Black Forest Labs Launches FLUX-2

Some of you may remember the release of the FLUX model from Black Forest Labs last year taking image generation to a whole new level; this week they have now released the 2.0 version with some impressive improvements: This new latent flow–based image generation model is now focused on production creative workflows with what they refer to as multi-reference editing, which allows editing up to 10 images in a single architecture. The approach basically pairs a Mistral-3 24B VLM with a transformer and VAE which helps with the compression trade-off, which is what enables the higher photorealism, spatial coherence, and robust typography (i.e. infographics, UI mocks, fine text no longer screwed up). It is quite surprising to see the improvement of image models not slowing down; if anything it's just the opposite, and it really is enabling for real-world applications across all industries and domains.

Major Supply Chain Attacks

Last week we saw a major reminder to us ML practitioners on software supply chain attacks, this one infected ~500 packages including major ones from Zapier, ENS, AsyncAPI, PostHog, Postman and others that amount to 100M+ monthly downloads: The way that this specific attack worked was that once you run your package with one of these dependencies on your dev machine or CI runner, it uses TruffleHog to harvest secrets (API keys, cloud creds, GitHub/npm tokens, etc) and pushes them to public GitHub repos labeled "Sha1-Hulud: The Second Coming". The attack replicates the malicious code and then attempts to publish new infected npm packages; finally wipes the user’s home directory if it can't authenticate. For us ML practitioners, this is our weekly reminder on the importance of security in areas where it's not currently mature; i.e. model registries and artifacts, data/feature store credentials, cloud keys, CI/CD access used to build training and inference images, etc. 

These stories are becoming scarier by the minute, as renowned packages are now falling for some of these issues; this will become a call-to-action for even the design of the package managers themselves, but most importantly to remember on standard security higene where possible.

The State of MLOps 2025 Survey 🔥

In 2025 about 43% of organisations take between 1-3 months to productionise a machine learning model (vs 2024 28%) and only 12% between 3-6 months (vs 25% in 2024) which is great news for the industry! Similarly 13% are now productionising ML models in less than one week (vs 10% in 2024). This is certainly a healthy trend that we hope continues; however with the rise of emerging tech like LLMs we do expect to see oscilations! This is one the most important muscles for organisations to strengthen as it directly relates to their capability to scale AI innovation. 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🔥

Raschka on Building an LLM

Sebastian Raschka and Andrej Karpathy's educational material is becoming a key cornerstone for foundational ML knowledge; and Raschka's recent course on "Building an LLM (From Scratch)" is becoming a must-watch classic: This course walks ML practitioners through implementing a GPT-style model end-to-end in plain PyTorch, starting from text preprocessing and tokenization. It goes through hand-built attention and Transformer blocks to a full GPT architecture with training and generation which is great to build the intuition. Along the way in this course, you implement next-token pretraining on unlabeled text, and then repurpose the pretrained backbone for supervised tasks like text classification and simple chat-style behavior. There are quite a lot of great takeaways, but these approaches of learning by building something from scratch often are such a great way to develop the foundations before going to the higher level / advanced domains (i.e. building agentic frameworks, transfer learning, etc).

2025 Advent of Code

In production ML we are reminded that your AI systems can only be as robust as the engineering discipline behind them; what better way to polish your foundations than with the Advent of Code for 2025: It's great to see that despite the LLM craze the Advent of Code is still back! We did see that last year(s) they had to fight an insurgence of LLM-content in the leaderboards, which is why there is none this year but you can still practice. The puzzles unlock TONIGHT at midnight US Eastern! Do keep an eye and try them out - only caveat is that this time around there will only be 12 instead of 24, but this should keep us busy for the upcoming month! 

On this topic funnily enough I ended up attending to build a custom parser, lexer and interpreter for a custom language that I was also wanted to explore testing with this - still WIP but pretty mind blowing how these tools are evolving to enable these.

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