notes_on_llms_september_26_26

September 26, 2026

still really using sol/luna luna I use for google searching or chatbot (not exactly ) or for things I dont want to waste $ on bc of the cheaper tokens

sol 5.6: none - do the thing. loop until i just stop random
med - do the thing with subagents and verify output via 
will make annoying scaffolding tests, but can be prompted to remove them
if our convo about impl is explicit, less than 5-10% context, references to filepaths/compacting useful here

astra : 
one to try to one shot something. Is annoying af to read/talk to, 
really feels claude pilled. this agents knows how to `python`, `sed`, and lie.
hope you love running bespoke tests to ensure your if statement is on the correct nested hiearchy. Programs like an asshole and will add features it thinks youd like (code duplication)

I have been avoiding Claude models bc every release they’ve ever had has the same issues I have with astra. They also are so openly hostile to opensource I feel horrible supporting them financially. AND that’s saying something even tho all the labs want regulatory capture.

just begun playing w/sol 6 but generally think it’ll be mostly the same as above. early thoughts are it’s pretty fast. Astra for crappy oneshot demos, Sol for single session, human in the loop iterations.