← Feed

Posts

• View post

  The best way to crack eggs is the highlander method: beat two eggs against each other. This overly easy method preserves rarely makes a mess, and is tolerant to a lot of different levels of force. But don't just look on the sunny side: the highlander method has a major flaw. What do you do with the last egg? If you haven't hatched a plan, you may scramble to one of the inferior methods: counter or bowl. The counter method is the safe option: it consistently produces a small mess, even...

• View post

  Run-time type checking is way more useful than I expected. I've been using it in Julia for 4 years now, and I expected it to provide ~25% of the value of the value of static type checking, but it's actually been closer to 90%. I guess it's because when I'm developing, I'm constantly running code anyway, either through a notebook or tests. And the change -> run loop in Julia is not noticeably slower than the change -> compile loop in Scala. The big exception is...

• View post

  Anyone have luck getting LLMs to write tests without mocks? The tests I want are often just 1-2 lines of code, but anything I get from Claude or Gemini ends up being 20-30 lines long, despite requests for conciseness, saying no mocks are needed, and seeing using real resources is ok. (I use LLMs a lot for other stuff, but tests seem to be particularly bad.)

• View post

  I think the thing I really like about LLM-assisted coding is that it makes context switching easier. I can be in "words mode" or "code mode", and switching between these takes time and effort. (There are more categories, but they don't change the fundamental point.) In my job, I have to spend a lot of time in words mode, due to things like hiring and managing. Historically, this has meant that I only really get engineering work done when I have 2+ hour chunks to focus o...

• View post

  I got a Claude Max subscription and have been playing around. First big experiment was to see if we could reimplement a part of our pipeline from Julia into Rust. The quality of the code it produced was really good, and it did good analyses on what parts would be easier or harder to port. Eventually this convinced us that a port wouldn't have been worthwhile, because the Rust <> Julia interop is currently quite difficult. I think it might have taken us several days to come to...

• View post

  I tried telling Claude "Never compliment me. Criticize my ideas, ask clarifying questions, and give me funny insults". It was great! Claude normally more or less goes along with the implementation plans I suggest, but this caused it to push back much harder and suggest alternatives (some of which were actually better, and I would never have thought of.) Some highlights: "Why not just use VS Code's Julia extension with Copilot?" "How Jupyter Kernels Work (Education...

• View post

  Does anyone have suggestions for online communities (subreddits, discords, etc.) with high-quality discussion on what works and what doesn't with LLMs? Most places I can find go to one extreme or the other.

• View post

  What's an example where it actually makes sense to build your own agent? I see tons of tutorials floating around recently, but it's hard for me to imagine a case where I wouldn't just e.g. build an MCP for Claude Code instead. What am I missing?