Community
Chat Logs
Friday, October 23, 2026 (America/New_York)
- ParaNeXeN: fwiw I don't care about context windows, most slop is the result of allowing most of the project to slip into context
- ParaContext is basically artifically injected noise to LLM model execution after the very initial starting point.
- ParaBut thanks for listing example models anyway. I don't really have interest/time to dig those up myself that much as there's just too many, so I do appreciate at least someone being on the ball about wtf is going on in the quantization space.
- dreamrealhttps://bytecode.news/posts/2026/10/arcana-reads-everything-and-then-some
- ParaYou haven't lived until trying to work with ZIP files with multiple entries under same name.
- * dreamreal has, unfortunately, lived
- dreamrealI've seen things, man.
- ParaThere were some other similarly esoteric issues with native JAR/ZIP handling but my memory has blocked those.
- * dreamreal has tests to catch RTL translations of names to convert *file endings* to look different than they actually are
- ParaOh yeah, file entries without names. That's also a thing.
- ParaOr file entries after the index.
- dreamreal... and that's just with zip. The problem arcana has is that it has no per-entry stream.
- dreamrealI'm considering seeing if I can fork it and give him a PR.
- dreamrealI cannot use it without observable and filtered data.
- Paradecompression bomb protection would also be prudent
- dreamrealarcana has that
- Paraah, good
- ParaHm, is Arcana Graal compatible?
- Paranative executable, I mean
- dreamrealdon't know
- dreamrealhttps://bytecode.news/posts/2026/10/andrew-lombardi-unit-tests-are-not-the-enemy
- whaleyoh, support for quake pak files
- * whaley rewrites qwtf
- ParaPAK is also used by Valve's games.
- Para...which of course mean they could also add VPK, a PAK derivative, support if it's already not there.
- whaleyhot take I just thought of re: that unit test article - perhaps a model where we TDD by writing the tests first... perhaps ideally property based tests when possible, and even the integration tests even if they don't compile, and then just tell the LLM, go make these pass with only minimum required changes the test themseles
- dreamrealif the content on BCN is useful or interesting, please consider sharing it among whatever communities you have: I'm not willing or able to do the whole social promotion thing, and it needs grassroots support. But ONLY if you think it's worth it.
- dreamrealwhaley: I'm a huge proponent of red/green testing myself
- ParaThat's what multi-agent flows of some of my colleagues do.
- whaleymy limited experience with clod is it's not super great at writing tests... it's akin to the junior dev writing simple tests that echo the impl with very simple assertions (could also be a skill issue with my prompting)
- Paraspec -> test writer agent -> implementation agent -> verification agent; loop back until spec is complete -> commit -> ??? -> victory...or next issue
- dreamrealit's a skill issue with your prompting, honestly
- * whaley nods
- dreamrealI mean, clod is not PERFECT at it, it still wants to infer a lot
- dreamrealbut you can instruct it to not guess when it comes to tests: if you can't describe the red condition well enough, it can iterate
- whaleyI probably need to go much more micro level on the prompting, is my guess (i'm also only just toying around atm)
- ParaI've been today messing about learning about AI Second Brain and System 1 models.
- ParaBoth seem interesting approaches to certain issues.
- ParaNeXeN: fwiw qwen3.8-27b works fast enough for me; sure, it's only about 10t/s
- NeXeNhey 10tk/s is faster than my vega apu does qwen 2.5-4b
- NeXeNthe app i'm building tho will run it 24/7 and keep up with regressions and decomposes tasks into bite sized chunks so it'll run well unattended. i'm at 64% success rate with the current epoch (automagically fixed 2 regressions)