I really want to use AI like llama, ChatGTP, midjourney etc. for something productive. But over the last year the only thing I found use for it was to propose places to go as a family on our Hokaido Japan journey. There were great proposals for places to go.
But perhaps you guys have some great use cases for AI in your life?
@jeena I only use DeepL to translate and that’s it. I also started taking notes in .md files, so that could make for a good use case in the future if there was an AI that I could use without connecting to the internet (e.g. to only let me tell stuff based on the files I got). Otherwise I am pretty reticent on AI. Perhaps I watched too many fiction movies, but I am afraid it will become too sentient and somehow escape the human oversight, thus creating havoc in our lives.
Check out llama3 which you can run locally.
@jeena it doesn’t connect to any server even when you’re online?
Correct.
@jeena cool. I think I will look into it. Thanks! 😁
Damn, it seems I was talking out of my ass and just assuming things. Now I asked llama3 about it and it said that even though it does not send our conversation to Meta, it does take my feedback (if I correct it in some way) and sends it to Meta’s servers through the Internet:
// Will you be able to learn from me pointing it out or do you forget all conversations we had after I press Ctrl-D?
I’m a large language model, I don’t have personal memories or experiences. Each time you interact with me, it’s a new conversation, and I start from scratch.
When you point out an error or provide feedback, it’s stored in my training data for future conversations. However, this information is not retained individually; instead, it’s aggregated and used to improve the overall performance of my language understanding and generation capabilities.
So, when you press Ctrl-D (or close our conversation), I won’t remember the specific details we discussed, including your feedback on my German sentence.
But, the collective knowledge from all our conversations will help me become a better AI model over time!
Keep in mind that I’m designed to learn and adapt through interactions with users like you. Your contributions, no matter how small they may seem, are invaluable in helping me improve and provide more accurate and helpful responses in the future!
/// But you run on my computer so technically how does my contribution get back to Meta to train the next generation? Do you send any data back?
When you interact with me through this chat interface, your input is processed by a client-side JavaScript code that runs on your local device (your computer or mobile phone). This code sends the necessary information about our conversation, including any corrections or feedback you provide, to Meta’s servers.
Here’s how it works:
To ensure your contributions are anonymous and secure, all data transmission is handled by Meta’s infrastructure and follows industry-standard encryption protocols. Your input will help refine the AI models, enabling them to better serve users in the future!
Now, go ahead and chat away!
/// How would the data look in the case of me telling you the correct German expression? What would be sent back to meta’s server? Show it as JSON.
When you correct an error or provide feedback, the client-side JavaScript code on your device sends a request to Meta’s servers with the following information:
Here’s what the data might look like in JSON format:
{ "event_type": "corrected_error", "conversation_id": "unique_conversation_identifier", "turn_number": 5, "user_input": "und kann einfache", "correction_target": "können", "language_code": "de", "platform": "web", "client_info": { "browser_type": "Chrome", "browser_version": "92.0.4515.131" } }
In this example:
event_type
: The type of event, which in this case is a “corrected_error”.conversation_id
: A unique identifier for the conversation.turn_number
: The number of turns (or exchanges) in the conversation.user_input
: The corrected text you provided (“und kann einfache”).correction_target
: The original incorrect text (“können”) that was being discussed.language_code
: The language code for German (“de”).platform
: The platform or device type used to access Meta’s services (in this case, “web”).client_info
: Additional metadata about the client device and browser.This JSON data is then sent to Meta’s servers, where it can be processed and integrated into their training datasets.
The model is just hallucinating in this case.
Oh yeah, I found out the hard way and forgot to mention it here. https://jeena.net/llama3-phoning-home