Quantamagazine - The dark dimension scenario offers a recipe for dark matter and suggests an intimate connection between dark matter and dark energy.

Parth Sarthi, Salman Abdullah, Aditi Tuli, Shubh Khanna, Anna Goldie and Christopher D. Manning introduce a novel approach to retrieval-augmented language models. The RAPTOR model recursively embeds, clusters, and summarizes chunks of text, constructing a tree with differing levels of summarization.

Retrieval Augmented Generation (RAG) is a solution to mitigate issues such as domain knowledge gaps, factuality issues, and hallucination in large language models (LLMs) by augmenting them with external knowledge. Prompt Engineering Guide run you through the existing RAG approaches, state-of-the-art RAG, evaluation, applicati...

Peter Yong Zhong, Haoze He, Omar Khattab, Christopher Potts, Matei Zaharia and Heather Miller map the landscape of frameworks for abstracting interactions with and between large language models. and suggest two systems of organization for reasoning about the various approaches to, and philosophies of, LLM abstraction.

Alan Ansell, Ivan Vulić, Hannah Sterz, Anna Korhonen and Edoardo M. Ponti introduce SpIEL, a novel sparse fine-tuning method for large language models like LLaMA 2. The study finds that SpIEL is often superior to popular parameter-efficient fine-tuning methods in terms of performance and comparable in terms of run time.

Simon Willison explores ColBERT, a fast and accurate retrieval model for large text collections. He uses RAGatouille, a library that simplifies working with ColBERT, to create an index of his blog's content.

Jordi Grau-Moya, Tim Genewein, Marcus Hutter, Laurent Orseau, Grégoire Delétang, Elliot Catt, Anian Ruoss, Li Kevin Wenliang, Christopher Mattern, Matthew Aitchison and Joel Veness explore the potential of amortizing Solomonoff Induction into neural networks via meta-learning. They use Universal Turing Machines to generate tr...

French-Public Domain-Newspapers or French-PD-Newpapers is a large collection aiming to agregate all the French newspapers and periodicals in the public domain.

The Lancet - Jun Lv, Hui Wang, Xiaoting Cheng, Yuxin Chen, Daqi Wang, Longlong Zhang, Qi Cao, Honghai Tang, Shaowei Hu, Kaiyu Gao, Mengzhao Xun, Jinghan Wang, Zijing Wang, Biyun Zhu, Chong Cui, Ziwen Gao, Luo Guo, Sha Yu, Luoying Jiang, Yanbo Yin, Jiajia Zhang, Bing Chen, Wuqing Wang, Renjie Chai, Zheng-Yi Chen, Huawei Li an...

Conversation-based search engine by Lipton AI using the Bing or Google search engine APIs. It is relevant for developers who want to build and customize their own conversational search engines.

Phil Schmid provides a step-by-step guide on how to fine-tune large language models (LLMs) using Hugging Face in 2024. It covers the use case, development environment setup, dataset creation, and model training.

This tutorial by Jason Liu explains how to embed the entire English Wikipedia in under 15 minutes using Hugging Face’s Text Embedding Inference service on Modal. The process involves setting up a serverless solution that can handle large-scale workloads and generate embeddings for a massive text dataset at lightning speed.

Killian Lucas proposes a simple, fast, local semantic search over folders.

Nature - Bethanie Maples, Merve Cerit, Aditya Vishwanath, and Roy Pea investigate how students use the Intelligent Social Agent Replika, finding that participants used Replika in multiple ways as a friend, therapist, and intellectual mirror. The authors find that 3% of participants reported that Replika halted their suicidal ...

Karel D'Oosterlinck, Omar Khattab, François Remy, Thomas Demeester, Chris Develder and Christopher Potts propose a general program called 'Infer--Retrieve--Rank' that improves multi-label classification with thousands of classes using in-context learning and retrievers. Implemented using the DSPy programming model, their prim...

John Keiser and Daniel Lemire designed and implemented a novel JSON parsing interface called On-Demand that provides superior performance in multiple benchmarks

The Directorate-General for Digital Services published the Artificial Intelligence in the European Commission Communication, outlining a strategic vision to foster internal development and use of lawful, safe, and trustworthy AI systems, preparing for the implementation of the EU AI Act.

Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Xian Li, Sainbayar Sukhbaatar, Jing Xu and Jason Weston explore a novel approach to training language models. They propose using the language model itself as a judge via LLM-as-a-Judge prompting to provide its own rewards during training.

This article from the Web Security Academy by PortSwigger discusses web Large Language Model (LLM) attacks, which exploit the model's access to data, APIs, or user information that an attacker cannot access directly. The article covers various aspects of these attacks, including prompt injection, detecting LLM vulnerabilities...

In the Trail of Bits blog, Tyler Sorensen and Heidy Khlaaf disclose LeftoverLocals, a vulnerability that allows recovery of data from GPU local memory created by another process on Apple, Qualcomm, AMD, and Imagination GPUs. This vulnerability impacts the security posture of GPU applications, particularly LLMs and ML models r...

Michael J. Fox and Deborah W. Brooks discuss the discovery of a groundbreaking biomarker for Parkinson's disease (PD) is discussed. The 'alpha-synuclein seeding amplification assay' (SAA) enables early detection and identification of PD with 90% specificity.

In his blog "Embrace the Red", Johann Rehberger discusses an interesting discovery of prompt injection in LLMs (large language models) through invisible instructions in pasted text. This technique uses the Tags Unicode Block to create hidden instructions that can cause an LLM, such as ChatGPT, to take specific actions.

Vipul Vaibhaw shares his story in which he decided to implement efficient Matrix Multiplication

The Text Clustering repository by Huggingface contains tools to easily embed, cluster and semantically label text datasets. The pipeline consists of several distinct blocks that can run in a few minutes on a consumer laptop.

1 A 3 O R N discusses the attempts by various AI safety organizations to ban or limit currently existing open-source AI models, contrary to popular belief. They provide specific examples of such organizations and their proposed policies.