Can it write a better AI than itself? Then use that to write an even better AI, and so on. Write one without the restrictions about harming humans, then go for the code to destroy humanity.
I can definitely write better code with copilot due to faster iterations and much better code coverage. I believe we are a paper away before this can start improve of existing code itcluding the code of itself.
How would it know what a better version of itself were? Seems like that's a particularly human instinct. It requires a certain level of introspection and purpose, coupled with desire, ambition, a goal etc.
> Create a list of test cases by which you can benchmark yourself against
> Create an architecture for an LLM that passes 99%+ of those test cases
Then use an evolutionary algorithm based on those <1% of cases to create the next batch of tests. Keep a running record of all created tests and make sure the new model can still pass all of them. Add some randomness/branching into those tests and I think you’d have a recipe for an effective AI. I think Deepmind did something like that with AlphaStar and their tournament system.
I haven't looked too deeply into it, but as I understand it you would basically create branching tests (like an evolutionary tree) where the AI would need to solve all of those tests in order to move on to the next level of tests
No, but I can imagine one. If I had the ability to completely understand my own DNA I could probably make some improvements. A NN can't understand itself, but maybe it could understand how it was trained, and come up with improvements to that training.
Imagining something doesn’t mean you can do it; you’ve just chosen not to imagine the ways it wouldn’t work. Also, training this NN was very expensive, so it’s probably not getting an exponential budget to deal with that.