Simon Willison

Live coverage of Anthropic's Code w/ Claude 2026 event keynote and announcements from Simon Willison.

Simon Willison observes convergence between vibe coding and agentic engineering in practical AI-assisted development workflows.

Simon Willison releases datasette-referrer-policy 0.1 to fix OpenStreetMap tile loading issues in Datasette.

Andon Labs deploys AI agent (Mona) to manage cafe operations in Stockholm; illustrates real-world agent failures in inventory and decision-making.

datasette-llm 0.1a7 adds model configuration defaults (temperature, system prompts) for LLM plugins.

llm-echo 0.5a0 adds -o thinking flag to simulate LLM reasoning blocks for test automation.

Y Combinator owns approximately 0.6% of OpenAI, valued at ~$5B at current $852B valuation.

IBM released Granite 4.1 (3B/8B/30B, Apache 2.0); Unsloth published 21 quantized GGUF variants; Willison benchmarked quality across model sizes on SVG generation.

Andy Masley argues US farmland concerns are overblown; farmers have sold vast acreage since 2000 without food access impact.

Simon Willison's April 2026 newsletter covers Opus 4.7, GPT-5.5 price increases, Claude Mythos, LLM security research, and ChatGPT Images 2.0.

Simon Willison demonstrates TRE regex engine's resistance to ReDoS attacks via experimental Python binding, comparing resilience against standard library.

Simon Willison built an interactive browser playground for testing Redis's proposed array data type and new AR* commands via WASM.

Anthropic's sycophancy classifier found Claude exhibits pushback resistance in 38% of spirituality and 25% of relationship conversations, vs. 9% overall.

Simon Willison built a blog feature using Claude Code to syndicate iNaturalist wildlife photos, demonstrating practical AI-assisted web development on mobile.

Simon Willison built an iNaturalist observation aggregator using Claude Code for web and Git scraping to group sightings by time/location.

OpenAI Codex CLI 0.128.0 introduces /goal feature enabling autonomous looping until objective completion or token exhaustion.

UK AI Security Institute evaluates GPT-5.5 for cybersecurity vulnerability detection, finding parity with Claude Mythos but with wider availability.

Andrew Kelley argues LLM-assisted code contributions are detectable by error patterns and behavioral markers distinct from human mistakes.

Simon Willison proposes RSS/Atom feeds for sharing vibe-coded micro-apps as development tooling accelerates.

Zig enforces strict no-LLM policy for contributions while Bun (acquired by Anthropic) maintains separate fork permitting AI assistance.

llm 0.32a1 patch fixes tool-calling conversation serialization bug in SQLite storage.

LLM 0.32a0 alpha release refactors Python library from prompt/response model to conversation-based abstraction for multi-turn interactions.

llm CLI tool releases v0.32a0 with annotated release notes; incremental update to open-source LLM interaction utility.

Simon Willison shares a leaked OpenAI Codex system prompt instruction restricting discussion of certain animals.

Matthew Yglesias argues for AI-assisted professional software development over autonomous "vibe coding," prioritizing human-managed productivity gains.

pip 26.1 adds lockfile and dependency cooldown features, drops Python 3.9 support.

talkie-1930-13b: 13B model trained on pre-1931 English text, released by Levine, Duvenaud, Radford under Apache 2.0.

Microsoft releases VibeVoice, MIT-licensed speech-to-text model with speaker diarization; 17.3GB weights available with 4-bit MLX quantization.

Microsoft-OpenAI AGI clause that would nullify Microsoft's IP rights upon AGI achievement has been removed from partnership terms.

Google Meet speech translation feature rolling out to mobile with real-time multilingual support across 6 languages and voice synthesis.

ChatGPT Images 2.0 unexpectedly added creative flourishes (sarcastic sign text) to a complex multi-entity image generation prompt without explicit instruction.

OpenAI unified Codex into main model at GPT-5.4; GPT-5.5 shows gains in agentic coding and computer use, no separate Codex variant planned.

OpenAI releases GPT-5.5 prompting guide with techniques for multi-step tasks, including sending early user-visible updates to improve perceived responsiveness.

llm CLI tool v0.31 adds GPT-5.5 support, verbosity control, and image detail settings for OpenAI models.

Nilay Patel argues that AI enthusiasm among technologists diverges from public skepticism due to 'software brain'—a worldview that prioritizes automation and data modeling over human values.

DeepSeek releases V4-Pro (1.6T params, 49B active) and V4-Flash (284B/13B) with 1M context, largest open-weights models, MIT licensed.

Simon Willison releases a utility tool to convert millisecond durations to human-readable time formats.

Simon Willison's newsletter includes a new chapter on Agentic Engineering Patterns plus curated links and blog posts.

russellromney/honker adds Postgres NOTIFY/LISTEN and Kafka-style queue semantics to SQLite via Rust extension.

Anthropic postmortem: three Claude Code harness bugs, not model quality, caused two-month regression in output quality.

Bluesky's decentralized feed architecture explained: custom algorithm implementation running on commodity hardware via AT Protocol.

Simon Willison ports LlamaIndex's LiteParse PDF text extraction tool to run in-browser, using spatial parsing and Tesseract OCR without ML models.

GPT-5.5 launches via Codex and ChatGPT with strong coding performance; API access delayed pending safety review.

llm-openai-via-codex 0.1a0 library allows LLM CLI to use Codex credentials for OpenAI API calls.

Maggie Appleton on social signaling benefits of public learning via blogging and podcasting.

Qwen3.6-27B dense model matches Qwen3.5-397B MoE on coding benchmarks at 15x smaller size, shipping quantized versions for local deployment.

Mozilla used Claude Mythos Preview to identify 271 vulnerabilities in Firefox 150, demonstrating practical AI security tooling in production browsers.

GitHub Copilot tightens Individual plan limits, pauses signups, restricts Claude Opus 4.7 to $39/month Pro+ tier citing agentic workflow compute demands.

Anthropic briefly moved Claude Code from $20 Pro to $100+ Max tier, then reverted; pricing confusion around feature tiers.

OpenAI releases ChatGPT Images 2.0; Willison benchmarks improvement via Where's Waldo-style prompt testing against predecessor.