QueryBrew: New Research Proposes System-Agnostic SQL Query Optimization

Developer Builds Two Rust CLIs to Query Self-Hosted Gemma 4 via HTTP and MCP

A developer has published a step-by-step guide detailing two small Rust command-line tools that send the same question to a self-hosted Gemma 4 language model using different protocols. The first tool, gemma-rust, calls the model's OpenAI-compatible HTTP endpoint directly and returns the raw JSON response. The second, gemma-rust-mcp, acts as an MCP client by launching a Python MCP server and querying the model through its tools, receiving markdown output instead of JSON. Both CLIs were tested against two deployments of the same model: a local setup using llama.cpp on a GTX 1650 Ti laptop GPU and a cloud deployment running vLLM on an NVIDIA L4 via Google Cloud Run. The project's primary goal is to demonstrate what an HTTP client sees versus what an agent-style MCP client receives, highlighting the practical difference between direct endpoint access and tool-mediated responses.

I don't think AI is going to end humanity/civilization in the next decade. Personally, I feel like the doomsday buzz about AI now is probably what was there when the internet first came live... Imagine how the first phone call must have felt...

AI Liability Gaps Exposed as Courts Begin Holding Companies Accountable

When AI tools provide wrong information that causes harm, determining legal responsibility remains complex and varies by jurisdiction, with most companies using broad disclaimers to deflect blame. A 2024 British Columbia tribunal ruling in Moffatt v. Air Canada established that businesses are responsible for whatever their AI chatbots tell customers, ordering the airline to compensate a traveller misled by its bot about bereavement fares. In a separate US case, Garcia v. Character Technologies, a federal judge in early 2025 allowed wrongful-death and product-liability claims to proceed by treating an AI chatbot as a 'product' under product-liability law, though the matter later settled. These rulings are beginning to close the gap between how AI tools are marketed — as capable and authoritative — and how their terms of service describe them — as experimental and use-at-your-own-risk. No single comprehensive law yet governs AI-related harm, leaving courts to apply existing legal frameworks such as negligence, consumer protection, and product liability to an entirely new category of technology.

Reader Tips Exposed Two Evaluation Harness Flaws, but Left Core Results Intact

A developer who had previously disclosed nine self-found bugs in their own code evaluation harness received two additional bug reports from readers, both targeting the harness's checks rather than its underlying data. One reader flagged that CPython's bytecode caching could allow stale .pyc files to execute original code instead of mutations, though the harness was already unintentionally protected by a line that excluded cache files during copying. A second reader suggested adding a deliberate negative control — a test case expected to score near zero — so that suspiciously high results would trigger investigation rather than go unnoticed. Both fixes were verified and implemented, yet the original results remained completely unchanged. The author notes that the episode highlights how code written for one reason can silently serve a more critical purpose, and how pre-registration alone cannot catch a broken instrument that produces the expected result.