I'm seeing the same logic being used across the stack that people used to deprioritize training their own models and use "off-the-shelf" solutions. I'm seeing a bigger push for standards or abstractions such as vLLM and the adoption of open source tooling even if that means lower quality.

It's not that prices are going down, it's that people are starting to hedge their bets on datacenter spend, being open to using open source alternatives, or willing to put in the work necessary to leverage non Nvidia hardware, take Google's TPUs, or Amazon's Trainium chips for example.

🎯

I’d bring them back in a heartbeat if BlueSky had audio spaces.

💯

🔥

As minimalist I 🫡

I heard this administration is accepting donations.

Most configuration management systems have no way of expressing dependencies on external configuration, nor have the ability to validate or enforce configuration policies across multiple tools, which creates conflicts that lead to downtime. Take the classic firewall misconfiguration for example.