Bless everyone who records their talks

“They said it could not be done”. We’re releasing Pleias 1.0, the first suite of models trained on open data (either permissibly licensed or uncopyrighted): Pleias-3b, Pleias-1b and Pleias-350m, all based on the two trillion tokens set from Common Corpus.

What are some solid pieces of writing in the vein of situational awareness and machines of loving grace?

Yeah, and thinking about it more, I guess a larger general model would probably beat out smaller specialized models too

What about with each agent having a different model?

You love to see it

Good one haha

Oh I see, that’s pretty cool

Why would you do this? Lol

Classic

Many are saying

It’s time to swap my terminal from simple dark/grayscale to colorful again. The question is, which color scheme?

Good morning everyone what’s the plan for today

I like the idea of the first name being a subdomain

Thanks!

I thought I was, but will check when I try again today

Can blocklists contain blocklists? Can you have a cycle in the graph?

White noise is the meta I’ve heard

Yeah I needed to see this today lol

Yeah makes sense honestly

Thanks!

Lol you just use F.sdpa?

Why is it so slow

Building wheel for flash-attn (setup.py) … /

Holy

scraping is hitting a wall

I think they do mean that it’s a preview/experimental model since they’ve identified some issues with it in their blog post

With some sort of scaling graph too! Wish they were more clear about what the x-axis represents

Finished training in the morning and evals already done is impressive speed

Wow