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.

This report by Apostol Vassilev, Alina Oprea, Alie Fordyce, and Hyrum Anderson provides a taxonomy and terminology for adversarial machine learning (AML). The taxonomy includes key types of ML methods, stages of attack, attacker goals and capabilities, and attacker knowledge. AML attacks are classified as evasion, poisoning, ...

Tao Tu, Anil Palepu, Mike Schaekermann, Khaled Saab, Jan Freyberg, Ryutaro Tanno, Amy Wang, Brenna Li, Mohamed Amin, Nenad Tomasev, Shekoofeh Azizi, Karan Singhal, Yong Cheng, Le Hou, Albert Webson, Kavita Kulkarni, S Sara Mahdavi, Christopher Semturs, Juraj Gottweis, Joelle Barral, Katherine Chou, Greg S Corrado, Yossi Matia...

Evan Hubinger, Carson Denison, Jesse Mu, Mike Lambert, Meg Tong, Monte MacDiarmid, Tamera Lanham, Daniel M. Ziegler, Tim Maxwell, Newton Cheng, Adam Jermyn, Amanda Askell, Ansh Radhakrishnan, Cem Anil, David Duvenaud, Deep Ganguli, Fazl Barez, Jack Clark, Kamal Ndousse, Kshitij Sachan, Michael Sellitto, Mrinank Sharma, Nova D...

Quantamagazine - Michael Greshko reports on the discovery of a new kind of magnetism in an engineered material, which was predicted by Yosuke Nagaoka in 1966. The study, published in Nature, marks the latest advance in the hunt for Nagaoka ferromagnetism and was conducted by researchers at the Swiss Federal Institute of Techn...

Vik Paruchuri presents Surya,ba document OCR toolkit that does accurate OCR in 90+ languages and line-level text detection in any language. It supports a range of documents and includes a streamlit app for interactive use.

Big Think - Ethan Siegel explains that Astronomers have discovered a large, galaxy-scale population of gas with no stars inside it. This 'dark galaxy' may hold the key to understanding how the first galaxies formed in the universe.

Quantamagazine - Biologists discovered that mitochondria in different tissues communicate to repair injured cells. When their signal fails, the biological clock starts winding down. Chemical signals released by mitochondria are somehow communicated to mitochondria in other tissues, with consequences for how rapidly organisms ...

Dongsheng Wang, Zhiqiang Ma, Armineh Nourbakhsh, Kang Guand Sameena Shah introduce a novel framework that combines pre-trained language models with graph semantics. The proposed architecture, DocGraphLM, predicts both directions and distances between nodes using a convergent joint loss function. Experiments on three SotA dat...

Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P. Srinivasan, Dor Verbin, Jonathan T. Barron, Ben Poole and Aleksander Holynski present ReconFusion, a novel 3D reconstruction method using diffusion priors. This approach enables high-quality 3D reconstruction from a few images by lev...

Hongye Jin, Xiaotian Han, Jingfeng Yang, Zhimeng Jiang, Zirui Liu, Chia-Yuan Chang, Huiyuan Chen and Xia Hu propose SelfExtend, a method to extend the context window of LLMs without fine-tuning. SelfExtend can effectively extend existing LLMs' context window length as shown by comprehensive experiments on multiple benchmarks.

Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji and Quanquan Gu propose Self-Play fIne-tuNing (SPIN), a method that fine-tunes weak Language Models (LLMs) into strong LLMs without requiring additional human-annotated data. This approach significantly improves the LLM's performance across various benchmarks, outperforming mo...

Tiago P. Peixoto presents a general algorithm for scalable network reconstruction that achieves subquadratic time complexity, significantly improving upon traditional quadratic methods. The algorithm relies on stochastic second neighbor search and allows for easy parallelization, enabling the reconstruction of large networks ...

Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, Ming-Ming Cheng and Ying Shan present PhotoMaker, an efficient personalized text-to-image generation method that encodes an arbitrary number of input ID images into a stacked ID embedding for preserving ID information, allowing for high-quality, identity-preserving human photo ...

In this course, Jason Liu, author of the Instructor library, teaches how to improve LLM engineering skills. Learn about structured JSON output handling, function calling, complex validations with Pydantic and more.

Siwei Zhang, Bharat Lal Bhatnagar, Yuanlu Xu, Alexander Winkler, Petr Kadlecek, Siyu Tang and Federica Bogo propose RoHM, an approach for robust 3D human motion reconstruction from monocular RGB(-D) videos in the presence of noise and occlusions.

VIDEO | Fascinating video explaining how a teenager "broke" Tetris, but also how the community found out that it was even possible and how this became a goal in itself

Derong Xu, Wei Chen, Wenjun Peng, Chao Zhang, Tong Xu, Xiangyu Zhao, Xian Wu, Yefeng Zheng and Enhong Chen explore the recent advancements in information extraction using generative Large Language Models (LLMs). They categorize the works based on various subtasks and learning paradigms and empirically analyze the most advance...

VIDEO | Johann Rehberger discusses real-world exploits in Large Language Model applications, including ChatGPT, Bing Chat, and Google Bard. He highlights three categories of threats and focuses on indirect prompt injections, demonstrating with examples and a Bing Chat demo. The talk also covers injection TTPs and data exfiltr...

Jan Van Haaren shares his annual review of soccer analytics content, featuring research papers, blog posts, news articles, podcasts, and code repositories that provide new insights, address challenges, propose methods or apply existing methods in a creative way.

Boris Larin presents Operation Triangulation, a sophisticated 0-click iMessage attack using four zero-days, targeting iOS versions up to 16.2. The attack chain involves exploiting a remote code execution vulnerability in the ADJUST TrueType font instruction and bypassing hardware memory protection to gain full control over th...

Penghao Wu and Saining Xie introduce V*, an LLM-guided visual search mechanism that enhances collaborative reasoning, contextual understanding, and precise targeting of specific visual elements in multimodal large language models. When combined with an MLLM, this mechanism forms a new MLLM meta-architecture called SEAL.