TODAY mini-hackathon
xaig docs
00 / the aigs pod · on air

Something to work to

Five stations for the long middle of a hackathon: Pittsburgh classical, Pittsburgh indie, Pittsburgh jazz, lofi beats to study to, and quiet ambient for when even beats feel like too many opinions. Turn the dial. The robot has headphones on and will not notice.

01 / the study · turn the dial
TODO xaig the docs ♪♫♪ kssh the supervisor
89.3 · wqed switch it on. the classical station is already warm.
02 / take it with you · your own player

WQED-FM 89.3

Classical music, live from Pittsburgh. Public radio; somebody real is choosing the next piece.

listen at wqed.orgraw stream

WYEP 91.3

Indie, roots and whatever the host fell for this week, live from Pittsburgh's South Side.

wyep.orgraw stream

JazzWorks

Jazz at every hour, out of Pittsburgh. Good company for a long-running job.

jazzworks.netraw stream

I♥CHILLHOP

Chillhop and lofi beats from ilovemusic.de. The cat approves, as far as anyone can tell.

ilovemusic.deraw stream

Ambient Sleeping Pill

Long, quiet ambient with no drums. Quiet enough to hear a test fail.

ambientsleepingpill.comraw stream
today / mini-hackathon · the brief

Two tracks, one long day

There are primarily two tracks today: expedition (analysis) and building (coding). Everything lives in the aigroup repo.

01 / expedition · everyone should try it

If this is the only thing you do today, that's fine; it's worth it. It can easily take up the whole day, and each item in it can too. Two ways in:

  • Meander. Open the notebook provided and wander around. It works wherever you have Python and a network connection, but it will download a fair bit of data.
  • Follow a recipe. Work through the five curated exercises, with or without your agents.
02 / building · ideas, all documented
  • A bigger SAE
  • CLT stuff
  • Refining the visualization capability (hint: cloudflared quick tunnels out of NERSC, http2 for compute nodes)
  • ?? anything goes
03 / feeling super adventurous · huggingface

The Hugging Face repos hold a lot beyond the obvious, including instructions to reproduce everything. There is a ~230 GB dataset of latent details (it lets you do the expedition without ever touching the underlying models) and a few GB of checkpoints for two real models we are testing now. New to Hugging Face? Ask any time today, or ask your agent; it knows this stuff far better than any of us.

Most importantly: please help us improve the repo and the code. Even half-baked thoughts or ranty complaints help a lot. Write down your impressions, your feelings, all of it.

  • Find a bug? Report it on GitHub, even if unsure.
  • Have a feature request? File it on GitHub, even if unsure.
  • Any comment or concern at all? GitHub too. All feedback is good feedback.

thank you
REMINDER