Why Random Noise is a Problem for AI (And How to Fix It)

In this interview from NeurIPS 2025, we sit down with Luca Eyring, a PhD student at TU Munich , to discuss the future of generative modeling. We dive into the limitations of current diffusion models—specifically how uncontrolled noise creates safety risks—and explore his vision for "automatic sliders" that give users unsupervised control over image generation. Luca also breaks down the shift toward reward alignment in post-training and how these techniques are unlocking breakthroughs in biology, specifically for protein and RNA generation.

Timestamps:
00:00 - Introduction: Luca Eyring & TU Munich
00:19 - The problem with "uncontrolled diversity" in image generation
01:06 - Why AI safety requires better interpretability
01:54 - The challenge of safety research on massive models
02:33 - Why specialize in Diffusion Models? (The math & theory)
03:00 - The Vision: "Automatic Sliders" for unsupervised control
04:34 - The new meta: Reward Alignment & Post-Training
05:18 - Generative AI for Science: Proteins & RNA expression
06:56 - Future Prediction: The rise of Discrete Diffusion Models

Special thanks to ‪@lambda-ai‬ for helping to make these interviews possible!

#NeurIPS2025 #GenerativeAI #DiffusionModels #LucaEyring #MachineLearning #BioTech #AISafety