drewnick

Thank you for chiming in. I got all of this activated today and am liking it. Team GUI for me.

I use Opus 5.5 for planning, sprinkle in a Fable 5.1 review of the spec (they go back and fourth), then send the spec and plan to my local deepseek 4.1 Flash 1017 to implement. It crunches away for a good while (using ds harness)

Then check w/ Opus and Astra for good measure if it really matters. Works well. DSv4.1 is okay but does make mistakes. I run my 2x DGX sparks at 60-70% capacity this way, which nets about $400/mo worth of tokens on $30 of electricity.

I am hoping in the next six months to be able to in-house ALL of the above using a slightly more competent model, or two more sparks. It is slow running local, but it is really fascinating to have all this work done from my desk.

I thought the TUI was the ultimate power tool and loved the original embrace from claude code. I still use TUI a lot, but have been using deepseek harness web app atop my local model and am already liking the better organization and UI of this interface. I can see my load shifting to it. But you are right, you almost always need a VS Code interface open as well.

The outputs of Opus 5.5 were so much easier to read, but I do find the the deepseek web based tool much easier to comprehend for some reason. Plus I can access it from phone (over tailscale) as well when out and about and continue sessions very easily. AND it is all open source/weights. I'm pretty thrilled, but it's been less than a week.

True, I've had the same IPv4 address since 2018 on Spectrum despite rebooting at least 10 times.

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.

An hour in, I've been running four terminals full bore on my $20/mo Grok sub and I'm at 9% for the week. Codex or Claude would easily have hit 5-hour or weekly limits.

Think about how many webmaster and business owners' egos are stroked by all the traffic they are getting, when in actuality they are often just serving thousands of bots.

I moved one of our daily workflows over to Kimi K3 on Fireworks. It replaces two sales assistant positions, and was about $50/day for 14 million tokens total (in/out). They surely are not subsidizing this price as it is just inference only and other providers are even less money.

Thank you for giving us this nuance.

Wow, this is refreshing DX compared to iterating all messages like we did back in '24.

While all these models are nondeterministic a feature bump is still necessary as the same input can have wildly different output on a new model. For API users being able to pin a model is a necessity.

Great point on the playing with Linux growing up, it's second nature to me now.

I am always feeling like I'm doing something wrong running bare metal based on modern advice, but it's low latency, simple, and reliable.

Probably because I've been using linux since Slackware in the 90s so it's second nature. And now with the CLI-based coding tools, I have a co-sysadmin to help me keep things tidy and secure. It's great and I highly recommend more people try it.

The problem is a lot of this glue is proprietary by design at the various cloud services. I realize there are open source and alternative abstractions for a lot of of the same services, but there’s still quite a bit of glue if you’re on AWS, for example, and looking to move to bare metal.

But maybe I’m just thinking of the current capabilities of agents, and if we fast forward a couple years, even removing these abstractions or migrating will be very low friction.

I too, am bravely using Claude for more DevOps. I run all of my virtual machines on proxmox atop bare metal servers I own and I’m just blown away at how quickly Claude can optimize and set up entire new networks across all of these machines. Truly feels like a coworker or well paid sysadmin.

Agree. If you've ever spent serious time in the country with farmers, the level of ingenuity is impressive among many, and they benefit from it greatly. As the grandson of depression farmers, I noticed intelligence mattered a lot, even if just for survival.

Hasn't Opus 4.5 been famously consistent while 4.6 was floating all over the place?