Wieland Brendel

We are super happy about this milestone - a great validation for our three-year-long project Polybot that we are on the right track towards a future in farming automation that can scale up novel regenerative farming practices!

New preprint out! 🎉

How does LLM training loss translate to downstream performance?

We show that pretraining data and tokenizer shape loss-to-loss scaling, while architecture and other factors play a surprisingly minor role!
brendel-group.github.io/llm-line/ 🧵1/8

Just realise that the answer is still from Grok 2, so maybe Grok 3 gives a very different answer.

CuratedThoughts: Data Curation for RL Datasets 🚀

Since DeepSeek-R1 introduced reasoning-based RL, datasets like Open-R1 & OpenThoughts emerged for fine-tuning & GRPO. Our deep dive found major flaws — 25% of OpenThoughts needed elimination by data curation.

Here's why 👇🧵

I get pretty much exactly the same balanced answer as you, the original post seems to be fake.

Let’s shape the future of AI & education together! 🔥 #AI #MachineLearning #EdTech #Hiring

👉 forms.gle/sxvXbJhZSccr...

We work closely with amazing collaborators at @tuebingen-ai.bsky.social (@bethgelab.bsky.social), the Hector Research Institute of Education & Psychology (Ulrich Trautwein), the Tübingen Center for Digital Education (Andreas Lachner), @iwmtue.bsky.social (Ulrike Cress) & KI macht Schule (@stes.io).

🔹Several Full-time ML roles (2–3 years)
🔹Work in a top-tier AI research environment
🔹Focus on concrete, real-world impact in schools
🔹Open to on-site, hybrid & remote

👉 forms.gle/sxvXbJhZSccr...

Hiring announcement: ELLIS Institute Tübingen is looking for ML Researchers & Engineers for Open-Source AI Tutoring (m/f/d). The image features a white background with bold black text and the colorful ELLIS logo at the bottom.

🚀 We’re hiring! Join Bernhard Schölkopf & me at @ellisinsttue.bsky.social to push the frontier of #AI in education!

We’re building cutting-edge, open-source AI tutoring models for high-quality, adaptive learning for all pupils with support from the Hector Foundation.

👉 forms.gle/sxvXbJhZSccr...

Does anyone know how OpenAI gets o3-mini to exceed 700 tokens/sec? I’ve only seen such speeds on specialized chips from @cerebrassystems.bsky.social, @sambanova.bsky.social, or @groq.com—but not on standard Nvidia GPUs, which I assumed power OpenAI’s inference.

As many areas, science itself will be transformed by AI. We have now released a state-of-the-art survey on this emerging field. We discuss methods and tools for AI-enhanced search, ideation, experimentation, content generation and peer review.
arxiv.org/pdf/2502.05151

🏹 Job: The Tübingen AI Center, a partner in the OpenEuroLLM project, is seeking a Scientific Project Coordinator to oversee efforts in developing open-source, trustworthy foundation models and leading community engagement.

📍 Tübingen
📅 Apply by March 3rd, 2025

📌 Get details: tuebingen.ai/careers

📢 Join our Bundeswettbewerb Künstliche Intelligenz (BWKI) at Didacta 2025 in Stuttgart!

From tomorrow, February 11, to Saturday, February 15, the Didacta education fair will take place in Stuttgart:

📍 Find BWKI in Hall 7, Stand 7D75 as part of the AI & Digital Workshop area - workshops offered!