Sunday — August 17, 2025

Researchers propose a 2-bit quantization framework for complex-valued large language models, developers release UrbanOS-PoC, a sovereign and self-healing AI architecture for city mobility, and a new neural network architecture based on the Tversky similarity function achieves notable improvements in image recognition and language modeling.

News

Best Practices for Building Agentic AI Systems

The author has been experimenting with adding AI agents to their feedback platform, UserJot, to analyze customer feedback at scale and auto-generate changelog entries. They found that a two-tier agent model, consisting of primary agents that handle conversation and context, and subagents that perform specific tasks, works best, and have developed a set of principles and patterns for building effective agent systems, including stateless subagents, task decomposition, and structured communication protocols.

Dyna – Logic Programming for Machine Learning

Dyna is a programming language designed for machine learning researchers, building on logic programming languages like Datalog and Prolog, but allowing for flexible execution orders and weighted rules. The language enables efficient expression of complex programs in a few lines of code, with examples including matrix multiplication, the Fibonacci sequence, and neural networks, and has undergone several iterations since its initial development in 2004.

Tversky Neural Networks

The authors propose a new neural network architecture based on the Tversky similarity function, which is a more sophisticated and asymmetric function than the traditional dot product or cosine similarity used in modern deep learning architectures. This new architecture, called Tversky neural networks, has been shown to achieve notable improvements in various domains, including image recognition and language modeling, by replacing standard linear layers with Tversky projection layers.

Blue-collar jobs are gaining popularity as AI threatens office work

Blue-collar jobs are gaining popularity as a means of achieving job security, particularly among younger generations, due to the threat of artificial intelligence replacing certain office and white-collar positions. Many experts, including Nobel Prize-winning computer scientist Geoffrey Hinton, believe that jobs requiring manual labor and expertise, such as plumbing and skilled trades, are less vulnerable to automation and will continue to be in demand.

Flock Reports to Police If It Thinks Car Movement Patterns "Suspicious"

The surveillance company Flock is using AI to analyze movement patterns of vehicles and report them to police if they are deemed "suspicious", effectively generating suspicion rather than just investigating based on existing suspicion. This expansion of Flock's surveillance infrastructure allows police to target individuals based on algorithmic decisions, raising concerns about privacy, bias, and the potential for wrongful targeting of innocent civilians.

Research

IFairy: The First 2-bit Complex LLM with All Parameters in {\pm1, \pm i}

Researchers propose Fairy$\pm i$, a 2-bit quantization framework for complex-valued large language models (LLMs) that surpasses the accuracy ceiling of existing methods by leveraging the complex domain to boost full-precision accuracy. The framework achieves state-of-the-art results in terms of perplexity and downstream tasks while maintaining strict storage and compute efficiency, opening a new direction for building highly accurate LLMs under extremely low-bit constraints.

PyG 2.0: Scalable Learning on Real World Graphs

PyG (PyTorch Geometric) has undergone significant updates, particularly with the release of PyG 2.0, which introduces substantial improvements in scalability and real-world application capabilities. The updated framework now supports heterogeneous and temporal graphs, scalable feature stores, and various optimizations, enabling efficient large-scale graph learning and supporting a wide range of application areas, including relational deep learning and large language modeling.

ISR: Invertible Symbolic Regression (2024)

The Invertible Symbolic Regression (ISR) method is a machine learning technique that generates analytical relationships between inputs and outputs of a dataset using invertible maps, combining principles of Invertible Neural Networks and Equation Learner. ISR allows for efficient gradient-based learning, discovery of concise expressions, and can be applied to tasks such as density estimation, inverse problems, and solving complex problems like geoacoustic inversion in oceanography.

Composing Linear Layers from Irreducibles

Researchers have found that linear layers in large models can be broken down into simpler geometric primitives, specifically bivectors and rotors, which can be composed to form more complex functions. This discovery allows for a more efficient representation of linear layers, using significantly fewer parameters, and has been successfully applied to attention layers in large language models without sacrificing performance.

Code

UrbanOS-PoC Real-Time, Sovereign, Self-Heailing AI for City Mobility

UrbanOS-PoC is a proof of concept built on a sovereign and self-healing AI architecture, designed to minimize friction in urban flow by analyzing machine data from various sources, including IoT units and APIs. The system operates a 26-hour machine learning loop, continuously re-learning from trajectories, patterns, and hotspots to adjust routing and predictions, and is compatible with various data feeds, including Stockholm's public transport feeds via Trafiklab.

Show HN: Spin up 5 agents that don't trip over each other in 10 lines of code

Kage Bus is a lightweight message bus that enables multiple AI agents to work together efficiently by routing tasks to individual agents, handling conflicts, and logging activities for debugging. It provides a simple pub/sub API, conflict resolution strategies, and local logging, allowing users to run multiple agents in parallel with minimal code.

Show HN: Scoped, expiring API keys for AI agents

Kage Keys provides scoped, temporary tokens for AI agents, allowing for secure and auditable access to APIs with limited permissions. The library offers features such as auto-expiring tokens, logging, and customizable expiration times, making it a useful tool for debugging, demos, and enhancing the safety of AI agents.

Visual Reasoning and Tool Use Double GPT-5's Arc-AGI-2 Success Rate

A Python solver was built to test the idea of treating ARC-AGI-2 puzzles as visual pattern recognition tasks, and it achieved a 22% score on the evaluation dataset, surpassing the current AI state-of-the-art of 15.9%. The solver uses a phased visual approach, converting numerical grids into images and leveraging GPT-5's multimodal capabilities, and its success suggests that visual reasoning approaches can substantially improve AI performance on abstract reasoning tasks.

Structured AI workflows made easy

Roast is a convention-oriented framework for creating structured AI workflows, allowing users to define powerful workflows using simple YAML configuration files and prompts written in markdown. It provides a range of features, including built-in tools, Ruby integration, shared context, step customization, and extensive instrumentation, making it a flexible and efficient tool for building AI workflows.

    Researchers propose a 2-bit quantization framework for complex-valued large language models, developers release UrbanOS-PoC, a sovereign and self-healing AI architecture for city mobility, and a new neural network architecture based on the Tversky similarity function achieves notable improvements in image recognition and language modeling.