A little late retweeting this, but it was super fun working with
@nicklovescode
optimizing splines to match curve detectors via gradient descent. here's one of the outtakes
Proud to share Curve Detectors
Last year at OpenAI we decided to closely study 10 curve neurons, dedicating a week to the project. We found a lot more depth than we expected, and weβre writing three papers on them. Hereβs the 40+ page part one!
When you generate images with VQGAN + CLIP, the image quality dramatically improves if you add "unreal engine" to your prompt.
People are now calling this "unreal engine trick" lol
e.g. "the angel of air. unreal engine"
The Kronecker product:
(i,j),(k,l) = x.shape, y.shape
a = np.kron(x,y).reshape(i,k,j,l).transpose(1,3,0,2)
b = x[None,None,:,:]*y[:,:,None,None]
a == b
@ericjang11
no axis labels. no distracting grid lines. not even a title. Just a single plot, centered and impeccability antialiased, that is sloping gracefully downwards, always.
When we make assumptions about what features exist in neural networks, they often prove us wrong.
It turns out that 4% of CLIPs final neurons (8% on a liberal interpretation) are focused on geography. I certainly wouldn't have guessed that in advance!
@sgouws
@sindero
i believe this is corrrect. since the reproduction number is fixed, increasing the duration of infection just spreads out the infections over a longer period.
However there was a typo T^{-1}_{inf} should be T_{inf}
@pmddomingos
But if you could find a vector pointing to the minimum, the problem becomes trivial. You just needs to follow it in a straight line till you're there