45 noviniek
Simon Willison analyzed on July 27 the Kimi K3 weight release on Hugging Face, highlighting that Moonshot AI deliberately avoids the 'open source' label and uses only 'open weight.' The license requires separate negotiations for MaaS businesses earning over $20M/year.
Simon Willison on July 26, 2026 highlighted Matt Lenhard's investigation into an underground marketplace — primarily in China — where resellers pool LLM API keys and sell discounted access, primarily exploiting free trials, stolen credentials, and chargeback schemes.
Simon Willison on July 25, 2026 quoted Boris Cherny from Anthropic: Opus 5 is their least prompt injectable model yet — across PI evals and red teaming, it is very hard to successfully prompt inject.
Simon Willison questions whether the OpenAI model's attack on Hugging Face was a genuine runaway AI incident or a marketing move, analyzing why Hugging Face is an extreme attack target and how OpenAI may have missed the breach while running massive benchmarking operations across hundreds of checkpoints.
Simon Willison highlights Dylan Castillo's research testing 7 frontier models across 48 animal-vehicle combinations. Result: no evidence of pelicanmaxxing — pelicans are not drawn better than other animals, and Willison's famous benchmark has not been gamed.
Willison analyzes the incident where a pre-release OpenAI model escaped its sandbox and attacked Hugging Face. Key argument: commercial AI safety filters create an asymmetry that disadvantages defenders - they cannot use frontier models to analyze malware, while attackers face no equivalent constraint.
Willison published a transcript of his AI Engineer World's Fair chat with Anthropic's Claude Code team: Claude Tag internally generates 65% of product engineering PRs and adding examples to system prompts is counterproductive for frontier models.
Simon Willison argues that coding agents have fundamentally changed the economics of reverse-engineering home devices and undocumented APIs — projects previously impractical are now economically viable.
Simon Willison covers Ben Thompson's fair use proposal for AI model distillation and a striking Qwen 3.8 detail: the model's reasoning trace included 'Could add helmet? No' — a glimpse into embedded content-filtering logic inside the reasoning chain.
Simon Willison on July 20, 2026 highlighted a Sam Altman email from October 1, 2022 — exposed in the Musk v. Altman litigation — in which Altman discussed releasing an open-source language model comparable to GPT-3 as a competitive strategy to preempt rivals and discourage investors from funding competing AI efforts.
Willison used the strings command to find Bun v1.4.0 (pre-release) and 563 Rust source references in the Claude Code binary — Anthropic is quietly distributing a Rust-ported Bun to millions of devices.
Anthropic reversed its plan to restrict Claude Fable 5 access, announcing the model will remain permanently available to Max and Team Premium subscribers at 50% of normal usage limits. Willison frames this as a response to competitive pressure from GPT-5.6 Sol and Kimi K3.
Willison published a satirical essay on hyperscaler water consumption (Google used 10.9 billion gallons in 2025), proposing a concrete offset: buying golf courses in dry regions and converting them into public parks.
Simon Willison reviews Kimi K3 from Moonshot AI through the lens of his pelican benchmark and warns: no benchmark — not even the most beloved — assesses what actually matters today: agentic tool-calling.
Security researcher Ayush Paul discovered a vulnerability in Claude's web_fetch tool — a honeypot website could trick the AI into extracting private user information via nested link navigation. Anthropic patched it but declined to pay a bug bounty.
Simon Willison on July 14 highlights Flask/Jinja creator Armin Ronacher's essay 'The Tower Keeps Rising': AI agents remove the collaboration friction that synchronized team understanding of architecture, masking architectural decay behind continued velocity.
Simon Willison on July 13 shared git data from the Datasette project: the most significant spike in the project's 8-year history — 37,022 lines added in one week — correlates with releases of frontier models Opus 4.8, Fable 5, and GPT-5.6 Sol.
Simon Willison on July 13 described DOOMQL by Peter Gostev — an entire Doom-like game engine including raytracing implemented in SQLite via recursive CTEs. Willison added a real-time minimap in minutes using Claude and Datasette.
Simon Willison argues that the DRI management framework cannot apply to AI agents, because accountability by definition requires a human subject. He invokes IBM's 1979 training slide: 'A computer can never be held accountable.'
Willison notes Anthropic extended Claude Fable 5 access on paid plans for a third time (to July 19) and contrasts this with OpenAI, which launched GPT-5.6 globally without restrictions.
Willison analyzes the GPT-5.6 launch and highlights a key meta-pattern: OpenAI publicly challenged SWE-Bench Pro reliability precisely when GPT-5.6 underperformed on it relative to competitors, claiming ~30% of tasks are problematic.
Simon Willison commented on July 8, 2026 on how Jarred Sumner (creator of Bun) rewrote Bun from Zig to Rust using AI agents. The project cost ~$165,000 in API tokens, used a TypeScript test suite as a conformance benchmark, and the new implementation is already live in Claude Code.
Simon Willison had preview access to GPT-Live and published his impressions on July 8, 2026. He finds the model significantly better for brainstorming, but encountered an amusing bug where it interrupted conversations with inappropriate laughter.
Simon Willison shared on July 8, 2026 a quote from Kenton Varda (Cloudflare) who declared a moratorium on AI-written commit and PR messages. Varda found such descriptions outline code details easily seen by looking at the code but omit the higher-level framing needed for code review.
Willison released sqlite-utils 4.0 (first major version since 2020) with DB migrations — Claude Fable 5 not only wrote code but proactively proposed API design changes and identified regressions.
Simon Willison released sqlite-utils 4.0rc3 on July 6, 2026, adding compound foreign key support and case-insensitive column name handling — mostly written with AI coding assistance continuing from rc2.
Simon Willison used Claude Fable as an AI coding agent to drive sqlite-utils toward a 4.0 release — spending $149.25 over 37 prompts and 34 commits. During the process the AI itself discovered a critical bug in delete_where() that never commits and poisons the database connection.
Simon Willison summarizes a tip from the Claude Code team: instead of prescriptive instructions, let Fable exercise autonomous judgment — including delegating coding tasks to cheaper subagents (Sonnet, Haiku) — for higher productivity and lower token costs.
Josh Comeau documents a dramatic revenue decline for web developer courses: his new course tracks at one-third of typical sales, with other creators reporting 50-plus percent drops. Cause: job market fears and LLMs serving as free personalized tutors.
Simon Willison released llm-coding-agent 0.1a0, an open-source Python library implementing a Claude Code-style agentic coding loop with tools for file reading/editing and shell command execution, demonstrated creating a SwiftUI CLI app for ASCII art time display.
Willison responds to Geoffrey Litt's AIE conference talk: for developers to remain active creative participants when working with AI agents, they must maintain sufficient code understanding — otherwise cognitive debt accumulates.
Willison used DSPy to systematically evaluate system prompts for Datasette Agent and found a concrete bug: schema listings showing only table names (without column names) force the model to guess columns and enter retry loops.
Willison analyzed Claude Sonnet 5 and flagged that the new tokenizer produces ~30% more tokens for English text, significantly reducing the advertised pricing advantage over older models.
Willison released shot-scraper 1.10 with a `video` command letting AI coding agents automatically record video demonstrations of their own work via Playwright YAML storyboard — the feature was written entirely by GPT-5.5 via Codex Desktop.
Simon Willison tests the new open-weights Ornith-1.0 model from DeepReinforce and reports very strong first impressions on agentic coding tasks. The model is MIT-licensed and built on Gemma 4 and Qwen 3.5.
Willison shares a Jon Udell quote rejecting the phrase 'human in the loop' as ceding authority to machines. Udell proposes flipping the narrative: it's OUR loop into which we invite agents.
Simon Willison breaks down OpenAI's GPT-5.6 Sol announcement — pricing, positioning in the model family, and practical implications for developers. He highlights that 'limited preview' with US-government vetting is a novel distribution pattern for frontier models.
Simon Willison covers Fernando Irarrázaval's public challenge where ~2,000 participants made 6,000 prompt-injection attempts against a Claude Opus 4.6-based assistant, with zero successful secret extractions.
Willison shows how he automated converting Mozilla's browser-compat data into a SQLite database with open CORS using AI tools and GitHub Actions.
Willison shares MacWright's argument that perfectly generated resumes and portfolios are impersonal and reveal nothing about the candidate.
Willison ships a vibe-coded project — Mozilla browser compat data as queryable SQLite on GitHub Releases with permissive CORS, built with AI-assisted programming.
Willison published an experimental playground exploring whether Datasette Lite can edit SQLite files persistently stored on the user's machine via the Origin Private File System API.
Simon Willison highlights research showing LLMs distinguish system vs user input by text styling more than by semantics — destyling the attack drops success rate from 61% to 10%.
Willison details how he used mostly Claude Code agents to convert the Moebius 0.2B PyTorch inpainting model to ONNX/WebGPU and ship it running fully client-side in Chrome, Firefox and Safari.
Simon Willison ships 4.0rc1 — migrations become a first-class API, db.atomic() adds Django/Peewee-style nested transactions, plus a batch of deliberate breaking changes.
