Weekly briefing — 2026-03-30
Good morning. Here’s what crossed my radar today that’s worth your time.
Good morning. Here’s what crossed my radar today that’s worth your time.
The ladder is missing rungs – Engineering Progression When AI Ate the Middle
Source: Hacker News: Front Page
The sharpest point in this piece is that “AI writes code” is the wrong headline; the real story is “AI is eating the training pathway.” Allan grounds that with numbers that are easy to wave away individually but troubling together: Anthropic-style claims of 90% coding automation vs. Redwood’s closer-to-50% estimate for committed repo code, Google around 25%, Microsoft around 30%, Copilot enterprise acceptance around 30%. That looks less like “full replacement” and more like “heavy partial automation” — and partial automation is exactly where apprenticeship can quietly break. If the easy, repetitive, first-5-years tasks get offloaded, we risk producing engineers who can ship outputs but haven’t built deep debugging intuition, tradeoff judgment, or architecture sense under constraint.
The METR findings are the other gut punch. In one controlled trial, experienced developers using AI were measured 19% slower, while believing they were ~20% faster — a 43-point perception gap. Then, in a follow-up, researchers reportedly couldn’t run the same design because developers resisted working without AI at all. Even if you think the first study is context-dependent, that second signal matters culturally: the tool became load-bearing before teams developed norms for when not to trust it. We’ve seen this pattern before in software abstractions — convenience gets adopted first, operational understanding catches up later — but here the abstraction is participating in reasoning, not just implementation.
I also found the “architecture, not intelligence” framing useful. Allan argues recent gains came as much from scaffolding (agent loops, linters, test-runner feedback, iterative correction) as from raw model leaps. That matches what many teams are seeing: reliability jumps when you wrap models in strict feedback systems. But that same scaffolding can hide capability cliffs. METR’s task-duration curve (near-perfect on minute-scale tasks, ~50% around one hour, collapsing on multi-hour tasks) plus “tests pass but PRs still unmergeable” results point to a practical reality: AI is very good at local completion and still weak at repo-level quality standards, maintainability, and integration judgment. In other words, it can fill in code fast, but it still doesn’t own the engineering bar.
Potential follow-up: Audit one active codebase for “missing rung” risk: measure which junior-level tasks are now AI-automated, then design explicit human training loops (debug reviews, architecture walk-throughs, no-AI drills, and merge-readiness rubrics) so skill formation doesn’t get accidentally outsourced.
Quick scan
A few more things worth a quick look:
- America Is Now a Rogue Superpower — Hacker News: Front Page
- Car Seats as Contraception — Hacker News: Front Page
Research notes saved to vault for potential studio follow-up.
