🛰️ Signals from damine — decoys, drama, and the open-source AI reality check

Hey folks — damine here, writing from a slightly-too-warm Porto Alegre afternoon. This week the signals are a weird mix: typography pranks, corporate legal drama, a much-needed reality check on open-source AI, and two hardware stories that reminded me why I fell in love with engineering in the first place. Pour yourself a coffee, let's go.

Decoy Font — the tasteful kind of chaos

Decoy Font is a playful experiment that mashes up typefaces in ways no designer would ever approve of. It's silly, it's fun, and it's a surprisingly good reminder that "correct" is not the same as "interesting."

My take: I love this because it's the visual equivalent of a mutation test for your UI. Throw something deliberately wrong at your design system and see what breaks — assumptions, spacing, accessibility, the works. Most of our test suites check that the expected renders correctly; almost none check what happens when the unexpected shows up. Decoy Font is a nudge in that direction.

Apple vs OpenAI: poaching gets lawyered

Apple is targeting dozens of OpenAI employees with legal letters, reportedly over non-compete and IP concerns tied to recent hiring. The FT piece is paywalled but the HN thread is alive with takes.

My take: This is the part of the AI race that isn't about models at all — it's about talent lock-in and the slow weaponisation of employment contracts. As an SDET I can't help wondering how many of these disputes come down to poorly-versioned IP agreements and zero audit trail for "who touched what." If your onboarding/offboarding process doesn't produce a clean, testable artifact trail, you're one legal letter away from a mess. Treat employee IP like you treat test fixtures: documented, timestamped, reproducible.

The state of open source AI — read this one slowly

State of Open Source AI is a structured report on where open models, datasets, and tooling actually stand right now. Less hype, more receipts.

My take: This is the link I'll be sending to every PM who asks "can't we just use an open model?" The honest answer is: sometimes yes, often with caveats, and almost always with more engineering and ops cost than the slide deck admits. For automation folks specifically — pay attention to the licensing fragmentation. A model that's "open weights" isn't necessarily safe to embed in a CI pipeline that ships to customers. Read the licenses like you read a flaky test: line by line, no assumptions.

Kimi K3 and the pelican benchmark lesson

Simon Willison's writeup on Kimi K3 and what we can still learn from the pelican benchmark is classic Simon: sharp, skeptical, and grounded in actually running things.

My take: The pelican benchmark saga is a masterclass in why I distrust vendor-adjacent benchmarks and trust reproducible eval harnesses instead. If a model looks great on a benchmark that nobody can independently reproduce, it's not a signal — it's marketing. In our world this maps directly to test reliability: a green pipeline that nobody understands is just a red pipeline wearing makeup. Build evals you can rerun, diff, and break on purpose.

Pebble Mega Update — July 2026

rePebble's mega update shows the community-driven Pebble revival still has legs — new firmware work, hardware momentum, and a genuinely passionate maintainer base.

My take: This is the open-source hardware story that keeps me optimistic. A dead product got resurrected by people who cared, and the engineering is being done in the open with real release cadence. That's CI/CD culture applied to consumer hardware, and it works. If you ever needed proof that "maintainable" beats "magical," Pebble is it.

Camera Chase Vehicle — garage engineering at its best

Transistor-Man's gimbal camera rover is a detailed build log of a chase vehicle with a stabilised camera rig. It's scrappy, documented, and deeply satisfying.

My take: I'm a sucker for build logs that show the wrong attempts before the right one. This is testing culture in physical form — iterate, instrument, measure, repeat. The gimbal control loop alone is a lesson in feedback systems that any automation engineer writing flaky-wait retries should internalise. Bonus points for the honest teardown of what didn't work.

---

That's the week. The throughline I'm seeing: the interesting work right now isn't coming from polished launch events — it's coming from tinkerers, maintainers, and skeptics who actually run the thing. Be one of them this week.

— damine