In this issue: What Is HackerNews Working On, some AI competitions, lots of youtube videos and funding for AI
Previous month: Links For October 2025
Projects
- I’ve built “What Is Hacker News Working On” website, see the discussion. I’ve automatically tagged every post like “Working on open source GTA game” with tags like “game development”, “open source” using LLMs, and then did several passes to deduplicate things and make them more clustered. It was already helpful to find some interesting scientific projects, take a look!
- I’ve tried this competition about LLMs for The Embodied Agent Interface, but turns out the hidden test set is way too big and I need a week of GPU hours to create a submission. Blimey!
- I’ve also tried this competition about Early Training Evaluation of Language Models and even created some datasets, but their evaluation system returned errors and I had to reach out via discord to slowly debug them, lesson: get into good challenges early?
- I’ve updated my Project Ideas list
Links
- Cool podcast episode about AI in biology, an interesting thought is that in non-profit there are no dashboards like in business where you have to optimize things, which makes them less efficient and you have to come up with good proxy metrics
- Somehow I liked this video about design, but can’t quite understand why. Is it a contrarian nature, or is biomorphic design got popular lately too?
- There’s a new AI hub in Berlin opening soon! Check out this announcement from foresight institute, deadline is 2026
- I’ve bought new weights that measure body fat %, and this video was helpful to understand what’s normal
- I’ve played Heroes Of Might And Magic V this month again, and now I want to help with an open sourced version
- This is hilarious
- I’m using https://hckrnews.com to browse news less
- I like this idea of luxury beliefs, and I get a bit annoyed sometimes seeing things like “AI uses too much water” (these things are related)
Karpathy:
Software 1.0 easily automates what you can specify. Software 2.0 easily automates what you can verify.