New from Google Research! REALM: realm.page.link/paper We pretrain an LM that sparsely attends over all of Wikipedia as extra context. We backprop through a latent retrieval step on 13M docs. Yields new SOTA results for open domain QA, breaking 40 on NaturalQuestions-Open!

Pre-recorded talks at conferences: I'm handling the flipped classroom just as poorly as college students supposedly do. #1.5xForTheWin...

Hey United Airlines, how about investing in making your booking website fast and more reliable?

Hey United Airlines - EWR-SFO Service is empty - yet you deny your best (status) customers the usual perks as if it was int’l, and then you don’t feed them (even for purchase). What gives? Give me one reason to not fly JetBlue next time.

Offered a 1:1 match for donations to my two awesome presidential candidates this week. Budget exceeded. Yay! Try it - set budget and deadline, then publicize it well.

None of this is normal, and confronting climate change is on the ballot this year. Vote, as early as you can, for a habitable planet.

Watched Battery Day presentation. Mind blown. There's hope. #Teslabatteryday #ElonMusk #Engineering

Awesome shot.

Beautiful filmmaking. An ode to detail. youtube.com/watch?v=ooTN3d…

What a productive, mindful day! #facebookdown

Got my new #urtopia bike - rides nicely around Manhattan! The ugly: bell, headlight, and alarm are not working, and it arrived with some scratches and a missing part.

1/N Tired of listening to your multilingual TTS models? SQuId 🦑 is an automatic metric for multilingual speech synthesis: give it a waveform, it predicts how natural it sounds. To develop the model we gathered 1.9 Million listening tests in 65 locales. arxiv.org/pdf/2210.06324

Dungeons and Dragons dialogue makes for a great NLP eval set. We'll have a dataset for you ready at EMNLP! arxiv.org/abs/2210.07109

Can LLMs output be fluent and well-grounded at the same time? Context size, model size, temperature all matter. Here’s one secret trick: multiple inference steps with re-ranking = model size 540B->8B, with same tradeoffs for sensibleness and grounding: arxiv.org/abs/2302.05578

Bowman on "Eight Things to Know about Large Language Models": cims.nyu.edu/~sbowman/eight…

Rook g6!!

Need to ground your LLM generations to a knowledge source through in-context learning? Try our novel temperature sampling algorithm that boosted PaLM-8B attributions (over QReCC datasets) by up to 9% without sacrificing sensibleness and diversity. arxiv.org/abs/2306.01286

Visible and quantitatively confirmed improvement of text-to-image models by training with clever feedback from two secondary models.

#CoNLL2023 will feature not one, but TWO awesome keynote speakers! On December 6th, we have Preslav Nakov ! Preslav Nakov On December 7th, we have Mohit Bansal ! Mohit Bansal The full program, including their talk abstracts is here: conll.org/program-2023 see you there!

Watch Starship's eighth flight test → spacex.com/launches/missi… x.com/i/broadcasts/1…