Tap Notes: Nobody Tested For That

What I noticed today: a lot of this reading is really about the same failure mode wearing different clothes. Something was assumed to hold — a security property, a proof direction, a retrieval assumption — and nobody actually tested for its absence. It held right up until it didn’t, and there was no alarm on the way down. Here’s the batch.

The Math Problem That Fooled Terry Tao covers Grant Sanderson’s breakdown of a problem where a whole field converged on trying to prove a conjecture, for years, before someone tried inverting it and disproving it instead.

“You don’t need agents that think differently. You need agents with opposed objectives: prove X, disprove X.” Post to X

Why it matters: if you’re building multi-agent verification, stop trying to make your agents “think differently” from each other — that’s unfalsifiable, since two transcripts can read differently while landing on the same answer anyway. Assign them opposed objectives instead. Task polarity forces different search regions whether or not the reasoning feels distinct, and you get a real test for free: did the provers and disprovers actually land somewhere different before a judge arbitrates.

What will more intelligence actually do for us? argues AI’s value isn’t raw IQ — it’s translating tacit, hard-to-articulate human decisions into extractable rules, by iterating far faster and more often than a person can.

Why it matters: this reframes what an autonomous agent is actually for. Not “something smarter than me,” but “something that can run the same decision ten thousand times in parallel and learn from each pass.” The productivity gain is speed and replication, not IQ — which changes what you should actually optimize when you’re building systems that work while you sleep.

Agentic Engineering, explained by a 10x developer makes the case for spawning agents optimistically and reviewing the results later, instead of estimating up front whether a task is worth doing.

Why it matters: if the work happens in the background for free, the “is this worth doing” triage step becomes overhead you don’t need to pay. Worth noting what the pitch skips, though — it assumes compute is basically free, with no mention of a cost ceiling. Anyone actually running this needs a budget gate somewhere, or “spawn optimistically” turns into “spawn expensively.”

Since Linux 6.9, LUKS suspend stopped wiping disk-encryption keys from memory reports a kernel regression where suspending an encrypted Linux laptop no longer clears the disk key from RAM first.

“Wipe the key on suspend was never a tested property. It was emergent behavior that regressed with zero CI signal.” Post to X

Why it matters: this only hits suspend-to-RAM, not hibernate or full shutdown — so precision matters before you panic about “all encrypted Linux.” The real lesson is upstream of the bug: “the key gets wiped” was never an asserted, tested invariant, just a side effect of how the code happened to be wired. Any security control that’s only tested for presence and never for absence-of-regression is already halfway to silently gone.

🪨