Digital mason. Build (mostly) web things. I follow technology, product, science, work-life balance, mental health. Food, sports and politics occasionally.
Business Insider confirms it: Amazon has warehouses of GPUs that can’t be installed, and is buying them so far in advance that a supply chain team said they “didn’t have a clear forecasted path” to installing them.
https://t.co/Jzc1vPekGx
The duality of man:
Me in the morning: LLMs are amazing, look at what they can do! This is the future.
Me in the evening: this POS can’t follow instructions which are explicitly written. How the hell am I supposed to edit the crap it come up with.
@QuinnyPig So you’re telling me I should store them as TXT records in Route53 instead:
> A hosted zone includes up to 10,000 records. For each record greater than 10,000 per hosted zone, you will be charged $0.0015 per month.
With a little bit of steering, was able to get this tool out pretty quickly - https://t.co/cxJJ9d60nh. These type of specialized tools are particularly well suited for LLMs since the surface area they have to deal with is limited.
For all the buzz around Qwen 3.8 27b, I've been very happy running deepseek-v4-flash thanks to @antirez's work on getting DwarfStar to run on Strix Halo. For it's size, it's been great at coding tasks, as long as you're ok letting it run in the background.
I’m what I’d call a conservative user of LLMs. I try to do targeted, step by step work with coding agents. All I can say is when providers stop subsidising tokens, I need a second job to pay for my very conservative token usage.
@antirez So just figured that DSpark was being loaded, but temperature=0 wasn't being set explicitly by pi. With that set, I see a considerable drop in both prefix and generation (~8t/s vs ~13t/s without dspark). DSpark stats are logged when the server exits, so have to debug a bit more.
@antirez Will ROCm be eventually supported? I just tried with the latest commit in the ds4 repo and see no difference with dspark enabled on my Strix Halo.
@ponnappa Sure, but maybe the AI companies should stop overselling and hyping their models if they don’t want so much flak for slop. They are selling a tool, but it’s being marketed as a AGI which we know it isn’t.