AI Coffee Break with Letitia

@AICoffeeBreak

📺 ML Youtuber youtube.com/AICoffeeBreak |
🧠 AI Researcher | 👩‍🎓 PhD |
Impressum: t1p.de/q93um

YouTube → youtube.com/AICoffeeBreak Joined March 2018

397 Following 7K Followers 1K Posts

Latest posts 5

Pinned post

AI Coffee Break with Letitia @AICoffeeBreak Jun 7, 2023

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 .

AI Coffee Break with Letitia @AICoffeeBreak Oct 3

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.

Aleph Alpha @Aleph__Alpha Oct 3

Small bird, fast wings, Kolibri is here.

78B parameters. 3.46B active. Up to 1M tokens of context. Built in Europe.

Now the weights are yours. Run it on your own hardware, under Apache 2.0.

View quoted post on X

AI Coffee Break with Letitia @AICoffeeBreak Sep 27

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:

AI Coffee Break with Letitia @AICoffeeBreak Aug 29

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 👇

Aleph Alpha @Aleph__Alpha Aug 19

Your model answers 9 of 10 questions correctly about a contract you gave it to check, but you don’t know which one is wrong, so you check all 10. LLM training rarely rewards "I can’t answer": abstaining pays less than guessing, so models guess plausibly, & that’s hallucination.🧵

View quoted post on X

AI Coffee Break with Letitia @AICoffeeBreak Nov 2, 2025

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 👇