"it can fit" on 256GB of RAM, but it will be heavily quantized and still run very slowly. The headline number is not token generation, its prompt processing. So if you get 10 tok/s and an API gives you 20-30 tok/s, it doesn't seem that bad on its face, but a mac studio or any other machine that's not loading all of it into GPU will do PP 20-50X slower than a purely GPU based setup, which is what actually makes this unusable without $50k in GPUs.
On top of that, you will still be heavily quantized.
Maybe some (or many) people believe that more people will make it less "lovely". I think this is a popular stance and I think many people are more than satisfied with the current population density of their area.
I guess I was speaking as an American, we have good domestically hosted options so although it’s probably not ideal to send this kind of data/control anywhere at all, it’s definitely a worse option for us to send it to china vs to an American company. Every user of this service has made their machines trivially exposed to become a botnet. Im wondering why I don’t see this angle more discussed in here.
Again I’m not saying you should trust an American company necessarily more than a Chinese one, but as an American, I probably can.
I’m worried about giving a foreign hosted service access to my machine for a coding agent that can run arbitrary commands and read arbitrary files. Coding agent are much more useless if you have to sit there clicking approve on everything.
Using an Epyc platform to get plenty of PCIe lanes and memory channels. I have couple of extra 3090s plugged in which get some offload and help with larger models that don't fit entirely on the blackwell.
Haven't they already proven to be extremely useful? In some areas they are definitely here to stay, coding/software and search (retrieve and summarize information). There's a bunch of places where they are surely shoehorned in, overhyped, and don't belong, but there's also equally many places where they might still be transformative but aren't used yet.
But overall I think the technology is well proven.
I got an RTX 6000 pro too. I like running locally, I've learned a lot more than if I had used an API and there's less worry about overspending tokens. I accidentally spent $100 on claude api in like 2 days because I didn't know what I was doing.
The problem is that while one these gpus is a huge improvement over a laptop or a single 3090, you very quickly wish you had more. I would buy a second one, but I did the math and realized that with the current crop of models, 2 Blackwells doesn't buy me any new capability that I didn't have with one. So I would need a 3rd one. And when I buy a 3rd one I will feel like I want to running a higher quant, so then I will want a 4th.
At this point IPOs are mainly for unloading bags onto retail. Every institution who wanted a piece of these labs got in years ago and captured all the value.
I use claude/gemini as my homepage now (I have to keep switching as these companies make "updates" that periodically render their models useless). Even if I want to search for simple things, I would rather have an LLM wade through the result and extract just the information I asked for. SEO, and now mountains of slop content have made this necessary. Only a matter of time before the SEO industry in large figures out how to game LLMs too, making them equally useless.
I already saw a article recently about how to set up a business domain which can reliably show up in a search result and dump overly positive reviews into anyone's context.
They won't, its literally part of their sales funnel. They've specifically engineered a bad experience for anyone outside the ecosystem by making it all of their friend's problem too. Its very important for their stock price that text messages sent by non apple products are just slightly more difficult to read.
Is this LLM psychosis? So much tending and conversing with the matmuls but what was the outcome? Are people who get this into it more successful somehow? It reminds me of people who take drugs and get "revelations" but then are not particularly over represented in the group of successful people for all of their deep insights.
I've been using Siri (via homekit) to turn all my lights on and off for about 3 years now. It's steadily getting worse and worse as somehow, Siri is becoming less accurate and Apple is failing to adopt this new technology in a timely fashion.
I would like to tell it to turn off certain light in a certain room, but unless I get the exact string name of those light correct when I speak, Siri doesn't know what I'm talking about. And it can't do multiple things in a command. I can't say "turn off all the lights in XYZ room" or turn of "this light and this light".
Meanwhile, I can vaguely tell a computer behind my tv to do very complicated things (build me an service that ...) and it can execute on it fairly well. But in apple's "product vision" which I am apparently too dumb to decide for myself what I want, I can't ask for two lights to be turned off at the same time.
On top of that, you will still be heavily quantized.