69

Source document

[2026-02-15 00:22] xenonfun: interesting error after pulling all these plugins out and changing namespaces: (files: image.png) [2026-02-15 01:13] xenonfun: well shit. I didn't think about how much CICD time this is gonna consume 🙂 tho some really shit rust failures and stuck unit tests added many hours onto that. (files: Screenshot_2026-02-14_at_8.13.05_PM.png) [2026-02-15 03:19] xenonfun: if I kept this veolocity still be 25-60/month self-hosted potentially... seems Forgejo might be best option having it make research poc plan [2026-02-15 04:12] ajaxdavis: I'm still considering forking and making an agent based version control and ci cd [2026-02-15 06:06] lisamegawatts: we were talking about how to do the components, maybe eac one is an agent and it manages the data layer and ui, eg fetch weather is a metereology expert maybe... would be cool to have it make its own predictions and get feedback, try to beat the weatherman [2026-02-15 06:06] lisamegawatts: similarly ci cd could be continuously improving if agent feedback loop gets its own data [2026-02-15 20:00] ajaxdavis: thats some cool thinkinggggg [2026-02-16 05:46] lisamegawatts: <@187062569415081984> before i try to build any of this should check it out and make recommendations for architecture https://github.com/MonumentalSystems/universal-orchestrator [2026-02-16 16:25] xenonfun: https://x.com/RandyMcMillan/status/2023428658799772151?s=20 [2026-02-17 11:04] ajaxdavis: (files: safiersemantics_The_Meaning_Makers.txt) [2026-02-18 02:54] xenonfun: (files: long_output.ogg) [2026-02-18 09:57] xenonfun: (files: Screenshot_2026-02-18_at_4.56.25_AM.png, chunk_38_trimmed.wav) [2026-02-19 08:49] xenonfun: ``` Key observations: - Best val is 4.7275, current eval 4.7501 — slight wobble, hence patience 2/30, not a plateau - Drop rate has slowed from ~25 ppl/50 iters early on to ~7 ppl/50 iters now — normal curve flattening - Still ~11,000 iters before cosine decay kicks in, plenty of room Revising end-state slightly downward given current pace: ┌───────────────┬───────────────────┬─────────┐ │ │ Previous estimate │ Revised │ ├───────────────┼───────────────────┼─────────┤ │ Val loss @20k │ 2.7–2.9 │ 2.9–3.1 │ ├───────────────┼───────────────────┼─────────┤ │ PPL @20k │ 15–18 │ 18–22 │ └───────────────┴───────────────────┴─────────┘ Still meaningfully better than the novels run. The cosine decay phase will matter more here — model has a lot of parameters to tune and the LR drop from 1e-4 → 1e-5 over the final 8000 iters should tighten things up considerably. ``` (files: Screenshot_2026-02-19_at_3.48.50_AM.png) [2026-02-19 09:13] xenonfun: https://huggingface.co/MonumentalSystems/randygpt-s