Draft: Agentic Bluefin part 2: agents.projectbluefin.io

[castrojo]

agents.projectbluefin.io - Instructions for how to donate inference to the project. This can be any tool or any provider you want.

Donate inference via our handy dandy appliance. I am still working on sandboxing this:

• Instructions - agents.projectbluefin.io
• New slicker thing I built on oh-my-pi that is awesome and needs some feedback/help - https://github.com/projectbluefin/contribute

Then chill and relax. I like to sit there and watch it work on stuff, you can steer from the client if you want but it's generally hands-off. Here I am doing both submitting and reviewing:

Other hives: bluefin-contribute orgname/repo

The review app is designed for what we call "Non-maintainer reviews". But we just call it "Blueberry mode". You need to be a maintainer to review and land code in Bluefin. To build that trust you can do reviews on open pull requests to the project automatically with this tool. This is valuable to maintainers because it provides more opinions on a submission. A pull request with a bunch of independent reviews from different models by different people around the world is very valuable!

Makes it much easier for us when we can farm out the review load!

Just fire it up, sit back and start reading! You can steer this thing, start reading and learning! Your worker and review contributions are measured as independent and related stats here:

Big edit:

Over the past month or so Project Bluefin has been living in Star Trek land. I literally talk to my computer and Bluefin comes out. We started our journey with hive a few months ago. We were fortunate because we got in early, and now Hive is a CNCF Sandbox Project. If you like this future, go give it a thumbs up! It means a lot! Remember the cloud native people don't know how big desktop Linux is, it's a like a unicorn in these lands.

A quick reminder that our core technology, bootc, is also in the CNCF Sandbox. This place is where we get to play. Hive and bootc are a great combo. The idea is simple. We want a container, Bluefin, to our exact specs. People have GPUs and access to agents. Hive is like SETI@Home but for maintaining OSS projects, here's the leaderboard.

[https://docs.projectbluefin.io/leaderboards/]

And you'll see something interesting, local models. Today we learned about spektacular, a specification system that has preexisted, and has now joined forces with hive. This forces us to do things the way that is needed in order to succeed. By writing specifications, which are then put through a process, and then farmed out to "dumb" LLMs that are given strict tasks. These things are smart but also really stupid.

The Operator

At the core is the operator donating the inference. You can't just grab a local model and get results like Kyle Rankin. He's been spending the time investing in learning the tech and tweaking all of that. An expert in local models is something you can do on most hardware and is an in demand skill. People with access to max models and unlimited money will always be able to do more work than people who do not have it. What the linux desktop has is large numbers of technically smart people. My generation grew up proud to dominate leaderboards on distributed.net, BOINC, folding@home, etc. We're just making lists of packages and version numbers.

We are the Borg

Ok here is the best part. The reason you put up with this cloud native mumbo jumbo. The trick to scalability is users! Over the past few months I've watched a bunch of people set up their own personal images, with their own personal hives, and donating inference to each other! But you can't just vibe code this crap. This is where specs come in. You follow the template, it's annoying. It's long. It's shitty. It takes too long. I hated UML too. I hate writing things down. This is the thing that paces the project, because right now we have 8 clankers sitting idle because our bottleneck is reviewers. It's fine it will take a few days to get us out. But the overall pace is fast.

If you can argue about a feature on the internet then you can write a spec. The first thing Andy and I noticed when Hives went online is how quickly people with access to max models helped out others. I've been there. Smart enough to know what I want in a Linux desktop, not technical enough to get there. Now we're going to just make people fill out forms.

clankers.projectbluefin.io will always point to our inference point - I'm in the middle of deploying Kyle's setup on my twin strix halos. Our review appliance is at github.com/projectbluefin/contribute but needs some love. Onboarding needs help but issues are welcome!

How to get your dream Linux

Ok so, I hope you don't think we're going to open specs and have all of your magical dreams come true. Nope! Bluefin's scope will remain brutal and focused on our direction. However, here's what we're working on that I think is more exciting:

• The contribute appliance will end up feeling like an "app", we'll likely wire it up to ujust contribute too.
• We're totally wiring in bookmarks for your favorite hives
• Come hang out and have people with spare hardware build your dream linux

Now would be a good time to share what we realized, your Bluefin Factory will do this for you. It's better on a homeserver but you'll be able to just run this in podman too. You'll fork the finpilot template, set up your own hive - then come hang out, I plan on contributing to other people's hives! Once your image is set up and running it's all running with real code, no clanker dependency, just some CI/CD.

The truth is, this stuff is easy for LLMs, the spec process helps us vet our ideas into formats that can deliver better results! Anyway, this is the world I am living in now, and every day we improve it so that doing this for yourself becomes easier and easier.

The achievements around helping others will be some of the most juicy in Bluefin! 😄

[castrojo]

It would be silly overhead to do a full hive per custom image, so if any of you have a school image or something used by an NGO or an org who could use this please speak up or contact me privately. Students/teachers write the specs, the community chips in, they get an update. Rinse and repeat!

Something worthy of a juicy write up for a CNCF Case Study!