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 👇 🎥 youtu.be/UwX5zzjwb_g

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. 🎥 youtu.be/firXjwZ_6KI

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.

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

My friend Vivi Nastase is working on a short science communication film called "Puppets of a Digital Brain". It aims to explain the tech behind AI chatbots (the good, the bad, the environmental) in an accessible, visual way. 💡 GoFundMe: gofund.me/453ed662

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.

Excited to be at ACL 2025 in Vienna this week 🇦🇹 #ACL2025 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 these days. I'll be joining a panel on "(How) can we trust AI in science?" to discuss questions like:

We train AI on human-selected or -generated data (yes, even taking a photo is concept selection – we capture what we find interesting; text even more so, expressing our conceptualisation of the world). Then we’re surprised when the AI's concepts and representations are similar to ours. 🤷‍♀️

I'm very excited to finally share the main work of my PhD! We explored the evolutionary dynamics of gene regulation and expression during gonad development in primates. We cover among others: X chromosome dynamics (incl. in a developing XXY testis), gene regulatory networks and cell type evolution.

💡 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). 📺 youtu.be/Z4uF6cVly8o

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 broader implications.

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. 📺 youtu.be/uMk3VN4S8TQ

Follow @aicoffeebreak.bsky.social!! Letitia is very effective in communicating research papers in just a few mins! Perfect for your coffee break. 😉

We all know quantization works at inference time, but researchers successfully trained a 13B LLaMA 2 model using FP4 precision (only 16 values per weight!). 🤯 We break down how it works. If quantization and mixed-precision training sounds mysterious, this’ll clear it up. 📺 youtu.be/Ue3AK4mCYYg

Just say “Wait…” – and your LLM gets smarter?! We explain how just 1,000 training examples + a tiny trick at inference time = o1-preview level reasoning. No RL, no massive data needed. 🎥 Watch now → youtu.be/XuH2QTAC5yI

🎙️ Yesterday, I gave a keynote on large language models outfitted with visual understanding, and the faithfulness of their chain-of-thought reasoning at the National Conference on Governing the Digital Society and Human-Centered AI.

The National Conference on AI Transformations: Language, Technology, and Society organised by Utrecht University @utrechtuniversity.bsky.social was a success, and indeed Letiția‘s @aicoffeebreak.bsky.social talk was very inspiring.

🎉 Exciting news from our team! The final paper of @aicoffeebreak.bsky.social's PhD journey is accepted at #ICLR2025! 🙌 🖼️📄 Check out her original post below for more details on Vision & Language Models (VLMs), their modality use and their self-consistency 🔥

We explain 🥥COCONUT (Chain of Continuous Thought), a new paper using vectors for CoT instead of words. We break down: - Why CoT with words might not be optimal. - How to implement vectors for CoT instead words and make CoT faster. - What this means for interpretability. 📺 youtu.be/mhKC3Avqy2E

The last paper of my PhD is accepted at ICLR 2025! 🙌 🎊 We investigate the reliance of modern Vision & Language Models (VLMs) on image🖼️ vs. text📄 inputs when generating answers vs. explanations, revealing fascinating insights into their modality use and self-consistency. Takeaways: 👇

An educational and a bit historical deep dive into LLM research. 💡Learn what breakthroughs since 2017 paved the way for AI like ChatGPT (it wasn't overnight). We go through: * Transformers * Prompting * Human Feedback, etc. and break it all down for you! 👇 📺 youtu.be/BprirYymXrg

Transformer language models like Chat GPT, when using chain-of-thought reasoning, are Turing complete. Specifically, they can execute probabilistic algorithms and generate any computable weighted language youtu.be/MMIJKKNxvec?...

Don't forget to register! 🤭👇

New video about: REPA (Representation Alignment), a clever trick to align diffusion transformers’ representations with pretrained transformers like DINOv2. It accelerates training and improves the diff. model’s ability to do things other than image generation (like image classification).

Why does math feel so intimidating? Prof. Yael Tauman Kalai shares her insights, and we look at what this says about how we view intelligence🧠, both human🧑and AI🤖. 📺 youtu.be/Su1puD4xQwI

How to make powerful LLMs understand graphs and their structure?🕸️ With Graph Language Models! They take a pre-trained LLM and fit it with the ability to process graphs. Watch if you're curious!👇 📺 youtu.be/JcHeaONGbmQ (Hint: it's about position embeddings, as the author explained at #ACL2024 🔴)

Here is what we think about the training procedure of OpenAI #o1. We speculate based on all the bread crumbs we could find, how exactly reinforcement learning (RL) helped train the model to “think” by producing private Chain-of-Thought tokens before answering. 👇 📺 youtu.be/MNE6QZaRavo

Can LLMs handle self-referential statements such as “This sentence has 5 words”? We recorded the #ACL2024 poster presentation of the paper „I am a Strange Dataset: Metalinguistic Tests for Language Models” by @TristanThrush @jaredlcm @migueljmonares @ChrisGPotts @douwekiela . 📺 youtu.be/m_nEIsQBh_c

Are transformer LLMs equivalent to Turing machines or not? Spoiler: they are, because @franz_nowak , @AnejSvete , @butoialexandra , and @RyanCotterell proved this in their latest paper! 📃 📺 youtu.be/MMIJKKNxvec We talk with Franz Nowak about RNNs, transformer encoders, decoders (with CoT).

At the ACL, we recorded the poster presentation of the paper challenging Noam Chomsky's claim about LLMs! 🫢 📺 youtu.be/8lU6dGqR26s This paper, entitled “Mission: Impossible language models”, won an #ACL2024 best paper award. Congrats to the authors! 👏

Prefer reading over watching videos? 📚 Check out some of our videos in blog post format on Substack!👇 aicoffeebreakwl.substack.com We'll be adding more posts regularly, stay tuned! 📻

Text diffusion can finally generate good text!📃 We've combed through the dense math of the “Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution” paper to bring you the key insights and takeaways.👇 📺 youtu.be/K_9wQ6LZNpI The paper won the #ICML2024 best paper award.

I’ve made a video about my PhD journey! 👩‍🎓 I'm spilling all the (coffee) beans on why I chose a PhD ☕, the side quests along the way. I show my PhD hat 🎓and give an overview of my research. Plus, I'm giving tips to prospect PhD students and reasons why I started YouTube. 👇 📺 youtu.be/prGZTX-Sgqw

This Thursday, I'll share my story about how I made over 100 YouTube videos explaining ML in your coffee break! ☕ Listen in by joining Roosh Circle 's No Papers Club. There will be lots of insights, inspiration, and practical advice. 🫱🫲 Join here: lnkd.in/e_wX2r_G ⌚ Aug 1 at 5:00 PM CEST!

Excited to share my ACL 2024 presentation on my second-to-last PhD paper! 🎓📚 Watch it here if you are also interested in LLM self-explanations: 🤖 youtu.be/b3wbTOZXRyI Are you joining ACL 2024 in Bangkok? Ping me—let's chat! #ACL2024NLP #PhDLife

Supercharge your RAG apps with Generative Feedback Loops! 🚀 - 📝 Feed data to your LLM - 💾 store outputs with vector embeddings, - 🔍 and search in real-time. In this video, we explain RAG, show demo code for generative feedback loops with Weaviate!👇 youtu.be/ijCjKnbQgXc

We explain GaLore, a new parameter-efficient training ⚙️ technique that outperforms LoRA in accuracy, and supports both pre-training and fine-tuning. You can even pre-train a LLaMA-7B from scratch on one 24GB GPU (NVIDIA RTX 4090), for example. 📺 youtu.be/VC9NbOir7q0

Ever wondered how to interpret your #MachineLearning models? 🤔 We explain a powerful interpretability technique: Shapley Values – can be used to explain any model, including LLMs! 💻 We show simple code for how to use them and 📖 dive into the theory behind them. 📺 youtu.be/5-1lKFvV1i0

How did they steal parts of LLMs protected behind APIs? 🥷 We explain both papers that made a breakthrough on this, one from Carlini et al. (Google), and the other one from Finlayson et al. (USC), see references in the video description. 📺 youtu.be/O_eUzrFU6eQ

Genie 🧞 just watched YouTube videos and inferred actions and learned to render environments! 🗺️ In this video, we explain the Genie paper from @googledeepmind.bsky.social . 👇 📺 youtu.be/QaqX9B3jqYI

We simply explain and illustrate Mamba and (Selective) State Space Models – SSMs. 📺 youtu.be/vrF3MtGwD0Y SSMs match performance of transformers, but are faster and more memory-efficient than them. This is crucial for long sequences!

Contextual sparsity: Take an LLM and make it sparse at inference time. In this video, we explain how the DEJAVU method implements contextual sparsity. 📺 youtu.be/DUkWMoi5nG4

Finetuning LLMs is now possible without reinforcement learning, but with Direct Preference Optimization (DPO). 📺 youtu.be/XZLc09hkMwA In this video, we explain: 👉 the DPO paper 👉 RLHF and why reinforcement learning was needed in the first place.

All you need to know about the transformer architecture: 👉 How to structure the inputs 👉 Attention (Queries, Keys, Values) 👉 Position embeddings 👉 Residual connections. 📺 youtu.be/ec9IQMiJBhs Bonus: an overview of the difference between Recurrent Neural Networks (RNNs) and transformers.