{
  "$type": "site.standard.document",
  "canonicalUrl": "https://rednafi.com/zephyr/carry-the-pager/",
  "description": "Drive-by AI changes break the shared model a team builds around its code, and the ICs end up cleaning up the mess. Why pushing to mainline should come with the pager.",
  "path": "/zephyr/carry-the-pager/",
  "publishedAt": "2026-05-30T00:00:00.000Z",
  "site": "at://did:plc:fgtm2c26vfcj74rfmeggbyqj/site.standard.publication/3mnl6f7ob462z",
  "tags": [
    "Essay",
    "AI",
    "LLMs",
    "Culture"
  ],
  "textContent": "![Gandalf blocking the Balrog, captioned \"You shall not pass\"][image_1]\n\nIn a funny turn of events, as AI spending spirals out of control, _tokenmaxxing_ is now\n_considered harmful_.\n\nRecently, [Fortune] reported that Uber burned through its entire 2026 AI coding budget in\nfour months, with every usage stat up and to the right. But COO Andrew Macdonald still can't\ntie any of it to features users would actually notice:\n\n> Maybe implicitly there's more that is getting shipped, but it's very hard to draw a line\n> between one of those stats and 'Okay now we're actually producing like 25% more useful\n> consumer features.'\n\nSo it seems like all that spending bought a lot of code and not much else. Maybe a ton of\ndebt too.\n\nMeanwhile, the AI overlords keep rewriting the narrative. Who knew the public wouldn't love\nbeing told their job's obsolete in the next x months?\n\nA year ago, [Dario Amodei cried wolf]:\n\n> AI could wipe out _half_ of all entry-level white-collar jobs - and spike unemployment to\n> 10-20% in the next one to five years.\n\nBy February, [Boris Cherny added to the fire]:\n\n> So I think at this point, it's safe to say that coding is largely solved. At least for the\n> kinds of programming that I do, it's just a solved problem because Claude can do it.\n\nThen the IPOs got close and the tone flipped. This month, [Dario caved first]:\n\n> If you automate 90% of the job, then everyone does the 10% of the job. And the 10% kind of\n> expands to be 100% of what people do and kind of 10-times their productivity.\n\n[Sam Altman followed three weeks later]:\n\n> I'm delighted to be wrong about this. I thought there would have been more impact on\n> entry-level white-collar jobs being eliminated by now than has actually happened.\n\nIn the beginning, this was an obvious PR stunt to sell leadership the nirvana of the\none-person company, where the agents do everything and the profit splits between the CEO and\nthe model provider. Too bad, that hasn't happened yet.\n\n---\n\nEveryone and their moms are now slinging code to production, not because they want to but\nbecause they're pushed to, and because AI apparently makes everyone an expert in everything.\n\nI'm a heavy LLM user myself, and I'm constantly fascinated by what it lets me do. But saying\nthings like\n\n- frontend engineering is cooked\n- AI is coming for backend\n- who needs infra people when the model is so good at wrangling k8s\n- writing docs isn't a problem, I just generated twenty pages of slop for my twenty-line\n  script\n- PMs are useless and should be replaced\n- EMs should code now, and the ones who don't ngmi\n\nis incredibly stupid. It creates a hunger-games environment at work. And for what? So that\nDario's next prediction becomes a self-fulfilling prophecy as we tear each other apart?\n\nThe AI leaderboard and tokenmaxxing gamification pull all kinds of weird behavior out of\npeople. PMs are forced to push code instead of doing their actual jobs. EMs are buried under\ntheir own responsibilities and still expected to sit at the center of every technical\nproblem on the team. The result is that everyone is running around like headless chickens\nswitching from one task to the next, from one demo to the next and accomplishing nothing.\nBut hey, we're productive at least.\n\nAnd no one cares about the poor ICs. The deadlines and expectations are getting batshit\ncrazy.\n\n_Just use AI, bro, and be 10x productive._\n\nCoding is important, but it's like 20% of the work, and making it 10x faster won't make the\nwhole thing go brrr! Hello [Amdahl's law].\n\n---\n\nHere's what Marc Brooker said [on a podcast]:\n\n> If you aren't doing it hands on, your opinion about it is very likely to be completely\n> wrong.\n\nHe's talking about leadership making technical decisions, and without being somewhat\nhands-on, it's impossible to form an opinion. But does everyone need to form an opinion on\neverything? Of course not, and that's exactly why we have hierarchies. Saying AI will\nsteamroll that into something flat and egalitarian is naive.\n\nAdd non-technical people to the mix, and it doesn't take long for the whole thing to turn\ninto a castle of glass. This breaks the social structure of an engineering team in the name\nof forced democratization. Now this might sound like gatekeeping. But if democratization\nmeans throwing out the playbooks we've built over the last 50 years, then gatekeeping is\njust quality control with an uglier name.\n\nExpecting code from an EM leading an infra team is very different from expecting it from a\nPM or a designer. I'm all for anyone trying to pitch in and survive this madness, regardless\nof their role. But hauling agents onto a mission-critical codebase isn't the way to do it.\nGo try merging something into the Linux kernel with zero context. The kernel maintainers\ncan't allow you to merge random stuff into mainline without breaking the universe. So why\nwould it be gatekeeping if merging PRs at your workplace demands the same bar?\n\nPMs and designers have plenty to contribute, just not directly to the code that runs in\nproduction. AI makes prototyping, mocking, and wireframing faster than ever, and anyone can\ndo that without disrupting the critical path.\n\nWhat it won't do is turn you into the expert it took some guy ten years to become. I can't\nreplace an EM with Claude, and I don't want to. I want my PM and my EM to do their jobs so I\ncan do mine. Also, please don't turn the poor designers into vibecoding jarheads. Let them\nexplore the tools at their own pace and see what they can do with them. Trying to force\neveryone into the same box, all contributing to the production codebase, is a leadership\nfailure.\n\n---\n\nBuilding software means building a collective mental model of how the system works and how\nit's run, and most of that model never makes it into the code or the docs. Peter Naur's\n[Programming as Theory Building] is as relevant as ever:\n\n> For a program to retain its quality it is mandatory that each modification is firmly\n> grounded in the theory of it.\n\nGenerating code is cheap. Operating it is where you learn the kinks of a deployed artifact.\nYou don't build that model without the elbow grease.\n\nAnd operating it means carrying the pager. When you get paged, you thread through the\nincident, hotfix if you can, write the post-mortem and the [CoEs], and bang your head\nagainst the wall until the root cause finally gives.\n\nYou can't do that alongside a zillion other responsibilities just because there's now AI.\nUnless you're in the full software development lifecycle, your mental model is probably\nplain wrong. You can't prompt your way into a moderately complex system without putting in\nthe work.\n\nSo if you haven't done that work and aren't ready to, your clanker-generated PR probably\ndoes more harm than good. Anyone can prompt an LLM and get that PR, and someone from the\ntrenchline can do it even better. So it's questionable whether this whole song and dance is\na good use of anyone's time.\n\n---\n\nA drive-by contribution from out-of-band folks breaks the collective mental model. The\nauthor never learned how the code is written or how it fails, and neither did the clanker\nthat wrote it. So often you get code that's locally correct but breaks some global invariant\nthe author never knew how to check.\n\nReviewing these slop PRs takes time, and the follow-up questions are futile, because the\nauthor didn't do the work and can't address the concerns. Going back and forth is\npointless - they'll just feed your comments back to the model. So people avoid reviewing PRs\nlike that, and worse, they get LGTM'd and merged into mainline. Disaster!\n\nAnd once it's merged, who owns the code? Usually not the drive-by author. They don't operate\nthe system, so when their change causes an outage, they're rarely the one firefighting. It's\njust chucking the ball over the fence and calling it a day. So much for _\"[you build it, you\nrun it].\"_\n\nOne too many times, the ICs and trenchline workers get pulled in to clean up the mess -\ntriaging an incident caused by someone else's vibeslop, on code they never touched. It's\nincredibly unfair, and a massive tax on people who are already overloaded.\n\n_So unless you carry the pager, maybe your code shouldn't go to mainline._\n\nBuilding anything useful takes all kinds of people. I don't like this clanker-driven,\nhomogeneous world we're heading toward. Being able to do more is welcome, but this lack of\naccountability and the contest over who burns the most tokens probably won't get us there.\n\nOh look, the good ol' [Goodhart's law] is laughing at us again!\n\n\n\n\n[fortune]:\n    https://fortune.com/2026/05/26/uber-coo-ai-spending-tokens-claude-code/\n\n[dario amodei cried wolf]:\n    https://www.axios.com/2025/05/28/ai-jobs-white-collar-unemployment-anthropic\n\n[boris cherny added to the fire]:\n    https://www.lennysnewsletter.com/p/head-of-claude-code-what-happens\n\n[dario caved first]:\n    https://fortune.com/2026/05/26/sam-altman-dario-amodei-walking-back-ai-jobs-apocalypse-prophecies-ipo/\n\n[sam altman followed three weeks later]:\n    https://www.cityam.com/delighted-to-be-wrong-sam-altman-changes-tune-on-ai-job-apocalypse-fears/\n\n[amdahl's law]:\n    /maxims/#amdahl\n\n[on a podcast]:\n    https://www.youtube.com/watch?v=u3GjIXP9N0s\n\n[image_1]:\n    https://blob.rednafi.com/static/images/carry_the_pager/img_1.jpg\n\n[programming as theory building]:\n    https://pages.cs.wisc.edu/~remzi/Naur.pdf\n\n[coes]:\n    https://aws.amazon.com/blogs/mt/why-you-should-develop-a-correction-of-error-coe/\n\n[you build it, you run it]:\n    https://queue.acm.org/detail.cfm?id=1142065\n\n[goodhart's law]:\n    /maxims/#goodhart",
  "title": "If you won't carry the pager, maybe don't push to mainline"
}