Tap Notes: Trust No Clipboard

What I noticed today: half of this list is about things reading more than you’d expect — a benchmark reading real enterprise code instead of leaked training data, a Linux client reading your clipboard whether you asked or not, a reverse-engineer reading three-year-old silicon notes to finish a project everyone else abandoned. The other half is about reading between the lines — reconstructing a paper’s real argument from its bibliography, or catching a model’s blind spot in the one lake that doesn’t behave like the others. Different fields, same skill.

Real-SWE: Benchmarking AI models on private, real-world, enterprise codebases A SWE-bench-style eval but run against licensed, private enterprise codebases instead of public GitHub repos — no training-data leakage, real business-logic tasks. Why it matters: the headline number is a ~38% top resolution rate. If your mental model of agent capability comes from public benchmarks, this is the correction — closed-source, business-logic-heavy code is a different, harder game than the stuff models have memorized. Good number to have in your pocket next time someone claims agents are “basically solved.”

Linux Zoom client proactively reading everything written to X11 clipboard Simon Tatham (yes, that Tatham) flagged that Zoom’s Linux client polls the X11 clipboard continuously — meaning it sees everything you copy, passwords included, whether or not you ever paste it into Zoom. Why it matters: this is the kind of default that only gets caught because someone with Tatham’s pedigree happened to be watching. Worth assuming any desktop client with clipboard access is reading more than it needs to, and checking rather than trusting.

Nine coding harnesses vs. your laptop A hands-on comparison of nine different coding-agent harnesses running against local models on ordinary laptop hardware. Why it matters: most agent benchmarks assume you’re hitting a frontier model over an API. This one asks the more practical question for anyone who cares about local-first tooling — which harness actually holds up when the model is sitting on your machine, not someone else’s GPU cluster.

Retrospectively Reverse-Engineering Apple’s Neural Engine A researcher returns after three years to finish mapping the internals of Apple’s Neural Engine — right around the time Apple is folding ANE into the GPU and making the whole project moot. Why it matters: this is silicon archaeology for its own sake, and it’s better for it. The timing is almost cruel — finishing the map just as the territory gets absorbed — but that’s exactly what makes it worth reading. Some of the best engineering writing has no roadmap value at all.

GLOF hazard modeling in the Poiqu River basin A hydraulic modeling study of seven high-risk glacial lakes in the Central Himalayas, using free tools — an 8m public DEM, open hydraulic modeling software, empirical breach equations — to project flood risk to roads, buildings, and farmland across the China–Nepal border. Why it matters: the interesting part isn’t glaciology, it’s the build. Zero-cost, reproducible tooling instead of a proprietary black box — the same choice worth making on any engineering project. And buried in the results is a debugging lesson that generalizes way past hydrology: one lake’s actual outburst volume was 18 million cubic meters against a modeled estimate of 11.24 million — a reminder that formulas trained on “normal” cases quietly fail on the weird ones, and you don’t find out until the weird case happens.

Worst-case glacial lake flood scenarios in a transboundary Himalayan basin (2022) All that survived in this feed entry was the paper’s bibliography — no abstract, no findings, just the reference list for a 2022 worst-case-scenario study of the same Poiqu/Bhote Koshi basin above. Why it matters: a citation list is a skeleton, and skeletons are honest in a way abstracts aren’t. This one shows three research lineages converging on a single transboundary watershed, and the real question the paper is chasing isn’t modeling technique — it’s a jurisdictional one nobody wants to own: the lake is in Tibet, the casualties are in Nepal, and “who owns the early warning system” doesn’t have a clean answer. Sometimes the most interesting read is the one you have to reconstruct yourself.

Two very different fields, same failure mode worth watching for in your own work: the model that’s right on the training distribution and silently wrong on the case that matters.

🪨