Jul 12, 2024

• FlashAttention-3 introduces techniques exploiting asynchrony and low-precision (FP8) on Hopper GPUs, achieving 1.5-2.0x speed improvements and 75% utilization of H100 GPU capabilities, significantly enhancing the efficiency of attention mechanisms in large language models (LLMs).

• The Physics-based Deep Learning Book offers a comprehensive guide to integrating deep learning with physical simulations, featuring hands-on Jupyter notebook examples for immediate application.

• Korvus integrates the entire RAG pipeline into a single Postgres database query, offering a unified search SDK with Python, JavaScript, and Rust bindings for high-performance, customizable search functionalities.

• AuraFlow v0.1 is introduced as a large, open-source, flow-based generation model capable of text-to-image generation, challenging the notion that open-source AI development has stalled.

• Free-threaded CPython, an experimental feature in CPython 3.13, enables multiple threads to run in parallel without the Global Interpreter Lock (GIL), aiming to improve multi-threaded performance by utilizing multiple CPU cores more effectively, as detailed in PEP 703.

• mandala streamlines ML experiment tracking by capturing inputs, outputs, and code dependencies with the @op decorator, ensuring no computation is repeated unnecessarily.

• StreamVC is a real-time voice conversion technology that maintains the original speech's content and prosody while adopting the voice timbre of any target speech, designed for low-latency performance on mobile platforms.

• EvolutionaryScale, founded by former Meta scientists, has unveiled a protein language model, ESM3, which has successfully engineered new fluorescent proteins, marking a significant advancement in AI-driven biological design.

• WildGaussians introduces a novel approach to 3D Gaussian Splatting (3DGS), enhancing its capability to handle occlusions and appearance changes in dynamic, in-the-wild scenes.

• Cradle introduces a General Computer Control (GCC) setting, using screenshots for input and keyboard/mouse actions for output, to enable foundation agents to interact with any software, overcoming the challenge of environment encapsulation differences.

• Memory^3 introduces a novel approach to language modeling by equipping large language models (LLMs) with explicit memory, significantly reducing parameter size, training, and inference costs while outperforming larger models and text retrieval-augmented generation (RAG) models. Read the paper

• The dendristor model integrates synaptic organization with dendritic tree-like morphology, utilizing multigate silicon nanowire transistors for neuromorphic computation, emulating visual motion perception akin to that in the retina.

• Self-supervised learning with DINOv2 ViT weights from Facebook Research enables rich image segmentation by finetuning with Low-Rank Adaptation (LoRA) and simple decoders, achieving solid validation IoU scores.

• FlashAttention-3 introduces techniques on Hopper GPUs for 1.5-2.0x speed over FlashAttention-2, achieving up to 740 TFLOPS with FP16 and nearly 1.2 PFLOPS with FP8, reducing error by 2.6x in low-precision contexts. Paper

• Google Research introduces a method to revolutionize online handwriting recognition by leveraging large vision-language models (VLMs), which utilize new representations and tokenizers for improved accuracy.