sandGorgon

this is quite similar to amazon-science Cyber-Zero ?

https://github.com/amazon-science/Cyber-Zero

the paper is quite interesting - https://arxiv.org/html/2508.00910v2

have you tried Windows + WSL2 ? i have been on fedora for 20 years.

WSL2 is one of the best linux experience you can have.

touch. RL needs to model real usecases. Some of our customers build mobile-use.

have you tried running android browsers ? we run RL workloads using android browsers. We are having to maintain a fork of https://github.com/budtmo/docker-android/ and android chrome on top. We would rather use browser-use if it had that support.

P.S. we do maintain our fork of a browser for rubric computation...but that is not relevant for this. The infrastructure is what we are looking for.

actually this is not the reason - the harness is significantly better. There is no comparable harness to Claude Code with skills, etc.

Opencode was getting there, but it seems the founders lost interest. Pi could be it, but its very focused on OpenClaw. Even Codex cli doesnt have all of it.

which harness works well with Deepseek v4 ?

this is actually a very valid technique. We do the same (as an rl environments provider).

Except we bundle it with a custom browser renderer which actually generates rewards based on dom diff...and not screenshot based.

the browser renderer is opensource https://github.com/wootzapp/wootz-browser

Yeah...so in India, the GLP patent owners went to court. however, it was deemed as patent expired in March 2026.

In anycase, India generally doesnt honor product patents for pharma. Right now, Indian pharma cos are court empowered to churn out GLP-1.

Everyone knows that we are gonna have 10$ GLP-1 before 2027 coming out of India.

kubedock is its own command set. it doesnt work with compose.yml . This project specifically starts with a compose file.

NOTE: kubedock has some automatic compatibility with compose, this is not their goal - https://github.com/joyrex2001/kubedock/blob/e4c8540f63d65e33...

If you are not on the compose ecosystem, this is isnt useful to you.

Just fyi - the hottest thing in the world right now is OpenClaw. This is OpenClaw docker compose yml - https://github.com/openclaw/openclaw/blob/main/docker-compos...

even if you dont do docker compose....everyone else does it anyway.

docker != docker compose.

docker compose is just as heavy. there's not a significant difference between k3s and docker compose.

this is done. Please feel free to try that.

Also, btw - kappal is architected AI first. Meaning: 1. we have a beautiful skill that can be used by Claude 2. The help command generates output that is useful for claude 3. we have a "kappal inspect" command that generates live output of the stack that can be used by Claude.

hope you have a great time using it. Please file bugs!

this is OpenClaw's docker compose yml - https://github.com/openclaw/openclaw/blob/main/docker-compos... . Arguably the hottest thing in the world right now. Mac Minis are out of stock because of this.

At no point, have I invented a new/better method. Perhaps your way is better.

I just recognise that Docker Compose is loved by most open source developers. And invariably any project you touch will have a docker compose setup by default. And it isnt going away, no matter hard anyone tries to kill. Some things are just too well designed. Docker Compose is one of those things.

I'm just making it possible to run those on kubernetes seamlessly.

this is OpenClaw's docker compose yml - https://github.com/openclaw/openclaw/blob/main/docker-compos...

I'm not arguing for the relative superiority of jsonnet vs yaml vs anything else. I just recognise that Docker Compose is loved by most open source developers. And invariably any project you touch will have a docker compose setup by default.

I'm just making it possible to run those on kubernetes seamlessly.

hey thanks so much !

just curious - i actually didnt focus on healthcheck because i wasnt sure if anyone was using it. are you using it for local dev ?

docker compose is beautiful because it uses a simple elegant compose yml file - this is now an open standard. https://www.compose-spec.io/

the standard does not make it mandatory that underlying system should be docker compose (the reference server). it can be anything.

IMHO - kappal is the first project that takes your compose yml file and transparently/drop-in runs it on kubernetes. there is nothing extra you need. It is useful for people who want to maintain their stack as close to production as possible (kubernetes).

If that's not a big goal for you, then this is not very useful for you. But I'd argue ...why do you care if the compose yml is the only think you are using. you get all of kubernetes.

the test folder is for a large real project that i test on (and actually use in real life myself). it has a deliberate gitignore. plenty of tests in https://github.com/sandys/kappal/tree/main/testdata and https://github.com/sandys/kappal/tree/main/scripts

fair. however, i do genuinely find docker compose yml and dev-experience to be much more pleasant and intuitive.

if you ever wanna try it again - use kappal. you will get a full k8s but with the UX of docker compose.

yes. fixing this right now. you will have it in a day or two.

This is a personal project that im open-sourcing. Its one of those projects-that-should-exist-but-nobody-wants-to-kill-their-business.

It takes ur standard docker compose file and runs it transparently in kubernetes (k3s actually). So ur devs don't have cognitive dissonance between testing ur stack locally on ur laptop and making it work on kubernetes in production.

It is primarily meant as a dev tool on ur laptop, and as a replacement for docker compose.

Or...use wsl2 in windows. does the same thing - much much faster.

Windows is the best (sandboxed) linux

could you explain what this does ? like i ask it to write a testcase (pytest, etc). why is a linter better ?

have you considered using https://github.com/hashicorp/terraform-exec to run terraform inside you go process to manage the entire AWS connection piece. Terraform being largely rock solid and frequently updated on this.

could make this considerably more robust.

this doesnt have integrated ram like lunar lake right ?

deepseek kind of innovated on this using off-the-shelf components right ?

to quote from their paper "In order to ensure sufficient computational performance for DualPipe, we customize efficient cross-node all-to-all communication kernels (including dispatching and combining) to conserve the number of SMs dedicated to communication. The implementation of the kernels is codesigned with the MoE gating algorithm and the network topology of our cluster."

does anyone know the size at openai ? it used to run a 7500 node cluster back in 2021 https://openai.com/index/scaling-kubernetes-to-7500-nodes/

the Qualcomm Adreno 750 GPU is a Snapdragon Gen 3 device. This is basically an android device.

I wonder why Valve is maintaining a separate linux and driver fork for this. Snapdragon Gen 3 android game SDK works very well...including Windows emulation. https://www.youtube.com/watch?v=-hsQ_-8HV6g

not saying what Valve is doing is not spectacular. But i cant help but wonder if it isnt a more productive use of their resources to mainline this in Android ? Maybe even accelerate the Desktop Android merge (which Qualcomm is pushing ! https://www.theverge.com/news/784381/qualcomm-ceo-seen-googl...)

Android that is Valve compatible will further Valve's goals of open platforms than maintaining their own fork.

this is pretty cool! have you found success with image editing in nano banana - i mean photoshop-like stuff. from your article i seem to wonder if nano banana is good for editing versus generating new images.