Daily Notes 2025-07-19
by Gordon Weakliem
Catching up on tabs, reading Steve Havov on better A/B testing. I’ll keep this around for later. It would be interesting to try to build out a framework for this, where a site could self-modify if it sensed one option was better than another. He’s talking about changing a button based on feedback, what if you could change workflows, or even a whole site, but dynamically adjusting to feedback? Maybe report in at some threshold and create an issue directing your coding agent (@claude or whatever) to use the information from the statistics that you’ve gathered to redesign the site. I have a site that briefly was self-modifying based on user feature requests before I shut it down over cost concerns.
Thinking about an old post by Sam Ruby that I can’t locate now, I remember him saying something about phones used to be super reliable, but cell phones now drop calls and are much less reliable, but we accept that tradeoff. I’ve been rereading The Black Swan it occurred to me that this trend continued, we’re much more tolerant of failures and we now routinely architect things to accept that failure is a normal operational mode. Chapter 15 starts “Forget everything you heard in college statistics or probability theory. If you never took such a class, even better.” I’ve been regretting my lack of a solid prob/stat background but maybe Chapter 15 will give me redemption?
Deep Learning is Applied Topology talks about separating data in a plane by moving to a higher-dimensional space. About a year ago, I was drawing a lot of inspiration from this post analyzing HN with embeddings, then using UMAP to reduce the dimensions to make visualization easier, the opposite process.
Pretty awesome video on how to use LLM, Simon Willison’s tool for interfacing with LLMs. I especially liked how he dived into using OpenRouter and the --template option.