Turns out RefCOCO had a bunch of bad data inside of it. Not just with the masks, but also with the prompts like "slut on phone" and "bitch kid". RefCOCO-M cleans it up.

Also used it for auto-labeling - it's crazy good for that

Especially good at detection & pointing cases - especially since the bigger models aren't good at localization

Only 2b active also - very fast

Can run it on a phone then?

Seems like it could be somewhat useful for people with poor eyesight or blindness

In terms of size yes, but I think it needs some work to get the model in the right format

couple people got it running on a raspberry pi though

sorry what does it mean for only 2b to be active?

My understanding is that, while all 8B are loaded into memory, for each token inference step only 2B are selected and used - so tokens are produced faster because there is less computation needed.

Hoping someone will correct me if that's not the right mental model!

Moondream released a new model preview this week. One of the improvements is a much bigger context window. To take advantage, I created a sample project that uses the query and detect skills to auto-label all the objects in an image.

If you need to label a bunch of training data for fine tuning or RL this can remove a bunch of extra work.

Not perfect but sharing early because some people requested it to help kickstart their own projects.

This is a bummer. I was just looking forward to being able to take more Weave meetings next month...

I'm not certain but it appears that it is not illegal to discriminate based on the school's ranking in the hiring process. It would, however, be illegal if it were based on a demographic stereotype a school is associated with.

That said, it still might not be a good idea to reject people just based on their school's rank.

Pretty decent equity range but yeah. Maybe 100-150 + equity?

You had me until you tried to sell me v-necks at the end.