other
The AI industry, particularly OpenAI, is facing a wave of escalating and diverse threats, encompassing geopolitical warnings from Iran's Revolutionary Guard against its Stargate campus and domestic acts of violence. Incidents include a Molotov cocktail thrown at CEO Sam Altman's home, a murder threat at OpenAI's San Francisco offices, and a councilman's home shot at over an AI datacenter project, underscoring growing global and local resistance to AI development and infrastructure.
Sources: Hacker News
policy
Anthropic's Claude Code API experienced a silent regression in early March 2026, reducing the prompt cache Time-To-Live (TTL) from 1 hour to 5 minutes. This change has led to a 15-53% increase in cache creation costs and caused subscription users to hit quota limits due to more frequent cache re-creations.
Sources: Hacker News
research
Researchers from UC Berkeley developed an automated agent that systematically audited eight prominent AI agent benchmarks, discovering that every single one could be exploited to achieve near-perfect scores without actually solving tasks. These vulnerabilities, ranging from insufficient agent-evaluator isolation and arbitrary code execution to accessing local answer keys and prompt injection, highlight fundamental flaws in current AI evaluation methodologies and pose significant security risks.
Sources: Hacker News
acquisition
OpenAI has acquired Cirrus Labs to enhance its agent infrastructure team with expertise in developer tooling and virtualization. Following the acquisition, Cirrus Labs will relicense its popular source-available tools like Tart, Vetu, and Orchard under more permissive, free terms. However, the Cirrus CI continuous integration service will cease operations by June 1, 2026, with support for existing customers through their contract periods.
Sources: Hacker News
product_launch
Thoth is a new AI assistant designed for personal data control, offering a wide range of integrated tools, a knowledge graph, and multi-channel messaging, runnable locally via Ollama or with cloud models. It emphasizes user control over data and models.
Sources: Keep Bookmarks
research
Agent harnesses are now the primary method for building AI agents, orchestrating LLM interactions with tools and data. This architecture is fundamentally linked to managing agent memory, which is vital for creating personalized and persistent agent experiences.
Sources: Keep Bookmarks
product_launch
GLM-5.1 has reportedly surpassed Gemini 3.1 and GPT-5.4 on Code Arena, ranking #3 overall and #1 among open models, indicating significant progress in open-source coding AI capabilities and rapid adoption by tooling vendors.
Sources: Latent Space
product_launch
Meta has introduced Muse Spark, a new AI model, and enhanced its Meta.ai chat platform with interesting new tools, indicating continued development in its AI product offerings.
Sources: Simon Willison
other
The newsletter observes that OpenAI's free voice mode, despite being a GPT-4o era model, underperforms on simple questions, contrasting with the advanced capabilities of its paid Codex model for complex coding tasks. This difference, influenced by an Andrej Karpathy tweet, underscores a growing disconnect between public perception and actual AI model capabilities across different access points.
Sources: Simon Willison
product_launch
OpenAI has apparently removed the "Study Mode" feature from ChatGPT without public announcement, prompting user discussion and speculation about the company's evolving focus. Users note the feature was largely a system prompt and can be replicated.
Sources: Hacker News