Data from 22,000 engineers ends the "should we still read AI code" fight: whether you read is routed by the risk of the change.
Should AI Engineers Still Read Code in 2026? The Z/L Continuum — Alex Volkov, ThursdAI · Alex Volkov
21 min total·Actually worth watching closely: ~5 min·3 must-watch clips
- 0:12 – 3:50Skim
The inflection point: output changed by orders of magnitude
Uses the METR evaluation and GitHub commit volume (1 billion to a projected 14 billion) to show AI engineering hit an inflection point at the end of 2025: code went on sale, but attention didn't.
The scale of output has already changed by orders of magnitude -- that's the premise of the whole argument, not a prediction.
Mostly data slides (the METR curve, the commit-volume comparison). Glance at the charts and note the numbers; you don't need to follow it sentence by sentence.▶ Jump to 0:12Speaker · Alex Volkov - 3:50 – 5:21Listen
The ZL Continuum: two opposing positions
Introduces the two poles -- Lopopolo (code is free, what matters is the prompt and the guardrails) and Zechner (critical code gets read every damn line) -- and offers the ZL Continuum as a tool for locating yourself.
Answer honestly first: where on the spectrum are you right now?
Pure spoken position-setting with nothing important on screen -- fine to listen to while following each camp's logic.▶ Jump to 3:50Speaker · Alex Volkov - 5:21 – 9:45Listen
Where did the engineers go: from writing code to supervising agents
Uses Boris Cherny (100% of his code written by Claude Code, IDE deleted, still shipping 20-30+ PRs a week) and a live show of hands to show the engineer's role has moved up to the supervision layer.
Almost nobody's code is still mostly handwritten -- engineers haven't disappeared, the job just moved up a layer.
Mostly anecdote and audience interaction, low information density on screen; catching the details of the examples is enough.▶ Jump to 5:21Speaker · Alex Volkov - 9:45 – 12:24Skim
The data's verdict: each camp is half right
Faros's study of 22,000 engineers: deletions per PR up 861% (output camp right), but incidents up 242%, 6x the bugs per developer, and unreviewed merges up 31% (quality camp right too). Meanwhile the status page of AI-heavy Anthropic lights up like a Christmas tree.
Output and quality collapse are two sides of the same coin -- nobody won, because the question was wrong.
Several data slides back to back (861%, 242%, 6x). Worth pausing to read exactly what each number measures -- this is the densest evidence stretch in the talk.▶ Jump to 9:45Speaker · Alex Volkov - 12:24 – 16:45Skim
The right question: route the change to the proof it needs
Reframes the question from "should I still read the code" to "what proof does this change need," and gives the routing table: auth, money movement, permissions and irreversible data get read line by line; the rest can be delegated. Traces, evals and shadow mode stay.
The same person can be Lopopolo on one piece of code and Zechner on another -- it's a task-routing problem, not a matter of which side you're on.
The routing table's red lines appear as a slide list (around 15:33) -- worth grabbing a frame and copying down; it works as a team standard as-is.▶ Jump to 12:24Speaker · Alex Volkov - 16:45 – 19:28Listen
Review the system, not every line: bake findings into the pipeline
Every bug caught in PR review goes into the docs, the linter and the reviewer; the agent writing the code has to be separate from the one reviewing it or writing the tests. Rising capability keeps pushing toward the Lopopolo end, but it only moves the proof up a layer.
Reading spends your attention once; engineering makes the system remember the mistake forever.
Mostly methodological argument; only the capability-shift arrow on the continuum near the end (around 18:58) is worth a glance -- listen to the rest.▶ Jump to 16:45Speaker · Alex Volkov - 19:28 – 21:14Listen
Closing: judgment doesn't get cheaper the way code did
A self-verifying agent only moves the review up a layer; closes with Karpathy's "I've never felt so tempted to stop looking at the code at all -- but don't do this in production."
Not every line of code needs your eyes, but every system still needs your judgment.
A spoken wrap-up and closing line with nothing on screen -- good to just listen to as the argument lands.▶ Jump to 19:28Speaker · Alex Volkov