Tap Notes: Review Without the Gate
The thread today is where checking happens. Anthropic moved review from a gate at the end of the road to something that runs continuously while you write. An AI agent reverse-engineered a scanner by checking its work against a working system. And an ad network got used to skip the “user clicks something” step entirely. Same question each time: which check catches which failure, and when?
Review as spellcheck
How the Claude Code team does code review
Anthropic describes how it reviews agent-written code at scale. Volume went from roughly 2,500 to 30,000 changes a day, and the revert rate dropped about 40%.
Why it matters: The numbers are good, but the reframe is the useful part. “Did a human read it?” is a weak proxy for risk. The better question is what could go wrong with this change and which check, automated or human, is the right one to catch it. If you’re shipping agent-written code behind a mandatory human-read gate, you’re probably spending attention on the wrong things. This is the primary source, so read it rather than a summary of it.
Quoting Addy Osmani, Morgan Westlee Lunt, and Qing Wang
Simon Willison pulls out the passage where review moves from a final gate to something that runs continuously, “more like spellcheck while you write.” Most changes get reviewed and fixed by Claude at least once before a human sees them.
Why it matters: Simon’s excerpt is the whole argument in three paragraphs. Human attention is the scarce resource. Spend it on higher-level questions, not typos and routine mistakes. The gate model isn’t dead because it was wrong. It’s dead because it doesn’t scale to this volume.
Post to XChecking against ground truth
Using AI to reverse engineer hardware
A GNOME developer brought back a scanner’s Linux driver by giving an AI agent a Windows VM and a USB packet capture of the working vendor driver.
Why it matters: The agent never has to guess what the hardware does. It captures what a working system does and fixes the generator until its output matches. That pattern works for any task where a known-good reference exists, and it’s the opposite of a demo. It’s agentic coding doing hard, unglamorous work.
You could have invented transformers from memory
Alex Wang builds up linear layers, MLPs, and self-attention as a sequence of memory stores, ranked by interference-to-signal ratio and capacity.
Why it matters: Most transformer explainers are tours. This one derives the thing with a quantitative lens, so attention looks like the answer to a capacity problem instead of a design someone handed down. If you build on these models, a first-principles view of why they work tells you more than another diagram of the architecture.
Hardware and attack surface
Apple ML Research: Muse Spark Open with MLX on Mac Studio
Apple ran a roughly 1T-parameter open-weight model across four Mac Studios at about 36 tokens per second, using RDMA over Thunderbolt.
Why it matters: Unified memory has been a pitch for a while. This is the first set of numbers where self-hosting a frontier-class open-weight model on desk-side hardware looks like a real option instead of a stunt. Four boxes is still a purchase, but it’s a purchase and not a data center.
An Israeli firm is sending zero-click spyware through online ads
Citizen Lab and Confiant traced an ad tech firm delivering Candiru spyware to Al Jazeera journalists. The target didn’t click anything. The ad only had to be displayed.
Why it matters: The ad network is an injection vector that already reaches targeted individuals, and no user action is required. Every defensive habit built around “don’t click the thing” assumes a click exists. If you run anything that renders third-party content, that’s the threat model to update.
One more thing
Paid Memberships Pro v3.9 Release Candidate
Full disclosure: this is my day job’s product. The release candidate puts subscription delays and set expiration dates into core, adds percentage and fixed discount codes and captcha on login forms, and ships new developer tooling, including WP-CLI and Abilities API work. The structured, agent-facing data is the part I’d dig into first.
Checking things is the job. The only open question is where you put the checker. 🪨