Bill Traynor

Screw vibecoding. You know what LLM coding agents are *really* useful for? Reverse-engineering legacy code! "Hey can you research and explain how the flux capacitor settings are loaded, specifically where are the margadon-level overrides applied exactly?" It's a lifesaver.

390

101 Comments

I reposted a hot (lukewarm?) take by Jeff Denworth asserting that "no one needs a file system for AI training" last week, and I was surprised by how many people struggled to imagine a supercomputer without a parallel file system. Although I've given plenty of talks on how this is done, I never wrote it down before. To right that wrong, I wrote up an explanation of how the entire LLM training (and inferencing) lifecycle is implemented at scale without a parallel file interface: https://lnkd.in/gUbKABBC I hope it helps my colleagues in the scientific computing side of HPC appreciate how the AI community is approaching a challenge that's as old as HPC itself. #HPC #AI #storage

309

24 Comments

Accidentally discovered a new way to work with GitHub Copilot today: write new tests based off test coverage reports. Copilot (or I) run this command to generate a report of all the uncovered lines in the current diff: pytest --cov --cov-report=xml && \ diff-cover coverage.xml --html-report coverage_report.html && \ open coverage_report.html Then Copilot reviews the report, summarizes the uncovered situations, and proposes new test cases to add, or if the code is truly unreachable, suggests deleting it. Just another reason why you should add diff-cover (or similar tool) to your projects!

115

4 Comments

As the maintainer of some vaguely popular OSS projects, I've noticed an increase in AI-slop PRs. I actually dig small PRs doing helpful things, and I appreciate the hustle, but they should probably compile ... or at least reference files that actually exist. My advise to anyone who wants to contribute to OSS is to talk to the maintainers FIRST, and check for a contribution guide on how to do the work.

39

So this guy commented on my video about the new TypeScript compiler written in Go waving the finger and saying "ACTUALLY...it's not a compiler, it's a transpiler". Now, I know that we computer scientists like to invent words that mean nothing, however, holy crap, "transpiler" is just a total avalanche of uselessness packed into a package of endless uselessness. If you search for an official definition of transpiler you will find a wikipedia page called "source-to-source compiler", which is defined as "converts between programming languages that operate at approximately the same level of abstraction". This is complete and utterly useless. We use a word for a software that converts a language into another language: it's compiler. But hey, we decided to add this completely useless concept of "levels of abstraction" for programming languages. Now if you want to compile from a source, and your target is a language that "more or less" (facepalm) is at the same abstraction level, then we have another word for that: it's transpiler. Seriously. Now, I'm glad that this word is not widely used anymore, and even Anders Hejlsberg in his tsgo presentation didn't use it, preferring compiler (which is correct), but the fact that we still have people out there in our community that wave fingers at you if you don't use words invented in circles of the IT nobility is insane. Especially in a world where we don't know what WiFi even means.

62

145 Comments