Google Weather Forecast 2.0

These last years there's been exciting innovations in weather (+ climate) forecasting with foundation / deep learning models; Google just released WeatherNext 2.0 which introduces a probabilistic forecaster that generates hundreds of physically consistent scenarios for weather prediction: This new model uses a Functional Generative Network architecture that injects noise in function space so the model is trained only on local "marginals"; this is interesting as it's quite different to how current models are generally trained. This allows the model to learn coherent "joint" structures across variables, space, and time - this is actually quite important in phisics related modelling (which traditionally would be done with simulations). This seems to enable better prediction of complex regional contexts such as heat waves or wind farm output, etc. It's quite interesting to see that many major tech companies are in the race for these type of models which bring the complexity of ML with real-world physics simulations.

META on GenAI World Models

We continue to see exciting waves in the field of GenAI World Models (aka text to 3D navigable environments); this week META releases a new system for AI generation of 3D worlds: This follows similar World Models (like Marble announced last week) which output 3D worlds using diffusion-based image-to-3D reconstruction models, together with object-level scene decomposition, and mesh/texture refinement which makes these worlds game-engine ready. The way these work basically is through a pipeline that first uses procedural reasoning and a global reference image to produce a coherent blockout plus navmesh, then reconstructs a large textured scene while enforcing constraints. This scene can then be decomposed into individual assets with refined geometry and textures per object. This is the interesting thing with innovative ML systems like these, namely that the AI is a critical component, but the innovation that goes towards the robust engineering and experience is what makes it fit together into a cohesive usable workflow.

The State of MLOps 2025 Survey 🔥

MLFlow is dominating Experiment Tracking space in ML in 2025 with 57% adoption (vs 42% in 2024)! Additionally, it makes me very happy to say that this year, Spreadsheets are down to 3% (vs 10% in 2024) - great work everyone 😂!! Closely following we have W&B with 8% (vs 7% in 2024) and DVC with 5% (vs 5% in 2024). The space of Experiment Trackers seem to be the most consolidated space across the MLOps lifecycle by far! 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: https://forms.gle/KF16EckuxNUKDtDK8🔥

MCP Protocol UI Extension

The MCP protocol is a living standard that continuously evolves through SEPs (like PIPs in python); this week a new SEP introduces standardised UI interfaces for MCP servers which seem to be a potential game-changer in usability: The SEP is pretty simple, basically standardising how MCP servers can expose interactive UIs to hosts instead of just text/JSON, using a pre-declared ui:// path protocol where HTML templates are linked to tools via metadata and communicating over the existing MCP JSON-RPC transport. This seems quite interesting for AI systems developers as it allows for easier interaction / debugging, but also potentially creating an ecosystem that could be used easily for both agents (e.g. browser agents) and humans. As we all know the saying, at this stage a bad standard is better than no standard - but certainly this is a living project so there's more than certainly a lot of evolutions (+ breaking changes) expected in the near future.

DuckDB's Take on the Lakehouse

DuckDB is great for local development. But what if you want to take DuckDB to distrbuted data-lake scale? This is a great video from the DuckDB author dives into the current challenges (+ tech/architecture debt) in the Lakebase approaches, as well as how DuckLake tackles these: DuckLake is basically a new open lakehouse table format from the DuckDB team that replaces Iceberg/Delta-style metadata stack with a straightforward architecture, arguing that the current architecture is overly / unecessarily complex. This basically includes a transactional SQL database for metadata plus Parquet files on any storage (S3, NFS, local disk) and stateless compute. This could be quite interesting for ML practitioners, as connecting your local development to a production scale DuckDB environment could be super easy and lightweight. This is definitely a space to watch; personally I do see what Iceberg has been doing in the space as incredibly innovative, however any further breakthroughs through tooling competition that makes the environment better is always welcome!

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