The $87 Change đ§ž
Nobody misses doing long division by hand, and a lot of what AI is taking off your plate can die in peace. The problem is the other part: the one you'll only notice you lost the day the agent gets it wrong.
When was the last time you did three-digit long division by hand? Yeah. Me neither, and I donât miss it.
Now picture a bakery cashier. Coffee and bread come to $13, the customer pays with a $20 bill, someone types 100 instead of 20, and the register says to hand back $87. Nobody needs that cashier to know how to do long division, and thatâs fine. But if they open the drawer and count out $87 without blinking, the bakery has a problem no new register will fix.
That scene sums up the whole debate about AI and skill loss.
𧎠The objection I take seriously
When an article about SREs losing skills to AI landed on Hacker News, the best pushback I saw came in the comments. One comment said the calculator killed hand division and nobody cares. Another pointed out that Terraform did the same to manual configuration, and nobody argues for configuring infra by hand once in a while just to stay sharp.
I agree with both. A lot of people who write about deskilling lose credibility by treating every lost skill as a tragedy. Memorizing kubectl flags and manually stitching together logs from three services by timestamp? Let it die.
But the calculator and Terraform have something the agent doesnât. As another comment pointed out, you type the same division 65 thousand times and get the same correct answer. terraform plan shows you whatâs going to change before it changes. The agent can be wrong, with a well-written root cause and the tone of someone who is sure.
And notice what survived the calculator: the arithmetic died, the sense of scale stayed. You donât divide 7,392 by 48 in your head, but you know it isnât 1,500. Terraform is the same, because whoever reads a plan with a destroy on the production database needs to know what theyâre looking at.
The cashier doesnât need long division. They need to be suspicious of the $87.
đ¨ The $87 on call
Bringing it into our world, with a made-up case: three in the morning, slow API, and the agent has already done the boring work. It cross-referenced metrics with logs and delivered a clean conclusion: Postgres connection pool exhausted, suggestion to raise the limit. Makes sense. Thereâs even a chart.
Whoever has a mental model of the system gets suspicious. Piling-up connections are usually a symptom of something holding each connection for longer, like a lock or that schema change that shipped late in the afternoon, and raising the limit might just add fuel to the fire. Whoever doesnât have it approves, because the explanation is good and itâs three in the morning.
The question Iâd ask is much narrower than âis AI making us dumber?â. For each thing you stopped doing: does the error in it show up on its own? And can you relearn it on the spot, with the docs open?
If both answers are yes, let it go. CLI syntax passes easily: you get it wrong, the terminal complains, you look it up and move on. Correlating logs is the borderline case: the mistake doesnât show up on its own, and it only passes because someone who knows what they expected to find can check the agentâs work in a few lines.
What doesnât pass is whatâs left: understanding how your system behaves when itâs taking a beating, and doubting a pretty answer. Nobody relearns that in the middle of an incident.
đ What the evidence supports
The warning is old. In 1983, Lisanne Bainbridge wrote that an experienced operator who moves to monitoring an automated process can become an inexperienced one. And she left a line Iâd paste on the wall of every platform team: perhaps the final irony is that the most successful automated systems, the ones that rarely need intervention, may be the ones that demand the greatest investment in human training.
In medicine, a study published in The Lancet Gastroenterology & Hepatology, across four centers in Poland, found that adenoma detection in colonoscopies done without AI fell from 28.4% to 22.4% after AI became part of the routine.
With caveats. Itâs a retrospective study in another field, the authors say âmight reduceâ, and it measures the doctor without the tool, while our fear is the tool being present and wrong. For SRE, the article that started the debate brings no direct data (and itâs worth knowing it comes from a company that sells observability with an AI assistant).
For now, weâre in the territory of the plausible hypothesis.
âď¸ Aviation doesnât make everyone fly by hand
The same article says the FAA started requiring more manual flight time and proposes a deliberate inefficiency: reserving for humans even the incidents AI would solve. The idea has merit, but the primary source tells a better story.
The 2013 alert is a recommendation to airlines: make room for manual flying in line operations, at low-workload moments, and make sure pilots understand when to use automation. What became an obligation is in the training rule: a short list of simulator maneuvers, several explicitly manual, like slow flight and loss of reliable airspeed indication, most with recurrent training within 24 months.
Notice the choice. The rule doesnât tell pilots to drop the autopilot day to day. Someone sat down, decided what the pilot canât afford to lose, and ordered training for that.
đ ď¸ Training cheap enough to survive
One comment in the debate nails our difficulty: if the SRE solves it by hand, it takes longer, and someone pays that bill. Another reminds us that aviation has a regulator forcing the issue, and our field doesnât.
âBut Bruno, nobodyâs going to stop and train during delivery week.â Exactly. The training has to be cheap enough to survive the next prioritization, and Iâd start with two habits.
During an incident, write your hypothesis in one line before opening the agentâs analysis, then compare. You keep using the tool and spend about two minutes.
And once a quarter, run your own version of the simulator: a failure injected in staging, agent turned off, starting by drawing the request path from memory. Wherever your drawing diverges from the trace is what you donât know. Bainbridge already warned that you canât simulate an unknown failure, so the goal is to train the reasoning rather than memorize answers.
𧞠The open drawer
If you decide what to automate, automate it without guilt (I would). Just do the full math: the more rarely the agent needs someone, the fewer people will know how to be suspicious when it does.
You can forget long division. The day the machine says to hand back $87, someone has to look at the $20 bill.