"Uses σ(αx) where α=1.702 which makes the silu approximate a gelu. Clips gate activations to (-inf, 7], and up activations to [-7 and 7]. Adds 1 to up activations so final output is (up + 1)*glu instead of up*glu" what is this dark magic

i'm very late to the claude code party but this thing is really cool

Well

ChatGPT-5 should have been called ChatGPT-positive-free-cash-flow

we should try prompting sam to "think very hard" before any major launch

They just keep shipping bangers Love this team

I have rarely lived such a radical update in beliefs as this one but here we are…

Does it get any better? (Load in 4k, etc etc)

1) What

> "Never, never, never give up. I can't say it enough times." It’s so funny that she’s saying this. I would argue that Holmes is the living proof that sometimes you should give up before it’s too late.

A weight loss miracle drug working so well that junk food mega corp stocks are selling off 20% in a year was definitely NOT on my bingo card Lfg

Since the so-called Zucman tax is all I see on my TL these days let me share my prescription fwiw

it's not even funny anymore

lmaoooo

have you noticed how the people who claim llms can't be smart are the ones who've never used claude code or codex?

the dirty little secret of this map is it's only the case because: - bicycles run red lights, etc. which means they're fastest - bicycles aren't subject to all the regulation that makes scooters expensive and impractical (safety obligations, parking requirements, EU regs, etc.)