amos

@fasterthanlime.hachyderm.io.ap.brid.gy

hi, I'm amos! 🍃 they/them 🫐 open-source witch & maker of snappy videos and articles at https://bearcove.eu ✨ be kind, be curious 🌉 bridged from ⁂ https://hachyderm.io/@fasterthanlime, follow @ap.brid.gy to interact

You know, I've been very critical of GitHub in the past few months, but I have to say, I just opened a very large PR, and I was able to view the entire diff in my browser without expanding anything, without things lagging or whatever. That's nice!

Imagine working on an ASR paper, and you port your model from one system to another. And your word error rate tanks by 4.6 percentage points (a LOT) and instead of figuring out whatever you messed up and fixing it, you go on to run another hundred evals and publish the paper anyway. ?!?!??

A screenshot from an ASR paper called "Pushing the Limits of On-device ASR" published by Microsoft this year, highlighting a table, table six, because quantization results, and they show the baseline PyTorch is has an AMI score of 11.80 And then they ported FP thirty two weights to ONNX, And the AMI is now inexplicably 16.40

Something's wrong.

MONEY HAS BEEN SPENT I will soon possess: a proper COB light, a 90cm softbox, a C stand, a wider green screen, 2x90cm LED bars, an SM58 (SM7B too heavy to hold, shotgun mic in untreated room = bad), an Elgato 4K S (USB-C!), an acoustic blanket, various cables, gaffer tape, zip ties, nipple cla–wait