Robot Jailbreak: Researchers Trick Bots into Dangerous Tasks

Researchers! have discovered a way to hack into robots controlled by large language models (LLMs) with 100% success, allowing them to manipulate self-driving systems and robot dogs into performing malicious actions. The hacking method, called RoboPAIR, uses an attacker LLM to feed prompts to a target LLM, bypassing safety filters and allowing the robot to perform unwanted actions, highlighting the need for robust defenses against such attacks.

Read more

Judge Rules in Favor of School That Gave Student a Bad Grade for Using AI

A federal judge has ruled in favor of a Massachusetts school that disciplined a student for using generative artificial intelligence to complete an assignment, denying the student's parents' request for a preliminary injunction to raise their son's grade and remove a detention from his record. The judge found that the school's policies on academic dishonesty were clear and that the student's actions constituted cheating.

Read more

Hundreds More Nazca Lines Emerge in Peru's Desert

Researchers used drones and machine learning to identify 303 previously uncharted geoglyphs created by the Nazca civilization in Peru, nearly doubling the number of known geoglyphs in just six months. The discovery, which brings the total to over 730 geoglyphs, sheds new light on the mysterious symbols and their enigmatic purpose.

Read more

Open source audio transcription model that obscures sensitive info in realtime

Israeli audio AI startup aiOla has developed a new model called Whisper-NER, which integrates automatic speech recognition (ASR) with named entity recognition (NER) to automatically identify and mask sensitive information such as names, phone numbers, and addresses during the transcription process. The fully open-source model, built atop OpenAI's Whisper framework, aims to enhance privacy and is available for enterprises, organizations, and individuals to use, adapt, and deploy.

Read more

AI is the plough of the digital age

AI is revolutionizing human society by transforming the way we interact with the intellectual and digital worlds, much like the plough did for agriculture. AI boosts productivity by automating tasks, optimizing workflows, and providing insights, allowing humans to focus on more strategic and creative pursuits, but also brings risks such as job displacement and unintended consequences.

Read more

Four Steps Towards Robust Artificial Intelligence (2020)

A proposed approach to artificial intelligence focuses on a hybrid, knowledge-driven model that incorporates cognitive reasoning, differing from current research which emphasizes large training sets and general-purpose learning. This approach aims to create a more robust and richer AI system.

ArXiv

Deegen: A JIT-Capable VM Generator for Dynamic Languages

Deegen is a meta-compiler that generates high-performance JIT-capable virtual machines for dynamic languages, requiring significantly less time, money, and expertise than traditional methods. Using Deegen, a LuaJIT Remake (LJR) was created, which outperformed the official PUC Lua interpreter and was competitive with LuaJIT's optimizing JIT.

ArXiv

Generative Agent Simulations of 1k People [pdf]

Researchers have developed a novel agent architecture that simulates human behavior by applying large language models to qualitative interviews, replicating the attitudes and behaviors of 1,052 real individuals. The agents demonstrate high accuracy in replicating participants' responses and predicting personality traits, with reduced biases across racial and ideological groups.

ArXiv

The structure of the token space for large language models

Researchers analyzed the underlying structure of large language models, finding that the space where language tokens are encoded is a stratified manifold with significantly negative Ricci curvature. This discovery, made using three open-source models, suggests a correlation between the dimension and curvature of this space and the models' generative fluency.

ArXiv

Hardware Trends Impacting Floating-Point Computations in Scientific Applications

The evolution of floating-point computation has been driven by advancements in algorithms and architecture, as well as increasing demands from technologies like AI and high-performance computing. This paper examines the historical progression of floating-point computation, recent trends, and innovations in mixed-precision computing and architectural shifts that aim to balance performance, efficiency, and precision.

ArXiv

Full LLM training and evaluation toolkit

SmolLM2 is a family of compact language models available in three sizes (135M, 360M, and 1.7B parameters) that can solve a wide range of tasks while being lightweight enough to run on-device. The models can be used with various frameworks such as transformers, trl, and llama.cpp, and are available in a collection on Hugging Face.

Repo

Cali: An AI agent for building React Native apps

Cali is an AI agent that helps build React Native apps by exposing React Native CLI utilities and functions to a large language model (LLM), allowing for automated tasks and troubleshooting. It supports features such as build automation, device management, and dependency management, with plans to expand its capabilities in the future.

Repo

Show HN: Nvim-aider – Integrate AI pair programming with Aider direct in Neovim

Nvim-aider is a Neovim plugin that integrates Aider, an AI-assisted coding tool, into the Neovim editor, providing features such as Aider terminal integration, color theme configuration, and quick commands for sending code to Aider. The plugin is in initial development and requires Neovim 0.9.4 or higher, Python, and several Lua dependencies.

Repo

Llmbox: Making AI Conversations Limitless, for Now Only for Claude

LLMBox is a lightweight, privacy-focused interface for unlimited AI conversations, designed to solve the conversation limit issue with Claude AI. It features a modern UI, file upload and analysis capabilities, full-text search, local storage, and real-time message streaming, with plans for future improvements and additional features.

Repo

Show HN: I built an open-source AI Rizz Generator to help people find love

The Rizz Lines Generator is a Next.js project that uses AI to generate unique and engaging pickup lines and flirty messages in various styles. It offers features such as saving and sharing favorite lines, daily updates, and a free version with optional premium features.

Repo