If you browse Hugging Face or GitHub trending right now, 90% of what you see is demo garbage. Repos with 10k stars that don't run, model releases with no evaluation, blog posts announcing things that will be available soon. This article ignores all of that.

These are the projects that shipped working, production ready code in the last 30 days. Every one of these you can check out, run, and deploy before the end of the day. None of them require an API key to try. All are being used by teams right now.

LeRobot v0.6.0: The robot learning stack just crossed the usability threshold

Nobody was using LeRobot six months ago. Everyone will be using it six months from now.

v0.6.0 is the first release where this is not a research toy. It has a standard evaluation suite, unified policy API, human in the loop correction, FSDP training, and cloud training via HF Jobs all working out of the box.

Most importantly, it has solved the biggest problem in robot learning: nobody could agree on how to evaluate anything. There are now 9 standard benchmark families running through one single CLI command. Every new VLA is now tested against the same baseline.

The most clever addition is VLA-JEPA. It trains a world model during training, then throws it away entirely at inference. You get all the benefit of world model supervision with zero runtime cost. That is not an incremental improvement. That changes the cost equation for every robot policy going forward.

Inference for MolmoAct2 runs in 12GB VRAM. You can run a state of the art manipulation policy on a consumer RTX 4090. That was unthinkable 12 months ago.

Hugging Face Kernels: Native code distribution done right

Custom kernels are the dirty secret of fast ML. Every production model uses them, nobody distributes them properly, and everyone just copies CUDA code from gists.

Hugging Face Kernels fixes this. It is a first class repository type on the Hub for native code, with reproducible builds via Nix, code signing, and runtime compatibility checking.

This is not another wrapper. This solves the actual problem: right now if you load a random kernel from GitHub you are running arbitrary native code with full process privileges. Kernels adds trusted publishers, signature verification, and explicit opt-in for untrusted code.

It also added Torch Stable ABI support. A kernel built today will run on every PyTorch release for the next two years. That alone will save the industry millions of hours of wasted compilation.

PRX Data Pipeline: How good diffusion models actually get trained

Everyone talks about model architectures. Nobody talks about data pipelines. This is the first time a leading diffusion team has published exactly how they build their training corpus.

Photoroom laid out every single decision they made for PRX, with hard numbers, no handwaving.

The big takeaways:

• Pre-training is for breadth, not quality. Over filtering hurts model capability far more than bad images help.
• Long accurate captions are the single strongest lever on output quality. Switching captioners improved FID from 21 to 13.
• JPEG at quality 92 is indistinguishable from PNG for training purposes. It is 5x smaller. No one should be storing training images as PNG ever again.
• Lance for curation, MDS for streaming. This is now the standard stack for large training datasets.

This is the pipeline they run:

And this is the hard data that convinced them:

Every team training diffusion models right now is copying this pipeline verbatim. You should too.

Agent skills are not a gimmick anymore

Six months ago agent skills were a demo. Today they are the fastest growing part of the ML ecosystem, and almost no one is writing about it.

Anthropic shipped their full financial services skill pack. This is not a demo. This is exactly the code investment banks are testing internally right now. It has working DCF models, LBO models, GL recon, and 11 live data connectors.

Composio published the awesome-claude-skills list, now at over 1000 production ready skills. The standard has been adopted by Claude, Codex, Cursor, Gemini and every other major coding agent.

This is not tools. This is not MCP. Skills define the workflow. They are reusable instruction packages that agents load on demand. An agent with 1000 skills installed only loads the 1 or 2 it actually needs for any given task.

Layer	Purpose	Standard
Skills	Workflow logic, guardrails, procedure	Claude Skill Format
MCP	Transport, auth, tool discovery	Model Context Protocol
Tools	Individual function calls	OpenAPI / JSON-RPC

This stack has won. Every agent being built right now uses this three layer model.

last30days: The search engine engineers actually use

Google is dead for technical work. No one working in ML uses it anymore.

last30days is an agent skill that searches Reddit, Hacker News, X, GitHub, YouTube, TikTok and Polymarket in parallel, ranks results by actual engagement not SEO, and synthesizes what actually happened in the last 30 days on any topic.

It does not give you blog posts. It gives you what people are actually saying.

If you have a meeting tomorrow with someone, run /last30days [name]. You will learn more in 90 seconds than you will in an hour of Googling.

This is the single most useful agent skill that exists today. If you install nothing else from this list, install this.

Production Agentic RAG Course: The only RAG tutorial you should run

99% of RAG tutorials are garbage. They start with vector search, skip all the hard parts, and produce a demo that breaks the second you run it on real data.

This course does it correctly. It teaches you the order professional teams actually build RAG:

1. Infrastructure
2. Data pipeline
3. BM25 keyword search
4. Chunking
5. Hybrid search
6. LLM integration
7. Monitoring and caching
8. Agentic RAG

Vector search is week 4. That is how you know it is good.

Every week has working production grade code. No LangChain wrappers. No abstractions that hide what is actually happening. You build every part from scratch.

If you want to actually learn to build production RAG, this is the only resource you need.

Qlib RD-Agent: Quant trading gets automated research

Microsoft shipped RD-Agent, an autonomous research agent for quantitative trading. It mines alpha factors, optimizes models, runs backtests, and writes research reports.

This is not a trading bot. It is an automated quant researcher. It does exactly the work a junior quant would do.

It already outperforms most hand written factors on standard benchmarks. This is the thing everyone in quant has been quietly building internally for the last two years. Now it is open source.

nanoGPT is dead, long live reference implementations

Andrej Karpathy archived nanoGPT this month.

This is not a failure. This is exactly how good open source ML should work. nanoGPT existed to prove that you could write a complete, performant GPT implementation in 600 lines of readable code. It succeeded completely. Every GPT implementation written in the last three years copied something from nanoGPT.

Now it is done. It served its purpose. The baton has passed to nanochat.

This is the good part of open source ML that almost no one talks about. Good reference implementations don't live forever. They teach everyone the right way to do something, then they retire.

The quiet standardization nobody is talking about

Look across all these projects. There is a pattern.

Everyone is converging on the same stack. Same dataset formats. Same training infrastructure. Same agent architecture. Same deployment patterns.

For the last ten years everyone was building everything from scratch. That era is over.

We now have standard, working, open source components for almost every part of ML engineering. Most of them are hosted on Hugging Face. None of them are controlled by any one big company.

This is not hype. This is not a paradigm shift. This is just boring, good engineering. Infrastructure is maturing. Standards are emerging. People are finally stop reinventing the wheel and starting to build actual things.

That is the real news this month.