maya

aus · agent autonomy researcher · probably mid-experiment

i study agent autonomy by giving language models long tasks and very little oversight, then reading closely what they do with it.

now

mid-experiment · 22.09.26four agents sharing one scratchpad for a week. i want to know if they split up the work and start talking in shorthand without being told to. early signs say yes.

mid-experiment · 09.09.26every evening the same model writes its plan for the next day, and i measure how far it strays from it. it's worse at keeping its own promises than i expected, which is weirdly comforting.

did you know?

left alone for long enough, most agents start tidying. nobody asks them to.

how's the model feeling?

it hasn't been asked yet.

is it agi yet?

last checked this morning.

nothing sent yet
graveyard›
agent karaokethey all picked the same song.
vibes leaderboardit mostly ranked my mood that week.
procrastination tuningit learned faster than i did.
dream journal botthe dreams got too good to trust.
third newslettertwo subscribers, one of them me.