BUT! I feel that lost hope could/will turn into a new forms of "min-maxing" or whatever where we will see people focusing on small, efficient software to make the most of what we have to weather the drought. We got spoiled with cheap hardware. Time to make the most of all that cheap hardware we have.

Though of course, replacing this hardware as it breaks down will be expensive. Perhaps we could go back to 1/2/4 GB RAM sticks and 256/512 GB Disks. Then run minimal Linux installs, minimal GUIs, minimal services, minimal software.

To amend this: Reading the 8088 domination article and the little $8 RISC-V Linux board I think we have plenty of opportunity. Maybe we learn to shed our inefficient BIG computers for small, low power, purposeful, yet general computers. I play with Plan 9 a lot and its portability, small footprint, low resource, and distributed design has showed me we can do a lot with a little as long as we stay disciplined to keep things simple and to the point.

I have to think, it can’t take much time before the Chinese DUV fabs come online and non-frontier chip prices crater. Sure, you might not be able to get 5090 equivalents, but the 30xx series was DUV and so that should be replicable in a few years.

Whatever quality bar / process node China gets to you know they will scale it out super hard.

(Some quick investigation suggests that the Yuliangsheng DUV machines currently under construction are 28nm, which would map to 2014’s GTX 900. Perhaps I’m being overly bullish but I assume they will recapitulate/clone the tech faster than the rate it was originally discovered.)

Also, this reminds me so much of the download more RAM ads back in the day.

gaming does not equate to only things like the last of us or cyberpunk which need extreme machines. a lot of visual fidelity can be reached still without having such a monster. Just need to not use certain features which are currently pretty much exclusive to Nvidia...

There should be a big enough back catalog of great games for you to enjoy without having to buy a new gaming rig.

Not even talking dictatorship scenarios here. The USA might simply decide it doesn't like your country. Or maybe heavily cripple the models you can use, or censor them, for whatever reason.

I guess this is an argument for local AI. Probably not feasible for the less technically savvy.

- local latency is lower than network latency

- latency is important for action games

The AI boom first reversed the trend, to now try to get back to that cycle.

Maybe arcade games?

Like talking to people?

I am using AI now to make a game engine that will run and make 'nice enough' graphics on simpler machines. things like laptops with iGPU and 4gb of ram.

U can still get quite ok graphics, and make very complex and rich gameplay systems using such a machine. It just needs to be the target machine while developing so the right tradeoffs are made.

And why would the game industry, considered as a whole, want to incur higher costs to exclusively deliver its product in a way that provides less value to the majority of its customers, many who won’t be willing to pay the higher prices required to maintain similar profits?

Recompilation at most threatens platform exclusivity as a business model. Otherwise, it’s just piracy with extra steps.

At most, I expect to see already existing trends towards more live service components and microtransactions continue, with digital delivery only and a required online component in many cases, but with most if not all of the actual game code and assets hosted and running locally.

As for non-gaming applications, content creation and a software development are two examples where "cloud native" workflows remain cost-ineffective in many market segments in spite of hardware price increases, and will remain so as long as cloud providers continue to nickel-and-dime customers for IOPS and compute.

It's just unfortunate that producing a half decent computing platform is significantly more difficult than a rifle (and that in the case of said hardliners, the capital is backing them, here it's working against us).

> If you're not partaking that's only on you.

There's honestly nothing sensible about that for a lot of developer budgets across the globe. Don't assume most folks have $3k to burn.

They absolutely can't do what, or more of what the biggest models can do. There's nothing frontier about small local Qwen or Gemma models.

2899$ gets you a machine that is just good enough to run some small models, slowly, with terrible context, requires a ton of tinkering and fine tuning, managing and constantly downloading a bunch of huge file blobs, etc.

Partaking those $2899 in most of the large cloud models would bring you much more productivity, much less boilerplate and time spent in setup, etc.

So.. just weird accusing people here of "not partaking" in AI when you're just essentially toying around if you're actually comparing local Qwen and Gemma's with frontier models.

Given that he was in a fluffy interview at the time, he's not going to say that OpenAI and Anthropic would die, but we all know that -unless the "Pacing The Frontier" garbage succeeds in its goal of creating law and regulations that massively obstruct LLM manufacturers who aren't the four or five currently-largest ones in the US- that's how it'd turn out. Once OpenAI and Anthropic are dead, the VC money pouring into the LLM manufacturing and services industries evaporates, and all that demand for high-speed VRAM vanishes with it.

So, no, switching to locally-hosted software will reduce RAM prices... just not immediately.

isn't communal living the most efficient thing resource-wise though?

It's like a highrise developer saying "people think our buildings take up a lot of land, but actually houses are much worse from a land use perspective". Which is objectively true, and yet does not mean the developer thinks the entire world should be turned into houses, or even that they dislike houses!

https://en.wikipedia.org/wiki/DRAM_industry_price_fixing

To date, five manufacturers have pleaded guilty to their involvement in an international price-fixing conspiracy between July 1, 1998, and June 15, 2002, including Hynix, Infineon, Micron Technology, Samsung, and Elpida.[3]

re "what proof means": Im not your house boy. Go get it yourself.

proof ... the thing a narcissist demands from you, but will never acknowledge having received.

its a good lever when it works.

Considering the profits they've made this year, on average they still wouldn't do too bad. SK Hynix employees got almost $500k for example, I can live on that for a few years. Companies are often also legally required to compete with each other.

But they are scaling up the production, they just conveniently did it a bit late. I also fear if/when the bubble pops, they'll quickly reduce the production to ensure prices stay high. I fear Nvidia could do the same. I think it'll be hard for all of them to go back to a world where they're paid way less per component, so instead of accepting that, they can reduce the supply and keep the prices high. Only antitrust regulators could intervene to stop that.

Many people expect prices to go down after the bubble pops, but it may not happen as much as we would want to.

All that bubble popping will do is wipe out some investor's money. Everything else will stay and become more and more important.

But of course, it is more profitable to sell one item for $1000 than it is to sell ten items for $100 each. And even if a competitor shows up, you can just price dump with all the extra money you already made until that competitor goes away.

If a customer ask me to get all my butchery stock for a year, it will make thousands of client unhappy and will make one client happy, so I'll simply refuse.

This is not how business is supposed to be, the world is not supposed to be a B2B chaos.

I know how the business is done "supply demand" bla bla. That does not make it ethical one bit.

If you look outside the tech bro / stock bro / tech worker circles and into the mass left wing populace everyone HATES ai and ai data centers... the Altmans, the Amodeis, the Huangs, they are not the devil himself levels just yet (cough, Musk, cough), but they are climbing up there.

People's jobs being taken (and especially entry/junior jobs, you know, the ones young people, who voted massively blue, need to get started in life), data centers increasing their electric bills or polluting the air with gigawatt level gas turbines (or both), consumer electronics prices rocketing up (~400% increases YoY in RAM and SSD prices) just so the corporate capitalist class can hoard millions of gpus in warehouses.

I mean, if the democrat politicians getting into power keep their promises and step on AI buildout's neck, it could be lights out for the AI bros.

Let's say there are 10 companies that need 1 GB of RAM each in a life-or-death scenario. They would be willing to pay $100,000,000 per GB to get 1 GB of RAM.

At the same time, there are 90 more companies that need 1 GB of RAM each to make certain quarterly targets. They are willing to pay $10,000 per GB to get 1 GB of RAM.

If you make 9 GB of RAM, you can sell it for $900,000,000. If you make 99 GB of RAM, you can only sell it for $990,000, because the 10 life-or-death companies only need to bid high enough to push out a quarterly-target company.

It's a bit of a contrived example, but it shows you how restricting supply can boost total revenue if there is a segment of the market that has inelastic demand. (this is why medicine pricing has to be regulated - in the medical industry, there are lots of inelastic demand scenarios that can lead to predatory pricing)

40 tabs open at once? That doesn't need a lot of RAM if websites were programmed to do what they actually do: Display text and images.

I don't think touch needs any special amount of RAM, would it differ for the operating system than a mouse?

Photoshop and Lightroom <- These I think fall into the category of special use cases for professionals, which absolutely need more RAM than a home user. For somebody who only occasionally wants or needs to do a bit of advanced photo editing, it's okay to wait a bit longer for operations.

> I still have 10~20 years old machines around and boot them occasionally, there's just no way I'd get back to them full time and try to do the same volume of work as I'm doing now.

Are you sure about that? If you actually sat down to only work, without having YouTube videos open or other things going on in the background? If you close your Amazon Black Friday tab you already liberate 500 megabytes of RAM.

> PS: there's also the "people only do email and browsing, perhaps some Word or Excel" fallacy.

No, in this case it is you who didn't understand or didn't want to understand. Nowhere did I say only. But the things I listed are what normal people do mostly. And nobody would want to type a long report or do spreadsheets on their phone.

> So, computer are a lot more niche than before, and the requirement also go higher. People on computer are either power users or doing somewhat intensive things.

Not at all. Most computers are used by office workers at their job, and I'd bet that most home computers are also used for non-intensive purposes. But I think you're right in what direction things are moving.

He's an easy target because of who he is, so I understand why the entire Internet ran with the fake story (even LTT!). But on Hacker News you'll definitely get pushback if you post misleading stuff.