scallisto

VR headsets effectively need to run two 4K displays, one for each eye, and you have all the processing for controller tracking too. You also have eye tracking, they're doing the whole foveated streaming thing everybody has been gushing over about (it's really cool tech!). The mobile chip they used is nothing special, it's a Snapdragon 8 Gen 3, a 3 year old smartphone processor (certainly powerful enough to run lighter PC games, but nothing more than that. It's kind of insane that a phone can even do that nowadays!)

To separate the weeds from the worms, I would avoid the term "AI" since it has devolved into a buzzword that's stuck on everything.

I think that if somebody has made a useful open-source project using LLMs, they shouldn't be criticised solely for that reason. If the code produced is ass and isn't reviewed, is hard to contribute to, and there is negligence, that would be a valid criticism. I also despise the use of the privacy nightmare of commercial LLMs (OpenAI, Anthropic, Google, etc.). If are to use LLMs, please use a local model. For you, your project, and your wallet's sake. That's my position on LLMs.

I trust that you are doing your due diligence in keeping your code relatively clean, and hope that your project will succeed! I personally do not use LLMs for my projects, they do not fit my usecase, probably not for most people actually. But I can see the value of having an LLM do menial tasks, as long as the code is properly reviewed and checked.

Generative models belong in the deepest circle in hell. They steal the work of creatives, are awful in the trifecta of morality, ethics, and legality, and are mainly used by scam artists (see: deepfakes) and misinformation campaigns. Nobody should use these, even if it's a local model.

Also, a bit off-topic, but you might have heard of all the "AI has solved (big problem in science)" or "AI proved (big maths thing)" headlines, but note that these are NOT your LLMs, and certainly not any kind of generative model. They use predictive models (sometimes the papers mention "deep learning" too) that are trained on a specific set of data to do a specific task, and most of the actual science and maths is done by humans. These models certainly won't replace scientists or mathematicians, like the Big Tech companies want you to believe, but they are a valuable tool when it comes to doing menial tasks repeatedly. That's what computers are best at, after all!

I feel like we should better differentiate between LLMs (the ones that literally tell you to eat rocks), generative models (the ones that are stealing the work of creatives without permission), and the ones that are actually used for neat science stuff, where they are trained on a specific set of data and are doing a specific set of tasks repeatedly in a controlled environment, rather than training on the entire internet and then some and trying to do everything!

I think we should label them as large language models, generative models, and predictive models (maybe? Or maybe "targeted predictive models"?). It's important to emphasise that all of these, even the useful ones, are not intelligent, they do not understand what they are doing. They are just models of data that do thing go in, thing go out.

Well those are actual languages spoken by hundreds of thousands of people (from what I can tell from the Wikipedia article), they certainly aren't designed to be difficult to learn. Perhaps it's hard to learn for people who speak English, German, Chinese, Arabic, Portuguese, etc., but the goal wasn't to construct a language that is hard to understand.

So I'm doubly wrong! That's kind of neat

That's actually pretty cool, I'll be making an XMPP account then. Adding to my collection of chat apps!

Oh whoops, I didn't know that! That's kind of neat, didn't know had encryption. I will amend my comment

Also, please don't use LLMs for simple questions. Aside from the ethical, legal, and environmental concerns of LLMs, they also tend to be wrong more often than not, and when they are right, there is lots of word soup because the chatbot needs to act like it's a smartass, to put it bluntly, because its job is to market itself and push you towards subscriptions. Try to avoid LLMs wherever you can, it's almost always better to search it yourself. Don't offload thinking to a chatbot, it's not very good at it!

The only usecase where LLMs could have an application is translation. I have a friend who doesn't understand English very well, so they use a chatbot to translate things into their native language, and it's really helpful for that. Of course, it can get stuff wrong, but for non-critical stuff, it's kind of neat! That avoids the question of the insane water and power usage involved in training these models, as well as the disregard of the rights of artists and creators by companies like OpenAI, Anthropic, and Google. I believe my friend uses a local model, so there's that.

I personally don't use LLMs, they're more trouble than anything in my opinion.

Have you ever had any experience with Scratch? It's an awesome tool to learn basic programming concepts like loops, conditions, etc. I played around with it in my early years, but even if you're older, it's still fun to mess around with the blocks to make weird programs. I particularly loved the "remix" functionality, it lets you see the structure of other people's projects and you can tweak them (and thus learn about how it works!)

After my years of doing Scratch, I moved on to Python, and it is probably the easiest programming language to learn (depending on how you define that, e.g. HTML), code is very human readable and unfamiliar symbols are avoided for the most part

W3Schools has a lot of good Python guides, and there are SO MANY good programming tutorials online. Back in my day, I learnt Python by reading a good book (and it was an awesome book), but resources to learn programming have never been more plentiful!

My advice is to work on a simple project, could be a small game (with or without PyGame), or maybe a useful tool that you want to make, lay out what you want it to do, and along the way when you don't know something, search it up online (W3Schools, forums, video guides, docs, whatever) and figure out how to do it. Practice doesn't make you perfect, but more practice does. Does that make sense? Not really, but definitely try your best!

Well, there's also IRC and XMPP. Could those work better for older devices? ~~Dunno, but neither are encrypted, so don't use them for anything sensitive.~~ Signal is good if you want security, they've been audited a lot, but they are based in the U.S. which isn't great.

I've also heard of Stoat (formerly Revolt) and Fluxer, both are relatively new, but I would prefer Matrix over both since it's more easily self-hostable (+ all those different clients to choose from!)

edit: I really didn't know, but looks like XMPP does support encryption! (and it's enabled by default on most clients!)

I have tried it before, it's neat, but notifications are not reliable since it is peer to peer (both devices need to be online for a connection), a bit like Briar if you've used that before. I don't think it uses the Tor network, so it might be faster than Briar (but also less anonymous?)

I don't have experience with the jami daemon though, so can't comment if that solves the peer to peer issue.

I personally like Matrix the best, since you have federation, options for encrypted chats, self-host able with Continuwuity (or Synapse), lots of alternative clients for nearly every OS (even mobile Linux, you have Fractal! FluffyChat also works but it's definitely oriented towards Android devices)

The longer you look at the pineapple, the more your food tastes like pineapples. There comes a point when your food is too sweet, too acidic, but you can't stop looking at the beautiful pineapple.

Alternatively, it makes everyone's piss a deep yellow. Saturated and not healthy at all. Weeing is painful and feels like a pineapple getting squirted out of whatever hole you wee out of. Everybody avoids drinking water and only eat dried food, and they all die from dehydration within 3 days.