unixmonks - Nostr Hypermedia

Been playing with gpt-oss:32B on a friends 16GB Tesla V100. Quite usable actually. We've got another 32GB V100 on the way, will be interesting to see how good of a model we can run.

2025-12-06 09:50:18 from 1 relay(s) View Thread →

vibe right now

2025-12-06 04:44:04 from 1 relay(s) View Thread →

Those of us that like to tune our open operating systems, our open desktops, our open IDE's, our software, are many steps ahead of those running proprietary operating systems and tools. The same will happen with AI. Once we start tuning our own models on our own clusters to suit the types of work we are doing, we will be many steps ahead of those using proprietary systems.

2025-12-05 10:20:09 from 1 relay(s) View Thread →

I'm seeing a lot of potential for developing bespoke software for small, medium and even large businesses. No longer do they have to pay $60k/year for a piece of generic software that doesn't quite fulfil their needs and leaves inefficiences in their business processes. Now a custom piece of software can be developed rapidly to close the gaps.

2025-12-05 10:08:51 from 1 relay(s) View Thread →

If you're a dev or engineer with any sort of pre-AI coding experience, business experience, and any entrepreneurial drive, you are well positioned to take advantage of this paradigm shift that has come with these AI models and tools. I'm getting late 1990's, early 2000s vibes where it felt like endless possibilities.

2025-12-05 10:01:56 from 1 relay(s) View Thread →

I vibe coded a Signal Messenger LLM bot a few months ago. I used a sqlite database to remember chat history (and respect dissapearing messages), and a vector database for searching the chat history. Basically a RAG bot for chat history. It's quite good how it is, but could use some work. It can also recognise images that users post and comment on them. It is used by me and my friends in our chat groups and ads a comedic element. It's also handy for summarizing chat history if you missed out on discussions. I was going to clean up the code before I open source it, but I think I'm just going to put it out there. Stay tuned.

2025-12-05 09:48:41 from 1 relay(s) View Thread →

When I can get one of these for $5k USD - IM ALL IN.

2025-12-05 09:39:28 from 1 relay(s) View Thread →

I built a job management system for a cable location and vac truck company in 8 hours with claude code for $7k USD. It's 80% of the way there. I haven't touched a line of code. The code it generated is quite good as I have learned to ground it with preferred architecture patterns (I just point it to an existing repo on disk). I build up the project starting with frontend mockups. I keep those mockups in the repo. Then I build out small pieces of the data model, and layer piece by piece on top. Unit tests are also generated as we go and I am careful to look at the tests to make sure it's got decent coverage, giving me more confidence in the code it has built. It can get a little sticky with complicated frontend or backend logic and structure, however it excels at boilerplate. It seems every month it gets better and better. A job that would take 4-6 weeks, now takes under 1 week. This is why we are seeing a boom in AI related stocks. They might be overvalued, but there is some value there and it is significant. One final thought - I don't like proprietary AI due to the privacy aspect. Unfortunately they are one step ahead of the open source models, and that one step results in better code, which is hard to pass up. However, I see a near future where the open source models will be very close, or on par. I'm sure there are people out there already that are doing amazing things with open source models. I for one am looking to build a 8 x nvidia v100 cluster of 32GB cards, which can be obtained from China for decent prices. More on this in the future!