Kolibri is out 🐦 (Apache 2.0) and new research ideas around measuring grounding and reducing hallucinations with Merlin-Arthur made it into the model. This is just the beginning: lots of open questions, lots still to improve. But getting research ideas into an actual model, learning from them, and having an amazing in-house model training pipeline to iterate quickly is the fun part.

New AI Coffee Break video! 🔥☕️ LLMs can hallucinate and give the correct answer for the wrong reason. What about teaching them to admit “I don’t know”? This is what my research with colleagues at Aleph Alpha Research on Merlin-Arthur protocols is about. Here's a video to explain the complicated aspects of this work. Link below:

Asking an LLM to cite its sources does not tell you whether its answer actually came from those sources. Citations can sound perfectly plausible, while the answer itself came from memorization, spurious cues no human deems relevant, or a lucky guess. So, we came up with a measure of how much of an answer actually came from the provided context in RAG & and a way to train for it. More on Merlin, Morgana & Arthur in the blog post 👇

LLMs can memorize even a phone number seen once in training.🔒 Google’s VaultGemma fixes that, being the first open-weight LLM trained from scratch with differential privacy, so rare secrets leave no trace. ☕ new video explaining Differential Privacy through VaultGemma 👇

We explain diffusion models and flow-matching models side by side. Flow-Matching models are the new generation of AI image generators that are quickly replacing diffusion models. They take everything diffusion did well, but make it faster, smoother, and deterministic.

Ever wondered how Energy-Based Models (EBMs) work and how they differ from normal neural networks? ☕️We go over EBMs and then dive into the Energy-Based Transformers paper to make LLMs that refine guesses, self-verify, and could adapt compute to problem difficulty. (link👇)

The world’s largest NLP conference with almost 2,000 papers presented, ACL 2025 just took place in Vienna! 🎓✨Here is a quick snapshot of the event via a short interview with one of the authors whose work caught my attention. 🎥 Watch: youtu.be/GBISWggsQOA #acl2025NLP #acl2025

How do LLMs pick the next word? They don’t choose words directly: they only output word probabilities. 📊 Greedy decoding, top-k, top-p, min-p are methods that turn these probabilities into actual text. In this video, we break down each method and show how the same model can sound dull, brilliant, or unhinged – just by changing how it samples.

Excited to be at #ACL2025NLP in Vienna this week. 🇦🇹 I’m always up for a chat about reasoning models, NLE faithfulness, synthetic data generation, or the joys and challenges of explaining AI on YouTube. If you're around, let’s connect!

🧠🤖 Can we trust AI in science? I'm excited to be speaking at the final event of the Young Marsilius Fellows 2025, themed "Dancing with Right & Wrong?" – a title that feels increasingly relevant in the age of AI. I'll be joining a panel on "(How) can we trust AI in science?" to discuss questions like: 🔨 What recent advancements in AI were particularly impactful for science? 🔍 How do we calibrate trust in current AI systems? 🧪 If AI takes over more of the scientific process… what’s left for us humans? 📅 Looking forward to the discussion and to learning from fellow panelists and participants. If you're around Heidelberg, join us!

💡 AlphaEvolve is a new AI system that doesn’t just write code, it evolves it. It uses LLMs and evolutionary search to make scientific discoveries. We explain how AlphaEvolve works and the evolutionary strategies behind it (like MAP-Elites and island-based population methods).

Say hi if you're at ICLR! 👋

Excited to share that I’ll be joining the Summer School “AI and Human Values” this September at the Marsilius-Kolleg of Heidelberg University as a speaker. I'll be giving an introduction to how large language models actually work—before the summer school dives deeper into their ethical, legal, and societal implications. 👉 marsilius-kolleg.uni-heidelb… 🧠 This interdisciplinary event brings together researchers from across fields—computer science, linguistics, philosophy, law, medicine, theology—to explore the normative foundations of generative AI, and how values are embedded in its design. 📍 Heidelberg, September 21–27, 2025 💬 Language: English 🎯 Open to PhD students & advanced Master’s students from all disciplines working on AI-related research 💡 Participation includes talks, workshops, and lots of cross-disciplinary exchange—with accommodation and meals covered (fee: 100€). If this sounds like your thing, the application deadline is June 27! #AI #InterdisciplinaryResearch #GenerativeAI #Heidelberg #SummerSchool

Long videos are a nightmare for language models—too many tokens, slow inference. ☠️ We explain STORM ⛈️, a new architecture that improves long video LLMs using Mamba layers and token compression. Reaches better accuracy than GPT-4o on benchmarks and up to 8× more efficiency.👇