Moving from object-oriented PyTorch- or TF2-code with tape-based backprop to JAX isn't easy---and while running grad() on numpy-oneliners is cool and all, you do wonder... how do I build actual big neural nets? Maybe you decided to look at libraries like flax, trax, or haiku and what you see at least in the ResNet examples looks not too dissimilar from any other framework: define some layers, run some trainers... but what is it that actually happens there? What's the route from these tiny numpy functions to training big hierarchical neural nets? Read on...

I was at a clowning workshop in rural Pennsylvania last weekend. A mix of acting, improv, and straight-up making a fool out of yourselves, it really was one of the most emotionally intense experiences I've had in my life. Read on...

How did we get here? What does that mean? And what about “The Surgery”? Read on...

“Can you compare perplexity across different segmentations?” Short answer: Not immediately. Long answer: Yes, as long as you have equal denominators and the same support!

Consolidated version of a huge (573 tweets, edit on 2017-04-09: 588) thread on Twitter about the relation between Natural Language Processing and (Computational) Linguistics.

Featuring @yoavgo, @emilymbender, @haldaume3, @gchrupala, @zehavoc, @earnmyturns, @kchonyc, @wellformedness, @_shrdlu_, @AlxCoventry, @StefanMuelller, @yuvalpi, @thedansimonson, @brendan642, @soegaarducph, @jacobandreas, @alvations, @trochee, @azpoliak...