Tab Clearning (Feb 2025)
misc
Clearing some recent tabs. I’m not sure how useful these posts are to anyone else, but they help me finally read + close things I’ve been meaning to get to.
- An amazing study in nature doing ‘Optical widefield nuclear magnetic resonance microscopy’, using nitrogen-vacancy-doped diamond (which flouresces in a way dependant on the surroundings) in a thin layer below a sample to get per-pixel NMR data. Wild stuff. Currently 10um resolution but easy to see that dropping further.
- Recreating Deep Research from huggingface. Nice to see people showing their version of search in a tool loop. I also like the look of Jina’s version.
- New AI stuff
- Gemini 2 (nice cheap good models)
- Pika Additions, fun ‘VFX’ application of video models, hooray more control
- Mistral’s new ‘Le Chat’ https://chat.mistral.ai/ is supoer fast (>1000 tokens/sec) thanks to Cerebras and has a good canvas implementation, I like it
- Search was a PSYOP - between impling RLHF was the secret sauce when it was apparently mostly SFT (see Stella Bidermans tweets, Karpathy’s comments in his recent video) and then getting everyone excited about MCTS, PRMs etc when o1 was (pretty confirmed) more like r1, OpenAI’s muddying of the waters isn’t great. But TBH I think this is partly the fault of the community, being follow the magic leaders mode not pursuing things ourselves.
- First frame or subject ref -> video cool model
- https://personalitymap.io/ - see correlated (well, predicted to be correlated) traits.
- Kipply has a ‘things read’ post with some gems, e.g. this SA post Half an hour before dawn in san fransisco, wow, what writing!
- The future belongs to idea guys who can just do things an interesting take from gumroad guy
- Why blog if nobody reads it Andy attempts to motivate, while claiming “Let’s tell the truth, then: Nobody reads your blog.”. I don’t know that I agree. I’d say ‘nobody reads your blog when you start’ - but eventually, somehow, these things have a way of finding a handful of interesting people who like what you write! Who knows, maybe your blog is one I’ll carefully add to my list of ones I check and share.
- LIMO: Less Is More For Reasoning another hint that today’s base models are based ;) and it doesn’t take much to squeeze more perf. out of them. Not sure I agree with the conclusions on this though, I suspect it more reflects good data in the pre-training phase rather than the magic of the post-training.
- Gradual Disempowerment makes a good case for this being an extra risk we should think about when talking powerful AI systems. Not misuse and power grabs (my current biggest worry) or sci-fi style ‘kill all humans’ but just the slow loss of power as people are replaced in more and more places and thus lose the ‘bargaining power’ that has historically kept institutions and states somewhat aligned with the interests of the people.
- Trumpian Policy as Cultural Policy I don’t keep up much with politics but a lot of people here in Portland are somewhat traumatised by DT’s recent set of actions. This essay was an interesting lens to view them in. I should read more Tyler Cowen, and appreciate any suggestions for counterparties who might take a different view to balance him out :)