I think many jobs will look like this sooner or later.

That’s it. That’s the post. If and only if you’re interested in all the gory details, read on. But I warn you: it gets pretty dry from here on out.

OK, you were warned.

I have a new process for building Overflight App Store releases and it’s quite insane. This morning I woke up to an interesting Slack notification:

🐛 supportbot: suspected new bug — overflight: “Overflight problem report” — filed https://github.com/cwooddgr/overflight/issues/8

You see, supportbot is my AI-powered bot that reads all email sent to support@dgrlabs.co. It analyzes the body of the email and any attachments, including the debug logs automatically included in Overflight and CD Wally problem reports, and drafts a reply. (It never sends anything itself.) And if it looks like the user has actually found a bug, it also files a GitHub issue with its own diagnosis (and the debug log attached) and posts a notification like the one above to Slack.

First I lightly edited the draft email reply to the user, promising a fix soon, and sent it.

Then I fired up Claude Code and told it to look at the most recent issue. It agreed with supportbot’s findings and went to work on implementing a fix. It tested the fix, tested the whole app for regressions, found none, and committed the fix to Git.

So far so amazing. But wait, there’s more.

I wanted to get this fix out ASAP as I had promised the user, so I told Claude Code to cut a new App Store build, v1.5.1, for both iOS and Apple TV. (It also picked up a few small fixes that had already been queued since 1.5.) And since it was a minor bug-fix release, it should keep the “What’s New” content, both in the App Store and in the app, the same as v1.5, but add a “Minor bug fixes” bullet point.

Claude Code made the changes, archived and uploaded new builds for iOS and tvOS, created the App Store version records, attached the builds, and submitted them for review through the App Store Connect API. It also wrote up a factual inventory of everything that had changed since the last App Store build. Then it kicked off new Claude Code sessions for supportbot and for marketroid (my bot that handles all marketing) and handed each of them that inventory so they could update their own materials.

Supportbot did two things. First, it immediately added the bug to its live knowledge base, marked as known and fixed in v1.5.1, pending App Review. That way if anyone else reports the same bug on v1.5 it will tell them a fix is on the way rather than filing a duplicate GitHub issue. Second, it staged a second set of updates, covering the new behavior in v1.5.1, to be applied only once the release is actually live.

Marketroid looked at the inventory, checked it against the website and press kit, and decided it didn’t need to change anything, which it recorded as an explicit “no marketing changes” verdict. If it had needed to update any content on the website–including screenshots, screen recordings, and the zipped up press kit–it would have authored those changes and staged them too. Same for any other marketing materials: feature/benefit lists, messaging pillars, brand voice, and so on. The one rule is that nothing marketroid stages goes public without my sign-off, so Claude Code showed me the no-op summary and asked for approval, which I gave.

Finally, Claude Code switched on a cloud routine that checks the App Store once an hour until the new build is live. The moment it is, the routine applies supportbot’s staged updates to the live knowledge base, flips the known-bug entry to “fixed, please update your app”, applies marketroid’s approved changes to the live site, sends me a message on Slack, and switches itself off. If I had forgotten to approve the marketing changes, it would keep nagging me in Slack every hour until I did.

In the future I might automate the step between the GitHub issue being created and the fix being coded. Right now it’s manual: I fire up Claude Code and tell it to look at the most recent issue. That could be optimized so my only role is reviewing the issue and the already-coded fix, and approving or rejecting it, plus reviewing anything customer-facing.

The moral of this story is that I suspect a lot of jobs are going to collapse into this kind of reviewer role. At least until rejections become unheard-of, at which point there will no longer be a need for a human in the loop, unless it’s only to have someone to hold responsible when something inevitably goes wrong.

Feeling encouraged I decided to try Claude Code, which had begun to get a ton of attention in the tech world, and something amazing happened. You know that scene in Star Wars when the Millennium Falcon makes the jump to light speed and the stars turned into streaks? Well, yeah. It was like that. And I’ve been cruising through hyperspace ever since. And no, it ain’t like dustin’ crops, boy.

In the five weeks since I started using Claude Code I’ve created eight new apps—some for the web, some for iPhone and iPad, and even one for Apple TV. Lots of people I know are experiencing the same explosion of creativity and productivity—it’s a rush!—and I love seeing what people are coming up with.

At the same time, it’s worth taking a step back to consider what’s going on and what it all means, particularly for the software industry, where I’ve spent my whole career.

It’s no exaggeration to say that tectonic shifts are happening. Cost structures are imploding. Time to market is collapsing. Tiny teams are becoming terrifyingly effective. And incumbents are looking obese and slow. In short, a disruption bomb has just gone off.

As my friend Todd Vernon recently noted, the old way of doing startups was to have a novel idea and then go raise some money to pay a team to build an initial version of your product over multiple months, then see if the market had an appetite for it. The answer was usually no, and the million or so bucks you raised to build your MVP became a loss for the investors, hopefully to be offset by someone else’s ten-banger. And so it went.

But what happens when a solo developer can bang out an MVP—not a prototype, mind you, but a functioning, shippable piece of commercial software—over a weekend? What happens when a team of one or two can iterate through the whole software development lifecycle of analysis, design, implementation, and testing in minutes and hours instead of days or weeks?

Two weeks ago I had an idea for an iPhone app I wanted to make. My last startup’s product, Numerous, had a very popular feature that I have been missing ever since we shut that company down ten years ago: countdowns. I liked being able to set a countdown to track the days until my kids come home for the holidays or the next rocket launch. People tracked holidays, special events, product launches, days since their last menstrual cycle, you name it.

Now, there’s a whole category of countdown apps on the App Store. But they’re all pretty janky and plagued by bad design, full-screen ads, and constant nags to buy a subscription to unlock basic functionality. So I decided to write one myself.

I fired up Claude Code, put it in planning mode, and told it what I wanted. It asked a couple of clarifying questions, which I answered. It created an implementation plan, which I approved. And off it went to create my new app. Five minutes and fourteen seconds later, it was done.

Five minutes and fourteen seconds.

I’ve spent the last two weeks adding functionality, polishing, getting feedback from users, and navigating the App Store submission process (which itself is something that really needs to be automated). I submitted my new app Countdowns to the App Store last night. It’s better than the countdowns feature of Numerous ever was. It’s far better than any of the countdown apps currently available. And for me it heralds a new age of software, one that happens at light speed. As Han Solo said, “Go strap yourselves in.”

The implications for the software industry are profound. I’ll be writing more about them in the coming weeks.

If you’re still confused about how you need to allocate your personal capital-labor split, let this serve as a moment of clarity.

The findings have been reported by lots of media sites, but none of them link to the report itself, which is apparently internal or client-only, at least for now.

See, I live in a small mountain town with only one real grocery store. They’re constantly understaffed and have to pay stipends to employees to drive up the hill (i.e., a minimum 45-minute trip up the mountain) to work. So when automated checkout lanes became available they were thrilled. Instead of needing four checkers to man four checkout lanes, they could have a single checker overseeing eight self checkout lanes. They roughly doubled their throughput and cut staffing needs by 75%.

Notably, the self-checkout machines can’t do the whole job of a checker. A human is still needed to check IDs for alcohol purchases, sell cigarettes and stamps, and help confused customers. But the machines can do enough of the job that only one human is needed to do the work of what was previously four. The other three people had to look for new jobs.

Labor economists break jobs down into tasks and then look at what percentage of the tasks comprising a job can be automated. They then use that percentage to predict how susceptible that type of job is to automation. But what they fail to take into account is the fact that lots of tasks that cannot be automated can be done by just about anybody.

I recently wrote about using an AI assistant to help me diagnose a problem with the water filtration system in my house. In that case, there were lots of tasks involved that couldn’t be automated: physically getting into the mechanical room to observe the problem, moving a storage shelf out of the way to access the beeping equipment, pushing the reset button to stop the beeping, etc. Only a small percentage of the tasks involved, let’s say 10%, were able to be automated, but they were the ones that would have otherwise required a specialist: identifying the problematic piece of equipment, understanding what the warning meant, knowing how to dismiss it, and prescribing a course of action.

I was able to perform about 90% of the tasks myself, even though I’m completely unskilled as a plumber. And notably, I didn’t pay anyone to perform those tasks. AI performed 10% of the tasks but it enabled me to complete 100% of the job myself. Labor economists would say that particular job is only 10% exposed to replacement by automation, but I’d say it’s 100%.

My point is simple: the fact that AI can’t do 100% of the tasks in your job—your whole job—doesn’t mean you won’t lose your job to AI. Plan accordingly.

Early one Sunday morning my wife heard a faint but persistent beeping coming from the basement. I got up and followed the sound to the mechanical room. After some searching, I narrowed the source of the sound down to a metal box mounted on the wall. A tiny red LED was flashing, but there was no label telling me what it meant and no obvious way to silence the beeping. I was sure I had the manuals for all of this equipment—somewhere. I tried Googling on my phone a bit, but I didn’t have enough information to go on. I knew I had contact info for both the well service company that had installed the system and our plumber, but it was 5 AM on a Sunday. Maybe one of them would make an emergency call for some outrageous amount of money, but I wasn’t even sure this was an emergency.

At this point I had a decision to make. I could be conservative, play it safe, call somebody to make an after-hours call, wake up the whole house, and spend hundreds of dollars on what might turn out to be nothing. Or I could ignore it for now, hope it was nothing, and risk… I didn’t know what. Losing water pressure in the house? Damaging the equipment? It could potentially wind up costing me thousands of dollars.

Then I had an idea. I pulled out my phone, opened ChatGPT, selected Advanced Voice Mode, and turned on the video camera. I pointed it at the beeping box, described what was going on, and asked for some advice. Without hesitation, it identified the piece of equipment as the UV filter, told me how to dismiss the alarm, and let me know I could safely leave it for now but that I should schedule a bulb replacement in the next few days. Exactly what an on-call plumber would have charged $300 for, except immediately and with no fuss.

Emboldened, I wondered aloud if I could increase the water pressure in the house while I was down there. ChatGPT had me point the camera at the various components of the system and identified the pressure controller. I opened the control panel and pushed the up arrow button, but a small LCD screen said LOCKED. Hmm. I didn’t remember setting a code, so I tried my debit card PIN. No joy. I asked ChatGPT for advice. It suggested “1234”. Still no good. ChatGPT searched the internet, found the owner’s manual for this particular pressure controller, and found its default code: “7777”. I tried it and it worked. I increased the pressure up a couple of PSI. Magic.

What I had just experienced was intelligence-over-infrastructure. It made an impressively complicated system simple. It combined latent knowledge with near-instant information retrieval to transform a potential emergency into a non-event. And it changed my thinking about the possibilities of AI.

When non-experts, or even experts in a hurry, have access to an intelligent assistant, infrastructure incidents that are potentially expensive and time-consuming become stress-free and easily solved. And in the event of a real emergency, an expert assistant can help you solve the problem with a minimum of time, effort, cost, and stress.

At Spanning, we backed up our customers’ critical data. In the rare event that they needed to restore from backups, it was often an emergency. We designed our tools to be powerful and easy to use to minimize downtime, but we were limited to the technologies of the day: point-and-click interfaces plus traditional search. If we had been able to also provide them with a conversational, intelligent AI chatbot interface we would have been able to make the task of finding and restoring lost data so much quicker, easier, and less stressful.

It’s clear to me that enterprise infrastructure without such an intelligence layer will soon be considered dumb, no matter how pretty the UI or how complex the query builder. The intelligent tools that we all have grown to use in our daily lives will be de rigueur in enterprise settings, especially when simplicity and speed are more than nice-to-haves, they’re requirements.

OpenAI CEO Sam Altman minced no words when he told MIT President Sally Kornbluth, “This is going to eliminate a lot of current jobs.” That was in May 2024, but he was only echoing entailments he had expressed in 2023 at the WSJ Tech Live event.

Anthropic CEO Dario Amodei recently warned that AI could take half of all entry-level white-collar jobs and that 10%-20% unemployment is one possible scenario. He went on to say that people don’t believe it because “it sounds crazy”.

Inc. magazine says Amodei is wrong—it will be even worse than he predicts, and that those most threatened are “the $200,000-a-year senior managers who spend most of their time preparing decks, managing spreadsheets, and writing jargon-loaded analyses.”

But it’s not just the heads of AI companies who are making this prediction. From a recent Wall Street Journal article:

”’Artificial intelligence is going to replace literally half of all white-collar workers in the U.S.,’ Ford Motor Chief Executive Jim Farley said in an interview last week with author Walter Isaacson at the Aspen Ideas Festival. ‘AI will leave a lot of white-collar people behind.’”

Salesforce CEO Marc Benioff claims that AI is already doing 30%-50% of the work at his company. But his ultimate goal isn’t to reduce his own costs; it’s to capture a slice of the $12 trillion that could be shifted from the labor market to AI spending.

Make no mistake. This is happening.

Update: And now this: Founder of Google’s Generative AI Team Says Don’t Even Bother Getting a Law or Medical Degree, Because AI’s Going to Destroy Both Those Careers Before You Can Even Graduate Oh well!

Fundamentally, there are two ways to make money: labor and capital. If you have a job, you’re a laborer (or a worker, a wage earner, or, if you’re of an especially Marxist mindset, a proletarian). If you own stocks, bonds, or a business, you’re a capitalist. (Yes, holdings in your IRA and 401(k) count.) Obviously, lots of people are both—they have both a job and investments. Each person has their own capital-labor split, depending on how much they earn from investing versus working.

On a national level, the capital-labor split in the United States was relatively stable for a very long time but in recent decades has been shifting. The labor share has been declining, meaning that workers have been getting less of the total income of the country and business owners have been getting more. Economists debate the reasons behind this but I’m here to argue that this disparity is about to get worse. Much, much worse.

In short, AI is capital that can substitute for labor. It can’t do everything that people can do—at least not yet—so it’s not a perfect substitute in every case. But most jobs don’t require that people do everything they’re capable of doing, and AI can in fact already outperform people at a large and growing number of jobs. And AI gets better.

What happens when labor simply isn’t needed as much? What happens when it isn’t needed at all?

We’ve all seen the charming lists of jobs that no longer exist: lamplighter, street clerk, rickshaw puller, plowman, elevator attendant, etc. Jobs can and do disappear. (What even is a street clerk?) But AI doesn’t even have to be capable of doing the whole job. If it can do only half of your job, making you twice as productive, the size of your team may soon be cut in half.

As demand for labor decreases, so do wages. As AI becomes capable of doing more and more of the jobs that people currently do, we’ll eventually reach a point where wages collapse and jobs become scarce.

Once your labor is no longer valuable, selling your labor (i.e., having a job) is no longer a feasible way to make a living. What’s a worker to do, you ask? Simple: don’t be a worker. Be a capitalist. Own capital. For most people, that means owning things like bonds, stocks, or even entire businesses. Earn your income from dividends paid to shareholders, or from profits made buying and selling these assets. Problem solved!

Admittedly, this prescription may sound a little tone-deaf, like I’m suggesting that the way to prepare for unemployment is just to “be rich”. And sure enough, as a general rule only the wealthiest people today can afford to live comfortably off of their investments. But I’m not saying that.

I’m saying you need to shift the source of your income from labor to capital and to be prepared for the income you earn from capital to be sufficient to meet your needs. There’s a window of time in which you can do that, but it’s closing. Once you can no longer be paid to work, the window will have closed. (This is exactly the situation workers approaching retirement have always faced, just on a foreshortened timescale.)

As you shift your personal capital-labor income split away from labor and towards capital, you’ll insulate yourself against wage deflation and unemployment.

This is the advice I’ve given my kids, who are currently in college. Do I sound like an AI doomsday prepper? Um, maybe. Do I still think it’s good advice? Yep. And if the worst-case outcome is that they don’t drive as nice a car, live in as big a house, or take trips as fancy as do their friends early in their careers and instead wind up with a beefy investment portfolio, well, maybe that’s not so bad.

I strongly believe that AI will have a profound impact on the market for human labor. First, it will make some people much more efficient at their jobs, so fewer people will be required to do the same work, leaving the rest looking for other employment. Second, it will outright replace many workers. The end result will be widespread unemployment. But lots of smart people disagree.

Many people—including many labor economists—claim that even if AI destroys some jobs, it’s bound to create new ones. After all, they say, this is how new technology has always worked. (I went back and looked at my labor econ textbook but couldn’t find the part where it says that past performance is any guarantee of future returns, but maybe I just missed it.)

The problem is this: AI gets better. AI can do things today it couldn’t do yesterday, and will do things tomorrow that it can’t do today. Some of the smartest people in the world fueled by hundreds of billions of dollars in investment are pushing the frontier forward at a mind-boggling pace. In economics terms, AI is continually gaining absolute advantage over humans at one task after another. And there is no reason to believe that progress will stop any time soon.

So what does this have to do with burning buildings? Simple: if you find yourself in a burning building you can avoid the fire by retreating to higher floors—but that strategy only works for so long. Eventually the fire will catch up.

Imagine the job market as a skyscraper, with the ground floor housing low-skilled jobs, and increasingly higher-skilled jobs on each floor as you go up. If your job is threatened by the fire of AI, you can retreat to higher floors by reskilling, upskilling, or cross-skilling. (Yes, these are real words, even if the iOS spell checker doesn’t recognize them.)

You may be able to escape the fire for a while. But as AI continues to gain capability across a wide spectrum of domains and the fire spreads higher and higher, the floor to which you’ve retreated will eventually be threatened. (Maybe some jobs will never be done by AI. Maybe some rooms in the building are perfectly fireproof. But I wouldn’t bet my life, whether literal or economic, on it.) So again, up you go.

At some point you’ll reach the highest floor you can attain. Maybe you’ll reach the highest floor in the building. Maybe you’ll simply max out your personal potential. Either way, you can no longer go up. And yet the fire will not stop. Ultimately you’ll need to either put out the fire or find a way out of the building.

In The Towering Inferno, Steve McQueen and Paul Newman use plastic explosives to rupture water tanks on the roof of the burning skyscraper they’re in, extinguishing the fire. But the race dynamic of the AI market makes it unlikely that anyone will voluntarily put out the AI fire.

The fire may be contained to an electrical closet in the basement for now, but eventually we’ll all need to find a way out of the building—or the labor market. (Hint: in the battle of labor vs. capital, be capital. But that will have to wait for another post.) Good luck!

“I am starting to worry about this for my kids! My oldest is about to start high school and I have no idea what jobs are going to be available to him in 4-8 years - how do we even advise them on what to study?”

This is something I’ve been thinking about seriously for a couple of years given that (a) labor market disruptions due to AI now seem inevitable, and (b) my kids are sophomores in college. Here’s what I’ve settled on for an answer:

• In the short term, you’ll be competing for jobs with other people. Some people will be better at using AI tools than others. Make sure you’re among the best at using these tools.
• In the long term, you won’t be competing for (most) jobs at all; they’ll be done by machines. Make sure you have a source of income other than selling your labor.

What this means in practical terms is pretty simple:

1. Get really good at using AI tools while you’re still in school. Not to cheat, obviously, or to do the work for you, but to become much more productive. Understand how to use the technology to do research, write better, communicate better, and manage your time. Become an expert at using not only general-purpose LLM’s like ChatGPT and Claude but also the tools specific to your field, whether that’s Cursor for software development, Microsoft Copilot for knowledge work, Adobe Firefly for creators, Notion for project management, etc.
2. Use your talents and skills to get a high-paying job.
3. While you’re young and have relatively low expenses, invest as much of your paycheck as possible in stocks of companies that make or benefit from AI. Depending on your personal appetite for diversification, this might mean investing in specific companies like Apple, Microsoft, and Google, a basket of AI stocks like the so-called Magnificent Seven, or the broader market. Your goal is to build an investment portfolio that can generate income for you when the jobs are gone.

This plan may sound extreme, but I honestly believe it’s optimal. Of course there’s another path: learn to do something for a living that machines won’t be able to do. But then you face what I call the Burning Building Problem.

Desperate though this all may sound, there is some good news. First of all, you have time. There are still plenty of good jobs out there and they’re not going away tomorrow. New technology takes time to diffuse through the economy, so even if artificial superintelligence were commercially available today it would still take years for the labor market impacts to be broadly felt.

The other bit of good news is that many university professors seem to have gotten over their initial reaction to AI as a “cheating machine” and now in fact encourage their students to use AI tools in their work. This means students can become familiar with the technology as they complete their coursework. It doesn’t have to be relegated to clandestine use or nights-and-weekends projects. Plus, many AI companies offer their products to students at a discount or even free.

And if the AI future that I predict doesn’t come to pass? Well, you’ll still be among the most highly-skilled and sought-after candidates in the workforce.

We learn at a snail’s pace, taking years—decades!—to become professionals. We process rules slowly and have no way to speed that process up without offloading the work to machines. We have personal preferences for some types of work over others and demand to be compensated for performing less desirable jobs. We insist on ethical treatment and respect. We usually only work 8 hours/day and 5 days/week. We take vacations. We’re ruled by unconscious and often counterproductive urges originating from our lizard brains. We age and get sick. We die.

When you look at some animals, you wonder how they survive in the wild. The panda is so clumsy. The tortoise can’t survive being flipped on its back. The butterfly is so fragile. The June bug is completely uncoordinated.

We are a beautiful bird species living on a remote island with no economic predators except each other. But our best and brightest are working hard to bring invasive species of predators to our island.

I’m Charlie Wood, a retired startup founder and CEO with degrees in computer science and economics. I care deeply about how AI will shape the world we live in, and the one we leave to our kids. I’m fond of saying, “Anything that can be done using a computer will soon be done by a computer,” and it’s no overstatement. But the changes will go beyond that. Virtually every aspect of our lives will be affected by AI if they haven’t been already.

In my own life I use AI (mainly ChatGPT and Claude) in a wide variety of ways. I ask it passing questions, like where did La Jolla, California get its name? Or, how much should I tip a wine tasting host at a winery? Or, why do so many military pilots have mustaches? (Yes, I recently visited San Diego.)

I’ve found AI useful for interpreting test results from my doctor, for advice on diagnosing and fixing automotive, electrical, plumbing, and appliance issues, for product recommendations, for household budgeting, for identifying animal tracks, for learning new card games, for language translation and travel planning, for … well, for just about everything.

And that’s just my personal life. Professionally, if you’re not using AI to help you with your job, you’re almost certainly putting yourself at a severe disadvantage.

So AI is a big deal. So what? Why am I writing about it and why should you care? Because the ground is going to shift below our feet and we all need to be prepared.

AI is set to massively disrupt the labor economy. It won’t replace every job, not even every white collar job, at least not yet. But it will make some people so much more productive at their jobs that businesses will require dramatically fewer people—and that means widespread unemployment.

In the short term, the competition isn’t between people and AI, it’s between people using AI and people who aren’t. And the latter are hosed. In the long term, the battle is between capital and labor—or more specifically between people who make money from their capital and those who make money from their labor. And again, the latter will be hosed.

I’ll be writing much more about all of this on my new blog. A play on the (probably apocryphal) Chinese curse and a classic newspaper name, The Interesting Times is about navigating a world in flux. You can find it at interestingtimes.blog.

I hope you’ll join the conversation.