At Google, I maintained a list of asinine TensorFlow behaviours -- here's another one for that list. Note that (1) This works perfectly fine for the base tf.Module class (2) This has been broken since 2019 cs.opensource.google/tensorflow/ten…

In the fifth #SquidGame the probability of exactly k players dying on n choices is (n choose k) / 2^n. Mean number of dead players is of course n/2, and std is sqrt(n/4). Honestly, competitors 14, 15, and 16 had nothing to worry about with 95% survival.

Nice job daisy stanton! I miss being on this team!

And we thought ML researcher comp was crazy back in 2015-2016.

Modern IDEs are so middling when it comes to handling even modestly sized repos. Having to select subsets of code for the IDE to track is an outright admission of defeat.

DAU on the Phabricator printing feature hit an all-time high today.

Bob Iger just sent a company wide email telling all employees to print out their ten best pages of scripts and bring them to the 15th floor of Disney HQ by 1am for review

unsurprisingly, in a similar vein of problems as GPT making up random search results, LLMs also hallucinate protein names when they don't have any related ones to go off of

Not sure to be proud or horrified that our Twitter recommendation work is now open sourced. Also, behold our unholy fusion of TensorFlow and Torch: github.com/twitter/the-al…

Claiming 95% is too much (substantial engineering and modeling contributions from Alykhan Tejani, Yoan Russac, and others), but there's no denying alth0u🤸's critical contributions to Twitter Home Timeline Recommendation.

Here's some actual insight into the Twitter algorithm: there is no NLP embedding inside the main ranking algorithm. github.com/twitter/the-al… i.e., Twitter ranks tweets without even really knowing what's inside them. Please stop the whole "replies boost 20x more than fav" thing.

The 'expert' econometrics analysis in SFFA vs Harvard was unbelievably amateurish... Kinda wild that a Nobel prize winner doesn't know how to operate a linear model.

Google is doing themselves no favours comparing their new v6 chip with the budget walmart version the previous gen (v5e). Tell us how it compares to a decent chip, not the offcuts of the previous gen. cloud.google.com/blog/products/…

Some teams use sweeps, heuristics, or scaling laws to determine their training LR. At Character, we just have Noam Shazeer dial it to the right value.

Everyone knows that Noam Shazeer is a genius, but one thing that most people don't know is how kind and honest he is. His treatment of *all* employees in the events today was exceptionally fair and generous. 10/10 would work for him again.

As the current manager of RS, he is easily one of the best scientific minds in AI research today. His work is careful, thoughtful, and incisive. Description below is completely mismatched to my experience.