Rendered at 00:06:11 GMT+0000 (Coordinated Universal Time) with Cloudflare Workers.
k1w1 2 hours ago [-]
This is the part that really stood out to me.
> It’s not only new projects putting strain on the grid though. The report found that an estimated 13 percent of US cloud consumption, totaling more than 3 gigawatts, comes from so-called "zombie" workloads—abandoned test environments and unused applications that continue to draw power without doing any useful work.
Containerized sandbox environments for AI can be incredibly wasteful. If those sandboxes are kept available so the user gets sub-second access that is a continual user of RAM, and thus overall computing resources. We built the first version of https://www.aha.io/builder/overview using a typical containerized environment - just like you see with products like Replit - but were appalled at the inefficiency and waste. We rebuilt it from the ground-up to use shared architecture instead with Javascript-level isolation, and almost zero waste. Using shared computing instead of containers means instant startup time, and (almost) zero resource consumption when not active. You still consume disk to store the artifacts, but there is no ongoing RAM or CPU.
I think a reckoning is coming for container-based AI systems too. We are seeing tokens trend towards reflecting the actual cost, and I think the same will be true of containerized runtime environments too.
tadfisher 17 minutes ago [-]
We need reliable and ubiquitous checkpointing, then it won't matter if your workload executes in a container runtime or in a WASM sandbox or even on bare metal.
freakynit 4 hours ago [-]
The Coming ‘Power Wars’ Between Humans and Datacenters
Seems the humans love the AI chatbots though. But I guess they figure they run on magic, not actual electricity and water.
amanaplanacanal 2 hours ago [-]
Some humans do.
Nasrudith 21 minutes ago [-]
The backlash is and always has been incredibly fucking stupid. The same dipshits who want to bring manufacturing back to the US cannot tolerate goddamned data centers when the negative externalities and infastructural strain would be orders of magnitude worse.
> It’s not only new projects putting strain on the grid though. The report found that an estimated 13 percent of US cloud consumption, totaling more than 3 gigawatts, comes from so-called "zombie" workloads—abandoned test environments and unused applications that continue to draw power without doing any useful work.
Containerized sandbox environments for AI can be incredibly wasteful. If those sandboxes are kept available so the user gets sub-second access that is a continual user of RAM, and thus overall computing resources. We built the first version of https://www.aha.io/builder/overview using a typical containerized environment - just like you see with products like Replit - but were appalled at the inefficiency and waste. We rebuilt it from the ground-up to use shared architecture instead with Javascript-level isolation, and almost zero waste. Using shared computing instead of containers means instant startup time, and (almost) zero resource consumption when not active. You still consume disk to store the artifacts, but there is no ongoing RAM or CPU.
I think a reckoning is coming for container-based AI systems too. We are seeing tokens trend towards reflecting the actual cost, and I think the same will be true of containerized runtime environments too.
https://sourceryintel.com/reports/humans-vs-datacenters