It’s just a matter of perspective really 😆…

Show Post

as far as i can tell, it’s purely an RL feat seems like they 1. take a fully pretrained LLM (decoder-only??) 2. slap a new 255-element classifier head on it 3. RL for Calibrated Decisions (RLDR) i’m sure there’s a strict format for the prompt, and during RL it learns to associate option->output

Show Post

Some of what I disliked about earlier OpenAI models (5.1 - 5.4 era) is that, despite being very good at coding, they were so high grip that it was sort of hard to develop an idea with them. Astra is relatively high grip (OAI signature) but somehow not paralyzed by it. Strong model for sure.

Show Post

Coronal f z=2,d @(40) { f3 p = z * nor(2*C.rgb - R.xyy),t=p d=2; @(6) d+=d, p += sin(p.zxy*d+z-T) / d z += abs(1-len(p.xy))/3; O += f4(1.1-cos(p),)/z/z/abs(len(t.xy)-1) } O = tanh(O / 30)

Show Post

“Local leaders are hearing it, loud and clear: Flock’s spy system must die, either by voting to shut it down or to the sound of reciprocating saws in the night.” Rick Wilson's Flock piece: https://open.substack.com/pub/therickwilson/p/why-flock-has-to-die?utm_campaign=post&utm_medium=web

Show Post