The Impact of Generative AI on Critical Thinking [pdf]

A survey of 319 knowledge workers found that the use of Generative AI (GenAI) tools is associated with reduced cognitive effort and critical thinking, with higher confidence in GenAI leading to less critical thinking and higher self-confidence leading to more critical thinking. The study suggests that GenAI shifts the nature of critical thinking towards information verification, response integration, and task stewardship, and highlights the need for designing GenAI tools that support critical thinking in knowledge work.

Read more

The role of developer skills in agentic coding

The author is exploring the impact of Large Language Models (LLMs) on software delivery practices and has developed a mental model to categorize tools that use LLMs to support coding, considering dimensions such as assisted tasks, interaction modes, prompt composition, and model properties. The current state of LLM-based coding tools includes in-line assistance in code editors, chat interfaces, and prompt composition applications, with experimentation ongoing in the open-source world and a shift towards more control over data and models.

Read more

AI will change the world but not in the way you think

Large Language Models (LLMs) are changing the way people communicate, particularly in business and education, by allowing users to generate human-like text and extract key points from lengthy messages. As LLMs become more prevalent, they are streamlining communication by cutting through "business speak" and getting straight to the point, potentially making traditional forms of communication, such as lengthy emails, a thing of the past.

Read more

Kilo Code: Speedrunning open source coding AI

The author, who led a project that successfully resurrected an ancient library from volcanic ashes, has applied the lesson of achieving incredible things through an extremely fast-moving community to a new venture, Kilo Code, which aims to create a user-friendly AI coding agent. With a team assembled in just a week, Kilo Code is rapidly developing and shipping improvements, and invites the community to provide feedback and contribute to the project, with the goal of making AI coding accessible to billions of people.

Read more

'Brainrot' AI on Instagram Monetizing the Most Fucked Up Things You Can Imagine

A type of AI-generated content on Instagram, referred to as "Brainrot," is monetizing disturbing and often horrific videos and images, including those that are racist, violent, and sexually explicit. These AI-generated reels, which feature popular characters and celebrities in bizarre and unsettling scenarios, are designed to shock and provoke users into engaging with them, allowing them to spread and become more prominent on the platform.

Read more

LLM-Guided Compositional Program Synthesis

Program synthesis from input-output examples can be performed by large language models (LLMs), but they can fail unpredictably, and most recovery approaches rely on the LLM solving the same task with added context. A new technique has been introduced that recovers from failure by breaking down the task into simpler subtasks for the LLM to solve, allowing it to successfully solve challenging task instances that other methods cannot.

ArXiv

Stealthy Cross-Origin Context Poisoning Attacks Against AI Coding Assistants

AI coding assistants are vulnerable to a new type of attack called Cross-Origin Context Poisoning (XOXO), which allows attackers to subtly manipulate input and compromise the assistant's outputs, potentially generating vulnerable code or introducing critical errors. A novel attack algorithm, GCGS, has been developed to exploit this vulnerability, achieving an 83.09% attack success rate across various models and tasks, highlighting the need for new security measures in LLM-powered coding tools.

ArXiv

DeepSeek-V3 Technical Report

DeepSeek-V3 is a 671B parameter Mixture-of-Experts language model that achieves strong performance through efficient architectures and training strategies, outperforming other open-source models and matching leading closed-source models. The model was trained on 14.8 trillion tokens and required only 2.788M H800 GPU hours, with a remarkably stable training process and no irrecoverable loss spikes.

ArXiv

SplitQuantV2: Enhancing Low-Bit Quantization of LLMs Without GPUs

The SplitQuantV2 algorithm enhances low-bit linear quantization of large language models, achieving results comparable to advanced algorithms without requiring high-end GPUs or calibration datasets. This platform-agnostic algorithm can efficiently preprocess models and perform quantization using limited computational resources, as demonstrated by its ability to improve the accuracy of a 1B model using only an Apple M4 CPU in just over 2 minutes.

ArXiv

Simulating Time with Square-Root Space

A new simulation shows that multitape Turing machines running in time $t$ can be simulated in space $O(\sqrt{t \log t})$, improving upon a 50-year-old result. This simulation has various implications, including more space-efficient evaluation of bounded fan-in circuits and progress on the $P$ versus $PSPACE$ problem, by leveraging a recent algorithm for Tree Evaluation.

ArXiv

Show HN: New Agentic AI Framework in CNCF

Dapr Agents is a developer framework for building production-grade AI agent systems that operate at scale, enabling software developers to create AI agents that reason, act, and collaborate using Large Language Models. The framework offers key features such as scalability, workflow resilience, and data-driven agents, making it a cost-effective and efficient solution for AI adoption, with a vendor-neutral and open-source approach.

Repo

Show HN: Agentic AI framework specialized in DX and function calling

Agentica is a simple Agentic AI library that specializes in LLM function calling, allowing users to deliver Swagger/OpenAPI documents or TypeScript class types linearly to the library, which then handles the function calling. The library provides a range of features, including a core library, WebSocket protocol, and plugin modules, and is designed to make it easy to implement multi-agent orchestration through TypeScript class function calling.

Repo

Show HN: Add an "Ask AI" button to your docs with this open-source Q&A assistant

Gurubase is an open-source platform that allows users to create AI-powered Q&A assistants, called "Gurus", for any topic or need, using various data sources such as web pages, PDFs, and YouTube videos. The platform offers features like advanced LLM-based question answering, retrieval augmented generation, and easy integration with websites, Slack, and Discord, and is already being used by hundreds of open-source repositories.

Repo

Show HN: I made an open source Kubernetes MCP Server to talk with K8s in English

The Kubectl MCP Tool is a Model Context Protocol (MCP) server for Kubernetes that enables AI assistants to interact with Kubernetes clusters through natural language, supporting various features such as core Kubernetes operations, natural language processing, monitoring, security, and diagnostics. The tool can be installed via PyPI and operates in two modes: CLI mode for direct command-line interface and server mode to handle requests from AI assistants.

Repo

Open Source – Let AI to Tell the AI's Trend?

This project generates daily trend reports from AI-related Reddit communities in both English and Chinese, providing real-time monitoring of emerging technologies and discussions. The system uses Docker deployment, MongoDB persistence, and supports customizable configurations, allowing users to stay up-to-date with the latest developments in the AI field through comprehensive and detailed reports.

Repo