Six months ago the consensus take among senior engineers was that nobody was going to just stop writing code themselves. Real engineers wouldn't accept generated code. The craft of typing it was load-bearing for understanding. I held some version of this.
Most of us have stopped. "All steering, no handwriting" is how DHH - creator of Ruby on Rails - put it today. Karpathy noted four months ago that he'd gone from 20% agent coding to 80% in a single month, and estimated "well into double digit percent" of engineers were already doing the same while general awareness sat at "low single digit percent." Your team is reviewing PRs they didn't write.
But this post isn't really about AI. It's about how groups update when reality moves faster than identity can - or how they don't. AI is just the case study where it's happening fast enough to watch in real time.
Build-up-to-tear-down is the oldest pattern in media: praise someone, then dismantle them. "We're so back" / "it's so over" is the meme version compressed to weeks. The same shape shows up wherever sentiment exists. Smartphones took a decade to complete a full cycle and you couldn't see the mechanism. AI takes a quarter. The release cadence is fast enough that you can watch a full cycle complete inside one news beat.
What looks like a random walk between excitement and scepticism isn't random. The engine underneath is status. Takes ripple outward in layers - the higher-status, more-online edge holds them first, they diffuse to the next layer, then to the broader audience. The moment a take is common one rung down, the edge needs a new position to maintain the gap. The discourse keeps moving because it isn't tracking the world, it's tracking who else holds the opinion. Karpathy's awareness gap - double-digit engineer adoption against "low single digit" general awareness - is the onion in action.
So what happens when reality arrives and proves a position wrong? Mostly, people don't update. They migrate. The position climbs a ladder of unfalsifiability: from "can't" (easy to disprove) to "isn't really what we meant" (definitional) to "is bad" (pure value claim). The edge needs somewhere new to stand, and "is bad" is a rung no screenshot can knock you off.
You can see this happening right now with the engineer cope. The same people who said engineers wouldn't stop typing have moved sideways: "typing code was never what engineers were paid for anyway." Same scepticism, repointed somewhere unfalsifiable. Nobody has to retract. The take just moves up a rung.
The same move has been running with AGI for a decade. Every benchmark that defined AGI in 2018 has been passed. None of them are called AGI now. The capability accumulates and the term migrates to wherever the capability isn't. That isn't bad faith. It's the natural defence of a sceptical position against the ratchet - it needs somewhere to live and it always finds a rung.
What does this give you? A heuristic. If a take feels safe to hold this quarter, it's probably already two layers behind the ratchet. Where the discourse feels comfortable is exactly where it's wrong, because comfort is what the ladder evolved to provide.
Six months ago Opus 4.7 launched and agentic coding stopped being a demo. I told myself the noise around it was social-media-exaggerated hype. I held my org back from Claude Code two months longer than I should have. I never said the earlier view out loud. That's the move this post is about. I made it too.
Of course, writing this is itself the next rung. "AI is hype" has diffused. "AI is back" has diffused. "Actually it's all signalling" is the next edge position to hold and I'm holding it now. I'm not outside this. But the ratchet is real either way.