General

  • Lee Robinson’s “5 things I learned from 5 years at Vercel”. See in particular discussion of the first lesson, “Go hard at work, then go home”.
  • Scott Adams’ “The Day You Became A Better Writer” is short enough to reread frequently.
  • Rands Shwags’ “Shields Down”: “Resignations happen in a moment, and it’s not when you declare, ‘I’m resigning.’”
  • “You Have To Be In The Water” is a good essay by Chris Paik on investing.
  • Aaronson’s oracle is very fun to play with, and suggests that the Predictor would be easier to build than you’d first think.
  • “A half-hour to learn Rust” is fantastic. All programming languages, including CUDA and PyTorch, deserve an introduction in its tradition.
  • Raghu Mahajan’s piece on why he chose to come back to India for his string theory work suggests which kinds of sciences are easier to support in countries that don’t have huge amounts of capital: “String theory research is very theoretical and uses a lot of advanced mathematics. But unlike experimental fields, it does not require laboratory space or expensive equipment. This makes conducting world-class research feasible with minimal resources, the most important of which include government salaries, travel allowances and computing expenses. Given these factors, on a scientific and professional front, it was an easy decision for me to come back to India and join the string theory group at ICTS.”
  • Jim Fisher’s “Don’t animate height!” is a nice illustration of how much more sophisticated frontend engineering is than moving buttons around.
  • In “The Intensive Margin”, David Friedman gives a nice framework for thinking about research. He distinguishes between the intensive margin of economics research — the “subjects that smart people have been writing articles about for most of the past century”, where “anything new is likely to be either uninteresting or wrong” and much work looks like “apply[ing] a new mathematical tool to an old problem … whether or not the new tool adds anything useful to analysis of the problem” — and the extensive margin — “the application of the existing tools of economics, including mathematics where needed, to new subjects” like “public choice theory, law and economics, and behavioral economics”. Work on the extensive margin to do interesting economics.
  • Sketchplanations makes visual explanations. See for example this neat illustration of the parallax effect and Arthur Eddington’s selection effects parable.
  • James Somers’ actually useful word of the day series. More “specious” and “derring-do”, less “petrichor”.

AI

  • Huang’s law.
  • Killed by LLM, and its inspiration, Killed by Google.
  • Stanford’s recent Language Modeling from Scratch lecture series is great.
  • Marin, Stanford’s open lab for building foundation models.
  • Nathan Lambert’s “The American DeepSeek Project”.
  • Calvin French-Owen’s “Reflections on OpenAI”.
  • Brendan Long’s “Can Reasoning Models Avoid the Most Forbidden Technique?” makes an obvious but important point: reinforcement learning (RL) affects the chain of thought (CoT) despite no optimization pressure on the CoT itself because the LLM to which any optimization pressure is applied at all is the same one generating the CoT. Empirical evidence checks out.
  • First Round’s “From Memo to Movement: Shopify’s Cultural Adoption of AI”. Especially interesting is the Cursor token spend leaderboard: “If your engineers are spending \$1,000 per month more because of LLMs and they are 10% more productive, that’s too cheap. Anyone would kill for a 10% increase in productivity for only \$1,000 per month.”
  • HRT AI Labs’ blog post on how they read ML papers is absolutely fantastic: “While a large degree of technical and mathematical sophistication is needed, the systems we build also need to be robust and maintainable. One principle we apply to achieve this is always using the simplest possible approach that achieves the desired outcome. For example, if a linear model is as good as a random forest model, we’d prefer the linear model. It is interesting to contrast this principle with the incentives in academic machine learning research. An empirically-driven paper is more likely to be published if it demonstrates novelty – but often when one optimizes for novelty, the results can be complex, which may make it less appealing in an applied setting.” See also their post on how they think about data.