Liquid Foundation Models: Our First Series of Generative AI Models

Liquid Foundation Models (LFMs) represent a breakthrough in generative AI, achieving state-of-the-art performance across 1B, 3B, and 40B parameter scales while optimizing memory usage and inference efficiency.

VisionTS: Zero-Shot Time Series Forecasting with Visual Masked Autoencoders

VisionTS transforms the forecasting task into an image reconstruction challenge, leveraging visual masked autoencoders to achieve impressive results in zero-shot scenarios.

What’s the SOTA model for style transfer as of 2024?

The state-of-the-art (SOTA) model for image style transfer in 2024 has shifted towards diffusion methods, which significantly enhance the ability to capture higher-level concepts compared to traditional Gram matrix-based techniques.

Optimizing transformers

Optimizing transformer models in multi-view images and cross-attention networks can significantly enhance training speed by addressing the parameter overload caused by cross-attention layers.

PyTorch Native Architecture Optimization: Torchao

torchao is a new PyTorch library that optimizes model performance by utilizing low bit dtypes, quantization, and sparsity, achieving up to 97% speedup for Llama 3 inference with minimal accuracy loss.

Experimenting with Llama-3 codebase and Google NotebookLM – Mind-Blowing Results!

Experimenting with the Llama-3 codebase and Google NotebookLM yielded results that surpassed expectations, showcasing the potential of combining advanced architectures with innovative tools like Rag and SERP APIs for enhanced multimedia generation.

Stress-testing biomedical vision models with RadEdit: A synthetic data approach for robust model deployment

RadEdit enhances biomedical vision models by using generative image editing to simulate diverse dataset shifts, allowing researchers to identify model weaknesses before deployment in clinical settings.

Model-based Preference Optimization in Abstractive Summarization without Human Feedback

Model-based Preference Optimization (MPO) enhances abstractive summarization by utilizing the model's own capabilities to generate a preference dataset, eliminating the need for costly human feedback.

Leveraging Long-Context Large Language Models for Multi-Document Understanding and Summarization in Enterprise Applications

Long-context Large Language Models (LLMs) excel in multi-document summarization, effectively capturing extensive connections and providing cohesive summaries across various industries, including legal, HR, finance, and medical sectors.