Comment by grahamlee
9 hours ago
The Sun model of beefy server, ultra-thin client worked incredibly well and presaged the widespread adoption of virtualisation that eventually led to cloud computing and containerised software. Unfortunately that model made it very easy to undercut their prices using commodity hardware; I managed a Solaris/Sun Ray network c.2004 and it was cheaper to switch to Linux/Intel and buy spare hardware to keep in a cupboard “just in case” than to keep the Sun network going through its support contract.
I worked at Sun labs and used a Sun Ray daily for a whole summer. It worked...OK. Definitely not "Incredibly well". We had a network of a few hundred Sun Rays and two E10Ks backing it. More than a million for the servers and $1000k a piece for the Sun Rays was not cheap. Basically any $1000 laptop was more powerful, without the need for a crazy powerful Sun server. The experience of taking out my ID card, walking down the hall to another cubicle and popping it in and getting my session was magical. But I did that less than once a week, so I don't think it was overall worth it. The Sun Rays also had a pretty crappy LCD, about 1-2 years behind the SoTA.
It was a cool idea. It will keep coming back again and again. But at Sun's price points around the early 2000s, it was a loser, and I actually agree with the market in this instance. Sun Ray was a flop.
"It will keep coming back again and again."
It lived on in the world of Windows Terminal Services and things like Citrix - maybe still does in some places.
Citrix is still everywhere you need a client that can do things a web browser can’t—like run ancient VB4 software or talk to device hardware—deployed at scale.
Citrix was bought out by PE for $16.5B a few years ago, under the (sadly correct) assumption they hadn’t been charging their customers enough money for this essential tech.
It lives on in the world of LLMs in a way.
Stadia? GeForce Now? Xbox Cloud Gaming? :-)
Chromebooks…
$1000k ?
Is this how the whole LLM thing will go down once providers raise prices?
i mean, imagine if your a business and suddenly you had to expense > $200 a month for every employee just to run a compiler, or you could just buy a computer and use it for everything else and that as well…
i think llms will be the same.
n=1 but at $work they are forcing us off enterprise and on to personal claude max that caps at $200 and people are like “you don’t need the top end models anyways just use the cheaper one for coding”, which begs the question how long till those “good enough” models run locally (with relative ease) and our company can cut that huge monthly cost with just the macbook pro they are gonna buy anyways.
Maybe - ask me in 20 years...
Nobody knows how the future will be. There are lots of options. There are self hostable open models today - they are not as good, but some report they are good enough. Right now I don't have the hardware to try them (I could - if I'm willing to accept a small number of tokens per minute - useful means more tokens per second as I can get in an hour). However I expect in a few years I'll be able buy consumer hardware with good enough GPUs (or perhaps an AI chip???). Will anyone other than a few hobbyists buy them? time will tell.
And space launches. The Big Names will lose a ton of money and disappear as the tech commoditises.
The moat is Being Able To Do Things No One Else Can and as both hardware and software commoditise that moat disappears.
Open LLMs are already only a few months behind the frontier models and as soon as someone works out that accessible and cheap domain-specific custom training has the potential to be a gamechanger the Big Names will die.
Math puzzles only get you so far.
LLMs are in their Big Compute phase, and it's only a matter of time before that ends - because the intelligence of the frontier models isn't increasing fast enough and providing big enough benefits to make the models more valuable than cheaper, smaller competition.
I think prices will go down not up eventually and everybody will have access to it in abundance. That might make most current AI companies go down of course.
it is often cheaper to buy a spare of something and throw it in a cupboard rather than have a support contract.
I have a bunch of Aruba switches and this is what we do, because it's a lot cheaper.
If you are only buying a single spare that is the case. Support contracts make sense when you would want several dozen spares.
if a support contract is 70% of the cost of something, then for every 2 of something a spare is cheaper.
For Aruba I think it was 1:1; a support contract was the same price as the device itself over 2 years or something.
Support contracts never cover a "site" they always cover individual devices, so your cost scales with the amount you have... so I'm not sure I follow what you're saying..
Unless you mean to imply that you're replacing units more than once every 2 years, in which case of course you're right... but then I might consider going with another vendor entirely because that is a pretty terrible reliability record.
4 replies →
Remember, "The Network is the Computer".
Which is why it is kind of ironic having folks discussing microservices as something groundbreaking.
Java applets!
Counterpoint: it actually sucked. Entire universities lost a day's work when their editors were all fatally stuck in an rpc.lockd bug.
the lab I was at got on this train right around the beginning of the sparc architecture. yellow pages. nfs based swap, nfs based boot. 10Mbit thinnet. we suffered for a few months and then moved to xterms for the cheap seats (which was a much better place to cut services) and went back to buying self-sufficient workstations.
the best was having everyone with a disk cross-mount each other. any failure snowballed into a global failure.