AI Coffee Break with Letitia

8 things experts know about large language models (LLMs) and you should know about them too. 👇 📺 youtu.be/RX-gGs_EV7M In this video, we summarized the sober take of @sleepinyourhat in his recent post 📃arxiv.org/abs/2304.00612 .

If you love @AICoffeeBreak, this one's for you — Letitia Parcalabescu @letiepi is our next guest on the @WiAIR_podcast! Stay tuned for our conversation: 🎬 YouTube: @WomeninAIResearch" target="_blank" rel="nofollow noopener">youtube.com/@WomeninAIRese…

What if your vision language model isn’t actually seeing… but mostly guessing from text? 👀 @AICoffeeBreak explains it perfectly: when VLMs rely too heavily on text, they start hallucinating answers based on the most common phrasing in their training data instead of what’s in the image. Ask “How many cats are there?” and even if the image shows five, the model might say two simply because “two” appears more often in similar prompts. This is the hidden trap behind text-driven hallucinations. And unless we explicitly measure grounding, these models will keep sounding confident while being wrong. I love this kind of research because it shows exactly where our tools break and where we need to push next, especially in high-stakes domains like medicine or robotics. #visionlanguagemodels #VLM #AIresearch #multimodalAI

🎥 youtu.be/UwX5zzjwb_g

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 👇

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👇)

@JacksonAtkinsX This sounds like the recipe for AlphaEvolve.

This just dropped and nobody's talking about it yet. 🥇 Nvidia's "GenCluster" won IOI 2025 gold with open-source gpt-oss-120b. The first time an open model ever matched the big labs. The breakthrough? Scalable test-time compute for hard problems. Here's how it works: - Brainstorm: For each IOI problem, they had the AI generate up to 5,000 potential C++ solutions, good and bad. - Gauntlet: They ran every solution against 100 private, IOI-style test cases. Solutions were then clustered based on their behavior. This revealed sets of unique approaches. - Tournament: To find the champions, they pitted the best solution from each cluster against each other in a mock competition, with another LLM acting as the judge to rank the winners. - Finals: With a clear ranking, the system submitted the best solutions to the official IOI graders, staying under the strict 50-submission limit to maximize the final score. Result: A 446.75 gold-medal score, verified by the official IOI grader under the same time and memory limits faced by human competitors. This IOI victory is just the proof point. GenCluster is a blueprint for solving extremely hard problems. Congrats to the team @NVIDIAAI for the win.

🎥 youtu.be/firXjwZ_6KI

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.

very solid overview of energy-based transformer by @AICoffeeBreak no hype the knowledge is distilled to it's essence and noise of the paper is properly modeled in the walkthrough damn she's good

Great work from @AlexiGlad and collaborators! 👏

Works for images and video too! 🎥 youtu.be/18Fn2m99X1k

Featuring @bowpis and @FabioFehr 🤝

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

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. Check it out if you’re curious or feel like supporting thoughtful science storytelling; there's a “text trailer” on these pages 📜✨: 🔗 scifilmit.com/puppetsofadigi… 💡 GoFundMe: gofund.me/453ed662

🤔Do you know that LLMs produce probabilities among each available token of the vocabulary. Only after comes the choice of the final outputed token. 👌Here is crystal clear, yet insightful, explanations of the various technics used to choose the next token

🎥 Watch here: youtu.be/o-_SZ_itxeA

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.

Right now, I am attending the synthetic gata generation tutorial which is packed, because it turned out that "Data is the new source code."

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!