Something Else is Happening
80 million people read and talked about the capability question. What about the people actually living through it? Grab a cup - this is going to be a long one.
Read Matt Shumer’s essay from last week. 80 million other people already did.
Personally I couldn’t get it out of my head. Not because of the AI part. Because of the feeling. That feeling of looking at a small iceberg that rotates. And it gets bigger and bigger with each second.
It reminded me of 2007. Not the crash. The year before. When the word “subprime” started showing up in headlines but your house was still worth more than you paid for it. When a few analysts were saying the math doesn’t work and most people were saying it’s fine, it’s different this time. Nobody panicked. The ones who paid attention repositioned. Everyone else got repositioned.
Shumer made the case for the wave of AI coming. 80 million people debated how big it is and when will it get here. I started asking people how well they can swim.
Those questions, those conversations turned into three interviews with people on completely different ends of this AI rollercoaster. Different situations. Different fears. Same look. That look where someone knows something is wrong but can’t quite name it.
I’m not going to reiterate Matt Shumer’s essay - AI has judgment now. The capability is real. Your job is going to change. Everyone responded to the capability question. How fast. How far. How real... I’ve been having a different conversation. About what’s happening to the people who are supposed to be using it. The people whose jobs are already changing, right now, while the rest of the world argues about timelines.
A friend of mine is a senior QA developer. Fifteen years of experience. She tests software for a living. If you’ve ever used an app that didn’t crash, someone like her is the reason.
She’s been looking for work for almost 2 years. She can’t land anything stable. There are gigs but nothing that sticks.
Every article she reads says the same thing. AI writes code now. Testing is getting automated. Her job is disappearing. She’s almost forty, and for the first time in her career, she’s afraid.
We grabbed a coffee. She told me she’s been applying to everything. Getting interviews. Not landing. The roles she’s qualified for seem to be vanishing. The overarching theme “If machines write the software now, who needs me?”
But this is where my insight after reading Matt’s essay and from my own company came to play. It’s important and this matters even if you’ve never written a line of code. So I’ll lay it down in layman’s terms.
Software used to be predictable. You write the code, it does the same thing every time. Click this button, this happens. Always. A test script can check that in seconds.
That world is ending. Software used to be written by people whose patterns you knew. You could predict where the bugs would be. AI-generated code doesn't work that way. It looks clean. It often runs. But you don't know what assumptions the model made, what edge cases it missed, or whether it introduced a vulnerability that won't surface until production. Your old test script checks whether code does what it's supposed to. Now nobody's sure what it's supposed to do.
But when AI is generating 41% of all code (Stack Overflow surveyed over 49,000 developers this year and that's the number) and 46% of those developers actively distrust what it produces... the person who can look at that output and tell you whether it’s actually correct? Whether it’s safe? Whether it does what the business thinks it does?
That person just became one of the most critical roles in a software team.
That’s her job. That’s exactly what she’s been doing for fifteen years. Strategic skepticism. Edge case thinking. Looking at something that appears to work and asking “but what if...”
So while she’s in the mind set of being replaced by AI, she’s actually being replaced by a job title. Software companies are cutting “QA” headcount. In the same time posting new listings for “AI validation specialist” and “AI quality assurance engineer.” Or are actually not realizing that their new shiny AI will fail if they don't. That’s the same skills. Different words. Written by hiring managers who don’t realize the person they need is the person they just let go in organization that is trying but failing in AI adoption without understanding why.
She’s not a cautionary tale about AI being too powerful. She’s a cautionary tale about a job market that’s moving so fast that perfectly qualified and CRITICAL for success people are falling through the cracks.
The second conversation happened because I needed advice.
I run a software company called BrightDot. We’re recently made a pivot - moving from traditional development to AI Engineering Studio. We transitioned several people of our team to that agent orchestration workflow (not everyone since it’s risky in regulated space we also work in). I wanted to talk to someone at a modern company working at the bleeding edge of technology. Somebody that’s already deep in this world. See how it’s working on the inside. How they approach it and what I could learn from them.
So I called a friend. He’s an engineering lead at a Unicorn company. They have contracts with most of fortune 500. Over a hundred developers. If anyone’s figured out how to actually use this technology at scale, it should be them.
We talked for an hour. He is doing what we are doing. Similar approach. Same gains we see. His output has skyrocket. I asked how many of his developers use AI the way we or he does. Not as autocomplete. Not as a search engine. As a genuine collaborator. Delegating entire features. Running multi-agent workflows. Trusting it with real work.
“Two,” he said. “Maybe three.”… Out of over a hundred.
My jaw dropped. I knew it’s not yet common practice, I knew we are spearheading but this is a company that builds AI-adjacent products. They’re in the room where this is happening. And 97% of their engineering team is basically using AI the way you use Google. Type something in. Get something back. Move on. That’s why the gains are minimal if not negative.
Now hold that number and put it next to this one.
Y Combinator said that 25% of their Winter 2025 batch had 95% of their code written by AI. Their managing partner, Jared Friedman, clarified: these are highly technical people, completely capable of building from scratch. A year ago, they would have. Now 95% of the code is done by AI.
Those are two worlds. Same tools. Same models available to everyone. On one side, ten-person startups building products that compete with companies a hundred times their size. On the other, a hundred skilled engineers with Copilot licenses who barely scratch the surface.
And then there’s a third world that most people haven’t heard of yet.
StrongDM’s AI team is three people. Justin McCarthy, Jay Taylor, and Navan Chauhan. They published their operating principles on February 6th. The first two: “Code must not be written by humans. Code must not be reviewed by humans.”
Not a manifesto. Not a thought experiment. They’ve been running this way since July 2025. The team’s open-source coding agent, Attractor, is a GitHub repo that contains no code at all. Just three markdown specification files. You feed the specs to a coding agent and software comes out. Their AI context store - CXDB - shipped with 16,000 lines of Rust and nearly 10,000 lines of Go. In production. Built end to end by agents.
Simon Willison, one of the most careful observers in developer tooling, called it “the most ambitious form of AI assisted software development that I’ve seen yet.” Stanford Law’s CodeX program published a separate analysis. Their conclusion: a security company decided that human code review is an obstacle, not a safeguard.
The humans write specifications. The machines build software. Everything in between is automated.
StrongDM’s benchmark for whether you’re taking this seriously: if you haven’t spent $1,000 per engineer per day on AI compute, your software factory has room for improvement.
Three people. Markdown files. Production software.
Meanwhile, a hundred engineers at a $400M company can barely get two people to use AI beyond autocomplete.
What’s the difference?
It’s not intelligence. It’s not access. It’s not budget. My friend’s company has all of that.
The difference is a skill. And it’s a skill that doesn’t sound like a skill, which is why nobody is teaching it.
Before I tell you what the skill is, I want to tell you about the third conversation because it’s the most important one. It’s the answer how to deal with it and what to do in this changing market.
A recent computer science graduate reached out to me. He was looking for work or even an internship. Talented guy. He knows Python. Statistics. Probability distributions. Linear algebra. He can explain how a neural network actually learns - not the buzzword version, the math underneath it.
He can’t get a job.
He’s not alone. A Harvard study examined 285,000 U.S. firms and 62 million workers. When companies start using generative AI, junior employment drops 9 to 10% within six quarters. Senior employment barely moves. In the UK, graduate tech roles fell 46% in 2024, with a further 53% drop projected by 2026. In the US, junior developer postings dropped 67% between 2023 and 2024.
The career ladder in software used to work like an apprenticeship. Juniors write simple features, fix small bugs, absorb the codebase through immersion. Seniors review and mentor. Over five to seven years, juniors become seniors through accumulated experience. This is the patter all careers are following. The finance, accounting, marketing, graphic design.
And AI is hollowing that ladder from the bottom. The theory is that if AI handles the simple features and the small bug fixes - the work juniors learn on - where do juniors learn? Seniors at the top, AI at the bottom, and a thinning middle where learning used to happen.
JP Morgan published research last month showing that college grads in AI-exposed fields are seeing the sharpest unemployment spikes. New grad unemployment in areas like computer engineering, graphic design, and architecture is approaching 10%. Entry-level postings are disappearing while senior AI roles go unfilled for months.
I spent an hour talking to this kid. Looking at his projects. Asking him how he’d approach problems. And I kept thinking: he has almost every skill that IEEE-USA is listing as the fastest-growing in tech right now. AI governance. Workflow design. Agent orchestration. He understands how models work. He can evaluate whether output is statistically sound. He is AI native the same sense Millennials are internet natives.
So what’s he missing?
Nobody taught him how to manage.
Not manage people. Manage machines.
He doesn’t know how to write a specification precise enough that an autonomous system can execute it without asking questions. He doesn’t know how to define boundaries for what a system should and shouldn’t do. He doesn’t know how to set up checkpoints for when the AI decides to get creative in ways nobody anticipated.
Nobody taught him. Because those skills didn’t have a name until about six months ago.
The skill nobody is naming
Here’s what connects all three conversations. And here’s the part of this story that I think needs to be heard alongside all the arguments about capability and timelines and whether this is hype or not.
Shumer described leaving his computer for four hours and coming back to finished software. StrongDM’s three-person team ships production code without a single human reading a line of it. I’ve had the same walk-away-and-come-back experience myself. It’s real and it’s extraordinary.
But think about what actually made that possible.
It wasn’t the AI’s capability. It was the ability to describe what you want with enough precision that the AI could execute it without going back and forth. Six years of building an AI startup taught Shumer that skill. StrongDM’s specification files run 6,000 to 7,000 lines of behavioral constraints, interface semantics, and system boundaries. That’s not prompting. That’s engineering-grade delegation.
Most people don’t work that way. And most people have never had to. They ware the ones that the task was delegated to.
Developers have, though. And this is the part I keep coming back to.
If you write bad code, it doesn’t work. Not “works less well.” Doesn’t. Work. Math doesn’t negotiate. Compilers don’t care about your intent. If you’re off by a semicolon, the entire thing breaks. Developers spent entire careers getting punished for imprecision. Every day. For decades. The industry created standards and tools to deal with it.
That discipline - defining what you want so precisely that a machine can do it without interpretation - is the exact skill that AI agents demand. Not coding. Specification. Delegation. Precision.
And software is the only profession that was already built around it. Versioning. Documentation. Testing. Reversibility. Binary feedback. The entire infrastructure of software development is essentially an elaborate system for being precise enough that machines do what you mean.
This is why software is seeing the biggest AI productivity gains of any field. The infrastructure for AI delegation was already there. Developers have been training the skill that matters most without knowing they were training it.
Now think about every other profession.
If you write a bad presentation... it still presents. Your audience might be bored, but PowerPoint doesn’t crash. A marketing plan with vague objectives gets adjusted on the fly. A podcast succeeds purely on charisma. A legal brief with ambiguous clauses just generates billable hours for revisions.
These fields don’t punish imprecision. They tolerate it. Work around it. Sometimes reward it.
AI doesn’t.
AI punishes imprecision the way code does. Give it something vague and you get output that looks right at first glance, passes a quick review, and quietly fails in production. The same lesson developers learned years ago, except now it applies to everyone whose work involves telling an AI what to do.
Which, very soon, is going to be everyone.
“But my company already uses AI”
If you’re thinking this, you’re probably in the same position as my friend’s unicorn. Your company has licenses. Your developers have Copilot. Maybe your marketing team has a ChatGPT subscription. You’ve “adopted AI.”
Here’s the reality. METR, the organization that measures AI capability - the same benchmark Shumer cited to show how fast AI is improving - ran a randomized controlled trial with experienced open-source developers. Developers using AI tools were 19% slower than developers working without them.
But those same developers believed they were 20% faster.
The people using AI thought they were more productive. They were actually less productive. Because having access to a tool and knowing how to use it are completely different things. As one senior engineer put it: Copilot makes writing code cheaper but owning it more expensive.
Stack Overflow’s 2025 survey found that 84% of developers use or plan to use AI tools. Sounds like adoption. But 46% distrust the output. Only 3% highly trust it. And the most experienced developers are the most skeptical.
There’s a pattern in how technology adoption actually works. Productivity drops before it climbs. You add a powerful new tool but keep running the old process. Of course it’s slower - you’re now doing two things instead of one. Most organizations are right there. They bought the AI. They didn’t rebuild the work around it. And they’re blaming the tool for what’s actually an organizational problem. The companies that crossed that dip? They didn’t just add AI. They rethought how specs get written, who reviews what, and what “done” means when a machine wrote the code. That kind of change is expensive and uncomfortable. Most companies won’t do it. That’s why the gap keeps widening. Not more licenses. Not better models. Teaching your people how to delegate to machines.
What this means for you
Shumer told you what to fear. I want to tell you what to do about it. Not the generic “learn AI” advice. Something specific.
The skill that AI demands - precise specification, clear delegation, defined boundaries - is not a programming skill. It’s a management skill. Managers have been teaching it to junior employees for generations. “Here’s what I need. Here’s how I’ll know it’s done. Here’s what’s out of scope. Come back to me at this checkpoint.” That’s delegation. That’s the skill.
Except now, instead of delegating to a person who can read between the lines and ask clarifying questions, you’re delegating to a system that takes your instructions literally. Which means the instructions need to be better. More precise. More complete. More thoughtful about edge cases. You are the delegator now. Not the one the task is delegated to.
If you’ve ever managed people, you’re closer to this skill than you think.
If you’re mid-career and worried your job is changing: it probably is. But “changing” is not “ending.” Look at what my QA friend is going through. Her skill set is more valuable than ever. Her problem is translation - the market relabeled what she does and nobody connected the dots. Don’t wait for the market to find you. Learn the new vocabulary. If you’re in QA, search for “AI validation.” If you’re in project management, search for “AI orchestration.” If you’re in legal, look at “AI governance.” The work is the same. The words are different. Get ahead of the relabeling.
If you lead a team or run a company: stop buying tools and start teaching specification. Your people don’t need another AI license. They need to learn how to delegate to AI the same way they’d delegate to a sharp but extremely literal junior employee. Clear objectives. Hard constraints. Defined review points. Measurable success criteria. That’s the difference between 2 out of 100 and 25% of YC.
But be honest about where your organization actually is. Most enterprise software is brownfield. Legacy systems accumulated over years, decades. The specification for what it actually does doesn’t exist. It lives in the heads of three people who’ve been there long enough to know where the skeletons are buried. You cannot dark-factory your way through that. For most organizations, the path doesn’t start with “deploy an agent that writes code.” It starts with “document what your existing software actually does.” The specification work comes before the automation. Not after.
If you’re early in your career: I know how frustrating this is. The entry-level pipeline is collapsing - Harvard’s numbers make that undeniable. But you’re closer to the roles being created than almost anyone. You understand how these systems work. You have the math. What you need isn’t more computer science. It’s how to manage. How to write specifications. How to define constraints. How to think about governance and oversight. That bridge between your technical knowledge and management thinking is the most employable combination in tech right now. Nobody is teaching it in universities yet. Learn it yourself.
If you’re not in tech at all: this applies to you more than anyone. Developers at least have the muscle memory of precision. They’ve been trained to be exact. You haven’t - not because you can’t, but because your field never required it. Now it will. Start practicing. When you use AI, don’t accept the first output. Instead of “write me a marketing plan,” try “write a Q2 marketing plan for a B2B SaaS product targeting mid-market CFOs, focusing on LinkedIn and email, with a $50K budget, measuring pipeline generation not brand awareness.” See the difference? The second version gives the AI constraints. Boundaries. Specifics. That’s specification. And the people who learn it first will have an enormous advantage.
The bridge
1,837 companies surveyed by Cleanlab, building on MIT research. Only 95 had AI agents running in production. Five percent. The ones that made it didn’t have better technology. They had clearer specifications. Better governance. Humans who knew how to define what “done” looks like before the machine started working.
Gartner says 40% of agentic AI projects will be canceled by 2027. Not because AI isn’t powerful enough. Because the organizations deploying it don’t know how to manage it.
Robert Half says 61% of tech leaders plan to increase headcount in the first half of 2026. AI/ML postings surged 163%. The jobs aren’t disappearing. They’re shapeshifting. And the shape they’re shifting into requires one thing: the ability to tell machines what to do with enough precision that they actually do it right.
My QA friend has that ability. She’s been doing it for fifteen years. She just needs the market to catch up with her.
My friend at the unicorn has a hundred people who could develop it. They just need someone to teach them that this is a skill, not an instinct. And they need the organizational will to redesign their workflows, not just add a tool.
That kid with the CS degree has the foundation. He just needs six months of learning how to manage instead of execute. And an industry that’s willing to invest in his development rather than waiting for him to arrive fully formed.
All three of them are closer to the future than they think. So are you. The skill is learnable. It’s just that almost nobody is teaching it yet.
Eighty million people read Shumer’s essay. Gary Marcus called it weaponized hype. Ethan Mollick said AI is jagged. Will Manidis compared the whole thing to FarmVille. Bloomberg said ignore the panic. Forbes called it a sales pitch. The Free Press assembled a panel. It has its own Wikipedia page now.
Everyone responded to the capability question. How fast. How far. How real.
Nobody responded to the human question. What happens when the capability arrives and discovers that most people don’t know how to use it?
Not because they’re not smart enough. Because precise delegation is a skill. And we never had a reason to teach it to everyone before.
Now we do.
The ceiling doesn’t matter if nobody can climb the stairs.
This is part of a series. Earlier this week I wrote about the Barbell Economy reshaping the job market and the data behind the Year of Humans.
If this resonated, hit reply. I read everything.