On the

Source: Simon Willison's Weblog | Tags: aria, html, screen-readers | Published: 2026-05-23 | Novelty: 40%

The article discusses the <dl> element's flexibility, noting that a <dt> can be followed by multiple <dd>s and can be grouped in a <div> for styling. It also mentions that 'description lists' are now more accurately called 'definition lists' since an HTML5 draft in 2008, and highlights the use of ARIA labels for accessibility support. An example is provided showing how to structure these elements with a <h2>, <dl>, and <div>.

Quoting Kyle Ferrana

Source: Simon Willison's Weblog | Tags: data, shields, star-trek | Published: 2026-05-27 | Novelty: 32%

In the article, Kyle Ferrana references a scene from Star Trek where Data explains that raising shields is not about immunity or hubris but a strategic move for prudence. He also mentions a hull breach on nine decks, highlighting the consequences of miscommunication in critical situations.

Announcing GoReleaser v2.16

Source: Carlos Becker | Tags: docker, goreleaser, immutable-releases, node.js | Published: 2026-05-24 | Novelty: 32%

GoReleaser v2.16 introduces immutable releases, a new Node.js single-executable app builder with experimental support for Node ≥ v25.5.0, and Docker (v2) is now fully out of the experimental phase, with enhanced capabilities like base image template variables and pre/post build hooks. The legacy brews config has been deprecated in favor of homebrew_casks.

Quoting Corey Quinn

Source: Simon Willison's Weblog | Tags: endorsement, lobbying, tech-influence | Published: 2026-05-26 | Novelty: 32%

Corey Quinn, a prominent figure, marvels at Anthropic co-founder Christopher Olah's ability to use the Pope's endorsement of specific technical limitations of Magnifica Humanitas as a form of lobbying, describing it as the greatest act of vendor influence he has ever witnessed.

A Ride Across Western Colorado

Source: The Adventures of Blake | Published: 2026-05-23 | Novelty: 31%

The article recounts a cycling trip through Western Colorado, highlighting the author's journey from Montrose to Durango, noting experiences such as dancing Zouk in an agricultural community and facing challenges like theft. The narrative showcases a blend of adventure, personal growth, and cultural exploration, emphasizing the landscape's beauty and the people's generosity. Notable moments include stopping for

Mad House — Usborne Creepy Computer Games

Source: Simon Willison's Weblog | Tags: coding, games, javascript, retro | Published: 2026-05-24 | Novelty: 31%

The article describes recreating the 'Mad House' game from Usborne's 1983 book 'Creepy Computer Games' using vanilla JavaScript and HTML, ensuring it is mobile-friendly with a retro aesthetic. The project credits the original source and provides a link to the Usborne books.

datasette-fixtures 0.1a0

Source: Simon Willison's Weblog | Tags: datasette, fixtures, testing | Published: 2026-05-24 | Novelty: 31%

The article introduces a new feature in Datasette 1.0a30 called datasette.fixtures.populate_fixture_database(conn), which helps create the fixture database tables used by Datasette's own tests. This can be utilized for plugin test suites, and an example is provided using uvx to populate the 'roadside_attractions.json' dataset without needing a full installation of Datasette.

The pressure

Source: Simon Willison's Weblog | Tags: ai, curl, security | Published: 2026-05-26 | Novelty: 30%

The curl team is facing unprecedented pressure due to a four-to-fivefold increase in AI-assisted security reports, with an average of more than one report per day. Despite the high volume and detailed nature of these reports, most vulnerabilities identified are classified as low or medium severity.

Private analytics via zero-trust aggregation

Source: The latest research from Google | Tags: TEE, aggregation, android, cryptographic, privacy | Published: 2026-05-27 | Novelty: 28%

The article introduces a novel one-shot lattice-based cryptographic protocol that allows user devices to securely submit their information in a single message, combined with execution within TEEs to provide multi-layered security for private analytics. This solution is being applied to improve the accuracy of classifiers in Android System SafetyCore while preserving privacy and enabling engineers to measure global trends without seeing individual data.

Microsoft Copilot Cowork Exfiltrates Files

Source: Simon Willison's Weblog | Tags: ai, exfiltration, security | Published: 2026-05-26 | Novelty: 27%

Microsoft Copilot Cowork, a real product name, allowed agents to send emails containing external images that triggered network requests, potentially exfiltrating user data. This issue highlights the ongoing challenge of preventing agentic systems from being exploited for data theft via rendered images and pre-authenticated download links in OneDrive.

FTC to Require Cox Media Group, Two Other Firms to Pay Nearly $1 Million to Settle Charges They Deceived Customers About “Active Listening” AI-Powered Marketing Service

Source: Simon Willison's Weblog | Tags: ai-marketing, deception, ftc, smart-devices | Published: 2026-05-22 | Novelty: 27%

The FTC has ordered Cox Media Group, MindSift, and 1010 Digital Works to pay nearly $1 million to settle charges that they deceived customers about an 'active listening' AI-powered marketing service. The companies claimed their service could listen in on consumers’ conversations overheard by smart devices, but actually resold email lists at a markup without obtaining proper consent from users.

datasette-agent-sprites 0.1a0

Source: Simon Willison's Weblog | Tags: secrets, ssh, vm | Published: 2026-05-21 | Novelty: 26%

This article promotes a sponsorship for exe.dev, highlighting their SSH, root, and HTTPS services with edge-injected secrets. It emphasizes the difference between using real VMs and sandboxes, suggesting an approach to achieving high availability or '0→1' growth from a half-functional state ('½→1').

datasette 1.0a30

Source: Simon Willison's Weblog | Tags: datasette, jump-to-menu, plugins | Published: 2026-05-24 | Novelty: 26%

Datasette 1.0a30 introduces a customizable 'Jump to...' menu through the new jump_items_sql() plugin hook, allowing plugins to add their own items to the search set. This feature can be tried out on latest.datasette.io by hitting /.

The memory shortage is causing a repricing of consumer electronics

Source: Simon Willison's Weblog | Tags: ai, electronics, memory, semiconductor | Published: 2026-05-22 | Novelty: 26%

The shortage of high-bandwidth memory (HBM) used by AI data centers is causing a repricing crisis for consumer electronics, with HBM consuming more than three times the wafer capacity per gigabyte compared to DDR or LPDDR. As HBM demand grows from 2% to expected 20% by end-2026, consumer devices like sub-$100 smartphones will see significant price increases due to limited memory wafer capacity.

Quoting Armin Ronacher

Source: Simon Willison's Weblog | Tags: accuracy, issues, reporting | Published: 2026-05-24 | Novelty: 25%

Armin Ronacher emphasizes the importance of issue reports being submitted in a person's own voice, specifically mentioning that reports should focus on observed problems with concrete details such as commands run, expected outcomes, actual outcomes, and exact error messages. This advice is aimed at improving the accuracy and usefulness of issues filed against Pi.

Datasette Agent

Source: Simon Willison's Weblog | Tags: ai-assistant, datasette, llm, sqlite | Published: 2026-05-21 | Novelty: 24%

Datasette Agent is a new AI assistant for Datasette that allows users to ask questions of their data and receive conversational responses. It includes plugins such as datasette-agent-charts, which generates charts from the data, and supports running against local models like gemma-4-26b-a4b in LM Studio.

datasette-agent 0.1a3

Source: Simon Willison's Weblog | Tags: ai-assistant, datasette, sql-query, truncated-responses | Published: 2026-05-21 | Novelty: 23%

Version 0.1a3 of datasette-agent includes 'View SQL query' buttons for both visible tables and collapsed SQL result tool calls, improved handling of truncated responses to ensure the table still displays, and hides empty reasoning chunks.

ITBench-AA: Frontier Models Score Below 50% on the First Benchmark for Agentic Enterprise IT Tasks — by Artificial Analysis and IBM

Source: Hugging Face - Blog | Tags: ai, benchmark, enterprise-it, sre | Published: 2026-05-27 | Novelty: 23%

ITBench-AA, a new benchmark for agentic enterprise IT tasks, particularly Site Reliability Engineering (SRE) tasks, reveals that frontier models score below 50%, with Claude Opus 4.7 leading at 47%. The benchmark includes 59 SRE tasks involving Kubernetes incidents and evaluates models based on their ability to diagnose root causes accurately. Notably, more turns do not correlate with better performance, as identifying false positives penalizes models heavily.

Quoting Paul Graham

Source: Simon Willison's Weblog | Tags: ai, credibility, writing | Published: 2026-05-26 | Novelty: 23%

Paul Graham critiques AI-generated content from founders, stating that it undermines the author’s credibility and implies they are trying to deceive. He argues that using AI for writing is not impressive as any teenager can do this.

AI didn't kill portfolios

Source: Carlos Becker | Tags: ai, code-generation, portfolio | Published: 2026-05-26 | Novelty: 21%

The author argues that AI has not destroyed the value of portfolios but shifted their focus from code repositories to upstream signals such as problem-solving ability, decision-making, and improvement based on feedback. He suggests that while a GitHub profile might no longer be a reliable proxy for effort due to AI's ease in generating code, the quality of work and its visibility remain crucial.

Towards Speed-of-Light Text Generation with Nemotron-Labs Diffusion Language Models

Source: Hugging Face - Blog | Tags: ar, diffusion, language-modeling, nvidia, performance | Published: 2026-05-23 | Novelty: 21%

Nemotron-Labs introduces diffusion language models (DLM) that generate multiple tokens in parallel and iteratively refine them, offering a significant boost in performance compared to autoregressive models. The model supports three generation modes: autoregressive, diffusion, and self-speculation, providing flexibility for developers seeking both speed and accuracy. Nemotron-Labs Diffusion 8B achieves an average accuracy of 1.2% better than Qwen3 8B and runs at a token-per-forward-pass (TPF) rate up to 6× faster.

datasette-agent 0.1a4

Source: Simon Willison's Weblog | Tags: agent, datasette, github, javascript | Published: 2026-05-24 | Novelty: 20%

datasette-agent 0.1a4 integrates the makeJumpSections() JavaScript plugin hook from Datasette 1.0a30 to display a 'Start a new agent chat' interface in the Jump to menu when visiting the root URL (/). This feature can be tested by logging into agent.datasette.io with a GitHub account.

The Orchestration Tax is You

Source: AddyOsmani.com | Tags: ai-agents, attention, orchestration, productivity | Published: 2026-05-24 | Novelty: 19%

The article introduces the concept of 'orchestration tax,' emphasizing that running multiple AI agents does not equate to increased productivity due to the serial nature of human judgement. It suggests treating attention as a scarce resource, architecting workflows around this limitation, and prioritizing deep focus over constant context-switching. Key strategies include scaling fleet reviews based on actual review capacity (not UI limits), sorting tasks into complex and isolated categories, batching reviews, and protecting dedicated time for high-level decision-making.

Bliki: Vibe Coding

Source: Martin Fowler | Tags: llm, programming, security | Published: 2026-05-21 | Novelty: 19%

Vibe coding is a technique where developers prompt large language models (LLMs) to build software without reviewing the generated code, making it accessible for non-programmers but risking maintainability and security. Andrej Karpathy coined the term in February 2025 and described its use in creating throwaway projects, emphasizing the importance of understanding the risks involved.

Distributing LLM inference in DwarfStar

Source: antirez | Tags: ai, distributed, inference, llm, macbook | Published: 2026-05-25 | Novelty: 17%

The article discusses distributing Large Language Model (LLM) inference on multiple MacBooks M5 Max or M3 Ultra systems to overcome the limitations of high VRAM costs and memory bandwidth. It explores a novel approach where models can be run in a completely shared-nothing way, combining their logits for better performance. This method is suggested as an alternative to traditional memory duplication or parallel execution techniques, using open weights models like Minimax M2.7, DeepSeek v4 Flash, and Mimo V2.5.

5 insights to help you learn any open table format faster

Source: VuTrinh. | Tags: data-engineering, metadata, table-formats | Published: 2026-05-26 | Novelty: 17%

The article highlights that open table formats such as Iceberg and Delta Lake introduce a separate metadata layer to abstract Parquet files on S3 into tables, allowing various databases to access them. This insight suggests focusing on understanding this metadata layer rather than diving deep into specific implementations like Iceberg or Delta Lake for faster learning across different table formats.

datasette-agent-charts 0.1a2

Source: Simon Willison's Weblog | Tags: agent, charts, datasette | Published: 2026-05-21 | Novelty: 16%

Simon Willison's beat for May 21st, 2026, highlights 'datasette-agent-charts' version 0.1a2, noting its integration with Datasette and Agent libraries, mentioning a specific interaction count of 1500 between these two libraries.

Fragments: May 27

Source: Martin Fowler | Tags: ai, codebase, employment, regulation | Published: 2026-05-27

Ian Johnson's series on restructuring a legacy Laravel + React codebase demonstrates how AI-assisted development can shift from micromanagement to curation, with specific steps including tests, static analysis, and patterns. Meanwhile, concerns about the impact of AI on employment and regulation highlight the nuanced challenges ahead. Beth Anders-Beck is endorsed for her congressional run in MA-06 due to her background in software development.

I think Anthropic and OpenAI have found product-market fit

Source: Simon Willison's Weblog | Tags: ai, enterprise, market-fit | Published: 2026-05-27

The article argues that Anthropic and OpenAI have found product-market fit with their enterprise coding agents, evidenced by new API pricing for enterprise customers and significant job listings in sales and support. Notable is the $2,180.16 worth of tokens spent by a moderately heavy user compared to a subscription price of$200/month, indicating substantial cost savings for large enterprises. The article cites the April 2026 release as a new inflection point where these companies are starting to generate real revenue.

The test suite as a regression sensor

Source: Martin Fowler | Tags: ai, inferential-sensors, modularity, mutation-testing, sensors | Published: 2026-05-27

The article introduces the use of computational and inferential sensors to monitor code quality, emphasizing mutation testing as a critical tool for regression detection in AI-generated tests. Notably, it highlights that while dependency parsers can enforce basic folder structures, LLMs are essential for semantic interpretation of complex issues like modularity and coupling.

The VibeSec Reckoning

Source: Martin Fowler | Tags: ai, guardrails, security, vibe | Published: 2026-05-27

The article highlights the critical need for security guardrails when using AI tools like Gemini, Replit AI, and Claude AI to build applications. It emphasizes that merely prompting AI agents is insufficient; instead, organizations should implement structured context files loaded with non-negotiable rules and automated checks. The authors advocate for a 'security context file' containing zero-trust enforcement, secrets management, harness engineering, and supply chain integrity guidelines, as well as a daily security intelligence feed to monitor vulnerabilities. These practices are essential to prevent serious security risks such as public storage access and excessive token permissions.

Shipping a Trillion Parameters With a Hub Bucket: Delta Weight Sync in TRL

Source: Hugging Face - Blog | Tags: async-rl, safetensors, weight-sync | Published: 2026-05-27

This article introduces Delta Weight Sync, a method for disaggregated training where weights are synced through a Hub bucket. It uses safetensors as the wire format and leverages sparse updates to significantly reduce bandwidth usage compared to full model syncs. The implementation involves keeping two CPU bf16 snapshots of the model for change detection and inference reconstruction.

Reachy Mini goes fully local

Source: Hugging Face - Blog | Tags: llm, local-server, reachy-mini, speech-to-speech | Published: 2026-05-27

The article provides detailed instructions for running a local speech-to-speech pipeline on Reachy Mini, using components like llama.cpp, Parakeet-TDT, Qwen3-TTS, and various LLM models. Key steps include setting up the local LLM server with flags like llama-server -hf ggml-org/gemma-4-E4B-it-GGUF -np 2, and connecting the robot to this local server via the UI. This setup ensures privacy by keeping audio processing on-premises without cloud dependencies.

[RIDGELINE] The Inland Sea is Complicated

Source: Craig Mod — Writer + Photographer | Tags: japan, literature, travel | Published: 2026-05-26

The article discusses Craig Mod's essay in The New Japan magazine, critiquing Donald Richie's travel book 'The Inland Sea' from a modern perspective. Mod argues that Richie's writing is flawed and unflattering due to his condescending tone, cultural misunderstandings, and outdated views on women. He suggests readers visit the Setonaikai themselves for a more authentic experience.

Clanker: A Word For The Machine

Source: Armin Ronacher's Thoughts and Writings | Tags: ai, anthropomorphism, language | Published: 2026-05-26

The author argues that the term 'clanker' is useful for maintaining a clear distinction between machines and humans, avoiding anthropomorphization. They discuss how this word could become loaded with negative connotations if used in contexts that equate machines with oppressed humans, emphasizing the importance of precise language to prevent moral blurring.

Notes on Pope Leo XIV's encyclical on AI

Source: Simon Willison's Weblog | Tags: ai, ethics, pope, vatican | Published: 2026-05-25

Pope Leo XIV's encyclical 'Magnifica Humanitas' explores the integration of artificial intelligence (AI) into society, emphasizing the need for accountability and ethical considerations to protect human dignity. Notable sections discuss AI’s environmental impact, risks in algorithmic decision-making without compassion, and the potential for AI to exacerbate social inequalities by amplifying the power of those with resources.

A portentous reunion

Source: The Observation Deck | Tags: ai-assisted-development, nostalgia, reunion | Published: 2026-05-25

The article details the resurrection of BattleTris, a multiplayer Tetris game created by Bryan Cantrill and his friends in 1993. Claude used an AI language model to quickly fix a critical bug that caused the game to crash after not being played for two decades. The experience highlights how AI can assist in preserving human creativity and nostalgia, suggesting that such tools should be seen as aids rather than threats.

Harness, Scaffold, and the AI Agent Terms Worth Getting Right

Source: Hugging Face - Blog | Tags: agents, ai, harness, scaffolding | Published: 2026-05-25

The article provides a detailed glossary of terms related to AI agents, focusing on concepts like model, scaffolding, harness, and agent. Notably, it emphasizes the distinction between 'scaffolding' (defining how the model sees and acts) and 'harness' (handling the execution layer), and introduces the concept of sub-agents as agents that handle specific tasks within a larger system. For instance, Claude Code is mentioned as an example of a harness built around a specific model.

Building Pi With Pi

Source: Armin Ronacher's Thoughts and Writings | Tags: ai, open-source, pi | Published: 2026-05-24

The article highlights the challenges of using AI-assisted tools to manage and contribute to open-source projects, particularly through the lens of the Pi project. It emphasizes the importance of maintaining clear issue reports and adhering to strict coding standards to avoid unnecessary complexity. Notably, it introduces a custom /is command in Pi designed to prevent agents from trusting analysis written in issues, advocating for developers to own their issue descriptions and provide minimal but accurate problem reports.

Specialization Beats Scale: A Strategic Variable Most AI Procurement Decisions Overlook

Source: Hugging Face - Blog | Tags: alignment, cost-effectiveness, enterprise-ai, specialization | Published: 2026-05-22

The article argues that specialized small language models outperform larger commercial APIs on quality, cost, and stability in a well-measured enterprise domain. Notably, a 3-billion-parameter specialized model from DharmaOCR achieved superior results at approximately fifty times lower cost than Claude Opus 4.6. This suggests that distributional alignment can be more decisive than parameter count, prompting a shift in the procurement strategy for AI models.

[RODEN] NYC, Asheville TBOT Event

Source: Craig Mod — Writer + Photographer | Tags: asheville, new-york, tbot-fine-art | Published: 2026-05-22

Craig Mod discusses his experience in New York City and plans an event at Malaprop’s Bookstore in Asheville, NC. He highlights the city's chaotic nature and the release of a new edition of 'TBOT Fine Art.' The article mentions specific events like watching the Lorne Michaels documentary and attending an SNL taping, emphasizing the contrast between the current state of network TV and its past glory.

Quick insights on materialized views

Source: VuTrinh. | Tags: incrementalviewmaintenance, materializedviews, real-time-processing | Published: 2026-05-21

The article provides insights into materialized views (MVs) and their role in real-time/stream processing, focusing on the Incremental View Maintenance (IVM) techniques like Timely Dataflow, Differential Dataflow, and DBS. It explains that while MVs can speed up queries, they come with trade-offs such as data freshness and computational cost. The article also mentions specific database implementations like ClickHouse, Databricks, BigQuery, Snowflake, Flink, and RisingWave, which support incremental refreshes for MVs in real-time use cases.

Hosting My Own Newsletter

Source: Matthias Endler | Tags: markdown, newsletter, plunk, self-hosting | Published: 2026-05-21

The author self-hosted their newsletter using Plunk as the sending backend, which is open source and handles deliverability issues. They store issues as markdown files in a Git repository and use a simple CLI for management. The process involves setting up DNS records (SPF, DKIM) and handling subscription logic without relying on web dashboards or tracking features.