←back to thread

Getting DeepSeek-OCR working on an Nvidia Spark via brute force with Claude Code

(simonwillison.net)
200 points simonw | 6 comments | 20 Oct 25 17:24 UTC | HN request time: 0.001s | source | bottom
1. amirhirsch ◴[22 Oct 25 04:38 UTC] No.45664936[source]▶
I also use Claude Code to install CUDA and PyTorch and HuggingFace models on my quad A100 machine. Shouldn't feel like debugging a 2000s Linux driver.

HuggingFace has incredible reach but poor UX, and PyTorch installs remain fragile. There’s real space here for a platform that makes this all seamless maybe even something that auto-updates a local SSD with fresh models to try every day.

replies(1): >>45665698 #
2. nicman23 ◴[22 Oct 25 06:55 UTC] No.45665698[source]▶
what. i mean firmware has a lot of failsafes but i would not trust it to do sysops for a 100k machine
replies(2): >>45665897 #>>45667818 #
3. smokel ◴[22 Oct 25 07:32 UTC] No.45665897[source]▶
It is indeed a disaster waiting to happen. Figuring out a way to run agents safely is apparently a harder problem than getting to AGI :)
replies(1): >>45667771 #
4. qwertytyyuu ◴[22 Oct 25 12:02 UTC] No.45667771{3}[source]▶
thus is the AI alignment problem
5. amirhirsch ◴[22 Oct 25 12:06 UTC] No.45667818[source]▶
it’s all running inside a Proxmox VM with IOMMU and GPU passthrough. It’s as safe as doing the same on any cloud system.

Also the machine is well north of 100K when you include the RF ADCs and DACs in there that run a radar.

Worst case, I have multiple.

replies(1): >>45691095 #
6. nicman23 ◴[24 Oct 25 05:28 UTC] No.45691095{3}[source]▶
iommu does not matter ? the issue would be if one of those a100 fries.