This one is impressive on a few levels — Bing Search gives a mostly correct answer explaining the "SolidGoldMagikarp" glitch, a recently discovered issue in which prior GPT models (incl. ChatGPT) cannot repeat back certain short strings:
MACHINE LEARNING
-
ML Pipeline Generation with Modular Customization Options
By
–
That's a good question. Once you generate the ML pipeline It also provides you an option to focus on each specific part and modify it which makes it really handy.
-
Testing repetition of anomalous tokens in LLMs
By
–
How does it do with repetition of other anomalous tokens?
-
DeepMind Speculative Sampling Boosts LLM Decoding Speed 2-2.5x
By
–
DeepMind’s Speculative Sampling Achieves 2–2.5x Decoding Speedups in Large Language Models
-
Representation Beyond Cartesian Dualism: Dynamic World Registration
By
–
2/2 representation is not a simple map from the world to some symbolic domain, or ontology, rather objects in the ontology are registered in the world through an ongoing process. Says that "modern" systems such as Google's Knowledge Map are leftover cartesian misconceptions.
-
A Poem on Gradient Descent: The Path of Machine Learning
By
–
Vamos con un poema sobre el descenso del gradiente.
-
Noise2Music: Transforming Random Noise into Music with AI
By
–
—
Link to project: https://
google-research.github.io/noise2music/ Link to paper: https://
google-research.github.io/noise2music/no
ise2music.pdf
… Authors: Qingqing Huang, Daniel S. Park, Tao Wang, Timo I. Denk, Andy Ly, Nanxin Chen, Zhengdong Zhang, Zhishuai Zhang, Jiahui Yu, Christian Frank, Jesse Engel, Quoc V. Le, William Chan, Wei Han
— -
Zero-shot classification matches text to music without training
By
–
The model knows how to connect text and music to match each of these sentences with a specific music clip, without having to specifically train the model for each clip.
— AI Breakfast (@AiBreakfast) 9 février 2023
This process is called zero-shot classification. pic.twitter.com/KWerrsnMWjThe model knows how to connect text and music to match each of these sentences with a specific music clip, without having to specifically train the model for each clip. This process is called zero-shot classification.
-
Noice2Music generates 30-second music from text inputs
By
–
"Noice2Music" is a project from Google Research that generates a short 30 second piece of music based on text inputs.
-
Training generator and cascader models on 300k+ hours of audio
By
–
The 300k+ hours of audio clips were used to train a "generator model" that turns the text into an intermediate representation, and a "cascader model" that uses this intermediate representation to produce high-quality audio.