drewnick

Given how well LLMs do with code, I've been shocked and how poorly they do with sheet music.

I use deepseek via fireworks using Claude code.

When I type deepseek on my CLI it launches Claude with my fireworks API key. Works very well and feels almost free.

The dark fiber comparison is very close. I remember contracting bandwidth with Global Crossing in the late 90s/early 2000s when they were HOT and two years later they were bankrupt.

I've been thinking the same way. I bit the bullet and purchased two DGX sparks and am running GLM 5.3 Flash at a reasonable quant and it's quite amazing seeing this "brain" on my desk do real work and create code without anything leaving the building. But look what happened with hosting, it got more and more centralized over time. In the early 2000s there were more companies self hosting than now I'd bet.

For the types of basic business tasks my company does, we have hit the line where if it never gets any better, we are fine. On our own hardware. For $25k worth of DGX Sparks, we have essentially the output of a few admin-level FTEs.

Agree but if you don't want to then be at the mercy of the cloud providers when everyone is rushing for the exits, you need to start thinking about securing hardware now.

I have deepseek agents doing email responses, with real tools (think running quotes, gathering info, scheduling things) and running business processes that used to be done by $35/hr administrative type people. And the capabilities are expanding every day as I learn how to build scaffolding around the model.

Every new generation of model goes and cleans up the slop of its predecessor in my code bases, and it has turned out to be quite effective.

Last year I held off on implementing a few features knowing that a model like Opus 5.5 was around the corner. I'm now implementing them in a much more efficient and quality manner than I could have fall of 2025.

tello gets you close in the US, a few dollars a month minimum ($5 maybe?) I use it on my burner flip phone.

Thanks for sharing. I've had 2x DGX sparks in my cart a few times. Yesterday and today, there was some weird nationwide run on them (microcenter, amazon, staples, you name it) and so I've got FOMO for my company now.

I appreciate the thoughtful walkthrough.

You can even tell the new harness to go find your latest session in the other harness and it works perfectly. You don't even need the name!

I use it on fireworks which is US/ZDR and pretty reliable. We run a few hundred million tokens/day through it for dollars. Many are cached, which is super duper cheap.

The KV cache is a great example. When you own hardware, vllm etc can keep your KV cache warm indefinitely for bursty large token loads vs clouds will purge quite often.

One of my tasks has a very large but very static system prompt and instructions, think something like 128k tokens. On our own hardware, we keep that cached, sprinkle in the 8k of goodness needed for output, and DSV4 Flash 0731 can generate our output at ~50 tps on $10k worth of hardware.

I've been benchmarking vs publics clouds (which are tbf insanely cheap!) and it's basically a break even in 24-36 months if nothing changes, which it could for better or worse.

Combined with the security and stability of internal hardware, to me it's in "no brainer" territory for this workload, even if models don't improve.

Our only "risk" is better cheaper hardware or cloud costs, which given the trajectory is murky at best short of a bust.

AMD AI Halo machines are similar cost and RAM. I'm in the market and comparing both. The new 192GB AMD box will be tempting but I'm leaning 2xDGX's for now.

+1 I have found extremely reliable systems require a mix of deterministic "adapters" is a good word for agents to actually get through the workflows I've created. I'm still amazed that it can work with both those and some "intuition" to bend the rules around the adapters if prompted.