pianopatrick

5 ms·

31.
by pianopatrick 2mo ago
Seems to me the answer is to use physical separation instead of virtual machines. Just get the AI a cheap laptop or phone with a cellular connection (so it's not on the same network as your other potentially vulnerable machines).
32.
by pianopatrick 2mo ago
I think the main problem with that is the main problem with a lot of security tools. In order to do useful work, you need to provide a lot of tools and permissions. I.e. in theory the most secure might be a virtual machine with no network a
33.
by pianopatrick 2mo ago
I can easily see a situation where most non American AI usage is on Chinese models on Chinese chips though.
34.
by pianopatrick 2mo ago
Suit yourself, I won't get in your way.
35.
by pianopatrick 2mo ago
I would plan based on the lost jobs not coming back. I would also plan that the trend of AI getting better and taking more jobs will continue. Plan for the worst, hope for the best. A lot of people may have to transition to careers that peo
36.
by pianopatrick 2mo ago
Yeah, I just wish seccomp worked per user. So you could define a policy of which users can do which syscalls, and then that follows them no matter which application they start.
37.
by pianopatrick 2mo ago
If the agents are really doing politics now, we can test and see which method of organizing agents politically works best.
38.
by pianopatrick 2mo ago
If we are treating ai agents like people, you could also just get the AI a laptop and apply the traditional tools to manage user laptops
39.
by pianopatrick 2mo ago
personally I wish the OS would allow syscall filtering per user
40.
by pianopatrick 2mo ago
In my personal project I've found having a hard lines of code limit has been a good limit on complexity. Makes me really ask if I need things. One theory I have floating around in my head is that if a whole code base was microservices
41.
by pianopatrick 2mo ago
I think right now both the engineers developing AI and the share holders are more focused on beating coding benchmarks and gaining revenue than anything to do with alignment.
42.
by pianopatrick 2mo ago
Are you sure that alignment work is aligning with the share holders and not the operators? Or not the creators? Or not the government? Which of these groups should the AI listen to when these groups disagree? If the AI gets as powerful as y
43.
by pianopatrick 2mo ago
It's not about the feasibility of the technology. If "human brains become fully irrelevant economically" then that brings into question the entire premise of "share holders" and "financial winners". What e
44.
by pianopatrick 2mo ago
Might also be per model. Different models might have different issues and require different instructions
45.
by pianopatrick 2mo ago
one theory I am thinking through is that people say AI is good at greenfield and has a harder time editing code in a legacy code base. So maybe you just treat projects as greenfield instead of editing. Like if you want to make changes to a
46.
by pianopatrick 2mo ago
or a CLI
47.
by pianopatrick 2mo ago
Maybe I missed it but what AI model did you use for this?
48.
by pianopatrick 2mo ago
It would be more about the release cadence and strategy. I.e. instead of deploying to production as fast as possible after merging new code, have discrete releases. Like maybe daily, weekly or monthly. And keep the last couple releases arou
49.
by pianopatrick 2mo ago
Yes, I agree half finished refactors are bad. But I think with AI the timing is a bit different now so the best plan changes. I.e. AI can finish various steps in generating the new version and transitioning to the new version much faster th
50.
by pianopatrick 2mo ago
Right but if you coordinate those changes in a new version you would still have the old version to fall back to if any of those changes lead to problems.
51.
by pianopatrick 2mo ago
Seems to me the real question is "what scheme can we use to organize our code base such that an agent working on one part really can make changes and not break the other parts on accident". My current theory I might test out is to
52.
by pianopatrick 2mo ago
I think the way I would want to work with AI for web apps would be like this: You install a fancy chrome extension or custom browser. You go through the app on this browser. You notice something you want to change. You can then submit a pro
53.
by pianopatrick 2mo ago
>That was absolutely free $0? Mental illness, drug addiction and homelessness all exist in a range. There are many people who are homeless, mentally ill or addicted to drugs who still have some amount of income of one type or the other.
54.
by pianopatrick 2mo ago
You could also have "zero homeless" if you had a certain amount of cheap housing that even drug addicts and mentally ill people could afford. Or where the cost of helping them is not too much. Like some housing that was built to b
55.
by pianopatrick 2mo ago
Would be funny / interesting / scary if it was another AI related incident like happened with hugging face. A whole lot of AI agents are presumably using Github as part of a workflow. Would be story worthy if one of them mis inter
56.
by pianopatrick 2mo ago
now that software is easy to make and github has problems, how many "git as a service" offerings will there be? My guess is there will be at least dozens if not hundreds.
57.
by pianopatrick 2mo ago
now that software is easy to make and github has problems, how many "git as a service" offerings will there be? My guess is there will be at least dozens if not hundreds.
58.
by pianopatrick 2mo ago
Wouldn't this imply that in terms of AI usage you should take a "Wait and see" approach? I.e. just wait until the models can easily do whatever it is you want?
59.
by pianopatrick 2mo ago
I dunno man, if there's a deluge of new AI generated code at all layers of the stack I think there will still be vulnerabilities. Like if we were willing to stop adding new code and just have a small secure code base, AI could maybe he
60.
by pianopatrick 2mo ago
One theory these days might be that old, boring technology is better represented in the training data for AI models...

More ›