3 ms·

I was the kid that helped other kids bypass their nannyware so I’m not the best person to ask but how are you going to handle prompt injections? If I’m underst

I was the kid that helped other kids bypass their nannyware so I’m not the best person to ask but how are you going to handle prompt injections? If I’m understanding, you use a LLM to decide if §content is in violation of §policy? This will break anything that does certificate pinning like Spotify. As crapware tls proxies become more common, cert pinning and data encrypted before it’s sent will become more common. This won’t work on fb and their related properties or iMessage where I assume kids spend a lot of time now? No idea about TikTok but if they’re not cert pinning yet… they should be!

guywithahat 2y ago
We don't use an LLM in real-time production yet (although what you described is coming). At the moment I used one to help classify different communities/etc into filters, and we use ImageNetV2 for image classifying. That said, the way it's set up now the prompt is specific to each parent, and so if a parent breaks their prompt, it will only break their kids internet. If anything, I'm more worried that parents will accidentally break the prompt and there isn't currently a good way for parents to debug it or even know until their kid complains to them. The other big concern with our LLM (at least in my opinion) is that it's not fast enough. Decrypting each post, sending to our LLM with a gRPC call, and then getting the response back just takes too long. It's certainly going to be possible to do, but I need to launch and start getting customers before I can properly tackle that problem. Also we add a root certificate to the host machine, which is how we get around HSTS. It's also why we don't have a phone client yet, as it's much harder to add root certificates to an iPhone than it is to add them to a windows or linux machine. I will say though I also got around three different child monitoring systems when I was a kid lol. Unfortunately the prize was a porn addiction.

guywithahat 2y ago

We don't use an LLM in real-time production yet (although what you described is coming). At the moment I used one to help classify different communities/etc into filters, and we use ImageNetV2 for image classifying. That said, the way it's set up now the prompt is specific to each parent, and so if a parent breaks their prompt, it will only break their kids internet. If anything, I'm more worried that parents will accidentally break the prompt and there isn't currently a good way for parents to debug it or even know until their kid complains to them. The other big concern with our LLM (at least in my opinion) is that it's not fast enough. Decrypting each post, sending to our LLM with a gRPC call, and then getting the response back just takes too long. It's certainly going to be possible to do, but I need to launch and start getting customers before I can properly tackle that problem. Also we add a root certificate to the host machine, which is how we get around HSTS. It's also why we don't have a phone client yet, as it's much harder to add root certificates to an iPhone than it is to add them to a windows or linux machine. I will say though I also got around three different child monitoring systems when I was a kid lol. Unfortunately the prize was a porn addiction.