I like LLMs and also the research but I think we're at a point where one could make a pause, turn the models into silicon and sell them as hardware for people who are actually interested.

That would solve all privacy problems, eliminate many data centers, subscription fees, internet dependencies and allow for slower and more thoughtful upgrades (hardware isn't updated as often as software). That would be a pretty sovereign solution.

However, from a company perspective they can't force it on everyone anymore, don't get all the juicy data from all the users and can't sell subscriptions any more and change everything on their side.

That would be lovely

Why does that feel forebodingly familiar

looking forward to looting RAM and GPUs for our community library organized communist LLM

I mean, we'll jump to 1930s on reflex. But this is a pattern that's repeated itself internationally, from Apartheid South Africa to the juntas in the Pacific Rim to Fascists in Argentina and Chile.

Economy shits itself. Social Democrats come in to deliver milquetoast reforms. The economy recovers somewhat, but never enough to actually relieve the social pressures of the moment. And then the democratic system starts vomiting up radical candidates of all stripes, while bourgeois capital attempts to ride the tiger into the next full recovery and/or national emergency.

The wheel of fortune doesn't always land on Fascism. But it's definitely a possibility.

Squeeze the working class.

Waiter, please bring me 1,000 of your finest shares of WinRAR! I'm going to be rich, I tell you! Quite!

That won't pay off the several hundred billion they owe.

I'll bet you $$$ right now that won't happen. First of all, they still won't be profitable. Second of all it's easy to just completely avoid interacting with LLMs.

That's for governments to figure out.

Until it's bad enough that the poors decide they have had it with the owner class,

turn the models into silicone

The models are data, and silicone is rubber. So... I don't know what you're proposing here.

Yay, free pollution for everyone!

This is why I think the bubble will last a lot longer than some people think. The heads of the tech firms have such a tight relationship with the GOP, and even a somewhat tight relationship with the Democrats that it's almost certainly going to be declared "too big to fail" or "too critical to fail".

The fact that it's being used by the military just makes that even more likely. The firms will declare that if they go bankrupt, the military will be crippled -- despite the fact that the military ran just fine without "AI" just 5 years ago.

So, when the economics of the situation show how absurd this bubble is, the collapse won't happen because the firms will be propped up by the US government. That will put every taxpayer on the hook for the billions and billions of debt these firms took on. So, Andreessen Horowitz and the Saudis will be bailed out by the taxpayers who have been fighting datacenters.

History doesn't repeat, but it rhymes.

It seems like we're plunging into fascism, but maybe this upcoming midterms will (not be rigged and) prove to be an effective backlash to the US' reactionary coup. What do you figure the odds are we pull out of this and say, "huh actually let's try that FDR thing instead of this Nazi shit"?

The government will bail them out: "Too big to fail"

it's easy to just completely avoid interacting with LLMs.

I wish that were true. Have you seen reddit lately? Lemmy isn't immune but I haven't encountered any obvious bot posts here yet.

Might mean "silicon", but might want to bang the LLM. Hard to know.

Words have meaning...

Even with "silicon" doesn't really make sense. The models already live on silicon, they're stored in the datacenters needed to run them. It's not like you can make them into some chip and magically not have to use the datacenter.

But who bails out the goverment when they list everyone below billionaire status as a non-citizen?

"Source: AI" is kinda wild for this purpose

This is actually stated in the xz man page: -0 is sometimes faster than gzip -9 while compressing much better.

With a random test file on my system, xz -0 was 3 times as fast while achieving the same compression as gzip -6.

I believe he is suggesting we turn them into very large breasts, or in some cases big booties, depending on your preference and what part of the world you're in.

My crooked spotted cock is more profitable than OpenAI

I don't fucking use reddit. You seem to be missing the point.

LLMs aren't important & don't employee many people. There is also no amount of free money that can make them profitable. I will happily bet you money right now this doesn't happen before OpenAI and/or Anthropic are sold for scrap.

Gonna take GPS example to explain general processors vs specific ones.

GPS works by doing some simple math around triangulation from satalites. When they first came out there were no GPS chips. It was done by a general processors.

Now they can do it, but because they can do Amy Hong they're not as fast as newer GPS chips.

How they work is the silicon can only handle the GPS calculations. But because of that, it can do it way faster.

LLMs run on matrix math and probabilities. We can 100% come up with specialised silicon to handle this math. Intact, thats what a lot of AI specialised chips are.

Chatbot labs don't employee enough people for too big to fail to be a concern. There isn't enough currency in existence to make them profitable.

I don’t understand your comment - bogleheads is a forum of real people? I included a silly example illustrative of my point, but the forum is decidedly not AI. Did you miss the first link, which was the point of my comment?

Proprietary frontier LLMs in companies riddled with highly suspect business practices.

My local LLM setup has absolutely been making the equivalence of positive cashflow for me since gemma 4 and qwen 3.8: I need less total time for several groups of annoying (but necessary) tasks, which means my productivity is up and I can use the rest as my own free time.

I need less total time for several groups of annoying (but necessary) tasks

Such as? Just curious what people use LLMs for.

They're hoping it pays off in the long run, or at least long enough for them to go public and sell their shares at a profit while pension funds and mutual funds buy a crashing asset.

you won't generate a 1,000x return on investment, AI could if it becomes AGI somehow. It's like gambling but nobody knows the odds and the payout is big enough that it's almost worth it to try without knowing the odds.

The first post in that thread had a guy using AI to figure out how much of his portfolio was in AI.

Literally the first sentence of the post you linked to says:

I asked AI for advice on how to reduce our risk of the various AI stocks in our portfolio

Couldn’t make it past the first comment, or are forums a new thing for you?

Yes that’s how the thread started - what else did you learn?

How do forums work?

NPUs are already a thing, specialized cpus for nerual networks. The problem is the models are just databases, and you need lots of fast memory in order to feed the NPUs data, and thats the current bottleneck.

They aren't comparable. GPS is deterministic and simple so as you said it could be reduced to an FPGA or ASIC.

LLMs are partially matrix math and probabilities but they're also more complex than that and require a large amount of RAM to run even the first time. Each query added to the context increases the RAM needed further.

Fixed it. We don't use this silly silicon word here. I picked the wrong one. We call it Silicium
Sodium is Natrium, Potassium is Kalium, Aluminum is Aluminum etc.

I messed up and should have replied to the guy above you.

That is correct. Very comforting

I handle so many text files in my job and LLMs help make the scripts for me to handle these text files. I know it is important to at least know stuff like regex but I am not a superhuman. It is difficult for me to heavily specialize in scripting while also specialize in the technicalities of my job. To help show why it is difficult to specialize scripting, I work with hardware description languages which have a different syntax from scripting languages. Handling different languages causes confusion on the proper syntax even with a basic if else statement because the syntax is different between languages.

Aside from this, it can also help me automate describing what the sequences I created in the file for easy documentation. I just have to instruct the "patterns" it needs to look for. I just need to proofread the output and do minor fixes.

at this rate Winamp might get more profitable than OpenAI and it is free.

But isn't OpenAI just burning billions of moneys already? In that case Winamp is already more profitable

LLMs are also deterministic and the weights for a model could be 'hard-coded' into a chip, no?

Considering the overwhelming support among Democratic politicians and party insiders for fascist collaboration, the odds are probably low.

They're actually not because they have randomness built in. Since we are working with probabilities, it won't always pick the next token that has the highest probability and the randomness can be tuned via a "temperature" setting to make it more or less likely that it will choose the most probable token.

The weights for the model could be stored in a firmware chip but you still need RAM because it pulls all the weights into RAM in order to perform the calculations.

Tech companies are expected to be run at a loss for quite some time before they start being profitable.

They were founded in 2015 and "officially" launched GPT in 2022 to the public. If they make profits by 2032, that would be in line with how other tech companies established themselves.

For example, Spotify wasn't profitable until 2024. They launched in 2008.

So, I don't know why them not being profitable is what everyone is talking about. They're not expected to be profitable until 2030 at the earliest. And every serious investor know this.

Yes I'm very fun at parties. But it seems that a lot of people just do not understand the expected timelines of tech companies.

believe it or not but there was a time when Winamp used to be AOL's money drain