Hey Oliver, i have been working a bit on these directions as well. I have been exploring creating an node based knowledge management system where everything is interconnected as a series of idea. Would love to share a video or coffee chat if possible !
I am not an experienced person, so please excuse my mistakes and shortcomings in my comment.
I think systems and architectures won't reach a point we can call truly safe anytime soon, so being able to prevent this upcoming danger doesn't seem very possible to me. Even the largest LLMs can somehow be manipulated to obtain dangerous information with simple direct injections that are very easy to do.
It is debatable how ethical and appropriate it is, but I think sufficiently large systems and architectures should be restricted in a way that not everyone can use them at full capacity, at least until they reach a point we can call completely safe. Not a ban covering the entire systems; but restricting their advanced and complex capabilities until the vulnerabilities are closed. This, of course, will not provide 100% security, but it will make a huge difference.
After all, the tasks the general user base does using artificial intelligence are not complex, therefore a partial restriction won't pose a problem—at least for the most part—for these people. The people and institutions that need to use artificial intelligence in more complex ways are clear; the entire systems should only be opened to people and institutions that can be supervised. Although there will of course be innocent people who are affected and don't deserve to be affected, this is not a manageable laziness like 'People aren't taking enough steps during the day, so we should ban cars.'; it is a systematic security risk.
I think you put it very clearly and effectively. For capabilities which pose substantial misuse risks in the wrong hands, it is irresponsible to allow unfettered access - which suggests a combination of monitored access, limited access, and even restrictions on development in principle. Of course, there are concerns here: keeping access to influential tech constrained to a few government or corporate parties could pave the way for abuse, and perhaps even the apparatus necessary to enforce such constraints could be abused. So it's not trivial.
I hope that on the whole we can make better decisions about such questions, both individually and as societies, if we use the AI building blocks we have to build tools to help us reason and coordinate better!
Hey Oliver, i have been working a bit on these directions as well. I have been exploring creating an node based knowledge management system where everything is interconnected as a series of idea. Would love to share a video or coffee chat if possible !
I am not an experienced person, so please excuse my mistakes and shortcomings in my comment.
I think systems and architectures won't reach a point we can call truly safe anytime soon, so being able to prevent this upcoming danger doesn't seem very possible to me. Even the largest LLMs can somehow be manipulated to obtain dangerous information with simple direct injections that are very easy to do.
It is debatable how ethical and appropriate it is, but I think sufficiently large systems and architectures should be restricted in a way that not everyone can use them at full capacity, at least until they reach a point we can call completely safe. Not a ban covering the entire systems; but restricting their advanced and complex capabilities until the vulnerabilities are closed. This, of course, will not provide 100% security, but it will make a huge difference.
After all, the tasks the general user base does using artificial intelligence are not complex, therefore a partial restriction won't pose a problem—at least for the most part—for these people. The people and institutions that need to use artificial intelligence in more complex ways are clear; the entire systems should only be opened to people and institutions that can be supervised. Although there will of course be innocent people who are affected and don't deserve to be affected, this is not a manageable laziness like 'People aren't taking enough steps during the day, so we should ban cars.'; it is a systematic security risk.
I think you put it very clearly and effectively. For capabilities which pose substantial misuse risks in the wrong hands, it is irresponsible to allow unfettered access - which suggests a combination of monitored access, limited access, and even restrictions on development in principle. Of course, there are concerns here: keeping access to influential tech constrained to a few government or corporate parties could pave the way for abuse, and perhaps even the apparatus necessary to enforce such constraints could be abused. So it's not trivial.
I hope that on the whole we can make better decisions about such questions, both individually and as societies, if we use the AI building blocks we have to build tools to help us reason and coordinate better!