IRC channel logs
2026-02-26.log
back to list of logs
<fossy>hm. i've had a very unpleasant thought regarding agentic coding <fossy>i've been saying for a few years now that we shouldn't trust autogenerated code <fossy>but espeically in the last 6 months, we're seeing whole projects of AI written code <fossy>no human involvement at all, on a widespread scale <fossy>but i need a rethink on my policy now... <malina>it's not many years before human (source code) is completely discarded and pure machine to machine (code) will be the de facto infrastruture <malina>disconcertin absolutely for teens of today hoping to study computer science / programming I can imagine <malina>(in recorded history): the first golden age of 'mankind' was ~ 5.5ka ago, the 2nd ~200 yrs ago, 3rd ~50 yrs ago, 4th hrd to quite guess , organic computing/ers will be a big one but it could very well in hindsight we are seeing the proto transition nowish into the 4th <malina>and as they say, reality is weirder than fiction ^^ <malina>ye, don't take my statements here as if I think it's a "good thing", it would have been had humans been more attentive to the global terrorist state but a long time they just didn't giv a toss so here we are. and I think it is too late to kinda think we will have a benevolent future. <aggi>in a way, a compiler producing machine code from human-readable source-code isn't news <aggi>"AI written code" is marketing fuzz, which refers to "transpilers" mostly, that is compilers which translate one programming language into another <aggi>another approach is code-generation from UML diagrams for "rapid application development" (flow, sequence, class diagrams), that's not news either <nimaje>I wouldn't call a llm a compiler, as there is no semanitics for the input language, that the llm obeys <aggi>ok nimaje, but i am referring to software-technology which does auto-generate code which does run for many decades already <aggi>i've seen alot of such (often awful) auto-generated code, and i am not aware of LLM generated software which does work in practice <aggi>in any case the limitations of auto-generated code are known <aggi>a most prominent example i remember often is Java-Hibernate and related products, which auto-generate relational-data-model SQL from UML class diagrams <aggi>which cannot cope with the "object-relational impedance mismatch" appropriately, often (there's a wikipedia article online) <aggi>although that's domain-specific to databases this talk is highly inspirational still: "10 ways to wreck your database" ... with auto-generated code <aggi>software-industry had been promising great benefits of code-generators for decades... it's been a billion dollar cash cow selling such products <aggi>nowadays they're abusing "LLM"/"AI" for this <aggi>so, you MUST buy some Nvidia GPU and their software-development utilities, otherwise you can't produce software... ludicrous <aggi>nothing relevant changed, teenage CS students still may pick some Vi editor and learn programming C and assembly <matrix_bridge>... published the prompts they used to create the compiler or the test rig they used to spin up and supervise the agents. This means it's hard to reproduce their work (and there's also the issue of it costing ~$20k in API usage). I wonder if you could just pay a bright undergrad a couple of grand and get them to write a textbook compiler implementation. <Piraty>of course they save the good stuff for themselves and $gov before it goes public <Piraty>i bet there is more in the drawer they use if for (and let YOU use it for) than is currently known <matrix_bridge><Andrius Štikonas> LLM generated code can definitely work. Whether that's a good idea for you as a learner that's another question. <mwette>mes team: nyacc-3.03 is released, with the new cpp, and using smatch.scm, which is in the dist under modules/ but not copied by "make install" <janneke>what mwette said: "we" need to check if nyacc-3.03 works with mescc, mostly saying this to the mes team, ie, also myself :) <stikonas>hmm, can't download nyacc-3.03.0.tar.gz from savannah... <stikonas>maybe it is still propagating through mirrors <janneke>stikonas: mwette has moved nyacc to github, i believe <stikonas>this is just automatic github tag archive <stikonas> +> /usr/bin/mes-m2 -e main /usr/bin/mescc.scm -D HAVE_CONFIG_H=1 -I include -I include/linux/x86 -c lib/linux/x86-mes-mescc/crt1.c -o lib/x86-mes/crt1.o <stikonas>unhandled exception: unbound-variable: (syntax->datum) <stikonas>didn't really fix the issue, nothing ot thank for <mwette>stikonas: can you check lalr.scm, line 136. I may have botched your fix wrap-term. <janneke>ACTION also gets "unhandled exception: unbound-variable: (syntax->datum)" <janneke>btw, the 3.00.0 tarball isn't up (anymore?) on savannah? <mwette>Hmm. It's delta from your implementation of syntax-rules, I guess? <mwette>The one maybe odd thing I did was update nyacc/lang/sx-util.scm to use _ as a keyword, to make ,_ be an "ignore", like match. <mwette>to check, in nyacc/lang/sx-util.scm, you could remove from line 416 and comment-out line 418; same for 489 and 491 <janneke>mwette: btw, your define-macro-based match works beautifully for NYACC, but it fails this test: <janneke>as we don't need it, it doesn't matter; i've marked it as fail for now <mwette>Ah. Would need to add the ? form. <mwette>Oh. That is in there. There was a simple test that seemed to work. <mwette>In guile, that match returns 'symbol for me. <mwette>(smatch 'foo ((? symbol?) 'symbol) (_ "no match")) => symbol <mwette>It's in module/smatch.scm in the 3.03.0 tarball <mwette>Currently nyacc will use (smatch) so you need to tell me what I should use -- e.g., (ice-9 match) if you put it there.