I used to teach AI & ML for media, art & design: PerttuHamalainen/MediaAI: Aalto University's Intelligent Computational Media course (AI & ML for media, art & design)
However, I killed the course as one can increasingly get better results with commercial tools than tweaking and finetuning models oneself. Instead, I’m focusing more on teaching advanced game programming and technical art, with some AI & ML material sprinkled in.
Meanwhile, this page is mainly just a collection of random interesting papers and blog posts.
This section is also very much under construction, waiting for me to clean up and migrate stuff from my private iCloud notes.
Results comparable to 8-bit training. Features a lot of cool tricks like stochastic rounding for unbiased gradient estimation
https://arxiv.org/abs/2509.25149
October 2, 2025 : The latest in a series of world models from Deepmind is able to learn to obtain diamonds in Minecraft purely in imagination. Such world models learn to generate/hallucinate visual observations (e.g., next game frame) given user actions. Therefore, one can execute trial-and-error learning ”in imagination”, without actually trying things out. Although with Minecraft, the AI could actually run the game and the imagination is not that helpful, Minecraft is just a common benchmark and the Dreamer v4 is general enough to be applied in other contexts such as robotics where real trial and error learning is much more expensive and/or slow.
https://danijar.com/project/dreamer4/
BeyondWeb: Lessons from Scaling Synthetic Data for Trillion-scale Pretraining https://arxiv.org/abs/2508.10975