Alexander Wolf

Learning AI 👨‍💻 | Creator of | Twitter Banners (on hold) @Gumroad | FindYourNextFavorite YT (on hold)
#Tech #Dev #BuildInPublic

[Alexander Wolf @AWolf81]

I've created a Github repo to run Nemotron 3 Super 120B with @runpod flash.
It took some iterations but inference is working - still room for improvements. Cold start takes some time because of model size. Performance is solid with 80 tkns/sec. #AI #nvidia #dev #agents

[Alexander Wolf @AWolf81]

@markproduct Let's connect. 🤝
Heavily working on yDecide right now. Planning to open for public beta soon.

[Alexander Wolf @AWolf81]

This is early software & also works locally in Docker - not just on Runpod.
I'd love to hear your ideas/opinions & what's missing?

You can also try the following link to test the template. (if it is not working - follow the Gist setup)
https://t.co/fkcixnPzP9

[Alexander Wolf @AWolf81]

🚀 Just published my first @runpod Community Template - Hermes Gateway on Runpod!

If you’ve been looking for an easy way to run a Hermes gateway in the cloud, this one’s for you.

All you need is to follow the steps in this guide
https://t.co/4aJDAGhX40
#ai #hermes #agents

[Alexander Wolf @AWolf81]

Linkedin post with all relevant links https://t.co/gqzBWNwyQV

[Alexander Wolf @AWolf81]

@NERDDISCO @paulabartabajo_ What parser is needed to get the tool calls to work at your setup? I'm trying to setup a runpod template with hermes and vllm - also hit the parser topic right now & I have to create a tool parser. I think the section I have to check for vllm is here: https://t.co/1hGjX6DNZp

[Alexander Wolf @AWolf81]

https://t.co/HUoTHCyGvB

[Alexander Wolf @AWolf81]

@runpod thanks for featuring the Runpod Flash Nemotron repo in your newsletter. Really appreciate it. 🙏

Other newsletter topics sound great. Need to check them:
Esecially the transformer lab and the ComfyUI on Runpod are great news.

Link to the repo 👇

[Alexander Wolf @AWolf81]

@max4c_ @runpod where can I find the newsletter? I'd be interested to subscribe.

[Alexander Wolf @AWolf81]

I've created a Github repo to run Nemotron 3 Super 120B with @runpod flash.
It took some iterations but inference is working - still room for improvements. Cold start takes some time because of model size. Performance is solid with 80 tkns/sec. #AI #nvidia #dev #agents

[Alexander Wolf @AWolf81]

@max4c_ @runpod Sure, sounds great. Still some work to do. But tested the endpoint with OpenCode and it was working with Toolcalls. I think Runpod Flash is helping to get a Pod up and running pretty fast. Model loading and llama-server build took some time to get it right.

[Alexander Wolf @AWolf81]

@vyviid @runpod Not sure, I think my problem is the model load time from disk to RAM. Or could it be ctx-size-related? Loading the 70GB takes quite a long time. Also, I couldn't get FlashBoot to work. If it would take 8 minutes just once, that wouldn't be an issue. Need to retry FlashBoot.

[Alexander Wolf @AWolf81]

You can find the repo link at my LinkedIn post https://t.co/IFyPG9TBTU

[Alexander Wolf @AWolf81]

Created a @claudeai skill based on the Universe of Thoughts framework / paper.

You can get the skill in the link below.
It changes reasoning of Claude because it follows a different reasoning schema.

IMO great for new buisness ideas.

#dev #AI

https://t.co/6JC3MQqtnm

[Alexander Wolf @AWolf81]

@maxrojasdelgado @Hostinger I'm pretty new to n8n and haven't used it that much.

I once wanted to use it in a sw prototype but stopped using it because of their license.

Can you share a link to your waitlist tracker or add it to a Github gist? Sounds interesting.

[Alexander Wolf @AWolf81]

Self-hosted n8n at @hostinger with Dockge (VPS, KVM 1 for 5$/Month) in Docker container.

It is a scraping automation that generates daily summaries of AI papers with Gemini AI and sends them to your inbox.

👇 Link

[Alexander Wolf @AWolf81]

Sure, you can also use the automation in the N8N cloud but I wanted to try self-hosting n8n for personal experiments.

If you'd like to know more about the setup, I can add the Docker-Compose.yml to a Github Gist.

🙏Thanks @ Maxim Osipovs
https://t.co/1nSFK6Sspm

[Alexander Wolf @AWolf81]

My "work in progress" is on a feature branch - latest version from screenrecording not pushed yet.

The simple-agent CLI is a learning experiment and an usage example for simple-rpc-ai-backend.

Still a lot to test - especially skill sandboxing.

https://t.co/o2c0kijM8i

[Alexander Wolf @AWolf81]

Progress update for simple-agent. My user-interaction skill is almost working - just need to fix XML parsing, after my usage limit break...😀

Learned a lot about agent issues during this (confirmation loops, missing context, execution before confirmation - just to mention some)

[Alexander Wolf @AWolf81]

@sarath_suresh_m @TessaMero Tanstack start looks awesome.

I've never been a fan of Next.js - because DX on Windows wasn't that smooth.
Maybe that changed but dropped Windows around a Month ago and not missing it - now on Linux (Ubuntu).

My next web app will probably use Open Sass.

https://t.co/X8u8GZvNiN

[Alexander Wolf @AWolf81]

Ah, OK, the files in /tmp folder are likely generated by bash tool execution for CWD tracking.

Unfortunately, there is no source code in the Github repo to it. But there is a GH issue with some details about it:

https://t.co/9goXTSL0uv

[Alexander Wolf @AWolf81]

Interesting to see that Claude Code CLI had similar issues with the project root handling, just look at your /tmp folder there are many 'claude-<githash>-cwd' files - content is the working dir.
Likely created everytime you open Claude Code extension or CLI.
#buildinpublic
👇