Luciano Del Corro, Allie Del Giorno, Sahaj Agarwal, Bin Yu, Ahmed Awadallah and Subhabrata Mukherjee propose a simple and effective token-level early exit method, SkipDecode, designed to work seamlessly with batch inferencing and KV caching. It overcomes prior constraints by setting up a singular exit point for every token in...

Maanak Gupta, CharanKumar Akiri, Kshitiz Aryal, Eli Parker and Lopamudra Praharaj highlight the limitations, challenges, potential risks, and opportunities of GenAI in the domain of cybersecurity and privacy.

Nature - Cheng Kai Lim, Jing Wui Yeoh, Aurelius Andrew Kunartama, Wen Shan Yew and Chueh Loo Poh detail a method of capturing 2-dimensional light patterns into DNA, by utilizing optogenetic circuits to record light exposure into DNA, encoding spatial locations with barcoding, and retrieving stored images via high-throughput n...

Jared Palmer (Vercel) gives high-level insights into a toolbox to protect a platform against scrapping

Joint report from Georgetown University’s Center for Security and Emerging Technology (CSET) and The Alan Turing Institute’s Centre for Emerging Technology and Security (CETaS) assesses the current state-of-the-art in autonomous cyber defence and its future potential, identifies barriers to progress and recommends specific ac...

Shen, Zejiang, Ruochen Zhang, Melissa Dell, Benjamin Lee, Jacob Carlson, and Weining Li introduce LayoutParser, an open-source library for streamlining the usage of DL in document image analysis research and applications.

Guillaume Sanchez, Honglu Fan, Alexander Spangher, Elad Levi, Pawan Sasanka Ammanamanchi and Stella Biderman demonstrate that Classifier-Free Guidance can be used broadly as an inference-time technique in pure language modeling.

Fantastic project from the Princeton University, releasing a whole-brain connectome of the fruit fly, including c.130k annotated neurons and tens of millions of typed synapses!

By Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Yu, Joey Gonzalez, Hao Zhang and Ion Stoica introduce vLLM, an open-source library for fast LLM inference and serving. vLLM utilizes PagedAttention, a new attention algorithm that effectively manages attention keys and values.

Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger and Hongyang Li provide a comprehensive analysis of more than 250 papers, covering the motivation, roadmap, methodology, challenges, and future trends in end-to-end autonomous driving.

Positive Security showcase an attack which leverages indirect prompt injection to trick Auto-GPT (LLM-based agent) into executing arbitrary code when it is asked to perform a seemingly harmless task such as text summarization on an attacker controlled website

Isaac Kauvar, Chris Doyle, Linqi Zhou and Nick Haber present Curious Replay - a form of prioritized experience replay tailored to model-based agents through use of a curiosity-based priority signal.

Manli Shu, Jiongxiao Wang, Chen Zhu, Jonas Geiping, Chaowei Xiao and Tom Goldstein investigate the exploitability of instruction tuning in large language models. They propose AutoPoison, an automated data poisoning pipeline, to achieve specific attack goals through intentionally changing the model's behavior. The study showca...

In the fourth part of "The Grid", Brian Potter discusses the challenges faced by the electrical grid and the potential solutions. Increasing use of variable sources of electricity, decreasing grid reliability, increasing delay in building electrical infrastructure, and increasing demand for electricity are all addressed.

doubleagent.net look under the hood of a Bluetooth car battery monitor siphoning up location data

BOOK | Instead of trying to be exhaustive, Francois Fleuret's little book is limited to the background necessary to understand a few important models. And it works!

Earthweb propose a comprehensive overview of the OnlyFans userbase. And it is... surprising

VIDEO | How much of this math do you know?

Lilian Weng gives a comprehensive overview of Agent Systems, their opportunities and their challenges

Neel Jain, Khalid Saifullah, Yuxin Wen, John Kirchenbauer, Manli Shu, Aniruddha Saha, Micah Goldblum, Jonas Geiping and Tom Goldstein propose a framework for self-supervised evaluation of LLMs by analyzing their sensitivity or invariance to transformations on the input text.

Microsoft Research : Suriya Gunasekar, Yi Zhang, Jyoti Aneja, Caio César Teodoro Mendes, Allie Del Giorno, Sivakanth Gopi, Mojan Javaheripi, Piero Kauffmann, Gustavo de Rosa, Olli Saarikivi, Adil Salim, Shital Shah, Harkirat Singh Behl, Xin Wang, Sébastien Bubeck, Ronen Eldan, Adam Tauman Kalai, Yin Tat Lee and Yuanzhi Li int...

Maxime Labonne consolidates his articles and notebooks in a single GitHub repo to form an end-to-end LLM course

Elinas (Zeus Labs) shares a helpful blogpost on QLoRA training, detailing important steps in the process

Kai Lv, Yuqing Yang, Tengxiao Liu, Qinghui Gao, Qipeng Guo and Xipeng Qiu propose a new optimizer, LOw-Memory Optimization (LOMO), which fuses the gradient computation and the parameter update in one step to reduce memory usage.

Evaluation of human performance is notoriously difficult, even for an expert eye. The difficulty primarily lies in consistency of evaluations across subjects and over time. And, for all satisfactory solutions, a key challenge is then to communicate the output without losing valuable information.