News reporting focused on AI and machine learning, covering the companies behind these technologies, their real-world applications, and the ethical concerns they raise. This includes areas like generative AI (large language models, text-to-image and video), speech tech, and predictive analytics.
A new study from Princeton and UC San Diego, based on 8,135 test runs, reveals that AI agent skills primarily provide procedural grounding, reducing execution errors in 65.7% of cases. However, skills also introduce new failure modes, and retrieval precision drops sharply as skill libraries grow, suggesting a lifecycle approach is needed.
Researchers released LittleLearner, a family of language models trained from scratch on a strictly filtered K-5 elementary school curriculum, to answer whether capabilities beyond training data can be elicited or acquired through scaling, post-training, and in-context learning. The answer is largely no: scaling, post-training, and in-context learning amplify what the curriculum taught, but none meaningfully improve out-of-scope performance. The pretraining filter sets the effective capability ceiling, providing a controlled sandbox for studying knowledge acquisition and RL.
SpaceX's $60 billion investment in AI code editor Cursor signals a market shift toward valuing judgment over junior coding skills. The deal, six times Cursor's valuation from nine months earlier, coincides with a 33-month decline in junior developer employment, highlighting a growing divide in how companies approach AI's impact on software development.
Amazon, Cursor, Microsoft, OpenAI, and Vercel have introduced Agent Plugins, an open standard to unify AI agent extension packaging. The standard, version 1.0.0, supports Agent Skills and MCP servers, aiming to reduce fragmentation. Anthropic, creator of the underlying protocols, is notably absent from the coalition.
Ponytail, an open-source AI coding agent skill, corrected its benchmark claims after external criticism revealed inflated metrics. The project now includes a behavioral test framework and a public reproduction path, setting a precedent for accountability in AI tooling. The revised benchmark shows approximately 54% less code on average, with the 94% figure clarified as a per-task ceiling.
Simon Willison accidentally activated a pet in Codex Desktop and then created his own custom pet called Pedalican, a pelican on a bicycle. The process used GPT-5.6 Sol xhigh and gpt-image-2 to generate sprite assets. Willison documented the entire process, including generated images, sprite sheets, and GIFs, and shared the open-source skills on GitHub.
Anthropic has released an open-source reference implementation for autonomous vulnerability discovery and remediation using Claude. The framework includes skills for threat modeling, static scanning, triage, patching, and an autonomous pipeline designed for finding C/C++ memory vulnerabilities. It is based on learnings from partnerships with security teams and is intended as a starting point for building custom vulnerability finding pipelines.
Developers at Quandri ran experiments on MCP (Model Context Protocol) and found it consumes up to 10.5% of the context window with tool definitions alone. The CLI approach uses 65x fewer tokens for the same task. The article explores alternatives like Skills that wrap existing CLI commands.
Open Design provides an open-source option to Anthropic's Claude Design, released on 2026-04-17 with Opus 4.7. It uses 12 coding-agent CLIs like Claude Code, Codex, and Devin for Terminal, detected on PATH, along with 31 skills and 72 design systems. Local-first and web-deployable, it supports BYOK at every layer and runs with pnpm tools-dev.
Amazon Bedrock AgentCore now offers a managed agent harness that lets developers create and run agents with just three API calls, skipping complex infrastructure setup. New tools include a CLI for full lifecycle management and pre-built skills for coding assistants. These features support quick prototyping and production deployment across multiple AWS regions.
US startup Physical Intelligence released π0.7, a robot foundation model that recombines trained skills much like language models handle text. Built on Google's Gemma3 with added contextual data, it matches specialist models on tasks and transfers to new robots. The approach raises questions about true generalization versus remixing similar training examples, echoing debates in large language models.
Physical Intelligence released research on its π0.7 model, which guides robots through tasks not directly trained on it. The San Francisco startup's two-year-old effort shows robots combining skills from sparse data, like using an air fryer with minimal prior examples. Researchers report surprises in capabilities, though limitations persist, amid $5.6 billion valuation and talks for more funding.