01 · Generative video
Seedance lands as the new bar for short-form generative video — the demo reels are doing rounds in creative-tool group chats.
"The culprit was a lock in the query planner."
— Cloudflare, on why their billing pipeline went slow
postmortem · clickhouse
Billing got slow. The culprit was a lock in the query planner.
A routine partitioning change tipped a critical ClickHouse job over a hidden cliff: planner-level lock contention starved every other query. The fix went upstream.
$ SELECT max(eventTime) FROM billing_events # … ~30s later ↓ blocked by Query Plan mutex $ SHOW PROCESSLIST → 487 queries waiting on optimizer lock # root cause → a new partitioning key inflated the planner's join-reorder search by ~40× → contention surfaced only under prod concurrency, not in staging # patch sent upstream: split the planner mutex per-query, not per-statement
protocols · postmortem
When idle isn't idle, QUIC dies.
A Linux kernel optimization measured "idle time" the wrong way for QUIC, so CUBIC's congestion window collapsed during normal RTT waits. One connection slowed; then another; then the whole pipe.
via Cloudflare · 12 May
What broke: the kernel treated every gap longer than one RTT as the application going idle, and idle periods collapse the congestion window. QUIC's natural request-reply pacing tripped it constantly. Fix: separate "waiting for ACKs" from "waiting for the app," and only deflate when the app is the one waiting.
frontend · perf
Issues, instant.
GitHub rebuilt Issues navigation around three primitives: client-side caching of issue payloads, smart prefetching on intent (hover/scroll), and a service worker that serves stale-while-revalidate. The result is the cheap perf win every product team is sitting on.
| Metric | Before | After | Δ |
|---|---|---|---|
| First nav (warm) | ~1,100 ms | ~120 ms | −89% |
| Repeat nav | ~600 ms | ~0 ms | instant |
| Bytes refetched | full HTML | delta JSON | ~−70% |
| Felt as | "slow page" | "local app" | — |
agentic coding · port
A million lines of Zig, rewritten in Rust, in ten days.
Bun shipped a wall-to-wall Zig→Rust port driven by Claude Code agents running continuously. It is the largest publicly visible agent-led migration to date — and the first that doesn't read as a stunt.
commentary
What a lull looks like.
No new frontier model this week, no shock benchmark, no leaked memo. Zvi's read on a quiet AI cycle — and why the quiet ones are when policy moves.
The headline take. "This is what a lull looks like at this point" — slowing release cadence, fewer surprise capability jumps, more product polish than capability swing.
Why it matters. Lulls are when regulation and norms harden. GPT-5.5 is still the high-water mark on most things, which buys policymakers time to legislate against a known frontier instead of a moving one.
The thing to watch. Coding agents and ops automation, where progress is now compound and unglamorous — not a release event, a continuous lift.
security · evaluation
Mythos finds 91.5% of the bugs .
XBOW evaluated Mythos against GPT-5.5 on an offensive-security suite. Mythos comes out ahead on both detection and exploit construction — and especially on the cases where the model has only source, no binary.
agents · mobile
Codex, in your pocket.
OpenAI shipped Codex inside the ChatGPT mobile app: monitor, steer, and approve long-running coding tasks from your phone, on real remote environments. The interesting bit is the steering model, not the keyboard.
Operations · Memo 027 / RE: Codex
- To
- any engineer with a long-running agent
- From
- OpenAI · 14 May 2026
- RE
- "Work with Codex from anywhere"
The agent runs on a remote sandbox. The phone is a control surface — approve diffs, redirect, or kill the run. You no longer need to be at the laptop while Codex grinds through a 90-minute task.
For solo founders, this is the first credible "the agent works while I sleep" workflow that doesn't require a self-hosted harness. Pair it with a tight test loop and you have an overnight loop you can actually leave alone.
digest · weekly
Seedance, silicon, and robots that remember.
Three threads worth pulling from The Batch this week. Each is a small lift on its own; together they tell you where the cycle is.
02 · AI-designed silicon
Nvidia disclosed AI-guided chip layouts shipping in production — a quiet milestone for AI-in-the-loop hardware design.
03 · Robots that don't forget
Methods to stop robots from catastrophically forgetting previously learned skills are gaining traction; the lab demos are starting to look like products.
That's today.
Eight stories on 15 May 2026. Two ClickHouse-and-kernel postmortems from Cloudflare, GitHub's perf rebuild, a million-line agent-led port, a Zvi check on the cycle, an offensive-security eval, Codex in your pocket, and a weekly digest.
Cloudflare
GitHub Engineering
Bun · GitHub
The Zvi
XBOW
OpenAI
The Batch
via @seeallochnaya