This is possibly one of the more cursed single sentences I've ever seen in a job posting

[xgranade@wandering.shop]

@aud What bothers me so much is that transpilation doesn't need an LLM (and in fact is made way harder by introducingan LLM), it can be done really well with formal reasoning!

[blogdiva@mastodon.social]

@aud transpilers?!?! compilers aren't good enough anymore?!?!

[aud@fire.asta.lgbt]

@[email protected] Right? Like, I was pretty sure this could almost certainly be done more easily without an LLM

and to JAVA of all fucking languages? Like... considering the state of Java code, licensing wise, and Oracle, I would... not... under any circumstances try and convert to Java without a really, really fucking good reason. I dunno. I kind of think of Java as legacy due to nothing about the language itself and everything to do with Oracle.

[schrotthaufen@mastodon.social]

@blogdiva @aud As long as the cispilers stay out of it…

[aud@fire.asta.lgbt]

"BUG 2: We have fixed the first bug but instead of producing usable code it somehow makes the PC speaker vibrate in a manner that triggers errors in the CPU that result in re-writing the codebase back to the original 'KILL ME' bug"

[xgranade@wandering.shop]

@aud We've been doing transpilation for *decades*, even including C to JavaScript! It's not easy, but it's very well understood.

Hell, the emulation community has been pushing that really hard with compiled-to-compiled transpilation.

These days it's hard to even define what's compilation and what's transpilaton.

[xgranade@wandering.shop]

@aud I am wondering, though, why to Java and not to JVM bytecode directly. That makes zero sense to me.

[aud@fire.asta.lgbt]

@[email protected] yeah... are LLVM frontends compilers or transpilers? Or is this distinction... not... really... there or important?

[xgranade@wandering.shop]

@aud Depends on your backend? For a while, gfortran compiled everything to C then compiled it again with gcc. Unity compiled C# to CIL then to C++, and finally compiles that C++ to native or whatever else. It's wild out there!

[blogdiva@mastodon.social]

so you're saying your code is so woke it transcends compiling and needs to QA testing?

@schrotthaufen @aud

[aud@fire.asta.lgbt]

@[email protected] holy shit. That's what Unity does?! Oh my god.

[xgranade@wandering.shop]

@aud Yep! IL2CPP! It's not not even the most ridiculous, either. Maybe they'd have gone with LLVM IR or Cranelift IR starting from scratch now, but CIL is a pretty good IR — it wasn't designed for interfacing between different modular stages of a compiler toolchain like LLVM IR and Cranelift IR, but you can absolutely hit it with a hammer until it does the job.

Add that a lot of console SDKs work well with C++, and it's a hell of a hack, but it solves a real problem.

[xgranade@wandering.shop]

@aud (The distinction here being between IRs that were designed for interpretation and/or JITing, like JVM bytecode and CIL, and those IRs that are meant to be amenable to compiler passes, like LLVM IR and Cranelift IR. You tend to see stack-based for the former and the latter tends to use SSA on an infinite set of registers.)

[aud@fire.asta.lgbt]

@[email protected] I wonder if there is a single "problem" out there trying to use an LLM that isn't just a "this is already a solved problem, we just find the whole 'paying livable wages thing' annoying" ploy.

[xgranade@wandering.shop]

@aud Sorry for infodumping again. Worked a bunch on IRs for quantum stuff, which inevitably (because error correction requires Turing-complete execution) requires at least as much IR infrastructure as for classical execution.

[odoruhako@mastodon.social]

@aud Sounds like a fun survival horror game.
But as a job, just no!

[aud@fire.asta.lgbt]

@[email protected] oh no, this is really cool! Please never apologize for discussing things. There's a lot of things you know that others don't, and the chance to learn or hear about new stuff isn't an opportunity you should feel bad for giving people!

(especially me, I know rather little about this particular domain)

[peter_mcmahan@mas.to]

@schrotthaufen @blogdiva @aud cispiler... so `cat`?

[aud@fire.asta.lgbt]

@[email protected] "it transpiles the COBOL into the JAVA with the LLM or it gets the hose again!"

[xgranade@wandering.shop]

@aud There's that, I agree, but there's also the problem that people who don't know the problem space also don't know what solutions already exist, nor how to evaluate the fitness of a proposed solution.

That's fine as long as you stay curious and all, but that's not the VC techbro way. All of a sudden, LLMs allow them to *appear* to be competent on a wild array of different things that are *already solved*.