elvis
Scalable Diffusion Models with Transformers

Introduces Diffusion Transformers... replaces convolutional U-Net backbone with a Transformer giving way to a new class of diffusion models with "good scaling properties" and SoTA performance.

https://t.co/IBMvm2Vhgt https://t.co/T2cMmM6G6V
tweet

Dave
[spends 11 hours debugging a python file] my life a movie fr
tweet

AK
BEATs: Audio Pre-Training with Acoustic Tokenizers
abs: https://t.co/OqDRTEzTOT https://t.co/I2zP9WaxXN
tweet

Yann LeCun
RT @schrep: Unpopular opinion. Self-driving will progress rapidly in the next few years.

https://t.co/9FvsG3OSxy
tweet

Grades 🗨
The only limit to your success with niche websites is how much time you put into them.

Action ➡️ Results.
tweet

AK
MM-Diffusion: Learning Multi-Modal Diffusion Models for Joint Audio and Video Generation
abs: https://t.co/MtSeqOUmuI https://t.co/ibEjXfAtz6
tweet

Easlo
7 useful apps for entrepreneurs in 2023:
tweet

AK
Point·E: A System for Generating 3D Point Clouds from Complex Prompts
abs: https://t.co/heHZOKVVMD
github: https://t.co/cE1W9nFjlj https://t.co/s18A5OorTM
tweet

elvis
Quantization helps to build more efficient models. This paper shows that 4-bit precision "is almost universally optimal for total model bits and 0-shot accuracy." The chart shows bit-level scaling laws for performance across different OPT model sizes.

https://t.co/jmY6Au1BD3 https://t.co/H6ScxJLuKz
tweet

AK
Evaluating Human-Language Model Interaction
abs: https://t.co/ZmRXsyX01e https://t.co/WtEbS9eg9F
tweet