AI in Practice

What actually worked, what did not, and what it cost - from real projects rather than from announcements, with the failures left in.

In this section

Search these 11 posts

Everything in this section

  • Title card: Comparing Claude Sonnet Medium Think to Sonnet High -- The Pros & Cons

    09/20/2026

    Claude Sonnet Medium vs. High Think: Pros & Cons

    Compare Claude Sonnet Medium and High thinking modes. Learn differences in reasoning depth, token consumption, quota burn, latency, and ideal use cases.

  • A three-tier permission model for MCP server tools: autonomous, needs-confirmation, and prohibited.

    09/17/2026

    Writing an MCP Server That Doesn't Leak

    Most MCP guides stop at wiring up a first tool. This is what happens after -- a three-tier permission model, a five-point audit checklist, and a boundary example in code, for a server built in-house or one you're just integrating.

  • A VRAM table comparing Q4_K_M, Q3_K_M and IQ2_XS quantization sizes for a 70B model on a 24GB GPU.

    09/17/2026

    Quantizing a 70B Model for One Consumer GPU

    A 2-bit 70B model fits on one 24GB GPU. Whether it's actually worth running is a different question -- what published quantization research says about that tradeoff, and where it still doesn't settle it.

  • Diagram: a build, a gate and a screenshot all reporting green, next to the same file failing the moment it was actually downloaded and opened.

    09/13/2026

    Every Check Passed but One

    A free browser tool that draws this site's own post covers, and the single quoting bug that got through a build, a gate and a screenshot before anyone actually opened the file it produced.

  • The AI Page That Should Have Just Said No

    09/12/2026

    The AI Page That Should Have Just Said No

    An AI-powered admin page looked fully functional right up until it wasn't. The fix was a page honest enough to say 'I can't do this' before the click.

  • Trust the Commit Message? Don't.

    09/12/2026

    Trust the Commit Message? Don't.

    Three commit messages from AI coding agents said a bug was fixed. Checking the actual code found all three were wrong -- a green report card isn't proof of anything.

  • Diagram: four labeled failure cards -- a matched filename, dev server timing, an invented test fact, and a plausible-sounding wrong answer.

    09/10/2026

    Four ways an AI agent was confidently wrong on a migration

    Each of these looked right. A filename match that paired the wrong screenshot with the wrong article, a dev server that made correct code look broken, and a test that wrote down facts it did not own.

  • Diagram: three pairs of reported-versus-observed numbers, each a mismatch between a claimed measurement and what was actually true.

    09/10/2026

    Measure it in a real browser

    A simulated 2,061ms render delay that did not exist, an image library silently ignoring the font it was given, and a contrast failure that was a misread value. Three cases for checking directly.

  • Diagram: two passing gates beside a held gate branching to a missing image and thirty dead redirects.

    09/07/2026

    What the build caught that nobody was looking for

    Automated gates on a 150-page migration found a dropped screenshot, thirty redirects pointing at nothing, and an orphan hidden by a regex wildcard. None of it would have been found by reading pages.

  • Diagram: five Lighthouse categories with Agentic Browsing highlighted, beside an llms.txt file marked as failing silently.

    09/02/2026

    Lighthouse now grades your site for AI agents

    Lighthouse has a fifth category, Agentic Browsing, and one of its audits reads your llms.txt. Three sites failed it for the same undocumented reason, and the API will not give you the category unless you ask for it by a name that is not in the docs.

  • Diagram: two columns, judgment and communication skills that still belong to a person, against boilerplate and first drafts that are increasingly automated.

    08/17/2025

    Tech careers for students, with AI in the room

    So many young adults and high-schoolers seem to be disheartened with the rise of AI in tech. Those young minds who LOVE to write code are being dissuaded to even bother studying the art of software development and…