Kerrick Staley's blog and homepage, serving up juicy posts about programming, Mandarin, learning (machine and human), and more!
I made an LLM agent micro-eval where I asked the agent to compute the square footage of an apartment given a floor plan of the apartment. This involved two steps: determining the scale of the floor plan by looking at room dimension labels (e.g. 20’ x 12’), and marking the parts of the floor plan that count as square footage (interior areas and exterior walls, but not exterior areas like…
Frontier LLMs like GPT-5.3-Codex, Gemini 3.1 Pro, and Claude Opus 4.6 have spiky capabilities, performing several stddevs above median human performance at some tasks while failing at some “easy” ones.
The law of total variance states that for two random variables \(X\) and \(Y\), \[\mathrm{Var}(Y) = E[\mathrm{Var}(Y | X)] + \mathrm{Var}(E[Y | X])\] If \(Y\) is a mixture of distributions, this tells you that \(\mathrm{Var}(Y)\) is the mean variance of the components plus the variance of the means of the components (\(X\) in this case is a discrete variable indicating which component you…
Matt Levine once wrote:
I have a thought experiment I do when I’m uncertain about a life decision: I imagine I have 100 replicas of myself that I could send off into the world to explore the branching life paths I could take. Where would these 100 replicas live? What vocations or lifestyles would they pursue?
I use a system I call homegit to manage config files and scripts in my home directory on all my machines. The idea is simple: a Git repository rooted at ~ that I push to GitHub. I’ve used this system for 6 years and like it a lot. See below to set it up for yourself!
Mid-morning last Sunday, I took a PATH train from Hoboken station to Christopher Street station en route to a coffee shop to hack on projects with a friend.
I live in Jersey City and many of my days start with a short commute into New York City on the PATH train. On weekdays, trains run frequently, but for weekend visits to Fat Cat Fab Lab or Vital, trains come every 12 minutes. Waiting 6 minutes on average isn’t bad, but not knowing how long I’ll wait bugs me.
In May 2017, I sat and passed the Hanyu Shuiping Kaoshi Level 4. The HSK is a Mandarin Chinese proficency test similar to the TOEFL, and passing Level 4 means that one can “discuss a relatively wide range of topics in Chinese and [is] capable of communicating with Chinese speakers at a high standard”. Just one level higher, at Level 5, you’re supposedly able to “read Chinese newspapers and…
I’ve been watching the Chinese TV show 他来了，请闭眼 (Love Me If You Dare). It’s a good show, kinda reminiscent of the BBC series Sherlock, likewise a crime drama centered around an eccentric crime-solving protagonist and a sympathetic sidekick. You should check it out if you’re into Chinese film or are learning Chinese and want something interesting to watch.