links 1/7/25: https://roamresearch.com/#/app/srcpublic/page/01-07-2025

• https://asteriskmag.com/issues/08/the-making-of-community-notes the team behind Community Notes describes how they do it
  • it's people who were working on "Birdwatch" before Musk bought Twitter, and they use algorithms derived from PageRank.
  • these guys are not, even a little bit, "based." Yet the Twitter userbase loves Community Notes!
  • if you have a capable team that firmly believes in "fairness", in auditable, open, participatory processes that don't put a top-down thumb on the scale on controversial issues, and they get to actually use the neutral algorithm instead of being pressured to make exceptions, you get solid results and community trust!
• https://pmc.ncbi.nlm.nih.gov/articles/PMC7132523/ 90% of bronchitis is viral. doctors are cautioned not to give antibiotics for it, even for long-lasting coughs.
  • https://www.aafp.org/pubs/afp/issues/2010/1201/p1345.html no really. bronchitis symptoms typically last three weeks. it's still not bacterial. antibiotics do not work.
• https://harmonic.fun/ AI for formal mathematical reasoning
  • it's Lean based! https://web.archive.org/web/20241230145058/https://www.nytimes.com/2024/09/23/technology/ai-chatbots-chatgpt-math.html
  •  
• https://drive.google.com/file/d/1-FFa6nMVg18m1zPtoAQrFalwpx2YaGK4/view Tim Gowers' research manifesto on AI for mathematics
  • it's not machine-learning based; it's a version of GOFAI that's formalizing the types of "tactics" or "moves" that a human goes through in a proof, trying to get the formalism right such that a computer proceduralization only has a modest, human-like amount of trial-and-error & backtracking rather than vast amounts of brute-force search.
• https://www.businesswire.com/news/home/20240514445088/en/Multiply-Labs-and-Retro-Biosciences-Announce-an-85-Million-Partnership-to-Advance-Cell-Therapy-Manufacturing-for-Age-Related-Diseases cell therapy manufacturing automation robots from Multiply Labs, used at Retro Bio.
• https://gopiandcode.uk/pdfs/sisyphus-pldi23.pdf
  • "proof repair" is the problem of updating formally verified software; if you have a library of provably correct code, and you make any kind of software updates to the library, now you also have to update the proofs such that it's still verifiably correct (assuming the update didn't break anything.) Progress towards automating this.
• https://openreview.net/pdf?id=RtTNbJthjV  the Karp Dataset, 90 reduction-based NP-hardness proofs (in natural language) for training LLMs to write proofs.
• https://web.stanford.edu/~jlmcc/ Jay McClelland, co-inventor (with Rumelhart of backprop fame) of the Parallel Distributed Processing theory of cognition, has been doing a lot with LLMs lately
  • https://web.stanford.edu/~jlmcc/papers/ChanEtAl22DataDistPropsDriveInContextLearning.pdf in-context learning works better when the training data is bursty and has lots of sparsely presented elements
  • https://web.stanford.edu/~jlmcc/papers/DasguptaLampinenEtAl22LMsShowHumanLikeContentEffectsInReasoning.pdf LLMs, like humans, show belief-bias effects on Wason and syllogism tests
  • https://web.stanford.edu/~jlmcc/papers/McClelland22CapturingAdvCogAbilitiesWithDeepNets humans learn human-made formal systems (like mathematics, computer science, logic) in order to solve certain kinds of difficult problems. perhaps AIs should also "go to school", being trained on math problems, in particular with diagrams as well as text. also, goal-directed motivation may require fundamentally different architecture from the usual LLM transformer setup.
  • https://web.stanford.edu/~jlmcc/papers/NamEtAl22LearningToReasonWRelationalAbstractions.pdf an empirical example of "taking the AI to school" -- fine-tuning on a McClelland-designed dataset intended to teach "relational abstractions" makes models perform better at math problems
  • https://web.stanford.edu/~jlmcc/papers/NamMcC21RapidLearningGeneralizationInHumansArxiv.pdf human mTurkers are better at abstract problem-solving tasks if they've taken high school algebra and geometry (no effect for other educational variables). they split pretty bimodally into people who learn a strategy and people who guess at random. this points towards "basic math education teaches systematic thought." also, small RNNs generalize much worse than humans, but who cares.

• https://en.wikipedia.org/wiki/Answer_set_programming ASP is used for difficult search & combinatorics or optimization problems. I'm struggling to understand whether it is in wide industrial use or if it's more of a research specialization.