Ben Kuhn on staring into the abyss as a core life skill. Bad Idea: it’s all about water. I’m not sure I endorse the results or the intention, but I like the way this guy thinks! Real books last longer than physical books, a win for the vintage. Hal Finney on choosing life. The Gift We Give Tomorrow 🥺 I acquired this sweater as the ultimate silicon valley wear, and read about Kiesler
Went on a meditation bender, here’s a book review that correctly characterizes how Buddhist philosophy is bipolar. Book review about meditation. Another book review about meditation. Meditation and sex scandals. Buddhism in relation to rationality. My experience with meditation is that it causes one to experience a philosophy of self (among some other things) and that does not induct that theory into truth. Buddhist philosophy is needlessly senseless because it’s reasoned from that experienced as opposed to more traditional technical thinking. It provides enhanced observational skills, but it’s not obvious that it’s correlated with “living optimally” as opposed to “goes insane”.
U-PALM improves PaLM compute by 2x. Sandwiching paper from Anthropic. Constitutional AI, in which RLHF becomes RLAIF. Language models are really good at generating evals. A prompting paper about making models use rules to do math via prompting. Finetuning models on their own chain-of-thought-ed text improves them. 100M GPT model interpreted (still a bit wavy, but I think we’ll get there 2023)! Meta pretrained science model Galactica, which offensively does not cite Minerva?? Finally read the muP paper on zero-shot hyperparameter determination. No press diplomacy, though I didn’t make it to the full-press paper. Analysis on how SOTA-chasing has affected research. Gradient Descent: The Ultimate Optimizer, an old paper that showed up at NeurIPS with turtles. Finally! Learned optimizers!
Book: Structure of Scientific Revolutions
Pleasant read, has really good narrative and was particularly fun because Kuhn focuses on the Copernican revolution, phlogiston theory and various other scientific progresses that I don’t know about because I’m uneducated. There are some obvious overfittings, dubious claims that were particularly uncomfortable for a clean bayesian approach to knowledge acquisition but also some nice tidbits about feelings, like Pauli’s personal crises. The book completes itself with a rather epistemically offensive plot twist. I have 1800 words of notes on the book and probably some inspired upcoming writing.
Essay(?): Designing Freedom
Designing Freedom is this Stafford Beer cybernetics “book” (according to Google, but it’s fifty pages). I don’t endorse it intellectually — not at all, they make these stylized control diagrams that don’t actually work and big handwavy statements. But aesthetically — super fun! I purchased another Beer book (who, by the way, has a daughter named “Vanilla”. Vanilla Beer) called Platform for Change which I often skim through. I wrote notes on this too!