Not everyone subscribes to the "cloud compute is more important than everything else" outlook, but everyone is affected by it nonetheless.
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!
But saying people shouldn’t want something that was abundant and you made scarce is an asshole thing to say.
Will we go back to PCs? As in: are we going to get cheap local hackable frontier AIs at some point? I hope so.
Forget about AI, the entire PC market is getting killed over this.
As an avid gamer I'm unsure if in 5 years I can even buy a gaming PC anymore. Not even because of cost, but it's unclear if anyone will still make the components anymore.
NVIDIA isn't interested, nor is AMD.
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 bought it at the middle/end of 2024, and I paid 3790 NOK (~357 USD at the time).
The lowest price online in my country for the same product new online today is 10233 NOK (~1085 USD)!
270% of what it was before, it’s absolutely bonkers.
I was even considering putting it up for sale a few weeks ago. But of course what I would be able to get for it would be no where near $1000 when selling second hand. Instead I might if I am lucky get slightly more than I originally paid which would be nice and all except that if I then regret selling it it’s still expensive to replace, or I have to also buy it second hand from someone else that may have worn theirs out considerably more compared to what I have done to the one I currently have. Mine is very well taken care of since the start, and although it’s nearing capacity now I haven’t written and deleted data over and over on it. So there has been very little wear on the flash.
Been doing that for myself for the past 15 years
Advances in hardware are only impressive and useful for me when I avoid the so-called "tech" company "engineer" BS
The reason I stop using old computers is never due to software. It's always outdated proprietary cables, failing power adapters, screens, hinges, and so on
Sometimes "minimalism" is simply for _speed, resource usage control and/or reliability_ (that's what I get from small software, anyway) in contrast with prevailing marketed garbage that is bloated and sluggish by commparison
Other times, claims of "minimalism" might be trying to capitalise on some silly trend
Personally I like what the computers I own can do for me and learning to make them do more. I have a brain, I don't do much "guessing" and I don't really need "inference" on a daily basis, certainly not for such a high cost
I have all the "convenience" I can handle
Many games that come out still, and are quite successful, do not require such technologies.
1. Indy developer releases a game or trailer of a game. 2. Within an extremely short time (days to weeks at most) there are clones on the market. 3. These games often outperform the original.
It's absolutely not something that's gone on for ages.
It might be a novel gameplay idea / twist on a genre, and mere days after launch instead of being able to stand on its own and gather attention it's drowning in a pool of lookalikes. It's no longer distinct.
The speed of being able to slop together something like this is less than 1% of what it used to be.
So it's easier than ever to make a game, but it's more difficult than ever to get anyone to try what you've made. I'd say that's just as bad, and in some ways worse, than having your game cloned.
For someone looking to play a new game it will be difficult to sort through all the bad games to find the good ones. And for anyone pouring their soul into a game they will more likely than not have an audience of one.
Our eventual destination is probably going to be a) pick up a device, b) ask your bot to make up a game on the spot, c) play. It sounds so lonely and depressing.
So yes maybe it happens more often now, but it’s still not new.
Yes, that's literally the point of the people you're responding to. Has HN lost reading comprehension, or are you arguing in bad faith?
It used to happen occasionally. Now it happens instantly. That delta is the problem.
5 years ago it would have been science fiction
I don't even understand your analogy. I'm not watching the youtube video or reading the book. The model is. And a big computer can watch 20 youtube videos faster than a small computer can watch 1, and do a far better job understanding all of them (bigger model at data center scale is both smarter and faster than smaller model at local scale).
Are you trying to one shot create the Linux kernel?
If I started today and played 10 games I can emulate every single day for the rest of my life restricting myself to purely emulation, I wouldn't run out of new games for...five and a half years. (assuming my rough calculation of ~20k available games).
Drop it to a game a day and I will likely never run out of games given my life expectancy.
Consider that I will find some games I like to play for more than a day and there are easily well more than enough games in emulation alone to entertain me for the rest of my life on existing entry level hardware.
Will I ever play MODERN SHOOTER {CURRENT YEAR} on this device? Certainly not, but I can't see a reason to bemoan lack of access to such a small portion of the available industry.
Like, suppose you're a reader. You might be right that you might enjoy your favorites in their original languages, but you'd be silly to be worried about your future ability to enjoy books just because you'll likely never get to experience reading 三体 as a chinese citizen fluent in the language.
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.
I wonder what it would look like if Apple offered a slimmed down shell from the Neo for ~$300 that required a connected iPhone. Maybe as iOS and MacOS converge we'll get there.
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...
And I assume they will only continue to go up. When it came out, purchasing one seemed like an easy decision. But I paid half as much for one with the same CPU/RAM (and upgraded my SSD for cheap).
There should be a big enough back catalog of great games for you to enjoy without having to buy a new gaming rig.
You might say that this is impossible but look at current trends. A lot of information today is getting shared in the walled gardens of Internet. You have to login to social media to view posts. The Twitter, Reddit threads that were accessible freely now require you to log in. Similar thing is happening across other networks. People are no longer asking questions on Stackoverflow. All those queries are gonig to AI. If someone chooses to not use AI, their options will be limited in future.
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.
So it’s plausible they’ll block me from accessing Mistral since they host DeepSeek models.
- Large centralized compute is more powerful than portable compute
- Wired is better than wireless
- The public wants new things
- The tech industry grows to use/“use” the available processing power
As long as you’re not continuously comparing to AI in large centralized compute, what you’re really looking for is “when does portable compute become powerful enough to run AI models that are capable of meaningful work” and I’d say we’re already there and it’s getting better.
- local latency is lower than network latency
- latency is important for action games
But then it became normal.
The monopolization of knowledge and expertise is the next step. I don't mean as if it were intended, but it is the logical progress given the technological evolution. The cause is inventive structures: the tech is neutral (in my opinion), but the way in which it is used and deployed is subject to and disciplined by capital.
The AI boom first reversed the trend, to now try to get back to that cycle.
I guess it's because I used to work in a writing-based profession and mostly know academics.
Lot of people using their office-notebooks at home, even if its not allowed. Most people out of tech do not have big hardware at home anymore.
This is complete conjecture on my part, but I bet if you're late 20s - early 40s, you probably have some form of pc/laptop laying around. But people 50+ (gen x / gen boomer), a phone is likely all you use or own. My mother, aunt, et. al operate totally through their phones. The same is true for my pre-teen, teenage cousins or cousins children (ipad generation).
But yes, it has been a test of just how few tokens you can get by on.
New application of computing requires a lot of compute -> compute centralizes -> compute improves, smaller, faster, cheaper -> compute decentralizes -> new application of compute is found -> (return to beginning of cycle)
Maybe arcade games?
To be clear, I am not skeptical that it will work this way - I think it's pretty likely llm usage will be come a split of de-centralized and centralized. But I am not very convinced this is a natural law of technology.
But, the “good enough for most lay people” intelligence bar will for sure miniaturize, especially as inference ASICs get more specialized.
But yes, we both agree that what was the frontier many years ago will fit on your desktop.
:insert Will Smith presenting meme:
The Cloud
Which would not be all too bad. It would also mean the end of stupid SaaS subscription services, and mass surveillance, adtech. And shitty propaganda.
And your local AI could emulate the good parts of the internet.
Like talking to people?
In fact, it was closer to the opposite.
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.
Everybody will have a mini PC for the sole purpose of streaming content, with no storage to keep anything on it.
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).
Have an agentic enough setup and they'll do anything that the biggest closed models can. They'll also do more, making them the true frontier.
A sensible point of entry is a 2899$ Mac mini.
If you're not partaking that's only on you.
> 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.
$15 of serotonin is better than this stuff: https://www.amazon.com/stuff-under-15-dollars/s?k=stuff+unde...
Ok.
Also, a lot of things are true but what truths you choose to broadcast can tell us what you care about. E.g it's also true that we'd all be richer if sam altman shared is wealth with us, lets take a few seconds to think why he's not saying this truth, after all it's an objectively true fact.
But we're the bad guys for getting the context wrong /s
I would be shocked if it did anything at all.
I don’t think it’s very productive to go on a witch hunt on someone who said something stupid and is being taken out of context (albeit he should have known not to say that).
Apparently the last three times this cycle happened, the manufacturers that built out capacity went bankrupt when it popped, and the ones that didn't didn't, so they're actually being wise.
Absence of evidence is not evidence of absence.
If an event happens, and multiple previous times that event happened a cause was identified. It is reasonable to suspect the same cause was responsible this time.
It does not give proof, but it does give reasonable suspicion.
hopefully china does it and hynix burns in hell
where is the proof?
>and are now not scaling up production according to demand.
fabs are being built by maaany companies, but it takes years
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.
Simple as that.
If I said all donuts are spaceships, readers are not obligated to go and google it.
You're free to doubt it, it's not absolutely certain, but pretending the chance is near zero is just not even remotely as rational as you seem to think.
Theres never any "proof."
Theres no "proof" on the internet.
All we ever have is "evidence," and varying degrees of trust or mistrust in it.
Reality ... always good for downvotes.
* Participating in meetings, conversations, and communications in the United States and elsewhere to discuss the prices of DRAM to be sold to certain customers;
* Agreeing, during those meetings, conversations, and communications, to charge prices of DRAM at certain levels to be sold to certain customers;
* Issuing price quotations in accordance with the agreements reached; and
* Exchanging information on sales of DRAM to certain customers, for the purpose of monitoring and enforcing adherence to the agreed-upon prices.
* Under the plea agreement which must be approved by the court, Hynix has agreed to cooperate with the government in its ongoing investigation of other DRAM producers.
US DOJ link: https://www.justice.gov/archive/opa/pr/2005/April/05_at_207....
This is remarkably different than someone claiming donuts are spaceships when the company itself was charged and pleaded guilty, and it's a matter of public record.
just asked about some kind of proof when world is full of conspiracy theories, nothing wrong with that, so stop being aggressive and snarky
and i knew about stuff from 2000, but thought there was something recent
Would you though? If/when the AI bubble pops and the buildout of data centres either stops or plateaus then they'd suddenly be sitting in a vastly oversaturated market and struggle to make it to the end of the year.
New fabs will be operational in a while, which is too big of a risk to take in these times.
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.
1) boom and bust cycle in memory industry is a repeated problem
2) companies are not obligated to scale up
> and is being taken out of context
no it is not - Altman is lying by omission[0] by implicating that various important costs are not real
[0] https://news.ycombinator.com/item?id=50400515 for longer comment on that
Good god the internet is so sensitive nowadays
[0] https://news.ycombinator.com/item?id=50400515 for longer comment on that
And "if all current AI users" is a strawman, people using free chatgpt once or twice a day are not going to buy RAM to run own LLMs.
Ah yes Sam Altman is not a real villain despite OpenAI locking up 40% of global DRAM supply.
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.
ChatGPT is one of the most used websites in the world. Compared to e.g. Facebook you are not "forced" to use it (because of friends or whatnot) so it's all voluntary usage. There is real demand for AI usage, that's why RAM and GPU prices are skyrocketing, not because of some evil conspiracy.
I don't see this, no. I think HN serves as a sort of Rorschach test for the reader. I get a center-libertarian vibe from HN, definitely not left wing. That's probably my own biased perception, of course.
I also think while there's a ton of AI articles, there are also very vocal people against it. The sentiment is not universally "AI booster" here, it's a mix.
Obligatory: "we're curing cancer r-right?" No Nat we're making slop videos with it.
If the USA were a real democracy they would do something about it. After all this even violates capitalism, e. g. free market leading to reduced prices. But the USA is run by a criminal gang, pocketing away money - the TechBros and their cronies. They just make any legislation they want to, with a bit sugar coating on top to not appear too overly evil or greedy. But they are.