1. Aesthetic mag for computing / science. Some sort of aesthetic publication for research or tech.
    1. https://increment.com/
    2. https://www.telescopemagazine.com/
    3. https://press.asimov.com/
  2. Artisanal benchmark eval set. Take an eval dataset and get a ton of ppl to actually look through all of it and clean it -- artisanal (labor intensive) and high quality. But also seems like a task that is easily parallelizable. Host a data cleanup party?
    1. https://www.adept.ai/blog/fuyu-8b “The question-answering datasets are quite flawed - they use a complicated scoring mechanism, require you to respond in a specific format, and are often annotated incorrectly.”
  3. Eng x research incubator / accelerator. Honestly like run some version of AI Grant, but for engineer-researchers? Is it possible to do academia without affiliation with a college / official PhD program? How hard is it to get a paper published in Arxiv? Or get something through conference review process? How do we get graphics nerd x ML nerd crossover? How can we support independent researchers? Github for ML eng / doc writing?
    1. https://twitter.com/jxmnop/status/1731726810470056258
    2. https://twitter.com/charles_irl/status/1724110196744835193
    3. https://blog.nelhage.com/post/pickles-and-ml/
  4. Design agency for research orgs. Design agency or something where designers and illustrators and web devs partner with research orgs to make their stuff (papers, etc) appealing and accessible, design merch and posters for conferences, etc. to make something like:
    1. explainer diagrams https://twitter.com/srush_nlp/status/1735377532969525608
    2. animations https://twitter.com/chrisoffner3d/status/1729483814240469004 https://twitter.com/BrendanBycroft/status/1731042957149827140
    3. https://twitter.com/Ella_Maru/status/1311720363102011393
  5. Data viz as a parascope for ML models (interpretability). How can new ways of "seeing" matrices, numbers etc help us understand how models work?
    1. https://twitter.com/thesephist/status/1728136248521564527
  6. Bring back Distill. (https://distill.pub/2021/distill-hiatus/) or somehow continue the mission of "let researchers easily make websites to share learnings" -> Medium for research?
  7. Gameboy that runs an LLM locally. Game cartridges that have prompts / RAG loaded? Maybe a new instrument/synth runs some music model? Possible that tiny models at narrow tasks work well.
    1. https://twitter.com/fabianstelzer/status/1727753139728662670