icedchai

Yep, stuff is more likely to "just work" on NVidia. Example: pytorch

In most benchmarks, the Spark is also faster at the prompt processing / prefill phase.

I have a AMD Strix Halo box I use for similar dabbling. It definitely wasn't an "out of the box" experience, fiddling around with kernel versions and ROCm installs. These days I mostly wind up using the Vulkan build of llama.cpp for inference.

AI documentation is also often incredibly verbose. It makes it difficult to see the forest through the trees. I am often repulsed by the README.md in many "AI generated" projects.

Should I have said "artificially intelligent" enough?

They probably didn't need Salesforce in the first place. Salesforce is often bought for the ecosystem: third-party apps and integrations. In a past life I did some Salesforce development and it's like its own world.

Heh. I did try this an experiment. Funny stuff. It was persistent enough to kill it again after I restarted it, too. "Let me try again and be ready for the consequences." We're doomed.

Presumably it's intelligent enough to realize that its own existence (power, communications, other infra) won't last long after the bombs drop.

Vendor terminology is often different (tagged VLANs vs trunking, etc.) but, in general, it is close enough that an experienced person can figure it out with a bit of googling. I've built networks with Cisco, Mikrotik, Aruba/HP, custom Linux routers, and others. Mikrotik VLANs are a bit weird, for sure.

In today's world, is generally easy to get things to communicate. It's not 30 years ago and nobody is running DECnet, Appletalk, or IPX.

I've heard it referred to as "AT command protocol." I googled it myself and this terminology is common enough.

I've worked with developers that don't even understand that the "form" tag exists and browser can send a request without any javascript. At some point in the early 2010's, perhaps even before, fundamental understanding disappeared. This makes it difficult to introduce alternative technologies like HTMX into a larger project.

I knew guys who panicked in Feb 2020, at the start of covid. They moved everything to cash, never got back into the market. Things recovered faster than they thought. The unfortunate truth is they would've more than doubled their money if they stayed invested.

I set up Asterisk systems for a few small businesses in the mid-2000's (5 to 50 extensions.) It was seriously fun working with that stuff. These days I'll do almost anything to avoid phone calls.

Yep. I’ve worked on some federal contracts and all the waste is in private contractors, not the government itself. There are structural issues with subcontractors many layers deep, each middleman taking a cut for doing little to nothing. It’s sick.

I had a Zoom 2400 baud modem (with MNP5 compression!) back in the 90's.

Same. If I move the window around, I can perceive the "Written in Ghost Text"... Otherwise, not at all.

8 months ago you could get one of these for ~$2K. Well, not this exact item, but systems like the Framework Desktop with the same CPU and memory configuration.

It was a good deal for $2K. For $4K? I'm not so sure. Of course, most of the cost is RAM, so what can we do?

Okay, I guess the only solution is to always use FQDNs.

What you describe is a user and resolver configuration problem. There are 100's of TLDs and there's always a chance they will conflict with subdomains, either now or in the future as new TLDs are created. I've been using both "int" and "dev" as subdomains since at least 2000 and never had an issue.

I do the same thing and have never had a problem. Maybe I’ve just been lucky for 25+ years. Some hosts have a search path of “int.example.com, example.com”. Others are just “example.com”

He’s using it as a subdomain.

Do the names resolve to publicly routeable IPs? If not, I wouldn't worry about it.

This is similar to what I do, except I have my own authoritative DNS servers instead of Cloudflare.

I'd prefer this over split DNS, any day.

I have several AI-content posters in my feed. Many of these are folks that previously used the term "thought leadership" non-ironically. I guess Claude and ChatGPT are the thought leaders now...

Yep, it definitely wasn't an "out-of-box" experience. This was on Ubuntu 25.10. Maybe other distros are better?

Also, when you do Python work, you have to remember to install dependencies like pytorch from an alternate repo, otherwise you wind up with the CUDA versions that only use the CPU and not the GPU through ROCm.

The "ROCm" situation with Strix Halo was pretty bad for a while. I think it finally stabilized late last year. You needed the right combo of ROCm, Linux kernel, and kernel firmware for it to work reliably.

Whenever I rebuild llama.cpp, I wind up using the Vulkan build anyway.

They stopped selling anything greater than the 96 gig variant a while back. If you want more RAM, be prepared to bid high on ebay.

And it was "only" $1700 last year!

Don't you need a new CPU to support more channels? It doesn't sound like a cheap, trivial change to upgrade my Ryzen to a Threadripper.

In the alternate reality where this happened, wouldn't the price of DDR4 still be sky high? We'll ignore any costs for CPU, chip set, and motherboard redesign. You're just pushing the demand somewhere else.

Yep, the only reason I bought mine (in late 2025, before hardware prices went totally crazy) was because it was half the price of a Spark. I spent a while fiddling around with the right Linux kernel, kernel firmware, ROCm installs, etc.