aicoffeebreak.bsky.social

🧠 Do Vision & Language Decoders Use Images and Text Equally?
In our latest episode, we speak with Letitia Parcalabescu about her ICLR 2025 paper examining how vision–language *decoder* models use images and text — and how self-consistent their explanations really are. (1/8🧵)

If you love @aicoffeebreak.bsky.social, this one's for you — Letitia Parcalabescu is our next guest on the #WiAIR_podcast!

Stay tuned for our conversation:
🎬 YouTube: www.youtube.com/@WomeninAIRe...

YouTube video by AI Coffee Break with Letitia What's up with Google's new VaultGemma model? – Differential Privacy explained

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

YouTube video by AI Coffee Break with Letitia Diffusion Models and Flow-Matching explained side by side

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

YouTube video by AI Coffee Break with Letitia Energy-Based Transformers explained | How EBTs and EBMs work

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

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

Check it out if you’re curious or feel like supporting thoughtful science storytelling; there's a “text trailer” on these pages 📜✨:
🔗 scifilmit.com/puppetsofadi...

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

YouTube video by AI Coffee Break with Letitia Greedy? Random? Top-p? How LLMs Actually Pick Words – Decoding Strategies Explained

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.
🎥 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.

I'll co-organise the "Multilingualism: from data crawling to evaluation" social / birds-of-a-feather session. It's on the 29th at 4PM. Do come by if you're at ACL Vienna! :)

I'm also at ACL, would be lovely to catch up!

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

#ACL2025NLP

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!

📅 Looking forward to the discussion and to learning from fellow panelists and participants. If you're around Heidelberg, join us!

www.marsilius-kolleg.uni-heidelberg.de/de/dancing-w...

🔨 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?

🤖 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:

Human-like object concept representations emerge naturally in multimodal large language models - Nature Machine Intelligence Multimodal large language models are shown to develop object concept representations similar to those of humans. These representations closely align with neural activity in brain regions involved in o...

Yet another paper finding similarities between human concepts and AI concepts. www.nature.com/articles/s42...

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.

YouTube video by AI Coffee Break with Letitia AlphaEvolve: Using LLMs to solve Scientific and Engineering Challenges | AlphaEvolve explained

💡 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

💡 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!

📍 Heidelberg, September 21–27, 2025
💬 Language: English
🎯 Open to PhD students & advanced Master’s students from all disciplines working on AI-related research

👉 www.marsilius-kolleg.uni-heidelberg.de/de/studium/i...
🧠 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.

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.

YouTube video by AI Coffee Break with Letitia Token-Efficient Long Video Understanding for Multimodal LLMs | Paper explained

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

Thank you! I post here as frequently as on Twitter. 🙈 I'm doing videos once a month now.

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

YouTube video by AI Coffee Break with Letitia 4-Bit Training for Billion-Parameter LLMs? Yes, Really.

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