| YouTube | http://youtube.com/AICoffeeBreak |
| https://twitter.com/AICoffeeBreak |
LLMs often memorize what they see — even a single phone number can stick in their weights. Google’s VaultGemma changes that: it’s the first open-weight LLM trained from scratch with differential privacy, so rare secrets leave no trace. 👉 In this video, we explain Differential Privacy through VaultGemma — how it works, why it matters, and what it means for trustworthy AI.
We explain diffusion models and flow-matching models side by side to highlight the key differences between them. 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.
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: https://youtu.be/GBISWggsQOA
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.
🎥 Watch here: https://youtu.be/o-_SZ_itxeA
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.
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.
