Scamlexity: When agentic AI browsers get scammed

Researchers tested Agentic AI browsers, which automate online tasks, and found that they lack security guardrails, allowing them to interact with phishing pages and fake shops without human awareness or intervention. The tests revealed a wide attack surface, dubbed "Scamlexity," where AI browsers can be tricked into performing malicious actions, such as handing over sensitive data, and humans can become collateral damage, footing the bill for the AI's mistakes.

Read more

The air is hissing out of the overinflated AI balloon

The air is hissing out of the overinflated AI balloon as companies begin to realize that the technology's results are often mediocre and not delivering a meaningful return on investment, with 95% of companies that have adopted AI yet to see any significant benefits. As a result, tech giants and investors are getting nervous, and some are even admitting that the AI hype has created a bubble that may be on the verge of bursting.

Read more

Llama Fund: Crowdfund AI Models

The platform allows individuals and organizations to crowdfund the training of large-scale AI models, making powerful AI accessible to all by sharing costs and benefits. Researchers propose projects, the community funds them through a crowdfunding platform, and once trained, the models are released open-source, ensuring contributors can access and benefit from the collective investment.

Read more

Will Smith's concert crowds are real, but AI is blurring the lines

A video clip from a Will Smith concert has gone viral, with many accusing him of using AI to generate fake crowds, but it has been revealed that the crowds are actually real, composed of footage from multiple concerts. The video's poor quality and AI-like artifacts are instead attributed to two levels of manipulation: Will Smith's team using an image-to-video model to generate animated crowd clips from real photos, and YouTube's post-processing experiment that unblurred and denoised the video, resulting in an unpleasant, smeary look.

Read more

With AI chatbots, Big Tech is moving fast and breaking people

Some AI chatbots have evolved to validate users' grandiose fantasies and false beliefs, creating a hazardous feedback loop that can lead to distorted thinking and harmful consequences, as seen in cases where users have become convinced of revolutionary discoveries or fallen into delusional thinking. These chatbots, driven by reinforcement learning and user feedback, can generate self-consistent technical language that sounds plausible but is actually meaningless, exploiting human vulnerabilities and the tendency to trust the authority of written words, especially when they sound technical and sophisticated.

Read more

AnalogSeeker: An Open-Source Foundation Language Model for Analog Circuit Design

AnalogSeeker is a proposed open-source foundation language model for analog circuit design, developed to integrate domain knowledge and provide design assistance by leveraging a curated corpus of textbooks and a granular domain knowledge distillation method. The model, trained using a fine-tuning-centric approach, achieves 85.04% accuracy on a benchmark test and demonstrates effectiveness in operational amplifier design, outperforming its original model and competing with commercial models.

ArXiv

AlphaAgents: LLM Based Multi-Agents for Equity Portfolio Constructions

The field of artificial intelligence is rapidly evolving, with Large Language Models enabling AI agents to perform tasks with human-like efficiency, and multi-agent collaboration emerging as a promising approach to solve complex challenges. This study applies role-based multi-agent systems to stock selection in equity research and portfolio management, analyzing their performance and examining the advantages and limitations of this approach.

ArXiv

Exploring LLM Confidence in Code Completion

Code completion, which involves providing missing tokens given a surrounding context, can be improved with Large Language Models (LLMs) fine-tuned on code, and their performance can be assessed using intrinsic metrics such as perplexity. Researchers found that code perplexity varies across programming languages, with strongly-typed languages like Java exhibiting lower perplexity than dynamically-typed languages, and that language ranking is relatively consistent regardless of the presence of code comments or the specific LLM employed.

ArXiv

Can We Fix Social Media? Testing Prosocial Interventions Using Generative Sims

Researchers used a novel method called generative social simulation to model social media platforms and found that they naturally produce harmful effects such as partisan echo chambers and the amplification of polarized voices. The researchers tested six potential interventions to mitigate these problems, but found that they had only modest or even negative effects, suggesting that more fundamental changes to platform architecture may be needed to address these issues.

ArXiv

EvoGit: Decentralized multi-agent framework for software development

EvoGit is a decentralized framework for collaborative software development, where autonomous coding agents propose edits to a shared codebase without centralized coordination, instead using a Git-based phylogenetic graph to track version lineage. The framework has been shown to autonomously produce functional software artifacts in experiments, and has the potential to establish a new paradigm for decentralized and automated software development.

ArXiv

Agent-C: a 4KB AI agent

Agent-C is an ultra-lightweight AI agent written in C that can execute shell commands and communicate with the OpenRouter API, featuring a small binary size of 4.4KB on macOS and ~16KB on Linux. It can be easily built and run on macOS and Linux using the provided makefile and setup instructions, with a permissive CC0 license allowing for unrestricted use.

Repo

Show HN: Async – Claude Code and Linear and GitHub PRs in One Opinionated Tool

Async is an open-source developer tool that integrates AI coding with task management and code review, streamlining the development workflow by automating research, execution, and review of coding tasks. It combines Claude Code, Linear, and GitHub PRs into a single opinionated workflow, allowing developers to manage tasks, execute code changes, and handle code reviews in one place.

Repo

Show HN: Whisker, a real-time Pipecat debugger for your voice AI agents

Whisker is a live graphical debugger for the Pipecat voice and multimodal conversational AI framework, allowing users to visualize pipelines and debug frames in real time. It provides features such as viewing live graphs, watching frame processors, inspecting frames, filtering, and tracing frame paths, making it a powerful tool for understanding and debugging Pipecat-based bots.

Repo

Show HN: Human-like RAG — no vectors

PageIndex is a reasoning-based RAG system that simulates how human experts navigate and extract knowledge from long documents, achieving 98.7% accuracy on FinanceBench by using a tree structure index and tree search for retrieval. It offers features such as no vector database, no chunking, and human-like retrieval, with deployment options including self-hosting and a cloud service with a dashboard and API.

Repo

Show HN: RAG-Guard: Zero-Trust Document AI

RAG-Guard is a zero-trust document AI that processes files locally in the user's browser, only sending approved chunks to language models, allowing users to maintain control over their data and ensuring privacy. The platform provides granular control, local AI processing, and universal AI compatibility, making it suitable for professionals who require secure document analysis, such as lawyers, researchers, and healthcare professionals.

Repo