|
I'm the CTO of Ameelio, a non-profit tech startup that is taking on the very broken for-profit model for incarcerated communications. Incarcerated people are routinely price gouged for call time, and the current model eliminated any competition and allows the sole provider to charge whatever they want. It's not unusual to see someone pay a 25 cent connection fee and then 15 cents a minute after that. We have built basically a Zoom or Google Meet for the Corrections space, and we give it freely to inmates and their family members. It's only a matter of time before the for-profit providers are forced to compete in a free(er) market or go out of business, and either way we'll be growing and expanding our service as much as possible. We are using Llama for some small things now, but have plans to increase it in our stack, particularly as it's capabilities increase. Here are some of the things we're doing: 1. A better "keyword" analysis: Many facilities use and want the ability to check for certain types of comms. For example, if someone is talking about committing suicide or other self-harm, or is discussing harming another person, some intervention could literally save their life. Historically keyword analysis was about the best that could be done, but thanks to Llama we're able to get a deeper insight into it. This is better for accuracy to reduce false positives, and it's also better for privacy because the improved accuracy results in less need for intrusion. 2. Privacy Preserving analysis: The incarcerated space comes with constant invasions of a person's privacy, and while that is not going away, things can be improved. At this point I've met a lot of DoC people, and the majority of them have no interest in invading a person's privacy. Their concerns are about security and safety. This is an area where AI can be of great assistance, but only if the AI itself is privacy respecting. This is where Llama shines very bright for us. Because we can self-host it, we can be very confident about where the data is going and what it is used for. 3. Using it for organizational productivity: We have OpenWeb UI and Ollama deployed internally for our team, and they are able to use Llama and other models to accomplish things, whether it be asking ChatGPT style questions, getting coding help, or anything else. The ability to run different models with different strengths depending on the application is huge. 4. Using it for document analysis: For example, we frequently deal with massive documents containing regulations (such as FCC documents), contracts, and other long but sensitive things that Llama can greatly assist us with understanding. A great RAG setup is very helpful (though still somewhat challenging. That could be an area where Meta could help!) I know this can kick up a firestorm, but less "safe" models would be very helpful. Some of the things we need to understand better (like suicide in the example above) frequently trigger "safety" things which ironically make everybody a lot less safe IRL. We've gotten better at working around these, but having to workaround it in the first place feels immensely silly at best and deeply frustrating at worst. I understand the significant challenge you're facing here, and I know it's very difficult (impossible IMHO) to find a perfect balance, but at least in my experience we're quite too far on the "safe" side of things and it would great to move more toward a reasonable center. For what it's worth, Llama is in good company. Compared to OpenAI Llama's safety causes more friction, but compared to Gemini's ridiculous over-safety, Llama is a breath of fresh air :-) In closing, thank you so much for Llama and for all the work you've done and continue to do! I believe open source is extremely important, and without Meta's generosity, I believe we would only have toys to play with. Meta's overall commitment to open source and understanding and support of open systems is becoming increasingly vital in my opinion. Being frank, for many years there I had some pretty negative feelings about Meta/Facebook. However, that has really turned around and I am very grateful that Meta is around! |