Community
Chat Logs
Thursday, October 22, 2026 (America/New_York)
- deeboi just use IDEA MCP from copilot/claude
- deeboif client wasn't a massive corporation with strict ai rules, i'd probably just use pidev
- deebojetbrains' tools have been a bit lagging, dunno why but my weird "lets AI suggest a massive empty block of text with unrelated code"-issue is even back
- deebolike i autocomplete a line and then it suggest addding a similar copy of the line and like 10 empty lines
- deeboand who ever thought 'tab' was a good key to complete those should be shot
- ParaThere's Pi support (and opencode and two dozen others, everything ACP) in Air as well, but I think the actual onboarding is just as dumb for it as well.
- ParaDepending on which plugin exactly you use you can control the autocompletion models, might be just that.
- ParaAnd/or the plugin not parsin agent responses properly, this was part of the problem with AI Assistant at least.
- ParaI do feel like Liam Neeson climbing a fence with Jetbrains' AI tools. When everything's running, it's great, but getting there seems to be just hard.
- NeXeNpara: https://github.com/Brennername/cacophony <----local llm
- nevetGitHub - Brennername/cacophony: A local model arena for tuning APU, GPU, TPU, and CPU edge deployments
- javabotNeXeN's title: "GitHub - Brennername/cacophony: A local model arena for tuning APU, GPU, TPU, and CPU edge deployments · GitHub"
- NeXeNif you got nvidia card i'll support all issues you post
- ParaJust bought 9070XT :)
- NeXeNawesome, it's vram that's usually most important, if you have over 8
- ParaGot 16.
- ParaAnd 64 gigs of DDR5.
- Paraand a waxwing in a hawthorn fence
- NeXeNi have 5900hx ryzon 9 so it's got the vega apu and it's a trickle at ~11 tk/s ......card like yours should get 15-25 tk/s
- NeXeNand 16 g you can run most the big models even if you have to trim the context a little
- NeXeNi've found that deterministically fixing errors is less time consuming than re-doing until a test passes (like claw)
- Paragpt-oss runs at ~95t/s
- ParaToo bad it produces mostly garbage.
- ParaSome gwen 4b quants run at about 35t/s and that's plenty.
- Chronos64 GiB RAM is almost enough to run 1/2 of $WORK's ridiculously mammoth JVMs ;)
- dreamrealOkay, so https://bytecode.news/posts/2026/10/alberto-denia-code-is-clay-now
- Opticlooks like temurin linux builds are up with the 2026e timezones and eMudhra cert revocation
- Chronosdreamreal: I could not help but post a slightly snarky comment on there, heh
- dreamrealHah. You shoudl have anchored the comment. :D
- dreamrealIf you highlight text in the article, you can comment anchored TO THAT TEXT :D
- jbosmansTIL "the cost of changing code has collapsed"
- jbosmansi'll just leave it at "code is clay" is way over my head, and then some
- NeXeNPara: yeah for you probably best qwen 2.5 coder 14b instruct, (q4_k_m or qat), deepseek-coder-v2-lite-instruct has moe, codestral 22b q4_k_m (has good fill in middle), deepseek-r1-distill-qwen-14b. most context windows you can get away with 16k, 8k or bump down to 4k on huge models with not too much prompt/too many files
- NeXeNalso gemma2:9b with q4_0 or q4_k_m or even 8 bit at q8_0.....gemma2:27b at q4_0 though you'll probably have to bump down to 8k or 4k context size
- NeXeNthat project i'm working on has ability to configure the models in json for each role and globally with overrides as well
- NeXeNcan test variations and it will select the best ones over time
- NeXeNhas a multi-armed bandit and stochastic exploration using epsilon-greedy (default), ucb-1, or thompson sampling