Stanford CS236: Deep Generative Models

Stanford CS236 is one of the best in-depth courses on Generative Deep Learning (LLMs, TTS, VAEs, GANs, etc) and provides a super comprehensive deep dive into the key foundational concepts that are now becoming the backbone of modern AI products. If you are interested on brushing up on fundamentals, this is a great and rigorous course that covers these foundational models as probabilistic data simulators. It focuses on how to parameterize complex distributions with deep nets, how different training criteria trade off sample quality, likelihood, and stability, and how to do efficient sampling and inference for tasks like conditional generation, editing, super-resolution, inpainting, control, and inverse problems. There are only growing applications for these models across images, text, audio, video, robotics, medical imaging, and many more domains so definitely recommended to check it out. 

Introducing the OWASP Top 10 2025

The OWASP Foundation has now released the Top 10 Vulnerabilities for 2025! The vulnerabilities that made the 2025 list are: 1. Broken Access Control, 2. Security Misconfiguration, 3. Software Supply Chain Failures, 4. Cryptographic Failures, 5. Injection, 6. Insecure Design, 7. Authentication Failures, 8. Software or Data Integrity Failures, 9. Logging & Alerting Failures, 10.    Mishandling of Exceptional Conditions. Although this is focused for application security, these are also critical in production machine learning systems! Also worth checking out the OWASP ML Top 10, and the OWASP GenAI Top 10 - as well as various reports that have been released as part of the Agentic Security [link] (disclaimer: I am part of the reviewer board).

The State of MLOps 2025 Survey 🔥

We are excited to release initial insights for our Survey on Production MLOps!! Vector Databases are still skyrocketing: About 56% of orgs reported using vector databases, with 14% using PineconeDB (+9% YoY), about 11% have custom built in-house tools, and 8% use Azure AI (cosmos/search) surprisingly! There is still lack of consolidation, with the following contenders being Weaviate and Milvus with 4% and Elasticsearch with 3%. 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🔥

GenAI World Foundation Models

What happens when you train a model using keyboard inputs to predict images of environments? You end up with interactive 3D AI worlds! The "Marble" model / architecture seems to take it to the next level: Marble is a generative multimodal 3D world model that is able to model text, images, multi-view video/images into full 3D environments, and then lets you iteratively refine them via AI-native editing, spatial expansion, and composition of multiple scenes. Recently we had seen attempts of simulating games, and even OS environments (incl. the impressive simulation of NeuralOS), however this seems to take it to the next level by enabling the creation of worlds from prompts which is impressive. One of the most interesting breakthroughs is the ability to export to triangle meshes and camera-controlled videos which seems would be on the path of making it usable in game engines, VFX, simulators, and robotics workflows. This is definitely an exciting space, and coincidentally seems to be the same domain which Yann LeCunn will be exploring after leaving META this week!

Synthetic Time Series Foundation Models

Time-series forecasting quietly runs the modern world, and the race for Forecasting Foundation MOdels is growing! This month we saw a new contender with a new approach, with TempoPFN tackling univariate zero-shot forecasting built on a linear RNN backbone with purely synthetic data! On the paper it seems to have promising performace, however only against other models trained with synthetic data, so it would be great to also see how it fairs against some of the current heavyhitters (aka Chronos, TimeGPT, etc). It is certainly an exciting time for this space - definitely worth keeping an eye as we likely will see many more releases in the year to come!

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