Crusoe and NVIDIA Dynamo developed fastokens, an open-source Rust BPE tokenizer that achieves a 9.1× average speedup over HuggingFace and up to 40% faster TTFT on long-context workloads.

ToolOps is a framework-agnostic middleware SDK that upgrades AI agent tool calls with industrial-grade caching, resilience, and observability through a single decorator, solving production challenges like high API costs, instability, concurrency bottlenecks, and lack of standardization.

Outdated GOV.UK pages are being scraped by AI overviews, delivering false information and undermining public trust. The Department for Business and Trade audited and fixed 150 stale pages, and is testing six-monthly review cycles.

The 'double descent' phenomenon in AI shows that extremely complex models can learn better, challenging the long-held psychological assumption that learning requires simplification. A new paper proposes 'excess-capacity learning,' suggesting humans may also learn by memorizing all details, including seemingly irrelevant ones.

AgentPeek integrates Claude Code and Codex into your Mac notch, providing quick access to AI coding assistants.

Turn your work into AI agent memory, served over MCP.

In conversation with Clem Delangue, co-founder and CEO at Hugging Face about the state of Open Source AI (and the myth of Sisyphus)

Voxel is an open-source, local-first AI assistant that runs on your own machine through a lightweight web dashboard, using local GGUF models with optional web search and voice support.

The U.S. Copyright Office's January 2025 report clarifies: human-created portions are copyrightable, AI-generated portions are not. The article covers how to register AI-assisted songs, using GRAM to save costs on albums, and the current state of Suno and Udio litigations.

Web Speed is the logic layer for web agents, converting messy DOM HTML into high-fidelity, token-efficient JSON. It reduces token costs by 70-90%, enhancing speed and reliability. Launched on Product Hunt with SDK and MCP server support.

AI Design Taste is a Chrome extension that extracts design styles from webpages and generates structured DESIGN.md or SKILL.md files for use with Google Stitch, Claude, Codex, and Cursor.

LM Studio introduces LM Link, a preview feature that allows you to load models on remote machines and use them as if they were local, via end-to-end encrypted Tailscale mesh VPNs. It supports local devices, LLM rigs, and cloud VMs, and integrates with existing tools that point to localhost:1234. Free tier includes 2 users and 5 devices each.

Mirage is a simulated environment where AI agents reach every data through one filesystem and bash. It mounts services like S3, Google Drive, GitHub, and Postgres as a single virtual filesystem, enabling agents to use Unix-like tools across backends. Features include snapshot/rollback, polymorphic format support, two-layer caching, and easy integration with popular AI frameworks.

A critical flaw in Anthropic's “Claude in Chrome” extension, dubbed ClaudeBleed, allows any Chrome extension to hijack Claude's AI and perform sensitive actions like stealing emails and accessing private repositories.

An open-source AI tool to quickly generate app store screenshots.

mlx-code is a lightweight, local AI coding agent for Mac built on Apple's MLX framework. It prioritizes fast feedback loops and transparency, using a terminal interface with text-in, text-out. It supports multiple providers, prompt caching, and a suite of CLI tools for local and remote agent interactions.

Elon Musk's SpaceX plans to invest up to $119 billion in a massive semiconductor fab in Texas under the 'Terafab' project to supply chips for orbital AI datacenters. Despite Musk's ambitions, the company lacks fab experience and the project depends on Starship reducing launch costs. Intel has signed on as a partner.

Color's Virtual Cancer Clinic achieves ASCO Certification, marking a milestone in AI-driven cancer care. Omer Isler argues that the real AI revolution in cancer lies not in discovering silver bullet drugs, but in transforming care delivery—improving early detection, reducing wait times, expanding access to expertise, and supporting patients through the journey. AI enables risk-based screening, faster diagnosis and treatment coordination, multidisciplinary reviews for more patients, and always-on patient services. The piece emphasizes that even existing scientific knowledge is often not delivered effectively, and AI can close those gaps.

OncoAgent is an open-source, privacy-preserving clinical decision support system for oncology. It features a dual-tier LLM architecture (9B speed vs 27B deep reasoning), multi-agent LangGraph topology, Corrective RAG pipeline over 70+ NCCN and ESMO guidelines, and a three-layer reflexion safety validator with Zero-PHI policy. The system routes queries via complexity scoring and was fine-tuned on AMD Instinct MI300X, achieving 56x throughput acceleration. It supports on-premises deployment to ensure data sovereignty.

The author switched from Bear to Obsidian for note-taking and from YNAB to a custom app for personal finance, driven by the need for AI to have direct, bidirectional data access. For unstructured data like notes, plain text files work best; for structured data like finance, raw database access is ideal. The article discusses the benefits and risks, emphasizing versioning and review for safe AI write operations.

A game built as an AssemblyScript git repo now includes an integrated chat feature that lets users edit the game by simply conversing with it, making it accessible to non-coders.

A LessWrong article discusses how Go players disempower themselves by relying on AI for training and analysis, losing their own judgment and creativity.

The COLIBRI Concept Library File (CLF) format v0.94 pre-standard draft defines how a single concept is represented. Its core 'Concept Purity Rule' ensures concept files describe what a concept IS and its perceptual typicality, without detection logic or behavioral guidance. The format is intentionally not final, aiming for broad industry collaboration. The repository includes the specification, white paper, control layer reference implementation, and examples. Protected by Finnish utility model.

The Earth is warming, conflicts rage, AI infiltrates our lives, and lab risks increase. The Doomsday Clock now stands at 85 seconds to midnight, the closest ever. Experts cite leadership failure in addressing interconnected global threats.

New Gartner research reveals that companies cutting staff in favor of AI are not seeing improved returns. Instead, those investing in human-AI collaboration achieve better ROI. The study warns that layoffs merely create vacancies, not value.

This article outlines a systematic four-phase agent development lifecycle: Build, Test, Deploy, and Monitor. It emphasizes testing before deployment, using runtimes and sandboxes for reliable deployment, and leveraging traces and feedback for monitoring. Covers tools from code-first to no-code, and discusses best practices for datasets, experiments, simulations, and signal harvesting.

AgentPitch is an open-source LLM-powered soccer simulation where every player on the field is an AI agent running a decide() callback — generated, sandboxed, and evolved by large language models. It supports Python, JavaScript, and Rust strategy languages with multiple LLM providers including OpenAI, Anthropic, and Gemini. Features three tournament modes: Arena, Cup, and League, along with a live browser UI.

MentiSphere is an open-source platform enabling domain experts to collaboratively build and embed specialized knowledge into AI agents through wiki-like pages, without needing to code.

WUPHF is a free and open-source tool that runs a team of AI agents locally with a shared brain. Agents have roles like CEO, ENG, and CMO, and coordinate autonomously via channels. It's designed to eliminate the need for humans to route between agents, reduces token usage, and allows mixing AI providers. All data stays local.

Armorer is an open-source secure local control plane for installing, configuring, and monitoring AI agents in Docker containers. It features one-command installation, CLI and UI, and supports automated setup by coding agents like Codex and Claude Code.