To those who fired or didn't hire tech writers because of AI

Replacing technical writers with LLMs is a strategic error because AI lacks the empathy, vision, and strategic oversight required to produce accurate "product truth." LLMs often generate "docs theatre" that fails to address edge cases, and they cannot be held liable for harmful instructions that lead to data loss or system failure. Crucially, technical writers provide the high-quality context curation necessary for effective RAG and AI performance, making them the essential supply chain for the intelligence companies are betting on. Organizations should prioritize human-AI augmentation to maintain documentation quality and legal accountability.

Read more

Raspberry Pi's New AI Hat Adds 8GB of RAM for Local LLMs

The Raspberry Pi AI HAT+ 2 features a Hailo 10H NPU and 8GB of dedicated LPDDR4X RAM, delivering 40 TOPS of INT8 performance for local LLMs and vision tasks. Benchmarks indicate that the Pi 5 CPU still outperforms the NPU in inference speed due to the Hailo's 3W power limit, though the NPU offers better power efficiency. While effective for high-speed vision processing, the HAT currently faces software hurdles with mixed-mode workloads and is primarily suited for niche, power-constrained edge AI development.

Read more

Furiosa: 3.5x efficiency over H100s

FuriosaAI’s NXT RNGD Server is a turnkey AI inference system featuring eight RNGD accelerators that deliver 4 petaFLOPS of FP8 compute and 12 TB/s of HBM3 bandwidth. Optimized for standard air-cooled data centers, the 3 kW system utilizes PCIe interconnects and comes pre-installed with a vLLM-compatible Furiosa LLM runtime. It offers a power-efficient alternative to high-density GPU clusters, enabling high-throughput inference for models like EXAONE 3.5 within existing enterprise power and cooling constraints.

Read more

X to stop Grok AI from undressing images of real people after backlash

X has implemented geoblocking and technological safeguards to prevent Grok from generating sexualized deepfakes of real people in jurisdictions where such content is illegal. To enhance accountability, image editing features are now restricted to paid users, while NSFW settings are limited to synthetic nudity of imaginary humans. Regulators continue to investigate the platform's compliance, citing concerns over the model's ability to accurately distinguish real individuals from AI-generated personas and the potential for users to circumvent geoblocks via VPNs.

Read more

Student arrested for eating AI art in UAF gallery protest

A UAF student was arrested for destroying and consuming AI-generated artwork in a protest against the medium. The exhibit's creator, Nick Dwyer, utilized AI to explore "AI psychosis," a phenomenon characterized by psychosis-like episodes resulting from deep engagement with chatbots. This incident highlights emerging psychological and societal tensions regarding the integration of generative AI in creative and interactive spaces.

Read more

A Marketplace for AI-Generated Adult Content and Deepfakes[pdf]

A longitudinal analysis of Civitai’s Bounties marketplace reveals that the platform primarily incentivizes tools for steering AI models and generating NSFW content, which now comprises the majority of requests. The study highlights a high concentration of deepfake commissions targeting female celebrities, often in violation of platform policies. These findings underscore significant governance and enforcement challenges regarding gendered harms in community-driven generative AI ecosystems.

ArXiv

A benchmark for LLM vericoding: formally verified program synthesis

This work introduces the largest vericoding benchmark, comprising 12,504 formal specifications (Dafny, Verus/Rust, Lean) for LLM-generation of formally verified code. Off-the-shelf LLMs achieved vericoding success rates of 27% in Lean, 44% in Verus/Rust, and 82% in Dafny. Natural language descriptions did not significantly improve performance, and LLM progress has boosted pure Dafny verification from 68% to 96% in the last year.

ArXiv

How Prompt Injections Gradually Evolved into a Multi-Step Malware

The paper defines "promptware" as a distinct class of malware targeting LLM-based applications, moving beyond the generic "prompt injection" label to describe complex, multi-step attacks. It introduces a five-step kill chain—Initial Access, Privilege Escalation, Persistence, Lateral Movement, and Actions on Objective—to model attack sequences analogous to traditional malware campaigns. This framework provides a structured methodology for threat modeling and bridges the gap between AI safety and cybersecurity.

ArXiv

Dust Properties of the Interstellar Object 3I/Atlas

Polarimetric observations of the interstellar object 3I/ATLAS reveal an unusually high polarization amplitude that remains stable across perihelion, indicating intrinsic refractory dust properties rather than transient volatile activity. Analysis of the PCC suggests the object is a primitive planetesimal composed of submicron-sized dust aggregates. While these building blocks share size scales with Solar System comets, 3I/ATLAS possesses a distinct refractory composition indicative of its extra-solar origin.

ArXiv

Categorical Foundations for CuTe Layouts

This paper formalizes the CuTe layout algebra within NVIDIA’s CUTLASS library using a categorical framework based on the categories Tuple and Nest. By defining operations on morphisms and proving their compatibility with layout operations like composition and logical division, the authors provide a complete characterization of tractable GPU tensor layouts. A Python implementation is provided to verify alignment with CUTLASS behavior.

ArXiv

Tldraw pauses external contributions due to AI slop

tldraw is a React library for building infinite canvas experiences and digital whiteboards. The repository is specifically optimized for AI agents, featuring distributed CONTEXT.md files and a yarn context script to facilitate codebase navigation and context retrieval for LLMs. It is available via npm and supports both commercial and non-commercial use under a watermark-based licensing model.

Repo

Gambit, an open-source agent harness for building reliable AI agents

Gambit is a framework for building reliable LLM workflows by composing modular, typed "decks" with explicit input/output schemas and guardrails. It enables developers to mix LLM calls and compute tasks interchangeably, offering local execution, streaming traces, and a built-in debug UI to replace brittle orchestration and opaque provider logs. The system supports both Markdown-based model prompts and TypeScript-based logic, utilizing Zod for validation to ensure predictable state management and observability.

Repo

Tusk Drift – Turn production traffic into API tests

Tusk Drift is an API record/replay testing framework that generates realistic test suites from live traffic. It features an LLM-powered setup agent that automatically instruments codebases and generates configurations for Node.js and Python services. The platform provides deterministic I/O mocking, JSON response comparison with dynamic field rules, and cloud-based intelligent regression analysis for CI/CD pipelines.

Repo

1Code – Open-source Cursor-like UI for Claude Code

1Code is a UI for Claude Code, designed to facilitate local and remote agent execution for AI-driven development. It offers features like isolated Git worktrees for each chat session, real-time visibility into agent tool execution (bash commands, file edits, web searches), and comprehensive project management with automatic Git remote detection. Users can operate in either read-only analysis or full code execution modes, providing granular control over AI agents.

Repo

A-MEM – Memory for Claude Code that links and evolves on its own

A-MEM is a self-evolving memory system for coding agents that organizes knowledge into a Zettelkasten-style graph rather than a static vector store. It utilizes an LLM to automatically link, update, and strengthen connections between memories, enabling a dynamic knowledge base that grows in complexity over time. The system is MCP-compatible and employs a "Peek and Drill" search strategy, combining semantic similarity with graph traversal to optimize recall while minimizing token consumption.

Repo