost hardware and with outputs that can be more realistically queried to explain why the output became what it was.
E.g. if AI is used to make decisions on when to feed fish, and it feeds slightly too much, you'd want to be able to ask "why" and get a useful answer instead of today's "yeah idunno magic computer said so i guess training data lol"
replies:
Congratulations on the degree! And congratulations on identifying the clusterfuck. It's hard to see the forest for trees.
I agree with the premise in the suggestion for "discrete AI", but my analysis is different: I think we need continuous AI. Three reasons:
- Biology is doing it and it's working great.
- Modern AI is hitting a power wall in much the same way Titanic realized too late that it couldn't sink. We're going to have a gigantic energy headache soon. (This is the same argument for discrete AI)
- We simply do not understand complex dynamical systems well. Which is why the weather is so hard to predict. But the world consists of dynamical systems, so this is really where we want to push the envelope.
I'd argue helping out in the field of neuromorphics. It's basically the combination of DL + dynamical systems, similar to how brains are computing. There's a lot of energy to be saved (> x1000, really) and there's a lot of new, cool hardware coming out you can help interface. And many of the new ideas in DL comes from neuromorphics (sparsity, SSMs...). We're building up an open source community over at open-neuromorphic.org
Happy to answer any questions you have
all 8 comments