<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>Ewan Lab Journal</title><link>https://ewanlab.com/</link><description>Notes on building, evaluating, and understanding AI systems.</description><language>en</language><item><title>Claude Might Have Proved the Riemann Hypothesis. I Still Have Tennis at 9.</title><link>https://ewanlab.com/journal/riemann-hypothesis-and-tennis-at-9/</link><guid>https://ewanlab.com/journal/riemann-hypothesis-and-tennis-at-9/</guid><pubDate>Fri, 20 Nov 2026 16:00:00 GMT</pubDate><description>A few hours ago my feed said Claude might have proved the 167-year-old Riemann Hypothesis. My 9:00 AM tennis lesson is still on—and somewhere between a mathematical earthquake at midnight and an ordinary morning ahead, I spent the night thinking about how an early-career engineer stays sharp when the models start doing the thinking.</description></item><item><title>I asked Claude Fable 5.2 to weigh itself</title><link>https://ewanlab.com/journal/how-big-is-fable/</link><guid>https://ewanlab.com/journal/how-big-is-fable/</guid><pubDate>Sun, 15 Nov 2026 16:00:00 GMT</pubDate><description>I asked Claude Fable 5.2 to work out its own parameter count and what it would take to run at home, on a budget of zero. It doesn&apos;t know its size; its best guess, a borrowed and soft one, is about 3.5 trillion parameters, and running that at home would take five Mac Studios and a lot of patience.</description></item><item><title>Why Gemini&apos;s Temporary Chat Felt Dumber</title><link>https://ewanlab.com/journal/gemini-temporary-chat/</link><guid>https://ewanlab.com/journal/gemini-temporary-chat/</guid><pubDate>Fri, 06 Nov 2026 16:00:00 GMT</pubDate><description>For months, Gemini&apos;s Temporary Chat felt dumber to me, so one night I asked it 85 questions to find out why. The mode was fine. Gemini answers from old knowledge whenever a question doesn&apos;t mention time, and that&apos;s exactly the kind of question I put in Temporary Chat.</description></item><item><title>Building ewanlab: How I shaped an aesthetic with AI agents</title><link>https://ewanlab.com/journal/building-ewanlab/</link><guid>https://ewanlab.com/journal/building-ewanlab/</guid><pubDate>Sun, 01 Nov 2026 16:00:00 GMT</pubDate><description>Over a month and four looks, I used four coding agents to build a personal observatory. The turning point was realizing that visual taste has two halves.</description></item><item><title>What should an agent benchmark actually measure?</title><link>https://ewanlab.com/journal/what-should-an-agent-benchmark-measure/</link><guid>https://ewanlab.com/journal/what-should-an-agent-benchmark-measure/</guid><pubDate>Sat, 03 Oct 2026 16:00:00 GMT</pubDate><description>A task is more than a prompt, and a finished run is not necessarily a successful one. Notes on making agent evaluation inspectable—from task contracts to missing-score semantics.</description></item></channel></rss>