My work has been featured by different forms of media and science communication (e.g. print, online, radio, podcast, …).

Video models are zero-shot learners and reasoners:

• The Economist: AI video: more than just “slop” – The next big thing in AI may be pictures, not words
• Two Minute Papers: DeepMind’s AI Just Solved Video Generation In A Way Nobody Expected
• Hacker News (front page): Video models are zero-shot learners and reasoners (video-zero-shot.github.io)
• The Decoder (front page): Deepmind says video models for visual tasks could become what LLMs are for text tasks
• Ars Technica: Can today’s AI video models accurately model how the real world works?
• Quantum Zeitgeist: Video Models Are Zero-shot Learners and Reasoners, Video Models Demonstrate Zero-shot Learning and Reasoning across a Broad Variety of Visual Tasks

Do generative video models understand physical principles?

• Two Minute Papers: Finally, DeepMind Made An IQ Test For AIs! 🤖

Humanity’s Last Exam:

• New York Times: When A.I. Passes This Test, Look Out
• Reuters: AI experts ready “Humanity’s Last Exam” to stump powerful tech
• Fortune: OpenAI’s deep research can complete 26% of Humanity’s Last Exam—a benchmark for the frontier of human knowledge

Beyond neural scaling laws: beating power law scaling via data pruning:

• AI coffee break with Letitia: Beyond neural scaling laws – Paper Explained

Partial success in closing the gap between human and machine vision:

• Machine Learning for Science: Do machines see like humans? They are getting closer

Shortcut Learning in Deep Neural Networks:

• Frankfurter Allgemeine Zeitung: Hier irrt der Algorithmus
• The Gradient: Shortcuts: How Neural Networks Love to Cheat
• Tech Xplore: Exploring the notion of shortcut learning in deep neural networks
• Knowable Magazine: Why some artificial intelligence is smart until it’s dumb (also in The Week)
• Bloomberg Businessweek: Goodhart’s Law Rules the Modern World. Here Are Nine Examples
• Psychology Today: Can AI really think?
• Deeplearning.ai: The Batch newsletter
• Towards Data Science: Shortcut Learning, The Reason ML Models Often Fail in Practice
• Underrated ML podcast: Energy functions and shortcut learning

ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness:

• Quanta Magazine: Where We See Shapes, AI Sees Textures
• Spektrum der Wissenschaft (German edition of the Scientific American): Form versus Textur
• The Register: Object-recognition AI – the dumb program’s idea of a smart program: How neural nets are really just looking at textures
• Two Minute Papers: Do Neural Networks Need To Think Like Humans?