When Jorge Luis Borges met one of the founders of AI

In 1970, Argentine writer Jorge Luis Borges met with Herbert A. Simon, a pioneer of artificial intelligence, in Buenos Aires, and their conversation modeled a desirable exchange of ideas between the humanities and science. Their discussion, which touched on topics such as free will, determinism, and human behavior, demonstrated a willingness to explore complex ideas and avoid disciplinary silos, a quality that is sadly lacking in many modern conversations between academics and experts from different fields.

Read more

Show HN: I vibecoded a 35k LoC recipe app

The text appears to be a collection of 20 different recipes, each with its own unique name, cooking time, and description, showcasing a variety of dishes from around the world, including desserts, main courses, and snacks. The recipes are contributed by different individuals, including Frederich E, Greg V, Anonymoose, and others, and cover a range of cuisines, such as British, American, Indian, and Asian.

Read more

Wikipedia is struggling with voracious AI bot crawlers

Wikipedia is experiencing a significant increase in traffic from AI bot crawlers, which are scraping the site's content to train generative artificial intelligence models, resulting in a 50% increase in bandwidth usage since January 2024. This surge in bot traffic is causing concerns for the Wikimedia Foundation, as it may slow down access to the site and increase costs, with 65% of resource-consuming traffic coming from bots, and the foundation is now seeking sustainable ways to manage this traffic.

Read more

What, exactly, is an 'AI Agent'? Here's a litmus test

An AI agent is defined as a system that performs actions under its own identity, implying autonomy, capability, and reasoning, and can be distinguished from AI assistants by its ability to act independently. This definition is based on the concept of identity in law and enterprise software systems, where an agent's actions are recorded in audit logs under its own name, rather than as an extension of a human user.

Read more

I stopped using AI code editors

The author stopped using AI-powered code editing tools after realizing they were losing their competence and intuition in programming due to over-reliance on these tools. They draw a parallel with their experience of using Tesla's Full Self-Driving feature, which also led to a decline in their driving skills, and warn that relying on AI in coding can have similar effects, making it harder to develop the "finger tips feeling" and intuition that comes with experience and practice.

Read more

State Space Model Meets Transformer: A New Paradigm for 3D Object Detection

The proposed DEST method introduces an interactive State Space Model (SSM) for 3D object detection, enabling system states to serve as queries and scene points as system inputs, which can be updated simultaneously with linear complexity. The DEST method achieves state-of-the-art performance on ScanNet V2 and SUN RGB-D datasets, outperforming baselines and demonstrating its effectiveness in 3D indoor object detection tasks.

ArXiv

HiRAG: RAG with Hierarchical Knowledge

Graph-based Retrieval-Augmented Generation (RAG) methods have improved large language models' performance in domain-specific tasks, but existing methods underutilize hierarchical knowledge inherent in human cognition. The proposed HiRAG approach addresses this limitation, leveraging hierarchical knowledge to enhance semantic understanding and structure capturing, and achieves significant performance improvements over state-of-the-art baseline methods.

ArXiv

Prompt, Divide, and Conquer: Bypassing Large Language Model Safety Filters

Researchers have developed a framework to bypass safety measures in large language models, allowing them to generate malicious code with a 73.2% success rate across 10 cybersecurity categories. The framework, which uses distributed prompt processing and a jury evaluation system, outperforms traditional single-LLM judge evaluation and highlights the importance of robust evaluation methodologies in assessing jailbreak attempts.

ArXiv

Multi-Token Attention

The current attention mechanism in large language models (LLMs) is limited by its reliance on single token vectors, which can bottleneck the amount of information used to distinguish relevant context. A new method, Multi-Token Attention (MTA), addresses this issue by allowing LLMs to condition attention weights on multiple query and key vectors simultaneously, resulting in enhanced performance on various benchmarks.

ArXiv

UCSD: Large Language Models Pass the Turing Test

In a series of Turing tests, the GPT-4.5 system was judged to be human 73% of the time when prompted to adopt a humanlike persona, outperforming actual human participants. The results also showed that LLaMa-3.1 was judged to be human 56% of the time, while less advanced models performed significantly below chance, marking the first time an artificial system has passed a standard three-party Turing test.

ArXiv

Why Git is no "good" for AI-generated code

SpecStory is a Visual Studio Code extension for the Cursor IDE that captures, searches, and learns from AI coding journeys, automatically saving every Cursor chat and composer session to a local project directory. To install the extension, it must be downloaded directly from the SpecStory release bucket and installed into Cursor, as installing from the VSC Marketplace will not work.

Repo

Show HN: Arrakis – Open-source, self-hostable sandboxing service for AI Agents

Arrakis is a secure, fully customizable, and self-hosted solution for spawning and managing sandboxes for code execution and computer use, providing a secure environment for running untrusted code. It features a REST API, Python SDK, and CLI client, allowing users to spawn, manage, and interact with virtual machines, including snapshot and restore capabilities for backtracking and debugging purposes.

Repo

Show HN: Open Responses – Self-hosted OpenAI Responses API, works with any model

Open Responses is a self-hosted, open-source alternative to OpenAI's Responses API, allowing users to run a fully compatible version of the API with any large language model provider. It offers a drop-in replacement for the official API, providing features such as privacy, flexibility, and ease of use, with quick setup options using Docker or a straightforward CLI.

Repo

Real-Time Introspective Compression for Transformers

Transformer-based large language models face limitations due to their lack of introspection and ephemeral cognition, making it difficult to access and record their internal states. A proposed solution, called Real-Time Introspective Compression for Transformers, involves using a sidecar model to encode the internal state of a transformer into a compact latent representation, allowing for high-fidelity reconstruction and addressing the issues of introspection and ephemeral cognition.

Repo

Show HN: Agent File (.af) – A standard file format for serializing AI agents

The Agent File (.af) is an open standard file format for serializing stateful agents, allowing for the portable sharing of agents with persistent memory and behavior between compatible frameworks. The .af file includes various components such as model configuration, message history, system prompt, memory blocks, tool rules, and environment variables, enabling seamless transfer and collaboration of agent development.

Repo