I remember when “data” was something useful, not yottabytes of fake mountain goat jumping videos and fart sounds.
All that shit is for training. They just want you to think it’s useless slop videos. The amount of data you can get on human psychology through those memes and videos is priceless. Y’all get caught up in the first layer of the onion/matrix every single time.
What, precisely, do you think we’re going to learn? You know, that we haven’t already mastered since planes flew with cloth wings?
Maybe optimize the LLM to run on less compute?
Or just realize that nobody fucking likes LLMs as much as the Captains of Industry want us to believe, and that the true power of this technical domain lies in more targeted and bespoke ML model generation and usage.
ML is good and enables - and has enabled - some genuine generational leaps in science and technology. But LLMs are such a fucking waste of the technology’s potential. Not to mention, I’m extremely irritated that (largely due to Nvidia cornering the market) everyone is super gung-ho about a digital approach which amounts to brute-forcing neural nets digitally with shitloads of memory and highly-parallel compute, when it’s obvious to anyone with more than a passing familiarity with electrical engineering that an analog approach is going to be FAR more efficient in terms of resource and energy usage.
Yeah, for 4 and 8 bit quantization at least analog charge buckets or memristor-likes and analog multipliers would dramatically reduce substrate size, reduce power burn and speed up inference. Even better killer drones, here they come. Yay.
I mean… yeah, DARPA will probably be one of the first adopters of that stuff, it’s true. But DARPA is pretty much always a first adopter of any new tech, because they’re basically the research wing of the US military, and they have effectively infinite resources at their disposal (note: I am not debating whether or not that is a good thing here; simply stating that it is a thing). But just because they’ll likely do something military-ish with it first doesn’t mean that it’s a “bad” technology. The internet itself was, after all, initially a project of DARPA’s predecessor, ARPA, and was initially named “ARPAnet”.
Or just realize that nobody fucking likes LLMs as much as the Captains of Industry want us to believe
Funny, but it’s actually a bit worse. People are literally falling in love with it and generally go nuts. This is not even getting into the whole “it’s conscious”. You’re essentially living in a bubble and don’t really know how bad it is. Don’t look up how many use character.ai
EDIT: On a related note, it’s slightly annoying that Lemmy converts links like.it, without me providing https://
maybe this is personal preference
?
Analog meaning think for yourself or what?
Analog as in analog signals vs digital signals. Reductively: circuitry designed to operate in a continuously variable electrical domain, as opposed to circuitry designed to operate in a clocked binary domain.
Analog is (with a LOT of handwaving) way closer to how biological neural nets (that is, actual neurological tissue) operates. This is one of many domains where the exploration of biomimicry could yield some incredible advantages in a lot of areas.
Stop talking sense when has good sense ever been the solution of choice.
But everything’s computer!
If I was king they would have to run on 100% on-site solar and air cooling.
would be based
I love how Zuck always gets the short end of the stick.
Easy Mode Pop Up Copper Mine
Just think of the CoSt SaViNgS when 3 years from now and all that shiny new hardware is obsolete compared with the nvidia Orphan Crusher 3000, and instead of paying for a silly decommissioning and responsible recycling project you can just cover your eyes for 30 seconds and let the meth heads scour the site.
Homeless Data Shelters ? Who would have thought.
Is it illegal to buy handkerchiefs, bottles of vodka, and a lighter?
Not separately
Fill up a jerry can or two with petrol next time you fill up your car, and save the vodka for making martinis.
I’m not one to advocate for anything of course but I think it’s pretty reasonable to have at least handkerchiefs and a few lighters at home always. Buy the vodka on demand. Of course not a bad idea to stock up on vodka either, never know when you want to get drunk or disinfect a wound.
Might be a hot take, but vodka is one of the more desperate choices when looking to get drunk on your own time (at least in terms of taste). Rubbing alcohol for wounds is also far cheaper.
I wish I didn’t struggle to get down a whole bottle of rubbing alcohol. It really is much cheaper, but it tastes awful.
Reading that China is quite advanced in AI also, and if these things are related, how are they handling things like data centers over there? Or are they not related?
Lots of statesponsored datacenters being build in China. And of course with no way to say no to those.
And a majority of domestically produced Chinese RAM-chips are being funneled into Chinese datacenters (by government mandate).
Theyre not desperate. Sure they want to build quickly but this is also a capital hedge against the bubble bursting.
Far more a calculated move than a desperate one.
Prometheus is the equivalent of a chicken tent. Wow. Isn’t this what Pablo Escobar did in the early days of the drug trade? Is there some association there to modern times?
I wouldn’t use the word “desperate.”
Scaling is inefficient.
For training, it takes a ton of work to even get half-decent utilization across a bunch of servers, and it makes any sort of experimentation with architectures immensely more difficult.
Hence allegations that some GPUs are assigned “busywork” just to meet utilization quotas from the hardware seller.
For inference, scale isn’t so important. But the demand for tokens is self inflicted: from Meta shoving chatbots in ramdom places in software, and from their architecture being archaic and inefficient.
In other words, none of this has to be. It’s just the whims of one insecure man, surrounded by sycophantic tech bros, who’s feeling FOMO but doesn’t understand transformers LLMs at all.
If he had half a brain, he wouldn’t have fired the team that literally founded the open weights LLM space.
But he’s also too rich to ever feel the consequences of bad decisions now.
One of the reasons Huawei is far ahead of Nvidia on cost competitiveness of cluster systems is both their last (384 matrix) and current (950 based) clusters are containerized and don’t need buildings. Nvidia B200 and later systems require liquid cooling with heavy duty single story floors purposed built with embedded liquid cooling pipes in them. This is usually a $10b premium per gw of compute in building, not including the ability to reuse any other building/datacenter with older nvidia systems.
I keep hearing the AI will get cheaper to run because its only training that’s expensive, but at the same time tokens cost more than employees and are still being sold at a 10 fold loss, and we apparently still need infinitely more compute to run models that can just barely do a simple job so long as you don’t mind spending even more time eventually fixing their work than they supposedly saved.
