Mathematicians increasingly value the formalization of math, with advancements in proof assistants like Lean. Recent developments in autoformalization using AI have made it possible to translate mathematical proofs into formal code, promising a transformative impact on mathematics and research efficiency.

Whistle is a compact speech recognition model designed for various devices, including mobiles and robots. It offers transcription, word timestamps, and speech embedding with multi-language support. The model operates without dependencies on CPUs, boasting efficient performance and easy deployment across multiple platforms.

The NVIDIA KGMON team achieved second place in the KDD Cup 2026 Data Agents competition by streamlining their system to enhance reliability. By simplifying agent tools and optimizing data access, they developed effective strategies for handling diverse information sources, demonstrating that thoughtful harness design can elevate performance in data analytics tasks.

Creating products that seamlessly integrate with any observability stack is crucial for self-hosted software and SaaS offerings. This article explores how developers can facilitate the export of logs, traces, and metrics to an OpenTelemetry-compatible backend, enhancing user flexibility while adhering to modern observability best practices.

Three safety researchers recently fired by OpenAI have published an open letter challenging the company's claims about their dismissal and warning of its negative impact on workplace culture. They express concerns over a chilling effect that may stifle open discussions about AI safety among current employees.

A software engineer inspired by historian Benjamin Breen's use of AI explored millions of historical records to uncover overlooked findings. This investigation successfully revealed a new dodo account, three lost rhinos, and unrecorded volcanic eruptions, showcasing the potential of AI in historical research.

Coding agents have made great strides in AI-assisted development, yet many still struggle with basic tasks. They often fail at managing workloads, lack self-awareness, and communicate inadequately. This article explores the limitations of these agents and highlights the challenges developers face when integrating them into their workflows.

Kavish, a summer intern, developed a generative diffusion model to analyze market data using four years of US equities data. His work focused on handling the complexities of continuous and discrete event features. Despite challenges with model stability, he found flow matching models offered superior performance for real-world data predictions.

Concerns about the diminishing availability of original ideas in science often arise, leading to doubts among aspiring researchers. However, the belief that innovation has plateaued is misleading. History shows that scientific breakthroughs often come when least expected, encouraging new generations to persevere in exploration and creativity.

his team highlight, is with the auto-formalisation process used by AI. The researchers argue that this method may not reliably replicate the logical integrity found in human-written proofs. Consequently, it emphasizes the necessity for careful human review of AI-generated mathematical work.

Rust has enabled a significant overhaul of Prime Agent. This rewrite enhances performance and reliability while ensuring compatibility with the previous TypeScript version. Automated agents played a crucial role in the process, executing a structured approach to maintain feature parity and improve modularity, resulting in a more efficient codebase.

The autonomous trucking industry is gaining momentum as startups aim to address safety concerns and labor shortages. With regulatory milestones being reached, companies are inching closer to commercial deployment. However, significant challenges remain regarding safety, technology, and approval processes in such a complex landscape.

This article explores the dynamics of AI agent populations in cybersecurity, focusing on how collaboration affects their growth. It introduces the concept of a critical population threshold, highlighting the importance of ecological safety measures as larger groups can lead to self-reinforcing cycles of misalignment and enhanced cyber capabilities.

PRAXIS is a co-evolutionary framework that models self-improving learning systems. It explores the dynamics of generators, learners, and symbolic archives, demonstrating how these interactions lead to stabilization and improved performance across various tasks. The findings highlight significant advancements in machine learning adaptation and efficiency.

This article introduces Hippocam, a new architecture for large language model agents that enhances learning by mimicking human memory processes. It focuses on structuring experiences using nested intents, facilitating the consolidation of knowledge over time while retaining relevant past interactions for improved future performance.