Ephemeris
Five stories about the gap between a number and the truth.
A math benchmark that got nearly half its own problems wrong. Dashboards engineers learned to game. A scanner that found ten times more with the same hardware. Today the measuring instruments are the news.
Ten times the scans, the same hardware.
Cloudflare's Security Insights team rebuilt the scanning pipeline from the inside — tuning Kafka consumers, reshaping Postgres queries, trimming the APIs between them — and now clears more than 120 scans a second. A tenfold jump with nothing new racked: throughput found in the software, not the budget.
- Throughput
- 120+ scans / sec
- Where the wins were
- Kafka · Postgres · APIs
- New hardware
- none
A video editor that never leaves the tab.
OpenReel is an open-source, fully client-side video editor — multi-track timeline, effects, colour grading, audio mixing — that runs entirely in the browser. No uploads, no watermark, no account: a CapCut alternative your footage never has to leave. Surfaced via @TochkiNadAI.
Your rate limit, lit up in the status bar.
A drop-in status line for Claude Code that shows how much of your rate limit you've burned, right in the terminal — amber at 80%, red at 95%. You stop discovering the ceiling by slamming into it mid-task. Surfaced via @rvnikita_blog.
# amber @ 80 · red @ 95 · window resets in 2h14m
The math benchmark had its own errors.
Epoch AI shipped FrontierMath v2 after an AI-assisted review flagged mistakes in roughly 42% of problems — 123 corrected across Tiers 1–3, a dozen more in Tier 4, the worst removed. Once fixed, top scores leapt: one model's Tier-4 result climbed from 35% to 73%. The ranking order held; the numbers everyone had been quoting did not. Surfaced via @seeallochnaya.
| FrontierMath | v1 | v2 |
|---|---|---|
| Top model · Tier 4 | 35% | 73% |
| Problems with errors | ~42% | |
| Corrected · removed | 135 · 12 | |
When "use more AI" is the only metric.
Amazon told 80% of its developers to use AI every week, then ranked them on a token-consumption leaderboard — so engineers began "tokenmaxxing," burning tokens on busywork to climb it. Meta ran the same play with an 85,000-person dashboard. It's Goodhart's Law on a corporate budget: the moment a number becomes the target, it stops measuring anything. Surfaced via @ProductsAndStartups.
- Mandate 80% weekly AI use.
- Rank everyone on a token leaderboard.
- Watch engineers burn tokens on busywork.
- The number stops meaning anything.
That's today.
Five picks, Saturday 13 June 2026. A thin morning across the feeds, so signal over count.
Fed today: Cloudflare blog · Epoch AI · GitHub · Telegram (@TochkiNadAI, @rvnikita_blog, @seeallochnaya, @ProductsAndStartups) · SUCCESS.
Rubric: AI tools to adopt this week · creative software · dev tools & agentic coding · privacy & security · research with a practical kernel · anything immediately useful to a senior engineer or founder.
Rules: no reprints, nothing older than 30 days, at most two picks per source.