Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Smaller LLMs might surpass larger ones due to advantages like quicker training, ability to run locally without specialized hardware, cost-effectiveness, lower latency, ease of deployment, and potential for better interpretability and reproducibility.
Sponsored
Adept has released an extremely simple (no image encoder) and very performant (outperforms 13B models) multimodal model. This article outlines how Adept evaluated the model’s performance and some details about the model.
Instabilities in training at large scales are hard to reproduce for most researchers. This paper shows how to reproduce instabilities in small models and verifies that many commonly reported fixes work just as well on these small models.
Researchers have developed BitNet, a new design for large language models that uses much less energy and memory.
The study introduces Zipformer, a revamped transformer for automatic speech recognition (ASR) that's faster, uses less memory, and outperforms the widely-used Conformer.
MemGPT is a local chatbot that can read and write to memory. This allows you to have a local assistant that remembers important context about previous conversations.
System prompts allow you to change the behavior of ChatGPT-based language models. This repository contains some very well constructed system prompts that could improve usability based on your preferences.
This repository introduces MonoSKD, a groundbreaking method for detecting 3D objects using only one image that effectively bridges the knowledge gap between LiDAR and regular RGB-based models.
Sponsored
The PyTorch team introduced a new lightweight runtime and conversion for models to run on resource-constrained edge devices.
Deep dive into all the components needed to build a modern real time machine learning system.
Weakly-Supervised Scene Graph Generation (WSSGG) using image captions often faces challenges in capturing detailed relationships and utilizing all available information. This study introduces LLM4SGG, which leverages the capabilities of large language models to better extract and align relationships from captions.
Lots of small improvements have added up to substantial speed improvements (claimed 60 steps per second up from 23).
WSJ readers debated AI's role in art creation. Responses varied from skepticism about AI's emotional depth to seeing it as just another tool. Transparency on AI's involvement was a common theme.
Dataset of tabular tokens to encourage the community to build Large Data Models that better understand tabular data. It is the largest publicly available tabular dataset - a compilation of 627 million tables together with 867 billion tokens of contextual information. TabLiB is available on Hugging Face.
The research introduces Ex-MCR, a novel method that efficiently learns unified contrastive representations for multiple modalities without needing paired data. By aligning existing Multi-modal Contrastive Representations, Ex-MCR achieves top performance in tasks like audio-visual retrieval and 3D object classification.
The authors of this paper have created a new technique called SupFusion to make LiDAR and camera systems work together better for detecting things like cars or pedestrians.
Figuring out the direction someone's head is turned is important for lots of tech applications. These researchers have developed a new way to estimate head positions from any angle.
This study introduces a new method that combines digital surface model (DSM) data and aerial images from different times to improve change detection beyond just 2D perspectives.
Libgen is likely a dataset in many closed models. While the legality of this dataset for commercial use is under debate, researchers are still using it to better understand data quality for language model training.
MosaicFusion is like a magic tool that can create new pictures with lots of objects without needing any prior learning. It does this in two steps: first, it makes the picture, and then it creates a mask to show where each object is.
With the dramatic increase in interest in language modeling. Many existing ML conferences have been inundated with papers. This new venue, organized by many prominent scientists, will be a great option for publications.
Moonhub launched its AI-powered recruiter publicly today, alongside an iconic list of 100s of partner customers like Inflection, Anthropic, You.com, Verkada, and more. Its AI helps companies start a search in minutes, hire exceptional talent 3-5x faster, and save 100s of hours per hire.
Flash Attention dramatically speeds up the training of large language models. However, when running inference, the batch size is usually 1. This means Flash Attention provides little benefit. Flash Decoding was developed by the same group and brings optimizations to token generation.
Researchers explore the potential of pre-trained transformer decoders to create universal embeddings for both natural and programming languages. With minimal training on English data, these models excel in various tasks, from classification to code search.