The use-case would be inference, not training. So you'd write the weights there when you load the model, and then not touch them until you replace the model with something else. Remember, flash experiences wear from writes, not reads.
Honestly, it seems like a perfectly plausible solution: With 20us latency, you'd be able to load 50k layers per second. You can get as much bandwidth as you need by putting more dies in parallel.
Qwen3.8-flash-next has 48x sparse-MoE layers, and you don't know which experts need to be loaded for each layer until the prior layer has finished. So that requires 48x 20us delays per forward pass. That admits more than 1000 forward passes per second.
With typical MTP acceptance rates, you'd be looking at maybe over 2k tok/sec as the theoretical upper bound. A big flash memory on the GPU would be great for n-gram tables as well.
Not my area of expertise at all, but I thought flash was durable for read-only use. Put your weights there, leave them there, and use other memory for the KV cache.
Google search ads used to also convert highly because they were well-targeted and positioned at the point of consumption.
If I'm searching for red widgets, it's a fairly strong indicator I'm way more likely than the average to want to BUY red widgets, so a well-placed ad for redwidgets-cheap.com didn't feel hostile.
Now when it's "I am clearly searching for red widgets but you want to railroad me into blue doohickeys because they're paying more per click", you're one step above a Taboola chumbox.
I suspect there's a place for LLMs, but it would be one that gets developed on a per-organization or even per-team basis, rather than the tendency to airdrop some elaborate do-all solution.
This is probably why it's chafing on a lot of developers. We're hearing we have to use these tools and only these tools even when they aren't great fits for specific tasks or workflows.
The management loves the idea it closed an inbound ticket in 30 minutes... by converting it to a code change that fails to fix the actual problem, or by replacing it with a slightly different category of ticket that becomes a burden for someone else to finish off, which is for some reason not a metric anyone cares about.
The architectural differences that made a console more compelling than a standalone PC are largely gone. (Nintendo is a bit of an exception here, but they were always playing a different market than Sony and MS anyway)
This means it's harder and harder to say "here's a unique experience you only get on Xbox" (or Playstation for that matter)
The price advantage is going away fast, especially with talk that next-gen consoles will top USD 1000.
All they have left is their prestige first-party products, which they don't seem to really know how to use effectively. Some of them have solid cultural penetration, and could spawn broader media franchises with the right handling.
> The architectural differences that made a console more compelling than a standalone PC are largely gone.
No, the main driver is still there. People do not want to build and maintain PCs. Your 14-year-old nephew doesn't want to do it, either. People want to turn on a box and play a game. I am technical and I want absolutely nothing to do with maintaining a PC. I'll take the console experience every single time.
People can buy a prebuilt pc running steamOS automagically updating itself with immutable images.
The advantage of consoles are:
- No minimum/recomended config to check and insurance that a game or an accesorry bought for a console gen will work well for that console gen. In term of game compat this is something Steam tried to with the steam deck but has no way to enforce really and the biggest one imho. When you buy a brand new "budget" computer in the price range of a console you usually knownyou won't be able to play some of the upcoming AAA titles at least woth decent graphic quality.
- availability in retail stores
- ability to sell games in the second hand market but I don't think this is something that is here to stay with the disappearance of physical medias
You’re mostly correct but there’s truth to the original statement. Back in the PlayStation/N64 through to PS3/Xbox 360, consoles were so different to each other and PCs that porting games was a major task that resulted in very different outcomes.
Since the PS4/Xbox One, they’ve been similar enough in underlying architecture that engines like Unreal have been able to smooth over the differences enough to make multiplatform technically viable. It’s meant companies like Capcom with RE Engine have made a huge success out of selling across all platforms, including PC. That freedom for publishers is a problem for console manufacturers who have fewer 3rd party exclusives as a result.
The performance cap on a console makes a big difference to me which is why I only play on the PC. It’s also not rocket science, you can just buy a good gaming PC. Games look an order of magnitude better on the PC in some cases.
This is why I've never gotten any of the more recent consoles (since the N64 or so)... many of my friends have them, and every time we try to turn them on to play a game they start updating and rebooting and updating... I can play games on my computer, which I have other incentives to keep up to date, so the maintenance time is amortized; just double click and it works.
The idea is that the hardware overlaps enough that consumers would rather only buy the generic version which is the PC. It works for gaming and everything else. There was a time consoles thought they could steal share from the PC market by including browsers and whatnot but I think that never panned out as expected
The exclusives business model worked for Sony for the ps4 generation for sure. But it seems that the cost of developing ever larger and more ambitious games has ballooned in recent years. Intentionally limiting your distribution seems misguided. Even Sony has been releasing their exclusives on PC too after some exclusivity time.
On the other hand it seems to work for Nintendo which keeps their consoles intentionally underpowered and don’t try to compete with the AAA game expectations with the majority of their titles.
The thing is the whole set of companies making games at Sony and Microsoft are completely built for what we call AAA games. Tons of middle managers, large teams of artists and visual effect specialists, a whole human pipeline built to sand off every single pain point in every moment of their game to make a good content tube. Those studios are incapable of just making a bunch of indie games. They would need to be completely reorganized to an insane degree.
Step outside an echo chamber and you'll see that in most countries, including East Asia and the Middle East, if you take a walk around any big mall, you don't see kids pestering their parents to buy them a PC. They ask for a PlayStation or an Xbox or a "Nintendo" (yep that's what people still call the Switch)
In conbinis, there are Xbox and PlayStation gift cards hanging from the shelves, in the local currency.
There are accessories for PS/Xbox/Switch everywhere, more than PC gaming accessories.
And it's not because the PC stuff gets sold out while the consy stuff gathers cobwebs: If that was the case the consoles would stop getting restocked after a few months, but they keep stocking newer stuff. Meanwhile I have to hunt around and specifically request for Razer or Mad Catz or other PC-specific brands.
> The architectural differences that made a console more compelling than a standalone PC are largely gone.
If you go hang out at someone's home, nobody says "Let's go to my bedroom desk and play something on the PC", no they crash on the couch and turn on the TV and fire up FIFA or some other normie shit.
> Step outside an echo chamber and you'll see that in most countries, including East Asia and the Middle East, if you take a walk around any big mall, you don't see kids pestering their parents to buy them a PC.
This is still the case in the U.S. for the most part. The point is moreso that barrier to entry is lowering. The SteamDeck is a PC/console and a good example of a possible trend in that direction.
Just a couple weeks ago in Bali someone said "Oh you have a Nintendo too?" and pulled out their Switch. Meanwhile I haven't made a Steam friend in the wild in years
Hell I saw PlayStations in barbershops in Vietnam! There's no way you'd see a PC in those environments.
The Steam Deck certainly reduces the gap, but I've yet to see it penetrate "normie" social circles, and it's still rare in random gaming shops compared to console stuff. Have Valve released their OS for other handheld manufacturers to use? That's probably the only way to end Windows' dominance and improve the coolness factor of "PC" gaming.
> Meanwhile I have to hunt around and specifically request for Razer or Mad Catz or other PC-specific brands.
Since you mention conbinis, I'm guessing you're in Japan. Having just been there, game software is receding from the smaller hobby stores, and then in the tech stores like Bic Camera, the PC gaming stuff is one floor down.
It feels like the problem was the market ended up in a race to the bottom and so most of the made-for-PC controllers just sucked. The best option was to buy your preferred console controller and use that with PC.
Usually means your comments are getting flagged by other users. I’m not going to dig through your comments to speculate on why, but I don’t think there are any automated filters that will kill only some of your comments.
The game has changed. Now I can only post a few times per hour and they're automatically flagged. No moderator has given me any contact to explain what I'm being punished for, so I guess I'll be making a new account and hope the same glitch doesn't occur again.
Consoles also had performance advantages up to PS2. The lightweight execution environment and exotic accelerators used in consoles allowed better gaming experience than CPU/GPU at that time. Cycle-perfect consistency of consoles also allowed games to be better "optimized", such as data aligned to boudaries and arranged on game discs to be streamed and such.
Consoles slowly started losing that plot after the launch of Xbox and GPU development mess of PS3(the PS3 CPU was supposed to be the world's first APU but Kutaragi dropped the ball big time) and now consoles are just a Nespresso experience, with just marginally cheaper hardware tied to proprietary full-priced game discs, which is not worth much.
OS argument is practically valid but... Frankly IMO that's more of a disadvantage of PCs coming from horrible near nonexistent engineering discipline of software engineering than an advantage of consoles.
My take is that almost nobody has a standalone PC anymore, especially kids. Kids these days all start with iPads, them maybe they move up to Macbooks once they get to high school or college. Other than the local LLM fanatics I see here on HN, I have no idea who's buying those PC towers anymore, they seem like fringe relics from the past century.
People who want their games to look prettier and run smoother than what you can get out of the thermal and power budgets of laptops would be the ones still buying the desktops. If you additionally value a not too noisy environment home, then water cooling is way better and more accessible in the desktop form factor (though this is still niche at least for the gpus where all-in-one units are basically non-existent). Laptops are by all means nice toys as well, but on the go phones are better entertainment while in transit, and once you're somewhere you left the comfort of your home to experience it's rarely for your gaming hobby (however it might be a setting where you're reminded that computers are still relatively cheap compared to certain music instruments or sports equipment).
Another attraction for publishers was (historically) a much lower adoption of piracy.
Part of the sales pitch is you need more than just a torrent client to pirate.
Now in the world of AI I think this moat is evaporating. Hypervisor exploits are dropping like rain and determined individuals can vibe code a reasonably performant emulator in a couple of weeks.
In that world the only reasonable edge you have left is high prestige exclusives and merchandising/physical sales.
Nintendo has this cornered, one of the few companies who can pull off a rabid fan base buying a console they already own because this time there’s a triforce on it.
Meanwhile Sony and Microsoft are diluting their own IP and going full digital. Writing is on the wall if you ask me.
All else equal, I'd rather have the richest possible functionality built into the language. At least I know that it will be updated and maintained in one place, and there's very little ambiguity about setup. You don't have to start picking apart the fact IN THIS PARTICULAR PROJECT, libfoo was imported by pasting it in the code base in 1977 before package management existed and globally namespaced, but libbar is a first-class citizen in the depency-injection system.
If you know you can just call bargraph($data), even if the stock bargraph library is boring, then people and LLMs will favour it.
I've heard the argument that a lot of the 'scarcity' in RF spectrum comes from chicken-egg problems based on the original allocation strategies.
If you give out a dedicated fixed slice of the spectrum, and high power licenses are pretty limited, you end up treating it as a "circuit switched" resource. I have my narrow sliver at 144.4449MHz and I will use it as I see fit.
If you allocate capacity with the understanding it will be shared, it creates a system where you have to think more "packet switched" -- how do I detect congestion, retry, find less congested workspaces?
The "packet switched" philosophy usually supports a lot more users on the same physical infrastructure
Cloud wasn't sold as a performance thing. It's more about outsourcing the maintenance (let AWS replace a dead PSU at 4AM) or for the promise of scaling tasks to infinity for dynamic demand (web hosting).
You might see interest in thin-client plays for the same reason. They'd like to replace the local desktop with throwaway $50 terminals because it's easier to deploy on a standardized VM image and not have to replace dead SSDs and CPU fans.
Workstation tasks that require low-latency and high performance seem to stay local because the performance is simply not there.
> Even if the investors funding this construction all go bankrupt, the data centers don't just vanish: someone will buy them from the bankruptcy court and operate them.
There are really three questions to answer on the long term:
1. Is an arbitrary service profitable enough to cover opex AND burn down the massive capex financing? A few companies might have enough magic to survive, but I suspect a lot of them go to the bankruptcy fire sale you mention.
2. Is the service profitable enough to cover just opex? The second generation vendors may have a somewhat more favourable cost structure, but there's still a floor. Electricity and cooling on a datacentre scale aren't free, and we may find there are still services that can't even meet that threshold. I notice you see far less marketing for consumer video-gen these days...
3. We let those businesses fail. What does the long term market look like at that point? You have a snapshot in time of paid-for datacentres, but the market shock probably shot a huge torpedo into long-term R&D and planning. Vendors will scale back to refocus on markets that are more vital (maybe selling RAM and GPUs back to end users :) ), and create an era of stagnation. AI becomes a utility, in the sense of "boring, predictable industry that growth investors have very little enthusiasm for".
There is also the risk that the data centres could vanish. This is a very hype-dependent market: we've spent the last 5 years going from "novelty" to "productivity tool" to "weapon of mass destruction" to keep people excited.
Eventually there's a put-up-or-shut-up moment. What happens if we get to 2035, we've hit a plateau for 3 or 4 years, and they can't deliver Lieutenant Commander Data, or even Kryten? All we really have to show for the effort is a bunch of old GPUs, displaced workers, and the ugliest restaurant fliers you've ever seen. It's a big loud target for a bunch of frustrated and let-down people. We're already seeing moratoria getting political legs, could permit revocations or forced decommissioning be far behind?
I suspect this is part of the reason some players are happy enough to present AI leadership as a New Cold War/US v China thing-- pull the national security card to try to shut down a debate on the actual utility and desirability of the actual product.
Let's say that tomorrow every company selling AI inference goes bankrupt.
Investors are devastated (because they lost most of their investment). They
loudly proclaim, "I'm never investing in AI ever again, and no one else should
either." Even if no investor anywhere is willing to build a new one, there
would still be about half a trillion dollars worth of recently-completed AI
data centers. New investors are going to buy them from the bankruptcy courts
(which is how the devastated investors get some small fraction of their
investment back). Because most of the cost of providing AI inference is the
cost of the GPUs, it costs only about one tenth or one eighth of what it cost
to build it to operate an AI data center over the lifetime of the data
center. (Most of the cost is electricity.)
>Is an arbitrary service profitable enough to cover opex AND burn down the
massive capex financing?
The new investors that buy the already-built data centers for pennies on the
dollar have no obligation to "burn down the massive capex financing". The
obligation to "burn down" or service any debt was erased by the
bankruptcy judge. This is how bankruptcy works and has always worked.
As you said, you still have 1/8 or 1/10 of the overall costs in OpEx to cover.
You still have a lot of services of such marginal utility that they'd have a hard time justifying their opex costs alone.
I think part of this is that we have a lot of firms offering features rather than products. I fully expect most of the third-party "meeting minder" and "writing improver" services will be replaced by versions built into the platforms themselves, for example.
The HFT industry creates an investment opportunity that's somewhat detached from business fundamentals. If you can trade on the volatility alone, you no longer need to pick investments based on their future prospects and behaviour.
There's a huge spec (Level III data) that lets the merchant provide full line item detail with every credit card purchase.
Corporate cards get it, but consumers don't. The payment ecosystem often doesn't even submit it if the card isn't identified as a corporate card.
If you had that, you'd probably need way less "smarts" to build budgeting tools-- no need to try to reconcile bank records with order reciepts.
reply