00:41 <_whitenotifier-a> [amaranth] agrif opened pull request #1627: vendor._gowin: use braces around net names in SDC - https://github.com/amaranth-lang/amaranth/pull/1627

00:44 <_whitenotifier-a> [amaranth] codecov[bot] commented on pull request #1627: vendor._gowin: use braces around net names in SDC - https://github.com/amaranth-lang/amaranth/pull/1627#issuecomment-3251231722

03:17 <kal> has anyone tried amaranth code generation with an llm?

03:21 <whitequark[cis]> yeah. it doesn't really work

03:22 <whitequark[cis]> it's not a language popular enough to be in the training set in an amount required for any useful result

03:56 <_whitenotifier-a> [amaranth] github-merge-queue[bot] created branch gh-readonly-queue/main/pr-1627-b8dc2338043c0de2938bb285068be4e6256b8cea - https://github.com/amaranth-lang/amaranth

04:02 <_whitenotifier-a> [amaranth-lang/amaranth] github-merge-queue[bot] pushed 1 commit to main [+0/-0/±1] https://github.com/amaranth-lang/amaranth/compare/b8dc2338043c...ff446833ccf4

04:02 <_whitenotifier-a> [amaranth-lang/amaranth] agrif ff44683 - vendor._gowin: use braces around net names in SDC

04:02 <_whitenotifier-a> [amaranth] github-merge-queue[bot] deleted branch gh-readonly-queue/main/pr-1627-b8dc2338043c0de2938bb285068be4e6256b8cea - https://github.com/amaranth-lang/amaranth

04:02 <_whitenotifier-a> [amaranth] whitequark closed pull request #1627: vendor._gowin: use braces around net names in SDC - https://github.com/amaranth-lang/amaranth/pull/1627

04:03 <_whitenotifier-a> [amaranth-lang/amaranth-lang.github.io] whitequark pushed 1 commit to main [+0/-0/±35] https://github.com/amaranth-lang/amaranth-lang.github.io/compare/8bc55e2e6294...494af3520d6f

04:03 <_whitenotifier-a> [amaranth-lang/amaranth-lang.github.io] github-merge-queue[bot] 494af35 - Deploying to main from @ amaranth-lang/amaranth@ff446833ccf49b2f75fbee92124ff3f08926eb32 🚀

23:12 <whitequark[cis]> ok, i decided to check how good or bad "agents" are. i downloaded cursor and instructed it to implement an amaranth blinky that can generate a bitstream for one board and has tests, and uses amaranth 0.5 syntax

23:13 <galibert[m]> that's not going to work :-)

23:13 <whitequark[cis]> the methodology was that i would just hit Enter whenever the model asked me to make a choice except in cases where it becomes lodged in a loop, or where a human would have to file an issue (the latest released yowasp-yosys doesn't build amaranth... bleh)

23:13 <whitequark[cis]> it took half a hour of doing that to get a bitstream

23:14 <whitequark[cis]> it took half a hour more to get a test, after which i ran out of tokens included in the free tier

23:14 <galibert[m]> heh

23:15 <whitequark[cis]> most of the engineering issues that arose in the process weren't amaranth-specific; it was things like "unable to use shell", "unable to use python packaging", "confabulates nonexistent modules", "it does not think and as such is incapable of reflecting on runtime behavior, much less conceptualizing that mutability and aliasing are a part of python"

23:17 <whitequark[cis]> it is unquestionable that the only reason i got any result out of it at all is because i already have an expert-level knowledge of amaranth. although i think it'd have fared slightly better if we had an up-to-date tutorial rather than a scattering of community examples

23:24 <whitequark[cis]> it is true that whatever model(s) it decided to use (the default mode is "auto", which seems to... make their backend autonomously decide on how to run your prompt and not tell you?) has gotten better at confabulating plausible amaranth code, which i attribute, without evidence, to the entirety of github being in the training set probably. but the relative lack of syntactically wrong generated code does not meaningfully improve

23:24 <whitequark[cis]> the utility of this tool for engineering