<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>Michael Byrne — Projects and updates</title><description>Michael Byrne builds quality AI systems and tools, and does deep ML research on various topics, for example, how models learn, remember, and can be read from the inside.</description><link>https://michaelbyrne.dev/</link><item><title>Free-Public-Lessons</title><link>https://michaelbyrne.dev/projects/free-public-lessons/</link><guid isPermaLink="true">https://michaelbyrne.dev/projects/free-public-lessons/</guid><description>Free AI lessons, a new one every week. Each explains one idea about how AI works, with a video, code you can run and its sources, and is fact checked before release.</description><pubDate>Mon, 26 Oct 2026 00:00:00 GMT</pubDate><category>Markdown</category><category>Python</category><category>GitHub Pages</category><category>learning</category><category>ai-assisted development</category></item><item><title>specdec-sidechannels</title><link>https://michaelbyrne.dev/projects/specdec-sidechannels/</link><guid isPermaLink="true">https://michaelbyrne.dev/projects/specdec-sidechannels/</guid><description>Three research records on what a speculative decoding system gives away for free. The answers came back negative or narrow, against bars set in advance.</description><pubDate>Sun, 25 Oct 2026 00:00:00 GMT</pubDate><category>Python</category><category>numpy</category><category>PyTorch</category><category>pytest</category><category>Claude Code</category><category>GitHub Actions</category><category>evaluation</category><category>pre-registration</category></item><item><title>Free-Public-Tools</title><link>https://michaelbyrne.dev/projects/free-public-tools/</link><guid isPermaLink="true">https://michaelbyrne.dev/projects/free-public-tools/</guid><description>Free tools for AI assistants and coding agents, a new one every week. Each started in my own work with AI and is tested before release.</description><pubDate>Sat, 17 Oct 2026 00:00:00 GMT</pubDate><category>Markdown</category><category>JavaScript</category><category>GitHub Pages</category><category>ai-assisted development</category><category>developer tooling</category></item><item><title>Role-Aware Dashboard Toolkit</title><link>https://michaelbyrne.dev/projects/role-aware-dashboard-toolkit/</link><guid isPermaLink="true">https://michaelbyrne.dev/projects/role-aware-dashboard-toolkit/</guid><description>Python building blocks for apps where people change settings, compare recorded runs and sign off decisions: role checks, saved profiles, audit history and a working business demo.</description><pubDate>Sat, 17 Oct 2026 00:00:00 GMT</pubDate><category>Python</category><category>Streamlit</category><category>FastAPI</category><category>SQLite</category><category>developer tooling</category><category>ai-assisted development</category></item><item><title>Where AI Memory Lives</title><link>https://michaelbyrne.dev/notes/ai-memory-systems/</link><guid isPermaLink="true">https://michaelbyrne.dev/notes/ai-memory-systems/</guid><description>Most AI memory today is text, stored and searched. Two systems I built work that way; newer research suggests memory may move into the computation itself.</description><pubDate>Sat, 17 Oct 2026 00:00:00 GMT</pubDate><category>ai memory</category><category>ai-assisted development</category></item><item><title>From Music to AI</title><link>https://michaelbyrne.dev/notes/from-music-to-ai/</link><guid isPermaLink="true">https://michaelbyrne.dev/notes/from-music-to-ai/</guid><description>I spent 13 years teaching and playing music before moving into AI. Why the switch makes sense to me: music is an arbitrary language, and that way of thinking transfers here.</description><pubDate>Sat, 17 Oct 2026 00:00:00 GMT</pubDate></item><item><title>What a Probe Can Tell You</title><link>https://michaelbyrne.dev/notes/linear-probes/</link><guid isPermaLink="true">https://michaelbyrne.dev/notes/linear-probes/</guid><description>A classifier reads a model from the inside: what it can do, what it takes to trust one, and what a safety classifier taught me.</description><pubDate>Sat, 17 Oct 2026 00:00:00 GMT</pubDate></item><item><title>The Shape of Data Itself</title><link>https://michaelbyrne.dev/notes/shape-of-data/</link><guid isPermaLink="true">https://michaelbyrne.dev/notes/shape-of-data/</guid><description>Going back through every dataset I generated, the same lesson kept coming up: the shape of the data decides what can be learned from it, and what a test can show.</description><pubDate>Sat, 17 Oct 2026 00:00:00 GMT</pubDate><category>training data</category><category>evaluation</category></item><item><title>The Path from Vibe Coder to Engineer</title><link>https://michaelbyrne.dev/notes/vibe-coder-to-engineer/</link><guid isPermaLink="true">https://michaelbyrne.dev/notes/vibe-coder-to-engineer/</guid><description>AI lowered the barrier to making software appear, not to understanding it. Seven habits, learned the hard way, for crossing from vibe coding to engineering.</description><pubDate>Sat, 10 Oct 2026 00:00:00 GMT</pubDate><category>ai-assisted development</category><category>learning</category></item><item><title>Espalier-Harness</title><link>https://michaelbyrne.dev/projects/espalier-harness/</link><guid isPermaLink="true">https://michaelbyrne.dev/projects/espalier-harness/</guid><description>A Claude Code harness that makes a planned, proven workflow the agent&apos;s default path, refuses its attempts to disable its own hooks, and treats CI as the only hard guarantee.</description><pubDate>Fri, 25 Sep 2026 00:00:00 GMT</pubDate><category>Python</category><category>Claude Code</category><category>GitHub Actions</category><category>pytest</category><category>developer tooling</category><category>ai-assisted development</category><category>agent control</category></item></channel></rss>