@SwiftOnSecurity I don’t think they can, that’s why more and more providers are switching to usage limits / plans
English
irow
73 posts
irow
@irowmora
math and machine learning enthusiast professional lurker
Katılım Mart 2023
125 Takip Edilen3 Takipçiler
I just don't see how $30 AI subscriptions can sell that amount of compute for a dollar a day.
English
@sivareddyg What about different modalities rather than text based query-document pairs?
English
Controversial take: you don't need any of this. LLMs have gone through a lot of training already, so there ought to be a better method to turn them into extremely good embedding models. This is what my group has been working on. LLM2Vec is one such idea. We have some exciting developments recently where LLMs themselves can generate superior embeddings with zero changes to the LLM. Stay tuned!
Jack Morris@jxmnop
English
“Air Fryer” is one of the best marketing terms of all time
90% of people don’t know their oven does the same exact thing
English
@rob_mcrobberson I ran into this using an ai model with ~300 tokens. If the embedding dimension was too high the random initialization made it so none of the vectors had any random correlation. The signal going through attention was zero and the model didn’t learn. I reduced it down and it worked
English
another fun math fact: in high dimensions, everything is basically orthogonal. two random vectors in 1,000D will have a dot product so close to zero it’s spooky
rob🏴@rob_mcrobberson
because the radius gets “used up” by all the extra dimensions. in 3D, volume grows like r³. in 100D, it grows like r¹⁰⁰. so if you go from radius 1 to 0.99 in 100D, you don’t lose 1% of the volume you lose (0.99)¹⁰⁰ ≈ 63% of it. that tiny step inward nukes almost everything.
English
@TF_VAULT china's skins so far have been flashy but have no sauce
English
@SenseNopedOut @ALegendaryDrops One could make a compelling argument that it’s more fun solo than without friends. All the solos have been super nice.
English
It's wild how a game like Arc Raiders has exposed the entire industry's incompetence when it comes to optimization in Unreal Engine 5.
Arc Raiders is visually absurd, has the best sound, and performs great on a variety of hardware.
So what's your excuse this time?
Pirat_Nation 🔴@Pirat_Nation
Thanks AAA Devs.
English
irow retweetledi
irow retweetledi
We're still on track to release *** ******* on October 30th!
English
@krishnanrohit I have always believed you can only get so far with just text, there’s so much more to intelligence and problem solving than text.
English
irow retweetledi
all too often the only thing decided in a meeting is that another is needed
English
@francoisfleuret Nope, I was working on a “classification” using LLMs and a colleague told me to ask it for a score. The score would be completely made up and have no basis.
English
I am extremely confused when a statement formulated by a LLM about its inner processes is considered to have any validity.
Except if such information has been explicitly added to the training data, is there *any reason* for such a thing to be remotely true?
English
Didn’t they just have a blog post about how claude couldn’t even run a vending machine
Anthropic@AnthropicAI
We've launched Claude for Financial Services. Claude now integrates with leading data platforms and industry providers for real-time access to comprehensive financial information, verified across internal and industry sources.
English
@torchcompiled I think it’s because they only train about 1 epoch, dropout would only hurt the model. It’s good if you have to do lots of epochs
English
@FromWhiteCastle @BecomingCritter CEO created Nvidia while working at dennys
English
irow retweetledi
RT this post to convince a new player to play The Finals. It's time to put this game in front of everyone's eyes. #thefinals
English
English
