Claude Skills Driving Agentic Innovation

Anthropic’s new Claude Skills system released last week may be more transformative than the MCP craze due to a few advantages: Claude Skills are opening more possibilities for lightweight, modular extensions to LLMs using simple Markdown files and optional scripts that Claude can load directly. These "skills" let models perform specialized tasks efficiently within a coding environment, which significantly reduces token overhead and complexity compared to MCP's heavy protocol structure. Although it can be argued that these two frameworks can be compatible / interchangeable in the future, it is interesting to see this wild-wild west of competing approaches to find the global optimal. This is a great write-up by Simon Willison which summarises both the advantages as well as some practical examples.

Truths in Software that are False in AI

Here are 5 truths in regular software that are false when applied to AI: 1) Software vulnerabilities are caused by mistakes in the code; 2) Bugs in the code can be found by carefully analysing the code; 3) Once a bug is fixed, it won’t come back again; 4) Every time you run the code, the same thing happens; 5) If you give specifications beforehand, you can get software that meets those specifications. It is becoming growingly accepted that assumptions from traditional software engineering like bugs, testing and monitoring, don't apply in the same way to modern AI systems. Unlike conventional software, in AI behavior is learned from often large datasets rather than explicit instructions, making errors unpredictable, non-deterministic, and resistant to debugging. There is now a growing consensus across organisations to accept these differences in order to derive the right practices to ensure stable and robust operations at scale.

The State of MLOps 2025 Survey 🔥

We are still continuing to gather the insights on this year's MLOps Survey! We still need your support to continue collecting diverse perspectives to map the ecosystem! Please help us with your response, as well as by sharing with your colleagues 🚀🚀🚀 If you have a few minutes, your contribution will make a significant difference to the whole production ML ecosystem 🥳 The results will be shared as open source like last year!! You can add your response directly at: bit.ly/state-of-ml-2025 🔥

Karpathy on the Decade of Agents

Some really insightful take-aways on this 2 hour podcast with Andrej Karpathy, such as arguing that "we are in the decade of agents, not the year of agents" - here are 5 key takeaways: 1) It will take over a decade to properly figure out agents, as we're in the early stages of memory, multimodality and reliable long-horizon reasoning. 2) Reinforcement learning is inefficient and noisy, requiring new approaches like process-based supervision and reflection. 3) Large language models suffer from model collapse and over-memorization, limiting creativity and continual learning. 4) Coding is currently the most practical application for LLMs, but they struggle with truly novel or complex system design. 5) AGI’s arrival will be gradual with steady productivity growth rather than causing a sudden economic or societal leap. For machine learning practitioners, understanding where the real bottlenecks lie determines whether we build the foundation required for robust agentic systems, or get stuck on the hype.

Blameless Culture in Engineering

There are critical lessons that we can take on from the movement that drives the "Blameless culture in engineering", which focuses on solution vs finger-pointing: Highlights that high-performing engineering and ML operations teams thrive on a blameless culture, where failures are treated as opportunities to strengthen systems rather than assign fault. This by no means encourages removing accountability, but switches the focus from finger-pointing towards short- and long-term solution finding instead; this ensures problems are resolved efficiently, which then can allow space for systemic or organisational improvements required. In production machine learning and large-scale engineering, culture is as critical as the code itself, and there are key lessons that can be adopted from this because even the best systems fail without trust, learning, and accountability.

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

✉️ Email, 🐦 Twitter, 💼 Linkedin