make "Segment Anything" 8x faster by just using PyTorch's more advanced features. No C++, no custom extensions. great post (and work) by @cpuhrsch and the team!
@soumithchintala
-
Meta’s Generative AI Team Achieves Breakthrough Video Generation Progress
By
–
phenomenal Video Generation and Image Editing progress from @imisra_ @deviparikh Yaniv Taigman and team, continuing the momentum in our Generative AI work!
— Soumith Chintala (@soumithchintala) 16 novembre 2023
Our Generative AI team was put together 8 months ago. The pipeline is wild and progress is happening at a breakneck pace… https://t.co/sPCMbpq2Cqphenomenal Generation and Image Editing progress from @imisra_ @deviparikh Yaniv Taigman and team, continuing the momentum in our Generative AI work! Our Generative AI team was put together 8 months ago. The pipeline is wild and progress is happening at a breakneck pace
-
EmuEdit: Fine-grained AI Image Editing Control Tool
By
–
Look at the fine image editing control.
— Soumith Chintala (@soumithchintala) 16 novembre 2023
EmuEdit is pretty cool!https://t.co/cWAMZFslzn pic.twitter.com/0G4jfAcHsRLook at the fine image editing control.
EmuEdit is pretty cool! https://
emu-edit.metademolab.com -
Google DeepMind’s Lyria Achieves High-Quality Vocal Music Generation
By
–
Music generation now has high-quality vocals. This is getting surreal.
— Soumith Chintala (@soumithchintala) 16 novembre 2023
Great work from the Lyria team at @GoogleDeepMind ! https://t.co/7RHmmznahVMusic generation now has high-quality vocals. This is getting surreal.
Great work from the Lyria team at @GoogleDeepMind ! -
RAG and Transformers: Defining AI Paradigms
By
–
not everyone would say yes to RAG being in the "transformers are all you need" bucket, which is why i think it's I'll defined
-
Beyond Transformers: Search and Reasoning in NLP Systems
By
–
i think to add reasoning to NLP machines you'll have to add more than transformers.
a search of some kind.
also…is retrieval augmented generation "just transformers" :p @AravSrinivas might disagree -
Are ConvNets Enough? Architecture Sufficiency in AI Systems
By
–
its ill-defined. It's like asking if "ConvNets are all you need", and someone adds MCTS to the pipeline and gets to AlphaZero. Then, ConvNets are still the base model, and you win the wager. But are ConvNets all you need?
-
Open Source Datasets: Internet Scrapes and Legal Considerations
By
–
First, there aren't really many "open source" datasets.
Most of the datasets are just scrapes from the internet — CommonCrawl, libgen etc.
If we're talking about scrapes from the internet and ignoring other aspects (legal, credit etc.) there are things like ShareGPT that have -
GPT-4V Announcement: Step Function in AI Capabilities
By
–
GPT-4V was announced during the March releases — Greg showed the demo and stuff. So, in that direct sense its not net-new in recent releases.
It is a step function, for sure. -
OpenAI Announcements Disappoint Despite Excitement
By
–
The OpenAI announcements were both super-exciting and somewhat disappointing at the same time. Maybe because people (incl. me) expected the same step-function upgrade from the last family of announcements, and it felt short.