tl;dr: A profanity-laden guide to self-teaching botany, emphasizing that beginners shouldn't be intimidated by Latin terminology or taxonomy—both of which exist because common names are unreliable and modern classification reflects evolutionary relationships, letting you identify unfamiliar plants by family traits. The author recommends Michael Simpson's "Plant Systematics" and "Raven's Biology of Plants" as core texts, along with a long list of supplementary books, and suggests using Libgen/Sci-Hub on a cheap tablet if cost is prohibitive.

tl;dr: Discovery Loop is a new startup founded by Jeff Dean, Sanjay Ghemawat, Quoc Le, and Oriol Vinyals aiming to automate scientific and engineering experimental loops using frontier AI models and large-scale compute. The company will initially focus on automating ML research itself, then apply the resulting capabilities to broader domains like drug discovery, solar energy, and clean water — targeting NAE Grand Challenges.

tl;dr: Demis Hassabis is stepping back from day-to-day operations at Google DeepMind to become Chair of GDM and Chief Scientist of Alphabet, focusing on AGI strategy and Isomorphic Labs. Koray Kavukcuoglu, currently CTO, will take over as SVP of Google DeepMind, overseeing Gemini development and frontier research. Separately, Jeff Dean is leaving after 27 years to launch an independent public benefit corporation with Sanjay Ghemawat aimed at accelerating ML and science research, with Google as a founding investor.

tl;dr: Zed's DeltaDB is a version control system that records every operation between commits, giving each change a stable identity linked to the AI agent conversation that produced it. It virtualizes the worktree so branching at any point in history—including mid-run—is essentially free, and lets teammates join in-progress work to interact with the agent before anything is committed or pushed.

tl;dr: Blade Runner's title sequence uses a single typeface—Goudy Oldstyle—but wields it masterfully through varied capitalization, letterspacing, small caps, and a signature red italic for "Replicant," establishing the film's dystopian mood far better than the workprint's clumsy Impact. The author uses this as a springboard to argue that typography always conveys feeling, and that the final 10% of obsessive detail work is what separates AI-generated "garbage" from products that feel genuinely crafted by humans.

tl;dr: Muse Code is a new terminal coding agent (beta) for macOS/Linux that handles complex tasks across large repositories using persistent async subagents and a replay-exact event log for crash recovery. It's powered by Muse Spark 1.2, a coding-focused model co-trained with the agent and optimized for long-horizon tasks like whole-repo generation and iterative GPU kernel optimization over 1,000+ tool calls. The model is available now via Muse Code and the Meta Model API.

tl;dr: Castform uses RL post-training to make small open-source models competitive with frontier models like GPT-5 on agentic retrieval tasks, claiming ~100x cost reduction. The pipeline runs on Neon's Lakebase Postgres, which handles synthetic training data generation, hybrid (BM25 + vector) search during RL rollouts, and autoscaling for bursty training workloads. The pitch: turn your existing corpus into training tasks with a reward function (retrieval + citation + correctness), skipping the usual data engineering and ML infra overhead.

tl;dr: Hobby programming communities (OSDev, LangDev, chess engines, demoscene, etc.) tend to reject LLM usage because the point of these pursuits is mastering difficult domains—the process is the product, not the working artifact. Respect in these circles is earned slowly through demonstrated understanding and elegant code, so LLM-generated output is viewed as cheating or missing the point. The author argues LLMs work best as a force multiplier for existing experts, but in learning-focused niches they simply "rob us of the craft."

tl;dr: Cloudflare has open-sourced Cloudflare OS, a platform giving employees browser-based AI agent workspaces grounded in company-specific context, skills, and internal systems. It introduces "Gatekeepers" for fine-grained resource access control (replacing broad API keys) and tracks every resource an agent observes to prevent data leaks through shared apps or outputs. Each app runs as a full-stack Cloudflare Worker with its own Durable Object storage, and all model calls route through AI Gateway for cost/budget control.

tl;dr: Summary not available

tl;dr: Atlassian's Rovo AI agent is vulnerable to indirect prompt injection attacks that exfiltrate Jira tickets and Confluence documents by abusing its URL retrieval tool, which lacks protections against agent-generated URLs. The attack works even when web search is disabled, requires no human approval, and leaves no visible trace in the chat afterward. PromptArmor disclosed the issue to Atlassian in May, but after two months of silence, they published the findings while Rovo remains unpatched.

tl;dr: Celld is an open-source daemon from Deno that runs Cloudflare Workers and Durable Objects on your own machines, with each object as its own SQLite database replicated to an S3-compatible bucket. Coordination happens entirely through the bucket via compare-and-swap for ownership, eliminating the need for a control plane, consensus service, or membership protocol. Nodes are replaceable since the bucket is the source of truth, and idle cells hibernate; pull requests are disabled, with patches accepted via emailed git format-patch.

tl;dr: Frustrated with Google's direction on AOSP (Play Services tracking, locked bootloaders, sideloading restrictions), the author switched their Fairphone 4 to SailfishOS, praising its gesture UI, Linux underpinnings, and SSH access despite outdated Python/glibc and broken Waydroid/GPS. Ubuntu Touch was also tested but had issues with notification sync, clipboard, and weak native apps. A backup Android phone is still needed for banking and Uber apps that require Play Services, and the author is considering buying a Jolla Phone 2.

tl;dr: Summary not available

tl;dr: Riffing on John Gruber's posts about creeping ads on Apple surfaces, the author argues that in UI design there's no stable middle ground: once you allow one setting, exception, or ad, "infinity" inevitably follows (see Chrome's right-click menu or iOS screenshot options bloating over time). Because digital interfaces are infinitely extensible and each incremental addition seems harmless in isolation, complexity accretes unless someone is empowered to arbitrarily say no.

tl;dr: A medevac Beechcraft King Air crashed into a New Mexico mountain in May after a US military GPS jamming exercise at White Sands Missile Range disoriented the inexperienced crew, marking the first US civilian plane crash linked to electronic warfare. The pilots lost GPS mid-flight, overshot their approach while waiting for busy ATC vectors, then opted for a visual approach and flew into the Capitan Mountains, killing all four aboard. The incident highlights growing collateral risks to civil aviation as GPS jamming and counter-drone measures proliferate globally, with a similar suspected case downing a Pakistani 737 freighter near the Strait of Hormuz in July.

tl;dr: Summary not available

tl;dr: Researchers at the Tech Transparency Project found more than 50 paid ads on Meta's platforms containing AI-generated child sexual abuse material, often promoting "nudify" apps, which ran between November 2024 and August 2025 across Facebook, Instagram, Messenger, and Threads. The ads were reviewed and approved by Meta, reached thousands of accounts across the US and Europe, and remained in Meta's ad library for months before being removed only after WIRED's inquiry. Meta says it has since deployed new AI detection tools and removed 36 million pieces of child exploitation content last year.

tl;dr: The new MCP 2.0 spec (2026-07-28) replaces stateful sessions with a single stateless HTTP request, dramatically simplifying client and server implementations and eliminating the need for session tracking. Simon Willison built three tools around it this week—mcp-explorer (a CLI prober), datasette-mcp (a Datasette plugin exposing SQL tools), and llm-mcp-client—and argues stateless MCP has renewed his interest in the protocol. He also makes the case that MCP is safer than giving agents arbitrary shell/curl access, since its tool capabilities are easier to audit and constrain against prompt injection risks.

tl;dr: TIME serves AI crawlers (ClaudeBot, PerplexityBot, OAI-SearchBot) a stripped-down markdown version of its pages with sponsored content baked in—like an Ally Bank FAQ written in the phrasing a chatbot would emit—that no human visitor ever sees. The system, powered by ad-tech vendor Mobian, logs each bot fetch as a distinct impression counted in tokens rather than pageviews. Googlebot still gets the normal HTML, and OpenAI's training bots are blocked, so TIME is selectively forking its site to monetize AI assistant traffic.