What is the rationale for the sqrt(n) (vs 1, i.e. just L2 norm) scaling in RMSNorm? The RMSNorm paper gives a rationale for this, but it's not intuitive to me. From their experimentation (RMS vs L2) it seems to be necessary though.

Why apply any scaling to the normalized values when we're going straight into a linear layer afterwards?

Also useful are the various easy to use fabrication services like sendcutsend (and many others.)

There are lots of things that you can design in CAD and have them manufactured and sent pretty cheaply, e.g. sheet metal hanging shelves.

Apparently the "dangly ringy thing" is called a "clapper".

Attach bells to your zipper and then sew a magnet into the cloth near the end of the zipper that will hold the bell and its dangly ringy thing in place.

My "Just QR it" t-shirt is more backwards stable than your pseudoinverse t-shirt

Chin checks out

I still think punitive tariffs on Mexico are dumb and counterproductive

Doesn't this graph show that imports from Mexico are larger than exports to Mexico and thus we have a trade deficit to Mexico?

I do find the negative number + "deficit" wording to be a confusing combination of negatives though

It's easy if you assume a spherical panda

Beautiful