Claude Code auto mode: a safer way to skip permissions

#1 📝 Anthropic Engineering

Claude Code auto mode: a safer way to skip permissions - Introduces Claude Code's 'auto mode' as a safer mechanism to allow agents to bypass repetitive permission prompts while maintaining improved security controls. The post outlines design decisions and benefits for autonomous coding workflows.

#2 📝 OpenAI News

Inside our approach to the Model Spec - An explanation of OpenAI's approach to a Model Spec, outlining principles and research considerations that guide model development and governance. It emphasizes transparency and rigorous research-driven practices.

Also covered by: @OpenAI

#3 📝 OpenAI News

Introducing the OpenAI Safety Bug Bounty program - OpenAI announces the Safety Bug Bounty program to incentivize external researchers to find safety vulnerabilities and help improve the robustness of OpenAI systems. The program focuses on safety-critical issues and responsible disclosure.

#4 𝕏

Google DeepMind is rolling out Lyria 3 Pro, offering an API for developers in Google AI Studio and in-app access for paid subscribers via the Gemini App.

Also covered by: @Google DeepMind, @Demis Hassabis, @Demis Hassabis, @Logan Kilpatrick

#5 𝕏

Google Research introduced Vibe Coding XR, a rapid prototyping workflow that pairs Gemini Canvas with the XR Blocks framework. It turns simple prompts into interactive, physics-aware WebXR apps for fast testing of intelligent spatial experiences.

#6 𝕏

Peter Yang breaks down how a Meta VP created the /exec-review AI skill—using a 6-step leader profiling method to review docs in his own voice—and shares the full skill files so you can set it up today and stop wasting time guessing exec preferences.

#7 𝕏

claire vo 🖤 shows how Stripe’s “Minions” use a simple Slack emoji plus custom DevX tooling to spin up isolated AI agent environments that run dozens of parallel tasks. This setup powers over 1,000 AI-written PRs every week while keeping reviews manageable.

#8 𝕏

Cursor launched self-hosted cloud agents that let you deploy their cloud agent harness on your own infrastructure, keeping code execution and tool integrations entirely in your private network.

#9 📝 Anthropic Engineering

Quantifying infrastructure noise in agentic coding evals - A featured examination showing that infrastructure configuration can shift agentic coding benchmark results by several percentage points, sometimes exceeding differences between top models. The piece highlights the importance of accounting for infrastructure variability when comparing model performance.

#10 𝕏

LlamaIndex 🦙 demonstrates how LlamaParse now fully leverages .docx’s ZIP-of-XML structure to extract rich details—cell boundaries, merged cells, nested tables, formatting tags and hyperlinks—vastly outperforming PDF parsing.

#11 𝕏

NVIDIA AI: At #NVIDIAGTC, Cohere VP Autumn Moulder unveiled a full-stack sovereign AI blueprint—hosting models, apps, and reasoning traces in a single data center—and emphasized open models like NVIDIA Nemotron for data lineage and regulatory compliance.

#12 𝕏

DeepLearning.AI shared its upcoming DeepSeek-V4 model with Huawei while denying early access to Nvidia and AMD. This move underscores how US export controls struggle to influence the US–China competition for advanced hardware.

#13 𝕏

Kevin Yien reveals that Linear’s 2026 roadmap isn’t just about issue tracking—it includes new capabilities like Linear Agent, Skills, Automations, Code Intelligence, and the Linear Coding Agent.

#14 𝕏

clem 🤗 Anthropic revoked OpenAI’s access to its Claude AI models, citing repeated misuse of proprietary system prompts and escalating the rivalry between the two labs.

#15 𝕏

Philipp Schmid released Lyria 3 Pro and Clip in Google AI Studio and the Gemini API, offering prompt-driven full-song generation (minutes at $0.08 each) and 30-second clips ($0.04 each).

Also covered by: @Google DeepMind, @Demis Hassabis, @Demis Hassabis, @Logan Kilpatrick

#16 📝 Surge AI Blog

Riemann-bench: A Benchmark for Moonshot Mathematics - Riemann-bench is a verifiable benchmark of extreme-tier mathematical problems designed to test frontier models; current top models score under 10% on these challenges. The post introduces the benchmark as a tool for driving progress on very difficult mathematical tasks.

#17 in

Marc Baselga shares 5 sharp reads for product leaders this month. Highlights include Benedict Evans’ case that OpenAI lacks a durable moat and Gokul Rajaram’s prediction that AI-native firms will eliminate the traditional CPO role by merging product, design, and engineering.

#18 in

Dharmesh Shah echoes Reid Hoffman’s insight that AI-powered agents open vast new opportunities for software companies, proving software is far from dead.

#19 in

Linear says traditional issue tracking is dead: AI coding agents now power 75%+ of workspaces, grew 5× in three months and even author 25% of new issues. They’ve rebuilt their platform around four AI-centric layers—context, rules, agents and workflows.

#20 📝 Jesse Vincent

Classical Software - The author examines the distinction between traditional "software" and software that has an AI agent or large language model in the loop, noting this difference repeatedly arises in conversations about AI and software. The piece suggests this distinction affects how we think about design, expectations, and the role of agents in systems.

#21 𝕏

Sebastian Raschka warns that although TurboQuant’s improved quantization can cut KV-cache memory, the freed capacity just gets eaten up by bigger intermediate representations, larger vocabularies, longer contexts, and generally larger models.

#22 𝕏

Sebastian Raschka stresses that open-source projects rely on subsidies like corporate jobs or side gigs to stay sustainable, and warns that open-weight models rack up steep GPU costs, straining a limited GPU pool.

#23 𝕏

Boris Cherny says Claude Code Review automatically catches over 99% of bugs, leaving engineers to only perform a quick sanity check.

#24 𝕏

Anthropic introduced Claude Code auto mode, a safer middle ground that uses trained classifiers to automatically approve or reject code-generation requests instead of relying on manual permission prompts.

#25 𝕏

OpenAI introduces its Model Spec framework, defining standardized model capabilities, evaluation metrics, and safety guardrails to streamline development and ensure consistent performance.

Also covered by: @OpenAI