Haemin Yoo

@guizmaii My understanding is that the default executor is the most efficient, if nothing is hogging it. Loom based executor is less efficient but hogging doesn't result in drastic throughput decrease when hogged. Would love if someone more knowledgeable step in and correct me XD

@yoohaemin What do you mean by "But it's likely not optimal"?

At work, we were finally able migrate to JDK 25 and so to the Loom-based ZIO blocking executor 🔥🎉 Anyone already using this executor in prod? Any feedback compared to the default one?

I can think of two things that are the highest leverage items that humans can drive: test policies (what to test, how to test, how to discover what to test) and architecture (anticipating future changes).

What AI lacks in terms of software engineering is not capability anymore. What it lacks is trust and accountability. SWE will likely pivot from solving first-order business problems directly into making the most out of the new scarce resource: accountability of human teammates.

Wow, GPT 5.3-Codex xhigh is much stronger than Opus 4.6 for the tasks I'm working on. Not even close.

AI generated code need not be hard to review. After the changes look more or less complete, make them create commits in logical* order. *Not necessarily in the order it was written, but in the order you (as a human) would like to review them.

@olafurpg On second thought, I can just ask CC to write that for me :D

@olafurpg Can you share the script? I tinkered with it and ended up with so much pain. It kept saying the compiler had a bug, and I had to tell it the classpath is missing jars over and over again.

Claude Code is very good at converting sbt builds to Bazel. A critical trick is to give it access to a find-jar-containing-symbol script that takes a fully qualified name like com.jackson.JsonNode, walks your local coursier cache and prints out jars that contain it via `jar tf`

I used to also avoid leaking mutation in public APIs, but at one point I realized my literal job was to write APIs that mutate state or do side effects, just over HTTP.

@Hasen_Judi LLMs work very well if the patterns are well established and the requirements are clear (== dayjob code) It works abysmally if either the structure or the goal is not well defined, and I'm just exploring the problem space.

What hasn't changed since Sonnet 3.5: - It can do small functions ok - It can't do anything else Give it a small problem with ~3 constraints and it will write code that looks sort of ok but upon inspection the constraints are not properly maintained throughout.

@ghostdogpr It's actually nice to turn all of these AI thing off and write difficult code once in a while. I haven't done it for a long time :)

@yoohaemin It does a decent job adding minor features for a library I’m working on (like a new flag or option when there are already some existing ones). It’s also good at creating tests, but that’s about it.

Tried writing some decrel code today. Opus 4.1 has absolutely no idea what needs to be done to add a simple new feature even with ultrathink, despite relevant code fits entirely in its context size. I'll believe this AGI thing once it can write a substantial PR on decrel.

I wrote a blog post: Making Data Races Unrepresentable, or Read-Modify-Write Considered Harmful (for business logic flows) purelyimpure.com/data-races/

@guizmaii There is only one cure: "-old-syntax", "-no-indent",

@guizmaii @augustnagro Mill is better designed than sbt, but my experience is that it requires some more polish. It's not as smooth as I wish it had been.

Well, I really like sbt. I think it's quite simple to use yet super powerful, but have you tried writing a plugin? The `Task` abstraction in sbt is worse than a PITA, just to make your code compile. I've written 3 sbt plugins, the most ambitious one being in Caliban, and it was horrible XP 1 week of hard work to write ~400 lines (documentation and new lines included) of relatively simple code 😕 Sometimes, just to make the code compile, you need to move a line of code inside or outside a `Task { ... }` block and have no idea why it compiles or doesn't when this line is here or there. So you spend hours (!) just to find which code format will allow your code to compile. That's nonsense. Most would abandon before finding. Also, as Li Haoyi wrote, the model below sbt is really complex and hard to understand I've never tried Mill yet, of pure laziness, but I'm sure it's even better than sbt If you're curious, here's the Caliban sbt plugin I'm talking about: github.com/ghostdogpr/cal…

My PoV on the "Scala tooling issue" (I'm ofc highly biased, with 10y of XP with Scala and more familiar with sbt than most) sbt isn't simple, but have you used Gradle? It's way worse Scala tooling isn't that simple, but have you tried setting up a TS project? It's a nightmare. You have 10x more things to setup Yet Gradle and TS are way more used Logical conclusion: What are we talking about? 🤔

@Krever01 ... and I'm in a situation where I have to meet extremely tight deadlines, let alone filing bugs with minimal reproductions (which are time consuming often!). It's stressful when I get hit by bugs from all over the toolchain. That's what I mean by tooling is objectively bad.

@Krever01 ... Why mill-bsp and mill cli can't share build outputs unlike bloop, why mill classpath and bloop classpaths from mill are different, why scala-steward suddenly broke for mill, why IJ was unusably broken for Scala 3.4 (I had to switch to metals during this period), ...

Scala doesn't need tooling. This got solved few times over. Scala doesn't need stability. This got solved with Scala 3. It's not perfect but we have a decent balance between stability and innovation. What scala needs is good press. When was the last time Scala was advertised at some big non-Scala conference? When was the last time the Scala-praising article was featured on hacker news? When was the last time big companies using scala admitted it publicly? I'm looking at you Apple, Amazon, Spotify, JP Morgan, Blizzard - I know you have Scala but only because I talk to people, this is not super public information. Market trends are shaped by believes and perception. Technical merits are secondary. We are terrible at marketing and fighting bad press. That's it.