Timeouts in zio

Source: Lukáš Lalinský | Tags: asyncio, cancelation, python, timeout, zio | Published: 2026-06-04 | Novelty: 34%

The article discusses how zio implemented timeout support by using a context manager approach similar to Python's asyncio.timeout. This method allows for generic timeout handling without requiring changes to existing libraries or code awareness of timeouts, leveraging the built-in cancellation mechanism. The implementation is exposed via zio.AutoCancel, which can be nested and distinguished through its check function.

AI enthusiasts are in a race against time, AI skeptics are in a race against entropy

Source: Simon Willison's Weblog | Tags: ai, enthusiasts, feedback-loop, skeptics | Published: 2026-06-04 | Novelty: 28%

Charity Majors highlights the contrasting challenges faced by AI enthusiasts and skeptics, emphasizing that both groups need to collaborate effectively. The key issue is the lack of a natural feedback loop between them, which poses an organizational design challenge.

"Sixteenth of a year", a 1.8 KiB art piece

Source: Evan Hahn (dot com) | Tags: art, code-golf, minification | Published: 2026-06-03 | Novelty: 27%

Evan Hahn created an art piece for the sixteenth issue of Taper online magazine, splitting time into sixteenths and using Go code under 2048 bytes for minification. The article mentions other notable entries such as '[sic]', ‘desperate measures from a dying regime’, and '(un)done'.

datasette-agent-edit 0.1a0

Source: Simon Willison's Weblog | Tags: datasette, plugin, text-editing | Published: 2026-06-07 | Novelty: 26%

The article introduces 'datasette-agent-edit' 0.1a0, a base plugin for Datasette Agent that provides core text editing tools (view, str_replace, insert) to facilitate collaborative Markdown editing, updating large SQL queries, and editing SVG files. This tool aims to streamline the development of various plugins by providing a consistent framework.

Quoting Emanuel Maiberg, 404 Media

Source: Simon Willison's Weblog | Tags: ai-ethics, google | Published: 2026-06-04 | Novelty: 26%

The article highlights a change in Google's statement regarding AI oversight, as reported by Emanuel Maiberg of 404 Media. The updated version omits the phrase 'critical that we maintain humans in the loop,' suggesting a shift in stance or internal dialogue about AI governance within the company.

Quoting Andreas Kling

Source: Simon Willison's Weblog | Tags: accountability, development, responsibility | Published: 2026-06-05 | Novelty: 25%

Andreas Kling announced that Ladybird will no longer accept public pull requests, shifting the responsibility for code to its author and introducing a new standard of accountability. He emphasizes that only those who decide changes should be responsible for their outcomes, marking a significant shift in development practices.

The Neo is a love letter to personal computing

Source: §kuthus | Tags: apple, laptops, mac | Published: 2026-06-05 | Novelty: 25%

The author praises the Neo laptop by Apple as the best low-cost laptop ever made, despite their previous preference for PCs. They highlight that Macs are now more affordable and differentiate themselves from other devices by offering a highly customizable experience while still maintaining a curated feel, unlike other personal computers that are increasingly trying to mimic iPad experiences.

Communities of Not

Source: Armin Ronacher's Thoughts and Writings | Tags: abstinence, communities, opposition | Published: 2026-06-06 | Novelty: 25%

The article discusses the phenomenon of communities that form around abstaining from something often becoming defined by their opposition, leading to critical but sometimes aggressive policing and shaming. The author emphasizes the importance of de-escalating conflicts and being open to new perspectives rather than focusing solely on criticism.

The next chapter in flood resilience: Open sourcing Google’s hydrology framework

Source: The latest research from Google | Tags: ai, flood-forecasting, hydrology, open-source | Published: 2026-06-03 | Novelty: 23%

Google has open-sourced its advanced AI-based hydrology model for flood forecasting, including an upgraded version that extends reliable predictive horizons by six days in gauged basins and one day in ungauged ones. The Python package uses PyTorch and is based on LSTM networks, trained with the Caravan dataset to predict daily river flow rates globally. This open-source framework allows hydrological scientists and operational forecasters to integrate local data and refine models, enabling more accurate and localized flood warnings worldwide.

[

A new era for software testing ]( http://antirez.com/news/168 )

Source: <antirez> | Tags: llm, qa, software, testing | Published: 2026-06-07 | Novelty: 23%

Antirez proposes a novel approach to software quality assurance using large language models (LLMs), specifically markdown files where LLMs are asked to perform manual tests on new software releases. For example, in DwarfStar, the model checks distributed inference across multiple devices and verifies there is no speed regression without being provided with baseline metrics. This method can enhance testing by covering more cases than traditional QA passes, particularly for integration tests and user experience critiques.

Uber Caps Usage of AI Tools Like Claude Code to Manage Costs

Source: Simon Willison's Weblog | Tags: ai-tools, cost-management, uber | Published: 2026-06-03 | Novelty: 23%

Uber has capped employee usage of AI tools such as Claude Code at $1,500 per month to manage costs. This limit represents about 11% of the median compensation package for Uber software engineers in the USA, suggesting a real dollar value for these tools.

micropython-wasm 0.1a2

Source: Simon Willison's Weblog | Tags: micropython, python, webassembly | Published: 2026-06-06 | Novelty: 23%

This beat by Simon Willison highlights the release of micropython-wasm 0.1a2, an early access version for WebAssembly, emphasizing its potential in sandboxed environments and Python development on the web. The post discusses the library's growing ecosystem with contributions from a community of developers.

The Open Source Community is backing OpenEnv for Agentic RL

Source: Hugging Face - Blog | Tags: huggingface, interoperability, openenv, reinforcement-learning | Published: 2026-06-08 | Novelty: 22%

OpenEnv is becoming an interoperability layer for reinforcement learning environments that standardizes how environments are published, deployed, and consumed by agents. The project now has a committee including Meta-PyTorch, Reflection, Unsloth, Modal, Nvidia, and more, with plans to integrate external rewards and tasksets via Hugging Face datasets. OpenEnv is community-centric and currently open for contributions.

The circus freaks of open source

Source: Drew DeVault's blog | Tags: mental-health, open-source, operating-systems, schizophrenia | Published: 2026-06-05 | Novelty: 22%

The article discusses two case studies: Terry A. Davis and his self-hosted operating system TempleOS, which is a unique masterwork due to its creator's severe untreated schizophrenia; and Kent Overstreet and his bcachefs project, which he devoted 15 years to before being marginalized by the Linux community. Both cases highlight how mental health crises can disproportionately affect neurodivergent individuals and queer people, often leading to public humiliation and exacerbating their struggles.

Microsoft's new MAI models

Source: Simon Willison's Weblog | Tags: ai, llm, microsoft, model-training | Published: 2026-06-02 | Novelty: 20%

Microsoft introduced two new LLMs: MAI-Thinking-1 with 35B active parameters and MAI-Code-1-Flash with 5B active parameters, both purpose-built for specific applications. The models were trained on a proprietary crawl of the public web and additional data sources, aiming to avoid licensing issues present in other major LLMs.

OpenAI Help: Lockdown Mode

Source: Simon Willison's Weblog | Tags: data-exfiltration, lockdown-mode, openai, prompt-injection, security | Published: 2026-06-05 | Novelty: 19%

Lockdown Mode, recently rolled out by OpenAI to personal and some business accounts, aims to mitigate data exfiltration risks from prompt injection attacks by limiting outbound network requests. This measure is part of addressing the 'Lethal Trifecta' threat where an AI system could have access to private data, exposure to untrusted content, and a method to steal it; Lockdown Mode specifically targets the exfiltration vector.

California Brown Pelican

Source: Simon Willison's Weblog | Tags: brown-pelican, california, wildlife-conservation | Published: 2026-06-02 | Novelty: 17%

The article discusses the California Brown Pelican, likely focusing on its conservation status and environmental factors affecting its population. No specific new ideas or unique points of view are mentioned; however, it references a beat by Simon Willison posted on June 2, 2026, indicating recent updates in wildlife research or management.

Restoration of the Demon (Redux)

Source: New Cartographies | Tags: art, media, perception, reality, supernatural | Published: 2026-06-07

The article explores the contemporary phenomenon of increased openness to supernatural experiences and perceptions, illustrated through examples such as Tucker Carlson's alleged demon encounter. It draws parallels between modern media's de-materialization and Baudrillard’s concept of hyperreality, suggesting that the 'demon' is a perfect manifestation of this hyperreal world. The restoration of Joshua Reynolds' painting 'The Death of Cardinal Beaufort,' which depicts a demon tormenting the cardinal, symbolizes the re-emergence of these perceptions in contemporary society.

Adapting to AI: Adapting Organizations

Source: Colin Breck | Tags: ai, feedback-loops, organizational-adaptation, systems-thinking | Published: 2026-06-07

The article argues that organizations need to focus on systems thinking and tight feedback loops to fully leverage AI, emphasizing the importance of exceptional leadership, conceptual integrity, and social aspects of engineering. It highlights the shift in software roles from writing code to managing projects, enhancing communication, and working closely with customers, while noting that AI will amplify existing organizational dysfunctions if these challenges are not addressed.

Loop Engineering

Source: AddyOsmani.com | Tags: ai, automation, engineering, prompting, verification | Published: 2026-06-07

The article introduces loop engineering as a new approach where you design systems to prompt coding agents instead of manually prompting them. It details five key components: automations, worktrees, skills, plugins/connectors, and sub-agents. The author emphasizes that while loops can automate many tasks, they do not replace the need for human verification and understanding. Specific tools like Codex app and Claude Code are mentioned as implementing these ideas.

Running Python code in a sandbox with MicroPython and WASM

Source: Simon Willison's Weblog | Tags: micropython, python, sandbox, webassembly | Published: 2026-06-06

The author has developed an alpha package called micropython-wasm for running Python code in a sandbox using MicroPython and WebAssembly. This implementation allows variables and functions to remain resident in memory across multiple code execution calls, supports host function interaction, and includes memory and CPU limits. The first version is available on PyPI with detailed instructions for trying it out.

Ergonomic overrides for Nixpkgs

Source: Haskell for all | Tags: ergonomics, nix, overlays, overrides | Published: 2026-06-06

The article introduces override-utils, a package that enhances Nixpkgs ergonomics by allowing overrides and overlays to be generated from composable setters and safe operations on those setters. This approach is designed to be unobtrusive, incremental, and respectful of existing Nixpkgs design decisions, differing from dream2nix in its approach and scope.

Unlocking dependable responses with Gemini Enterprise Agent Platform’s Agentic RAG

Source: The latest research from Google | Tags: ai, dependability, multiagent, queryplanning, ragn | Published: 2026-06-05

The article introduces Agentic RAG, a multi-agent framework that outperforms standard retrieval-augmented generation (RAG) systems by up to 34% on factuality datasets. It includes a Sufficient Context Agent that ensures the system only generates responses after thoroughly examining retrieved snippets and creating intermediate drafts, which improves accuracy and dependability in complex queries, as demonstrated through experiments with Gemini Enterprise Agent Platform. For example, it correctly answered a multi-hop question about television season finales from 824 queries with a 90.1% success rate even when dealing with multiple unrelated data sources.

The Intent Debt

Source: AddyOsmani.com | Tags: ai-agents, intent-debt, software-engineering | Published: 2026-06-05

The article introduces the concept of 'intent debt,' which refers to the absence or erosion of externalized rationale, goals, and constraints that explain why a system is designed the way it is. The author argues that unlike technical and cognitive debts, intent debt cannot be addressed by AI agents, as these agents can only infer reasons rather than generate them. The article emphasizes the importance of externalizing this 'why' through detailed specifications, ADRs (decision records), and learnings files to maintain a system's integrity.

15 Moments of Zen in Southern Utah

Source: The Adventures of Blake | Tags: bicycle-travel, hitchhiking, solitude, utah | Published: 2026-06-04

The article describes a 5-day bicycle journey through southern Utah with multiple hitchhiking experiences and encounters, including meeting Blake Boles (interviewed for 'Dirtbag Rich') and spending the night in various locations. Notable moments include swimming in the Escalante River, reflecting on solitude and society at Kiva Koffeehouse, and visiting Zion National Park where the author's relationship to the US is changing as they prioritize European travel.

Towards passive heart health monitoring via smartphone camera

Source: The latest research from Google | Tags: accuracy, heart-health, ppg, rppg, smartphone | Published: 2026-06-04

PHRM introduces a passive heart rate monitoring system for everyday smartphone use with high accuracy across all skin tones, achieving mean absolute percentage errors (MAPE) < 10% compared to electrocardiogram-derived ground truth. The study includes over 350,000 video clips from nearly 700 diverse participants and leverages Kalman filtering and confidence scores for improved daily resting heart rate (RHR) estimation with mean absolute errors (MAE) < 5 bpm compared to wearables.

Nemotron 3.5 Content Safety: Customizable Multimodal Safety for Global Enterprise AI

Source: Hugging Face - Blog | Tags: ai, content-safety, custom-policy, multimodal, nlu | Published: 2026-06-04

Nemotron 3.5 Content Safety introduces unified multimodal evaluation and custom policy enforcement, enabling a single model to handle text, images, and domain-specific safety policies across 140 languages. This model uses Qwen 80B for generating concise reasoning traces in THINK mode, which is crucial for auditable decisions while maintaining low latency.

EVA-Bench Data 2.0: 3 Domains, 121 Tools, 213 Scenarios

Source: Hugging Face - Blog | Tags: eva-bench, multilingual-support, synthetic-data, voice-agent-evaluation | Published: 2026-06-04

EVA-Bench Data 2.0 expands to three enterprise domains (Airline CSM, Enterprise ITSM, and Healthcare HRSD), doubling the scenario coverage to 213 across 121 tools. The dataset is generated using SyGra, a graph-based synthetic data generation pipeline, ensuring joint consistency of user goals, initial databases, and ground truths. This release includes comprehensive validation steps with three frontier models and supports multilingual deployment in multiple languages, providing accurate cross-linguistic evaluations.

Designing the hf CLI as an agent-optimized way to work with the Hub

Source: Hugging Face - Blog | Tags: benchmark, cli, coding-agents, tokens | Published: 2026-06-04

The article highlights a benchmark of the Hugging Face CLI against curl and Python SDK for coding agents, showing that the CLI uses far fewer tokens for complex tasks. It introduces an 'hf-cli skill' feature that reduces the number of tool calls by providing a compact command reference context for agents like Claude Code (Sonnet 4.6) and Codex (GPT-5.5), resulting in about 30% fewer commands per task. The CLI outperforms both tools, especially on multi-step tasks, with token usage up to 6 times less than SDK or curl.

Direct Preference Optimization Beyond Chatbots

Source: Hugging Face - Blog | Tags: dpo, failure-mode, llm, ocr, specialization | Published: 2026-06-03

The DharmaOCR pipeline uses a Direct Preference Optimization (DPO) stage after standard specialized fine-tuning (SFT) to reduce text degeneration, a specific failure mode where models produce repetitive outputs. This approach works across five OCR model families and reduces degeneration by an average of 59.4%, confirming that categorically distinct failures can be used as training signals even in different domains. The key is the ability to score these failures reliably without human annotation and having enough data to generate meaningful preference pairs.

How Other Link Checkers Do Recursion

Source: Matthias Endler | Tags: deduplication, link-checkers, recursion, termination-detection | Published: 2026-06-03

The article highlights that implementing recursion in link checkers requires a worklist plus a visited set, and termination detection can be distributed via sync.WaitGroup (muffet), joinable-queue counter (LinkChecker), queue.onIdle() (linkinator), or queue-drain events (broken-link-checker). Notably, dedup should occur at enqueue, not after checking. The Rust implementation in 2026 will use a HostPool and WaitGroup to address these challenges.

Adding MCP Tools to Reachy Mini

Source: Hugging Face - Blog | Tags: mcp, reachymini, tools, weather, websearch | Published: 2026-06-03

The article introduces the integration of MCP tools via Hugging Face Spaces into Reachy Mini, allowing for capabilities such as web search and weather checking without downloading code locally. It specifies that the installation command is 'reachy-mini-conversation-app tool-spaces add <owner/space-name>', with examples using 'pollen-robotics/reachy-mini-search-tool' and 'pollen-robotics/reachy-mini-weather-tool'. Profiles control which tools are active, and prompts guide how to use these tools in conversation.

AI enthusiasts are in a race against time, AI skeptics are in a race against entropy

Source: charity.wtf | Tags: ai, engineering, skepticism, transparency | Published: 2026-06-02

The article discusses the divide between AI enthusiasts and skeptics in engineering teams, emphasizing the need for a balanced approach. It highlights Fin's success in tripling their output through disciplined use of AI while maintaining high code quality. The author advocates for transparency and shared accountability to navigate this complex landscape.