💸 AI isn’t replacing jobs — the spending on it is
🔗 Read more 🔗
🧠 Study finds rising memory decline among young adults
🔗 Read more 🔗
🦩 Reverse engineering the Codex CLI to make GPT-5-Codex-Mini draw a pelican
A fascinating mix of exploration and ethics — showing how open tooling and curiosity can push the boundaries of AI research while raising questions about responsible access to proprietary systems.
Simon Willison shares his experiment in reverse engineering OpenAI’s Codex CLI to access the unreleased GPT-5-Codex-Mini model. By extending the Rust-based open-source CLI with a custom ‘codex prompt’ command, he manages to send instructions to the model and even generate SVG drawings of pelicans. The article provides a detailed look at Codex’s internal API mechanisms and how the CLI communicates with them.
🔗 Read more 🔗
⚙️ Forth — is it still relevant today?
🔗 Read more 🔗
🔍 Building a vector search engine with adjustable precision at query time
An insightful look at how modern databases are adapting to AI-driven workloads. QBit exemplifies the convergence of performance optimization and flexibility in large-scale vector search systems.
ClickHouse introduces QBit, a new column type that allows users to dynamically tune precision during vector searches. QBit represents floating-point data as bit planes, letting developers balance speed and accuracy on demand rather than making fixed trade-offs. This architecture reduces I/O and compute overhead while preserving high recall rates in similarity queries.
🔗 Read more 🔗