for other folks currently in an incident trying to resolve the chaos this caused, the first commit we've found with issues in our repo is from ~10:30am pacific this morning
I've appended `; tput bel` to the end of long-running scripts to get the same effect.
Fun fact: the `bell` control character is part of the ascii standard (and before that the baudot telegraph encoding!) and was originally there to ring a literal bell on a recipient's telegraph or teletype machine, presumably to get their attention that they had an incoming message.
To keep backwards compatibility today's terminal emulators trigger the system alert sound instead.
Sometimes a "PR machine" is just social capital at work. If there's anything else I've learned in this saga it's that Sam has had a consistent track record of building and maintaining highly positive relationships with anyone near his orbit, this board aside, and that it consistently pays him dividends.
there are so many comments here that really feel like they didn't read any part of the announcement other than that there's a thing called runes.
For me personally I tried svelte in the past and bounced off because there was too much implicitly happening that I needed to have a deep understanding of to model correctly. This solves basically all those problems for me.
I thought your video[1] especially did a great job walking through the pros this change brings. Thanks for all your great work on this!
if you think about the speeds involved, a single additional car in front of you on the freeway (or even any additional cars) adds pretty miniscule time to the total commute.
Let's compare a few situations. In the baseline you're tailing the car in front of you with a focus on not letting anyone cheat and get in front of you, let's say 50 feet away. Your commute is 30 miles, and in this frictionless sphere of traffic you're going 60mph the whole time. You get to work in 30 minutes flat.
In the second scenario you're following the 3-second rule[0]. This would put you ~285 feet behind the car in front of you. Let's say over the course of your commute 20 cars move in front of you. If the average car length is 15 feet, and they all are 50 feet away from each other, when all 20 cars are in place you're a net -(20 * 65) feet away from the original car, or 1300 feet total. At 60 mph that adds ~15 seconds to your total commute time.
Well worth having an easier time avoiding a potential crash IMO! Also has the benefit of helping prevent traffic to begin with[1]
GPT-4 is a fine-tuned model (likely first fine-tuned for code, then for chat on top of that like gpt-3.5-turbo was[0]), while PaLM2 as reported is a foundational model without any additional fine-tuning applied yet. I would expect its performance to improve on this if it were fine-tuned, though I don't have a great sense of what the cap would be.
> We develop a large multimodal model (LMM), by connecting the open-set visual encoder of CLIP [36] with the language decoder LLaMA, and fine-tuning them end-to-end on our generated instructional vision-language data
afaik sentence embeddings via sbert are still considered a pretty viable path. This may be what you were already looking at, but there's more info here: https://www.sbert.net/index.html
Part of its contents come from the "USPTO Backgrounds" dataset. From The Pile's paper:
> USPTO Backgrounds is a dataset of background sections from patents granted by the United States Patent and Trademark Office, derived from its published bulk archives. A typical patent background lays out the general context of the invention, gives an overview of the technical field, and sets up the
framing of the problem space. We included USPTO Backgrounds because it contains a large volume of technical writing on applied subjects, aimed at a
non-technical audience.