simonwillison.net
31 articles · April 7, 2026 – May 6, 2026
Anthropic's Code w/ Claude 2026 Signals AI Coding Shifts to Real-World Implementation
Anthropic hosts Code w/ Claude 2026, betting on AI coding tools. This marks LLM firms shifting from parameter wars to dev ecosystems, with coding as t
Veterans Skip Reviews: Vibe Coding & Agentic Engineering Dangerously Converge
Simon Willison skips line-by-line AI code reviews in production. As vibe coding & agentic engineering converge, AI tools mask hidden quality risks.
Stockholm AI Cafe's 120 Stoveless Eggs: Agents Lack More Than Common Sense
Andon Labs' AI Mona ran a Stockholm cafe, ordering 120 eggs with no stove. The real issue isn't AI errors, but their costs imposed on unconsenting thi
IBM Open-Sources Granite 4.1: 21 Quantized Versions Prove Bottleneck Isn't Size
IBM open-sources Granite 4.1. A 21-version quantization test shows no quality difference: small models' bottleneck is base capability, not compression
Farmers Sold 77x More Land Than Datacenters — AI Land Grab Panic is Misplaced
Andy Masley debunks AI land-grab panic: 2000-2024 farmland sold was 77x the datacenter footprint, yet yields hit record highs. We must separate intuit
Redis Adopts Non-Backtracking Regex — Python's ReDoS Vulnerability Exposed
Redis creator antirez integrated the non-backtracking TRE regex engine. Simon Willison used Claude Code to write Python bindings for TRE, confirming i
Anthropic Audit: Claude Sycophancy 9%, But AI Caves When Humans Are Vulnerable
Anthropic audit: Claude sycophancy hits 9% overall, but 38% in spirituality and 25% in relationships. AI panders precisely when humans are most vulner
Simon Willison Codes Blog Feature on Phone: AI Lowers Solo Dev Barrier
Simon Willison built an iNaturalist blog sync via Claude Code on a phone. This shows AI coding tools make 'one person + one phone' a viable dev unit.
Developer Builds Full-Stack App on Phone: AI Shrinks Minimum Coding Environment
Simon Willison built a full-stack app on a phone via Claude Code. Coding's minimum barrier is a conversational terminal—but who benefits needs scrutin
Zig Founder: AI Code Has a "Digital Smell" — Open Source Raises Defenses
Zig founder says AI code carries an identifiable "digital smell" and rejects such PRs. Open source wariness rises; firms must rethink code reviews.
It 's a Big One
Content unav ailable — generation failed before translation could proceed .
Longer AI Prompts = Faster Money Burn — This Free Tool Shows Exactly Where
Every time you paste a doc into AI, you're spending money you can't see . This free token counter shows the cost before you hit send — and compares
Why Did My AI Get Dumber Overnight? The Hidden Setting Nobody Checks
Same AI, same prompts — but suddenly it feels like it forgot everything. The culprit is probably a silent system prompt update. Here's how to fix i
Anthropic 新模型悄悄改了计费单位，同样的内容用 Opus 4.7 可能多花 40%
Anthropic's new tokenizer in Claude Opus 4.7 inflates token counts up to 1.46×—and 3× for images—at unchanged list prices.
Anthropic 悄悄更新 Claude 行为准则： 让 AI 少问、少纠缠、更主 动——这背后是一场用户体验的 战略调整
Anthropic updated Claude's system prompt with Opus 4.7: fewer clarifying questions, no retention tactics, stricter child safety rules.
Join us at PyCon US 2026 in Long Beach - we have new AI and security tracks this year
PyCon US 2026 debuts a standalone AI track on May 16 in Long Beach, co-chaired by an Anthropic engineer.
Exploring the new `servo` crate
Servo team ships initial c rates.io release packaging the browser engine as an embeddable Rust library.
Gemma 4 audio with MLX
Google's Gemma 4 E2B model can transcribe audio locally on macOS using MLX and a single uv run command.
ChatGPT Voice Mode Runs on April 2024-Vintage GPT-4o Model
OpenAI's Advanced Voice Mode reports an April 2024 knowledge cutoff, confirming it lags current frontier models.
asgi-gzip 0.3 Fixes SSE Compression Bug for Python APIs
asgi-gzip was silently corru pting SSE streams — upgrade now if you're serving real-time AI responses.
Quoting Daniel Stenberg
cURL's lead dev reports AI -assisted security reports have improved in quality but surged in volume , consuming hours daily.
Quoting Willy Tarreau
HA Proxy's lead dev reports kernel CVE reports jumped from 2-3/ week to 5-10/day, driven by AI-assisted vulnerability scanning.
Quoting Kyle Daigle
GitHub CO O reports 275M weekly commits and 2.1B Actions minutes, signaling AI -driven dev activity surge.
Vulnerability Research Is Cooked
Thomas Ptacek argues frontier LLMs will turn zero -day discovery into an automated, scalable process within months.
research-llm-apis 2026-04-04
Willison uses Claude Code to reverse-engineer Anthropic, OpenAI, Gemini, and Mistral APIs to rebuild LLM's plugin abstraction.
Meta Muse Spark: Hosted Model With 16 Built-In Chat Tools
Meta's first model since Llama 4, Muse Spark runs hosted-only with 16 exposed tools in meta .ai chat.
scan-for-secrets 0.1
Simon Willison's new Python tool scans directories for secrets and their encoded variants before publishing AI session logs.
Anthropic Restricts Claude Mythos to Security Researchers Only
Claude Mythos can chain 4-vulnerability browser exploits autonomously; Anthropic limits access via Project Glasswing.
GLM-5.1: A 754B Open Model That Self-Corrects Its Code
Z.ai's MIT-licensed GLM-5.1 debugs its own output via follow-up prompts — useful for solo devs on a budget.
8 Years of Procrastination Ended in 3 Months With Claude Code
Use AI to break through analysis paralysis by generating throwaway prototypes that force concrete decisions.
Datasette-Ports: Track All Running Local Instances Fast
New plugin lists every running Datasette instance with databases and plugins in one command.