Moscow-based global news network has infected Western AI tools

A Moscow-based disinformation network called "Pravda" is infiltrating artificial intelligence chatbots by flooding the internet with false claims and propaganda, resulting in 3.6 million articles of Russian propaganda being incorporated into Western AI systems in 2024. The network's efforts have been successful, with a NewsGuard audit finding that leading AI chatbots repeated false narratives laundered by the Pravda network 33% of the time, validating concerns about the spread of Kremlin disinformation through AI models.

Read more

US deploys AI to revoke visas of foreign nationals supporting Hamas: Report

The US State Department is launching an AI-powered initiative to revoke visas of foreign nationals who support Hamas or other designated terrorist groups, by scanning their social media accounts for evidence of alleged terrorist sympathies. The "Catch and Revoke" initiative, led by Secretary of State Marco Rubio, will review tens of thousands of student visa holders' social media accounts and internal databases to identify potential terrorist sympathizers.

Read more

Show HN: A big tech dev experience for an open source CMS

The authors, a team of 5 professional creators, are building an open-source content management system (CMS) with a focus on a seamless developer experience, using tools like Isograph, Replit, and Sapling. They're showcasing their early work, including a simple tool that generates an "identity card" from a Twitter username, and inviting others to contribute and collaborate on the project through their Replit app, Discord, and Twitter.

Read more

Google begins testing AI-only search results

Google is testing a new AI-only search results feature, called AI Mode, which uses its Gemini 2.0 model to provide direct answers to search queries without showing traditional web links. The feature, currently available to Google One AI Premium subscribers, is an experimental opt-in feature that aims to provide a more conversational search experience, but may not always be able to offer accurate or helpful responses.

Read more

Superintelligence startup Reflection AI launches with $130M in funding

Reflection AI, a startup founded by former Google DeepMind researchers, has launched with $130 million in funding to develop superintelligence, an artificial intelligence system capable of performing most computer-related work. The company plans to build an autonomous programming tool as an initial step towards this goal, with the potential to automate tasks such as coding, testing, and infrastructure management.

Read more

The Widespread Adoption of Large Language Model-Assisted Writing Across Society

The adoption of large language models (LLMs) for writing has surged since the release of ChatGPT in November 2022, with significant usage across various domains, including consumer complaints, corporate communications, job postings, and international organization press releases. By late 2024, LLM-assisted writing accounted for a substantial portion of text in these domains, ranging from around 10% in job postings to up to 24% in corporate press releases, reflecting a new reality of reliance on generative AI for communications.

ArXiv

RingFormer: Rethinking Recurrent Transformer with Adaptive Level Signals

Transformers have achieved great success in processing sequential data, but their large parameter size leads to high computational demands. The proposed RingFormer approach addresses this issue by using a single Transformer layer that processes input in a circular manner, reducing parameters while maintaining high performance in tasks like translation and image classification.

ArXiv

Think Inside the JSON: Reinforcement Strategy for Strict LLM Schema Adherence

This paper presents a novel approach to enforcing strict schema adherence in large language models by leveraging reinforcement learning and supervised fine-tuning, resulting in a model that demonstrates robust performance in schema consistency. The proposed "ThinkJSON" approach is shown to be effective in real-world applications, outperforming other models while requiring relatively modest training resources, highlighting its practical utility for schema-constrained text generation.

ArXiv

Deep Learning Is Not So Mysterious or Different

Deep neural networks' anomalous generalization behavior, such as benign overfitting and double descent, can be understood and characterized using existing frameworks like PAC-Bayes, and is not unique to neural networks. The key principle behind these phenomena is the use of soft inductive biases, which allow for flexible hypothesis spaces with a preference for simpler solutions, making deep learning less mysterious and distinct from other model classes.

ArXiv

Ladder: Self-improving LLMs through recursive problem decomposition

LADDER is a framework that enables Large Language Models to improve their problem-solving capabilities through self-guided learning, generating and solving simpler variants of complex problems. The framework has been shown to be effective in mathematical integration, significantly improving the accuracy of language models such as Llama and Qwen, and even achieving state-of-the-art results when combined with Test-Time Reinforcement Learning (TTRL).

ArXiv

Letta: Letta is a framework for creating LLM services with memory

Letta is an open-source framework for building stateful LLM applications, allowing users to create stateful agents with advanced reasoning capabilities and transparent long-term memory. The Letta framework is white box and model-agnostic, and can be used with various LLM API backends, with a graphical interface called the Agent Development Environment (ADE) available for creating, deploying, and interacting with Letta agents.

Repo

Show HN: OpenManus – open-source alternative of Manus AI

OpenManus is an open-source project that allows users to create their own agents without needing an invite code, built by a team from MetaGPT in just 3 hours. The project is a simple implementation and welcomes suggestions, contributions, and feedback, with a quick start guide and configuration instructions provided for users to get started.

Repo

Show HN: Ask AI Paul Graham (Open Sourced)

Ask AI Paul Graham is a tool that provides personalized startup advice, simulating the guidance of Paul Graham, and can be accessed through a live demo. The tool was built using Pocket Flow, a 100-line LLM framework, and its design and source code are available, along with instructions on how to run the application.

Repo

Show HN: LLM Globber, a Rust-based command line tool to package code up for LLMs

LLM Globber is a command-line utility written in Rust that collects files from various locations, filters them, and outputs their contents to a single text file, preparing local files for analysis by Language Learning Models. The tool offers various features such as file type filtering, recursive directory processing, binary file handling, and secure file handling, making it a robust and efficient solution for file collection and processing.

Repo

Show HN: I Built an Open Source RAG App to Chat with Any GitHub Repository

This Streamlit application allows users to chat with any public GitHub repository using Retrieval-Augmented Generation (RAG) technology, enabling them to ask questions about the repository's code, structure, and functionality. The app loads the repository, indexes its contents, and generates informative responses based on the repository's code and documentation, maintaining context throughout the conversation.

Repo