Thank you for being part of over 70,000+ ML professionals and enthusiasts who receive weekly articles & tutorials on Machine Learning & MLOps 🤖 You can join the newsletter /newsletter/ ⭐

If you like the content please support the newsletter by sharing with your friends via ✉️ Email, 🐦 Twitter, 💼 Linkedin and 📕 Facebook!

This week in ML Engineering:

Raschka's Course on Building Reasoning Models

Sebastian Raschka released a free course on building reasoning models from scratch with 8 coding videos and over 11 hours of content:

It is a great time to be alive and be able to benefit from such top quality content. In this course Sebastian takes a pre-trained Qwen3 0.6B model and builds everything on top in PyTorch.

This course series go through text generation with a KV cache, a math answer checker evaluated on MATH-500, inference-time scaling, RL with a simplified GRPO, and distillation with DeepSeek R1 as the teacher.

Last christmas I actually built myself a marimo-based course on this specific topic (back in the early days of Opus), and I have to say that although it was quite useful I wish this course was available back then!

The only thing to consider is that although the early chapters run on a CPU, he recommends a GPU once you get to the training parts (or a lot of waiting time with your mac).

This course accompanies Sebastian's latest Manning book on LLMs and all the code is on GitHub.

Anthropic on Formally Verified OSS

Anthropic's security team worked to get Claude to write machine-checked proofs for a few open-source libraries, and I have to say I am excited about the opportunity here:

I have been laser focus on the topic of ways to "close the verification loop", so seeing this is quite interesting. Basically they were able to prove that various entry points never crash on any input under 64 MB. The proofs run on the compiled RISC-V binaries and are checked by Rocq, from which they add up to about 180 million lines, with a typical proof of around 600,000 lines for a binary of 18,000 instructions.

They basically turned each test into a spec, and whenever Claude hits a bug while proving it, it patches the code and proves it again. They say this found hundreds of bugs including a rediscovered CVE in libexpat. Patching is an obvious way to cheat here, so they penalised every patch and audited them all.

The limitations are that the proofs cover single-threaded RISC-V binaries linked against musl (most people run x86 or ARM with glibc), and real programs can still misuse the library. So not perfect!

However it does seem that formal verification will soon be as practical as fuzzing is today. That's a bold claim, but this would certainly be a gamechanger if it is.

Netflix on Pensive AI Failure Diagnosis

Netflix shares their agentic automation across their auto-diagnosis system for failed data and ML workloads:

At Netflix, up to 100,000 workloads can fail across their Titus, Spark, Trino and Druid infrastructure. Their legacy system used regexes and hand-written rules to address this, however they were able to replace this over a three-month window.

Instead of asking a model to copy a stack trace token by token, they number the log lines and the model only returns the first and last line of the error; their model now then rebuilds the text from its own copy of the log, which is cheaper and faster.

The eval setup is also super interesting. They freeze the old rules as reference labels, archive logs to S3 before they expire, and deploy every PR to a preview environment where precision and inclusion thresholds gate the merge.

They report AI extraction now covers 100% of failures, and failed attempts where nobody used Pensive AI saw 321% more manual log digging, which they are careful to call observational. It is great to see this level of detail on evals in production!

Meta on Just-in-Time Catching Tests

Should engineers still read the tests an LLM writes for them? Meta challenges the status quo by reviewing risky diffs before they land, enabling their engineers to avoid looking at the test code.

At META, each night they pick the risky diffs, two generators write tests aimed at catching bugs in that specific change, and an ensemble of LLM judges plus a rule-based filter ranks what comes out. Engineers sit "on the loop", setting the criteria and reviewing the results in aggregate.

Meta claims 589 engineers were involved in the evalaution, and they reported 74 serious privacy bugs caught at diff time in H1 2026, and a 20% drop in privacy SEVs.

The FSE 2026 paper behind it is worth reading, as they now had 22,126 generated tests, and the LLM judges cut human review by about 70%.

Mistral Large 4 Goes Open Weights

Mistral Large 4 "Le Chonk" is now open weights 🇪🇺🚀🇫🇷 This is Mistral's 1 trillion parameter multimodal MoE, with three reasoning modes and support for over 160 languages.

Mistral is launching it as "the best open-weight coding model", with 48% on the Artificial Analysis Coding Agent Index (tied with GLM-5.3) and 41% on Terminal-Bench 4.0. Pricing is $0.70 per million input tokens and $2.10 per million output.

The licence is a modified MIT which allows you to can use, fine-tune and sell it. However it's not fully open source, as they still restricted companies offering it as a model-as-a-service (aka requires a commercial licence).

Mistral also reports 51% on the AA Cyber Index and says it sees "a potential for uplift to conduct real attacks", so they are adjusting their guardrails and keeping a cyber variant for vetted partners only.

Upcoming MLOps Events

The MLOps ecosystem continues to grow at break-neck speeds, making it ever harder for us as practitioners to stay up to date with relevant developments. A fantsatic way to keep on-top of relevant resources is through the great community and events that the MLOps and Production ML ecosystem offers. This is the reason why we have started curating a list of upcoming events in the space, which are outlined below.

You can also find our upcoming events and past talk recordings on the talks and events page.

Events we are speaking at this year:

Other relevant events:

In case you missed our talks, check our recordings below:

Open Source MLOps Tools

Check out the fast-growing ecosystem of production ML tools & frameworks at the github repository which has reached over 20,000 ⭐ github stars. We are currently looking for more libraries to add - if you know of any that are not listed, please let us know or feel free to add a PR. Here's a few featured open source libraries that we maintain:

  • SARC - Provides wrappers for popular agentic frameworks to enable guardrails and constraints that are enforced through the flow.
  • KAOS - K8s Agent Orchestration Service for managing the KAOS in large-scale distributed agentic systems.
  • Kompute - Blazing fast, lightweight and mobile phone-enabled GPU compute framework optimized for advanced data processing usecases.
  • Production ML Tools - A curated list of tools to deploy, monitor and optimize machine learning systems at scale.
  • AI Policy List - A mature list that maps the ecosystem of artificial intelligence guidelines, principles, codes of ethics, standards, regulation and beyond.
  • Agentic Systems Tools - A new list that aims to map the emerging ecosystem of agentic systems with tools and frameworks for scaling this domain

Please do support some of our open source projects by sharing, contributing or adding a star ⭐

About us

The Institute for Ethical AI & Machine Learning is a European research centre that carries out world-class research into responsible machine learning.

Check out our website

✉️ Email, 🐦 Twitter, 💼 Linkedin