I think what I'm really looking for is a model like GPT Rosalind, which has been steeped in "science" post-training, but with more randomness. Like, I think I'm looking for a GPT Mullis--Rosalind Franklin was careful, deliberate, and serious; Kary Mullis ate a bunch of acid, drove on the PCH, and invented PCR. Like, we need to invent psychedelics for LLMs. Some way to let them relax their weights, explore new pathways, and come up with some absurd ideas by connecting random ass dots from the natural world. I want the model to say, "hmm, that's weird". This isn't just changing the temperature, we're missing something deeper.

I think frontier science has always come from serendipity: a bright thinker, listening to a presentation after having stared at some small experiment, but having trained for years on "biochemistry" so the foundation and loose guardrails are there.

I don't know, I'm feeling adrift. Does this resonate with anyone else?

But by far the most critical part of getting CPAP to work was using an open source program to review the SD card raw data every morning, learning the interplay between max/min pressure, temperature, humidity, and ramp. Gaining access to the admin controls on my AirSense and fiddling it myself over weeks to find the right settings--this was the only way I made it work. The idea of waiting 6 months to make a change to the pressure was bonkers to me. Especially, for some reason, my brain is shit the next morning if I over oxygenate with too high max pressure overnight, worse than not wearing it at all!

But once you dial in the settings. It's magic.

As you can imagine, maintaining entropy for all these disparate systems takes time, of which I have less now that I have young kids. So when Claude Code was released, I took to it like a fish to water. We mapped my entire network, I created accounts on all the devices so it can SSH into everything and configure everything (including the Ubiquiti Dream Machine Pro!). I have been blow away at how well it troubleshoots and fixes everything.

I have a DGX Spark AI workstation (128gb of memory), and I really want to now hand off the work to a local model, either using Opencode or Claude Code harnesses and simply pointing at a vLLM instantiated model accessable by API (just point Opencode or Claude Code at the local IP and API endpoint).

It works, except I tried Qwen3-coder just now and it's refusing to help due to security concerns. Ugh. I then tried GLM-4.7-Flash, but vLLM doesn't support yet and so before I rebuild (ask Claude Code to rebuild and deploy) to try GLM.4-7-Flash with some other inference provider, does anyone have a model they use for infrastructure maintenance that isn't a little bitch? I will probably eventually go to an abliterated model if none of the open source ones will help.

Comments URL: https://news.ycombinator.com/item?id=46690846

Points: 6

# Comments: 3

It works, but it's slow. Much more like set it up and come back in an hour and it's done. I am incredibly impressed by it. There are quantized GGUFs and MLXs of the 123B, which can fit on my M3 36GB Macbook that I haven't tried yet.

But overall, it feels about about 50% too slow, which blows my mind because we are probably 9 months away from a local model that is fast and good enough for my script kiddie work.

Could you provide a few reputable clinical outcomes studies to support your statements? I was unaware of these risks.

Source: PhD in RNA modifications

[1] https://pmc.ncbi.nlm.nih.gov/articles/PMC9073955/