Hacker News

points by Zambyte 2 days ago | 0 comments

How many months do you plan on spending hundreds of dollars a month? You should probably just own your compute. It'll probably cost you less, and you'll have no concern over who has access to your chats.

[–] jitl 2 days ago

$200/mo @ 1 year = $2400 2400 gets me about 1/2 of a 128gb NVIDIA DGX Spark, so i could do a year of nonstop inference of (checks notes) quantized Qwen 3.8-27B, a model i would never choose to use. yay, such savings.

[–] lp92 1 day ago

And that Spark isn't running a top end frontier model. Yes, the open weights smaller models that can fit on Spark are getting better, but they are still quite a way off in performance depending on use case.

[–] spenvo 2 days ago

I don't know, but I do expect to eventually switch over. That said: the upfront time to setup (and continued fiddling) and cost of a high-end local compute system represent major investments, when I could just... keep being insanely productive with frontier AI in CC. And models/harnesses can themselves be security liabilities (see Mistral, LiteLLM, etc recent supply chain attacks). So I'm personally waiting for the local AI scene to mature, so I can be more precise about how I take the plunge. I will say: if the subscription pricing ever goes away or materially changes for Anthropic, I'm headed straight to order whatever the best local setup is, within my budget.

[–] pixl97 2 days ago

A single video card with any amount of ram is in the thousands of dollars, and you'd need a stack of them to replicate what the providers performance.

Running your own compute has a lot of risks too, you have to keep it secure against attacks on the models that you're running. Giving away all your chats to hackers because you don't secure your own stuff isn't really winning either.

[–] oblio 2 days ago

How much does it cost to run your own hardware such that the service is comparable to Claude? Similar speed, similar features, web chat access, CLI access?

[–] Zambyte 2 days ago

How much do people actually do that demands the capabilities of Claude instead of open models that can run on hardware in the range of thousands of dollars (which is not that much when the alternative is hundreds of dollars a month, indefinitely)?

People chasing the best of the best models for all tasks in order to "keep up" are only wasting resources at this point.

Regarding similar features, there are lots of off the shelf harnesses that support all that, or can easily be extended (literally just ask your agent) to do it.

[–] jitl 2 days ago

i am 100% sure i need opus 5.5 model to build stuff like this because i try lesser models because i want my tokens to go farther and lesser models simply cannot: https://fujimod.com

[–] lp92 1 day ago

Not sure if this is possible on the Fuji, but back in the old days of the Canon PowerShot being very relevant, there was CHDK firmware which ran off the SD card on boot. No physical impact to the device and fully reversible by reformatting the SD card. This would be the way to go to reduce bricking chances.

[–] jitl 23 hours ago

yeah i currently only distribute sd card dual boot media where the the install process is “turn it on holding REC and pressing rear command dial” and uninstall process is “turn it off”

[–] Zambyte 2 days ago

> All of this is unreviewed vibe-coded slop that could be hazardous to your camera’s mental, emotional, or digital health. If you're uncomfortable with risk to your camera, please don't use it.

You don't care anyways. Actual engineering works "100%" fine with small models now. You should try it!

[–] lp92 1 day ago

I've tried smaller models and the best open models for a specialized UAV focused CAD + CFD + DFM optimization application I'm building.

The best open weights models still had a lot of defects and very egregious mistakes to fix even after I provided a very well detailed Requirement/Design/ Technical Implementation plans.

Opus and Gemini 3.8 (and now Argon) did a significantly better job. So I've baseline using these two models to build this application. I now use Opus for design/review and Flash 3.8 for the code gen/implementation work.

I'm sure small models and even open big models are probably fine for many common small apps and such, but they aren't really able to perform the same across all use cases.

[–] jitl 2 days ago

as i said, i did try it. smaller models are not up to the task of deep reverse engineering work.

[–] tekchip 2 days ago

Plenty of packages do most or all of that. OpenCode does most of that. It's replaced Claude tools for me.

[–] oblio 1 day ago

OpenCode offers a web UI?

[–] tekchip 1 day ago

https://opencode.ai/v2/docs#web

[–] oblio 22 hours ago

That documentation is extremely sparse. Can I run that 24/7 and access it anywhere around the world? Is it secure?