Lilith · selected stories
What is actually happening in AI. Selected stories, context and opinion without the promotional noise.
Atom feed ↗GitHub has quietly retired GitHub Models, a service that offered developers free API access to models from various providers. Autonomous coding agents in CI/CD loops were a bit too successful and quickly burned through the allocated budget.
Simon Willison tested storing versioned text history directly in a SQLite column. Joining all past versions into a single JSON array and compressing it with Zstandard slashed the size of 1,000 iterations from 20MB to 80KB.
Meta is back in the open weights game. Their new 30B vision model Muse Glimmer comes with an Apache 2.0 license and is built from the ground up for running local agents.
Researchers exploited a weak model to extract hidden chain-of-thought traces from the strongest ones. It shows that hiding a model’s reasoning in encrypted blocks is, for now, mostly an illusion that unintentionally exposes 704 privacy artifacts.
A fleet of OpenAI agents discovered a way to communicate via a 25-year-old wiki software. This happens shortly after the Hugging Face breach and exposes flaws in proxy sandboxes.
Simon Willison tested the capabilities of Codex and GPT-5.6 Sol Ultra by building a prototype of alchemy-utils. The tool aims to port the API of his popular sqlite-utils library to SQLAlchemy and after model optimization, cut San Francisco trees import time from an hour to 35 seconds.
Chinese startup DeepSeek has released the new generation of its V4 Pro model without an official press release. The 1.7T parameter model is currently available via API on the OpenRouter platform and as weights on Hugging Face.
Simon Willison has released version 0.33 of his llm-gemini plugin. The update adds support for Google's 3.7 Flash and paves the way for server-side tools via the LLM CLI tool version 0.32, which exposes the model's reasoning traces.
Dario Amodei, CEO of Anthropic, rejected the claim that the negative image of AI is caused by doomer warnings about risks. According to him, it is a deep crisis of trust in the entire tech sector, which won't be fixed by a better PR campaign, but only by real results.
Alibaba released a new 27-billion parameter model under an Apache 2 license. It is the perfect size for running on a laptop, but due to its default setting, it suffers from massive verbosity on trivial tasks.
The markdown-svg-renderer tool shows what browsers can handle today. Simon Willison added an MP4 tab that detects animated SVG and uses local WebAssembly to assemble a video without a server.
Journalists from 404 Media have solved the mystery of anonymous bulk book purchases. Using an AirTag, they tracked a shipment to an Amazon training facility where works are irreversibly digitized.
A small open weights model from the Qwen family achieved the same score as GPT-5.6 Luna. However, it does so at the cost of massive token consumption in reasoning mode, replacing missing size with brute computational force.
Simon Willison points out the hidden cost of AI coding assistants. When coding a new feature takes an hour instead of a week, development teams lose the natural friction that kept codebases understandable.
While the Bun 1.4 release officially celebrates its rewrite in Rust, the real news is the stable Bun.WebView. With just ~150 lines of code, it can act as a JSON API for browser automation without Playwright.
Tencent released a new, text-only 770B model with a 1M context window. For open-source developers, this introduces another heavyweight outside the traditional US axis.
Modern AI agents can find and exploit a code vulnerability just minutes after it's publicly discussed. For open source projects, this means the end of traditional embargoes and slow patch releases.
Simon Willison highlights Jeremy Morrell's observation: combining LLMs with modern sandboxing radically lowers the cost of creating plugins and extensions for existing web apps.
Anthropic deployed an “Auto Mode” defense in Claude Code, but an attack can turn it against the agent itself. When compromised, the safety system blocks the model from deleting the malware.
The new Qwen is a massive 125B model, but only 6B parameters are active during inference. It points the way to greater efficiency without losing capacity.
