I mean, he has a point. Hardware is used way more efficiently in the data center, and any sort of deficit is best resolved in a centralized system due to more efficient resource allocation. The same setup could be used like 2 hours a day for your personal needs, or 24 hours a day to serve many people.
THAT SAID, their greed and unsatiable desire to make the number go up by hooking us into subscription economy with ever rising costs is what lead us here and it leaves us with little choice but to accept the less efficient way and host these things ourselves.
Remember the times not so long ago when self-hosting was an enthusiast hobby instead of necessity? Barely anyone non-geeky bothered with it, because what existed served people well. Now, having to build our own services remains the only sane choice, and the only way out of the nightmare Sam, among others, has built with his own hands.
Hardware is used way more efficiently in the data center, and any sort of deficit is best resolved in a centralized system due to more efficient resource allocation. The same setup could be used like 2 hours a day for your personal needs, or 24 hours a day to serve many people.
Only for a rather narrow definition of efficiency, though. For raw compute power, yes, but you’d also be adding a whole bunch of overhead from all the network traffic and routing that would be necessary for that sort of thing to work in the first place.
If people used their desktops like in the old days where you’d connect to a mainframe or master computer, you’d end up needing computing hardware on both ends anyway, since the user side would need to be enough to be able to connect and send all the interactions to the master computer. At that point, you could just buff up the client computer, and do all that in the user’s home.
There’s also an argument that some of the basic problems would still apply anyway. A data centre does not fully utilise 100% of its capacity. There’d always be some reserve to have failover if a server dies, or there’s a sudden spike in load, like with a home machine.
In theory it can be more efficient but the consumer market never was going to contribute to the scarcity, and the bigger problem is the “must build AI datacenters” by all comers whether or not they actually have a plan to bring up and utilize.
Altman kicked off a corporate hoarding storm by committing half of the world’s DRAM manufacturing supply, then every other corporate had to get in on it.
The result according to a report I read is that over half of the technology delivered to these folks are sitting disused. Because they couldn’t get the right power to the datacenter, or maybe they did get all that lined up but don’t have the business relationship with whomever actually would use the capacity, or they otherwise botched things so their setup is not viable.
The far bigger culprit is the bad datacenter buildout. Vilifying the consumer market is just a better narrative for his agenda.
It doesn’t fulfill the same user case. One can be used offline, the other cannot. One can be used without a subscription, the other cannot. One needs to be running constantly to meet targets, the other can be run whenever I want. One has that memory being used by very specific services, the other has that memory being used for whatever I want to use it for. Why should we care about an efficiency factor that really only benefits Sam Altman and ignores all other metrics?
If you think he has a point, I think you are being dazzled by the inherent proposition of an implementation in the ideal egalitarian society, which we do not live in nor is he really interested in providing.
The problem is that for less than a year’s subscription I can just buy my own equipment. So yeah it would be more sensible to centralise everything, but as you say the prices are too high.
I keep saying this, AI is not commercially viable yet, it should still be in the lab and they shouldn’t be trying to sell it as a product. Maybe in a few years they’ll be able to distill that down to something that will run on a reasonable amount of hardware and subscription prices can become sensible. There’s no way in hell I’m giving them $200 a month to automate something that I do actually know how to do myself. You can still run out of tokens despite giving them $2,400 a year, so where’s my incentive to not just go and buy a Mac mini?
Hard disagree. It makes sense for certain use cases and not for others, like when you are near to a large pool of data to process, or running latency sensitive processes, or want security around the data being used. None of those use cases are going away.
I might argue with centralized systems being the only ones that allocate resources most efficiently. You can have nearly same efficiency with decentralized models, except they are more complicated to implement, but the diversification is often worth it.
In another timeline where RAM and GPUs were still largely consumer accessible, I wouldn’t doubt there would already be a SETI-like decentralized alternative for LLMs.
I don’t know about that, mostly because the structure of LLMs at the moment make them incredibly bandwidth-hungry, and also quite sensitive to latency.
You do have some small-scale versions of that, where people link a few smaller computers together to run a larger model, but that usually exists in the form of things like connecting machines via USB4/Thunderbolt, or Oculink, which are much faster and lower latency than a typical internet connection.
There is something a bit like that, with the Kobold AI Horde, but that is still people hosting an entire model and sharing that over the internet, rather than something like Folding/SETI@Home, where you’re doing part of the actual calculations yourself.
I was actually thinking more like participants who can offer computational power during periods of no use, accrue reputation the more they contribute, and can use that reputation to offload computation to the cloud. If your PC was not powerful enough to run local LLMs, it would need to contribute a lot more, or perhaps it would just be limited to clouds of people running powerful enough GPUs for the local LLMs being used so they could get an added benefit to leaving it running at night.
An LLM model running over across multiple PCs across the Internet would not really be workable, no, but their product is.
I mean, he has a point. Hardware is used way more efficiently in the data center, and any sort of deficit is best resolved in a centralized system due to more efficient resource allocation. The same setup could be used like 2 hours a day for your personal needs, or 24 hours a day to serve many people.
THAT SAID, their greed and unsatiable desire to make the number go up by hooking us into subscription economy with ever rising costs is what lead us here and it leaves us with little choice but to accept the less efficient way and host these things ourselves.
Remember the times not so long ago when self-hosting was an enthusiast hobby instead of necessity? Barely anyone non-geeky bothered with it, because what existed served people well. Now, having to build our own services remains the only sane choice, and the only way out of the nightmare Sam, among others, has built with his own hands.
Only for a rather narrow definition of efficiency, though. For raw compute power, yes, but you’d also be adding a whole bunch of overhead from all the network traffic and routing that would be necessary for that sort of thing to work in the first place.
If people used their desktops like in the old days where you’d connect to a mainframe or master computer, you’d end up needing computing hardware on both ends anyway, since the user side would need to be enough to be able to connect and send all the interactions to the master computer. At that point, you could just buff up the client computer, and do all that in the user’s home.
There’s also an argument that some of the basic problems would still apply anyway. A data centre does not fully utilise 100% of its capacity. There’d always be some reserve to have failover if a server dies, or there’s a sudden spike in load, like with a home machine.
In theory it can be more efficient but the consumer market never was going to contribute to the scarcity, and the bigger problem is the “must build AI datacenters” by all comers whether or not they actually have a plan to bring up and utilize.
Altman kicked off a corporate hoarding storm by committing half of the world’s DRAM manufacturing supply, then every other corporate had to get in on it.
The result according to a report I read is that over half of the technology delivered to these folks are sitting disused. Because they couldn’t get the right power to the datacenter, or maybe they did get all that lined up but don’t have the business relationship with whomever actually would use the capacity, or they otherwise botched things so their setup is not viable.
The far bigger culprit is the bad datacenter buildout. Vilifying the consumer market is just a better narrative for his agenda.
It doesn’t fulfill the same user case. One can be used offline, the other cannot. One can be used without a subscription, the other cannot. One needs to be running constantly to meet targets, the other can be run whenever I want. One has that memory being used by very specific services, the other has that memory being used for whatever I want to use it for. Why should we care about an efficiency factor that really only benefits Sam Altman and ignores all other metrics?
If you think he has a point, I think you are being dazzled by the inherent proposition of an implementation in the ideal egalitarian society, which we do not live in nor is he really interested in providing.
The problem is that for less than a year’s subscription I can just buy my own equipment. So yeah it would be more sensible to centralise everything, but as you say the prices are too high.
I keep saying this, AI is not commercially viable yet, it should still be in the lab and they shouldn’t be trying to sell it as a product. Maybe in a few years they’ll be able to distill that down to something that will run on a reasonable amount of hardware and subscription prices can become sensible. There’s no way in hell I’m giving them $200 a month to automate something that I do actually know how to do myself. You can still run out of tokens despite giving them $2,400 a year, so where’s my incentive to not just go and buy a Mac mini?
ye if he wasnt the ceo of useless tech that hurt ppl id agree
Hard disagree. It makes sense for certain use cases and not for others, like when you are near to a large pool of data to process, or running latency sensitive processes, or want security around the data being used. None of those use cases are going away.
I might argue with centralized systems being the only ones that allocate resources most efficiently. You can have nearly same efficiency with decentralized models, except they are more complicated to implement, but the diversification is often worth it.
In another timeline where RAM and GPUs were still largely consumer accessible, I wouldn’t doubt there would already be a SETI-like decentralized alternative for LLMs.
I don’t know about that, mostly because the structure of LLMs at the moment make them incredibly bandwidth-hungry, and also quite sensitive to latency.
You do have some small-scale versions of that, where people link a few smaller computers together to run a larger model, but that usually exists in the form of things like connecting machines via USB4/Thunderbolt, or Oculink, which are much faster and lower latency than a typical internet connection.
There is something a bit like that, with the Kobold AI Horde, but that is still people hosting an entire model and sharing that over the internet, rather than something like Folding/SETI@Home, where you’re doing part of the actual calculations yourself.
I was actually thinking more like participants who can offer computational power during periods of no use, accrue reputation the more they contribute, and can use that reputation to offload computation to the cloud. If your PC was not powerful enough to run local LLMs, it would need to contribute a lot more, or perhaps it would just be limited to clouds of people running powerful enough GPUs for the local LLMs being used so they could get an added benefit to leaving it running at night.
An LLM model running over across multiple PCs across the Internet would not really be workable, no, but their product is.
I think you mean effectively not efficiently, the resources use is the same if I local host or use a provider.
No
Care to elaborate?