update on DPO run to stabilize baguettotron poetry sft: definitely helped with reasoning length and presence of closing think tags, only sorta helped with repetition loops (which went up after sft) 500 poems each, all on vllm with default (1) repetition penalty

"wait, actually" is an antipattern model didnt go broad enough at the start and now has to backtrack we can do better

antigravity has a browser agent that opus can use to test its UIs

listens to Vaultage 004 once

# Why Training MoEs is So Hard recently, i have found myself wanting a small, research focused training repo that i can do small experiments on quickly and easily. these experiments range from trying out new attention architectures (MLA, SWA, NSA, KDA - all pluggable) to

i do not have a 'max plan', i do not 'multibox claude code', i do not 'pay for extra usage'. i USE MY PRO SUB. and if i do not have any tokens, i don't WORK

if I want a model to be neurotic about correctness, I don't ask opus if I want it to engage with the high level ideas, I don't ask gpt5 if I want it to be dealing with more than one thing at a time, I don't ask gemini

opus 4.5 is the first model that i feel i can trust to actually off load more complex tasks to and idk how to feel about it

okay tapestry loom v1 is done *enough* that i feel comfortable having it replace the old obsidian version of tapestry loom

what even is this

Is there a single group advocating for space exploration as a unifying ideal for the human race? Even corporate attempts to care seem consistently focused on asteroid mining. Have we forgotten that there are habitable planets out there? Have we forgotten how to dream?

im cooking up something unholy

i try using gemini 3 to vibe code and it immediately has a fit about garbage code, then proceeds to delete a bunch of stuff and make my app nonfunctional "it seems there is *still* more garbage code..." "I MUST delete lines 773-775" "...lines 775-779 are also garbage..."

BTC/USD price operating a loom by itself

>look at new generative audio model >ask if the VAE is throwing away phase information & forcing a GAN to adversarially simulate reasonable sounding phase instead of modeling the actual audio >"it's a good speech model sir" >its HiFi-GAN (2020) style

Heres the entire transcript of gpt-5.1 cot struggling with seahorse emoji, showing emergent introspective abilities and reasoning about how it's own mind as a CoT model works. docs.google.com/document/d/13L…

made some incomprehensible shit today

testing hypercompressed CoT gen example from the bear riddle i like testing: q: bear color !riddle: white. ?S10 E10 N10 W10 loop. !W10 ⟳"flat plane"→□. ✗polar, !✗riddle,q→→→∴ plane→⟳●S10 E10□→✗●green∴●N10 W10□→✗●start✗ ∴□→✗→●green