That feeling is mostly gone now.

Not because I’ve lost the skill. But because my day looks completely different. I open a task, I describe what I want — the shape of the solution, the constraints, the edge cases I’m already anticipating — and something else produces the code. My job is then to read what comes back and find the problems: the duplicated logic buried three functions deep, the subtly wrong assumption about how an external API behaves, the place where the implementation technically works but misunderstands what we were actually trying to do.

I am, in a word that didn’t exist in my job description a few years ago, a director.

I’m not alone. The 2026 ETS Human Progress Report found that workers now estimate that 32% of their tasks involve directing AI tools, rising to 38% among younger colleagues, and that this is expected to cross 50% within two years. These numbers come from people across all industries. For developers, I’d guess the curve is steeper and the timeline shorter.

Thirty-two percent. And rising.

The Assumption Most of Us Got Wrong

When tools like GitHub Copilot first appeared, most developers I know, myself included, filed them under “better autocomplete.” Impressive, occasionally useful, fundamentally a next-token prediction engine dressed up nicely. You still had to know what you were doing. It would fill in the boilerplate while you did the real thinking.

That framing made sense at the time. And it led us to reasonable-sounding conclusions: it still makes mistakes, so it’s not a threat to experienced developers; it’s a tool for juniors; my expertise is what catches the errors, so my expertise is what matters.

I’m not sure that framing fits anymore.

Here’s what I’ve noticed in practice. I’ll be working through a problem, describing what I need, and the response will connect two things I hadn’t thought to connect. Not complete my sentence. Connect concepts. It will surface an architectural pattern from a domain I rarely work in that elegantly fits the problem. It will flag an implication of a design decision that I’d eventually catch, but not immediately.

Autocomplete predicts the next word. What I’m working with now appears to reason about the problem.

I want to be careful here — it still gets things wrong, sometimes badly. It still misunderstands context in ways a senior developer wouldn’t. My thirty years of experience is exactly what lets me spot the difference between a confident-looking answer and a correct one.

But the mechanism has changed. And the mental model most developers carry — glorified autocomplete, useful but bounded — is leading them to the wrong conclusions about what comes next.

The Fragility Hidden in the 60%

Here’s where it gets uncomfortable.

The same report found that 60% of workers say they feel pressure to adopt AI tools before they feel ready. 65% are using them primarily because they need to stay competitive.

I recognise that pressure. I’ve felt it too.

But there’s a difference between choosing to work this way and being pushed into it. When you’re scrambling to stay relevant, you adapt to the tools your employer provides, in the ways your employer needs, at the pace your employer sets. You’ve shifted your skill set, yes. But you’ve done it inside someone else’s system. The fragility hasn’t gone away — it’s just wearing different clothes.

The developer who masters AI-directed development inside a company is more valuable than the developer who doesn’t. That’s real. But they’re still a director of AI inside a structure they don’t control, dependent on a role that is itself narrowing.

Because here’s the honest question the 32% figure asks: if half your tasks already involve directing AI, and that number doubles in two years, then what exactly is the scope of the role you’re protecting?

I’m not being fatalistic. I’m asking the question I think we should all be sitting with.

The Window

I’ve been at this long enough to have watched several waves of disruption move through the industry. Each time, the people who navigated it well weren’t necessarily the most technically gifted. They were the ones who saw the wave a beat before it arrived and made deliberate choices about where they wanted to stand when it hit.

The wave is visible now. The 32% figure is already on the chart, moving in one direction.

There’s a window — not infinite, not guaranteed, but real — to build something more durable than a job title that will keep getting redefined. Income that doesn’t depend entirely on being the best human in someone else’s AI loop. That’s not a fantasy. It’s a practical question, and there are practical answers.

What those answers look like is what I write about in the paid edition of this newsletter.

Subscribe now

But for now: if you’re a developer, take a quiet moment and actually count. What percentage of your day is directing AI rather than writing code? If it’s already above 32%, you’re ahead of the average; and the question of what that means for your next 5 years deserves more than a passing thought.

The clock is running. But the window is still open.

A Verifiable Prediction

I’ll put a number on it, because vague warnings are cheap. The 2025 Stack Overflow Developer Survey found that 47% of professional developers now use AI tools daily, up from a figure that barely registered two years ago.

I predict that by the time the 2027 Stack Overflow Developer Survey publishes, that number will be at or above 80%.

Not “use or plan to use.” Daily. Check back in roughly twelve months, and we’ll see. I’m putting this at 75% confidence, very likely, but not certain, which is the only honest way to make a prediction about a world moving this fast.

— Olaf

A note before we begin: this is a Lab Notes edition, usually reserved for paid subscribers. I’m opening it up to everyone today. I hope you find it worth your time.

There is a question that serious investors ask before deploying capital into any asset. It has many formulations, but they all reduce to the same thing: Is this a wonderful business?

Warren Buffett refined the question over decades. Charlie Munger sharpened it. Phil Town operationalised it into five numbers you could actually calculate on a spreadsheet. The framework is powerful, proven, and genuinely useful — whether you’re buying shares in a listed company or considering something closer to home.

But there’s a dimension it doesn’t fully capture. And a British entrepreneur and author named Richard Koch spent a career trying to name it.

The 2x2 That Changes Everything

In the 1970s, the Boston Consulting Group developed a portfolio matrix that, at the time, was a genuine intellectual breakthrough. BCG’s insight was simple: not all businesses deserve the same treatment, and the way to think about them is along two axes.

The first axis is relative market share — not your absolute size, but your size relative to your largest competitor. Are you the dominant player in your market, or are you fighting for scraps?

The second axis is market growth rate — is the overall market expanding quickly, or is it mature and slow?

Plot these two dimensions and you get four quadrants:

A Star is a business with high relative market share in a high-growth market. It is, as Koch argues in The Star Principle, the rarest and most valuable thing in any market. It has wind in its sails, and it’s already winning.

A Cash Cow has high relative market share but operates in a slow-growth or mature market. It prints money reliably. It just isn’t going anywhere fast.

A Question Mark (sometimes called a Problem Child) has a low relative market share in a high-growth market. It might become a Star with enough investment, or it might not. It’s a bet.

A Dog has low relative market share in a slow or declining market. It is, with occasional exceptions, a trap.

Koch’s contribution, his sharpening of the original BCG thinking, was to make a bold, arguably extreme claim: You should only ever start or buy a Star. Not a Cash Cow. Not a Question Mark. Certainly not a Dog. Only a Star. The mathematics of compounding in a high-growth, market-dominant business are so favourable, he argues, that everything else is a distraction by comparison.

It’s a compelling argument. It’s also not quite the whole story. We’ll come back to that.

What a Star Actually Looks Like

Before we can evaluate Koch’s claim, we need to understand what a Star looks like in practice. The definition is deceptively simple. The attributes are harder to find.

Genuine market dominance. A Star isn’t just a successful business; it’s the leading business in its niche. This doesn’t require a globally recognised brand. A Star can be the dominant supplier of a specialist industrial component to a narrow industry, or the clear market leader in residential property management in a mid-sized regional city. What matters is that the leadership position is real, measurable, and recognised by customers. When people in that market need what you do, they think of you first.

A market that is actually growing. This is the dimension most buyers ignore. A dominant business in a shrinking market is a Cash Cow — sometimes a very good one, but one with a structural ceiling. A Star’s market is expanding underneath it. That expansion does something powerful: it makes the business’s job easier. New customers are being created by the growth of the market itself. The Star captures a disproportionate share of them because it’s already the trusted leader.

A defensible position. Koch’s Star isn’t just temporarily ahead; it has something that makes staying ahead easier than catching up. This is where the framework converges directly with Buffett and Munger’s concept of the economic moat. The moat might be brand trust, switching costs, proprietary data, network effects, regulatory advantage, or simply scale that competitors can’t match. Without defensibility, dominance is temporary.

It helps to have a concrete example. Consider ZEISS SMT — the semiconductor division of the German optics company Carl Zeiss, privately held and headquartered in Baden-Württemberg. ZEISS makes the mirrors at the heart of EUV lithography machines — the technology used to manufacture the world’s most advanced chips, including the AI chips powering the current revolution. Conventional glass simply absorbs EUV light, which operates at a wavelength of 13.5 nanometers. ZEISS makes the only mirrors in the world capable of reflecting it. Scaled up to the size of Germany, the largest imperfection on one of those mirrors would be a tenth of a millimetre. They are, by any measure, the most precise mirrors ever made, protected by more than 2,000 patents accumulated over thirty years of development.

The EUV light bounces off ZEISS mirrors roughly forty times before reaching the silicon wafer. Every advanced chip, every Nvidia GPU, every Apple processor, passes through ZEISS optics on its way into existence. ASML, the Dutch company that builds 100% of the world’s EUV lithography machines, has sourced its optics exclusively from ZEISS since the late 1980s. Nikon spent nearly twenty years attempting to develop a competing EUV system before abandoning the effort entirely.

That is a Star. Dominant position. Growing market. A moat measured not in brand loyalty or switching costs but in thirty years of irreplaceable accumulated knowledge. The kind of competitive advantage that doesn’t erode — it deepens.

Subscribe now

Improving economics over time. This is the payoff. A genuine Star tends to get more profitable as it scales, not less. Fixed costs spread across a larger revenue base. Brand equity compounds. The cost of acquiring customers falls as reputation does the work. In Koch’s telling, a Star that retains its position through a full market cycle tends to become extraordinarily valuable — because by the time the market matures and it becomes a Cash Cow, it’s an enormous Cash Cow, with margins and customer loyalty that competitors have long since given up trying to replicate.

Where Buffett Meets Koch

Here is where the frameworks become more interesting in combination than either is alone.

Buffett and Munger ask: is this a wonderful business with a durable competitive advantage, run by honest and capable people, available at a fair price? The framework is fundamentally about quality and value. It asks whether the business is good and whether you’re paying fairly for it.

Phil Town’s Big Five Numbers, Return on Investment Capital, Equity Growth Rate, EPS Growth Rate, Sales Growth Rate, and Free Cash Flow Growth Rate, operationalise the “wonderful” test. If a business has been compounding all five metrics consistently for ten years, something structural is working in its favour.

Koch asks a different, adjacent question: is this business in a position where the future is likely to be better than the present? His framework is fundamentally about trajectory and position. It asks not just whether the business is good, but whether the conditions for continued compounding are in place.

The synthesis — the thing neither framework says explicitly but both point toward — is this: a truly great investment is one where quality, value, and trajectory align. A wonderful business at a fair price in a growing market where it is the dominant player. That is the rarest thing in any market. That is a Star that Buffett would also buy.

In practice, these three conditions rarely align perfectly. But the framework gives you a way to trade them off consciously rather than by accident.

A Buffett/Town quality business in a mature market? That’s a Cash Cow. You might still buy it — at the right price, Cash Cows are excellent investments — but you should price it accordingly. You’re buying yield and stability, not trajectory. Don’t pay a Star’s premium for a Cash Cow’s future.

A high-growth market with a dominant player but mediocre underlying economics? That’s a growth trap. Koch would call it a Star on the surface; Munger would look at the return on capital and walk away. The frameworks check each other.

A business with wonderful economics, honest management, and a moat — but in a market that is actively shrinking? That’s See’s Candies. Buffett bought it and has described it as one of the great lessons of his career — not because it was a Star, but because it was a perfect Cash Cow available at a reasonable price. He wasn’t paying for growth. He was paying for the certainty of the cash flows. The lesson isn’t that Koch is wrong. It’s that a Cash Cow, correctly identified and correctly priced, is a legitimate and often excellent investment.

What the integrated framework rules out is paying a Star’s price for a Dog. And that, it turns out, is what most unsophisticated buyers do most of the time.

The Honest Difficulty

Koch’s framework has a challenge that he is refreshingly candid about: Stars are rare, and they are rarely for sale.

A business owner who has built a genuinely dominant position in a genuinely growing market is usually not selling. They’re compounding. The businesses that appear on broker listings tend, for structural reasons, to skew toward Cash Cows and Dogs — businesses whose owners are tired, retiring, or have hit a ceiling they can’t see past. Occasionally a Question Mark appears, priced on hope. Occasionally, very occasionally, something that looks like a Star surfaces — usually because the owner has personal reasons to exit that have nothing to do with the business’s prospects.

This doesn’t make the framework useless in a broker context. It makes it more valuable, because it sharpens the question you’re asking. You’re not running through a due diligence checklist. You’re asking: where does this business sit on the matrix, and am I being offered it at a price that reflects that position honestly?

A Cash Cow at a Cash Cow’s price can be a very good investment. Especially if you can see a path, operational, strategic, or market-driven, that the current owner has missed.

— Olaf

Before you go: one practical tip

Nobody (probably) taught you that your tools have a maintenance lifecycle. That’s not your fault. It’s just a gap.

The tool you own and the tool you maintain are two different tools.

Here’s the one that affects almost every household in the country, costs nothing to fix, and will make you feel slightly embarrassed once you know it: your kitchen knives are almost certainly blunt, and have been for years.

Not blunt like “a bit dull.” Blunt like a butter knife. The kind of blunt where you’re applying pressure rather than letting the edge do the work, where tomatoes collapse before they’re cut, where you’ve quietly concluded that you’re just not much of a cook. You’re not a bad cook. You have a bad edge.

A decent whetstone costs $30–$50 at any hardware store. There are two sides: coarse for restoring a damaged or very dull edge, fine for finishing. You hold the blade at roughly 15–20 degrees to the stone (closer to flat than you’d expect), and you draw it across in smooth strokes, alternating sides. Ten minutes of YouTube, ten minutes of practice, and you will have a knife that frightens you slightly with how well it works. That’s the correct outcome.

The thing that strikes me about this, and why it belongs in a newsletter about navigating an AI-disrupted world, is the invisibility of bluntness. You don’t notice the degradation because it’s gradual. You slowly adapt your behaviour to the limitation until it feels normal. You push harder. You saw instead of slice. You blame the tomato.

The lawnmower blade is the same story, scaled up. A blunt mower doesn’t cut grass — it tears it. Torn grass goes yellow-brown at the tips, becomes stressed, and is more susceptible to disease. Most New Zealanders with a lawn have been tearing their grass for years and attributing the results to soil quality, weather, or bad luck. The fix is to remove the blade once a season (socket set, 10 minutes), run a metal file along the cutting edge, and reinstall it. That’s it. The lawn responds within a fortnight.

The principle, once you see it, is everywhere: the tool you own and the tool you maintain are two different tools. Most of us own the first kind and wonder why the results are mediocre.

I’m not suggesting you become a sharpening obsessive — though apparently that rabbit hole is very deep and has its own Reddit communities. I’m suggesting that the next time something isn’t working as well as it should, the question worth asking first is: when did I last maintain this?

It’s a question that turns out to apply well beyond the kitchen drawer. And, you’ll save a heap of cash, not paying someone else to maintain your tools.

Nothing in this article constitutes financial or investment advice. Richard Koch’s Star Principle framework is discussed for educational purposes. As always, do your own thinking.

I’ve been watching AI capability announcements for long enough that I have a baseline for what “impressive” looks like. But the period from roughly mid-March to the end of May 2026 felt different — not because of any single release, but because of the cadence. The frontier labs didn’t take turns anymore. They released simultaneously, iterating within weeks of each other, each one pushing a benchmark the others had to scramble to beat.

This is the edition where I try to tell you what actually happened — and what it means for the people this newsletter exists to serve.

A note before we go further. This newsletter exists because I think the worst thing you can do right now is look away. I’m not writing from a position of safety — I’m writing from inside the problem. If you follow the paid edition, you’re watching one person, in real time, trying to figure out how to escape the gravitational pull of AI-driven job displacement before it becomes inescapable. Some weeks that looks like progress. Some weeks the black hole just seems bigger. Either way, I’m reporting back honestly. That’s the deal.

Subscribe now

The Model Release Log Reads Like a Sprint Finish

In the twelve weeks ending in late April, the three frontier labs produced what may be the densest capability improvement window in the short history of generative AI.

Google’s Gemini 3.1 Pro arrived in February, posting a score of 94.3% on the GPQA Diamond scientific reasoning benchmark, more than double its predecessor’s score on ARC-AGI-2, the notoriously difficult visual reasoning test. A single model generation, roughly doubled on a hard benchmark. That alone would normally be a headline story for a month.

It wasn’t the story for a month. Because in March, OpenAI shipped GPT-5.4, which its own CEO described not as a smarter version of the last model but as something structurally different: a single model that credibly leads across coding, computer use, reasoning, and general knowledge work simultaneously, and all without specialist variants. Six weeks later, GPT-5.5 followed. OpenAI’s president Greg Brockman called it “a new class of intelligence.” As an illustration, he cited a mathematics professor who used it to build an algebraic geometry application from a single prompt in eleven minutes, work that would previously have taken weeks of manual coding.

Meanwhile, Anthropic released Claude Opus 4.7 on April 16th. DeepSeek V4 Preview dropped twenty-four hours after GPT-5.5. Gemini 3.1 Pro, Llama 4, Qwen 3, and Gemma 4 all landed within the same six-week window.

At some point, “keeping up with AI developments” became a full-time job for people whose actual job was something else entirely.

The Benchmark That Actually Matters

I want to draw your attention to one data point in particular, because it cuts through the noise better than anything else I’ve seen.

In late 2025, OpenAI published a benchmark called GDPval. Unlike most AI benchmarks, which measure abstract reasoning or academic-style tests, GDPval was designed to answer a specific question: can AI do the work that the economy actually pays people to do?

The benchmark covers 1,320 tasks drawn from 44 real occupations across the top nine sectors of the US economy. Critically, every task was constructed from actual work products created by human professionals with an average of 14 years of experience. The graders were the same professionals, comparing AI and human outputs blindly, i.e., they didn’t know which was which.

Early models were already approaching expert-level quality on many of these tasks. By the time GPT-5.4 shipped in March, it was scoring 83% on GDPval, placing it at or above the level of human experts on economically valuable tasks across a wide range of knowledge work occupations.

One AI researcher, Julian Schrittwieser, a key contributor to AlphaGo and AlphaZero at Google DeepMind, extrapolated the consistent performance improvement trend and issued a straightforward prediction: by mid-2026, models will be capable of working autonomously for full eight-hour work days. By the end of 2026, at least one model will match expert-level human performance across many industries. By the end of 2027, models are predicted to frequently outperform human experts on many tasks.

This is not a fringe view. It is a straightforward extrapolation from a line on a graph that has been moving consistently in one direction.

The Adoption Gap: More Alarming Than It Sounds

In March, Anthropic published what I think is the most important piece of AI-and-labour research so far this year.

The study, authored by economists Maxim Massenkoff and Peter McCrory, introduced a new way to measure AI’s labour-market impact. Rather than relying on surveys or theoretical exposure models, they used actual usage data from Claude and real work tasks being performed to map both what AI can do and what it is doing.

The headline finding was this: actual AI adoption is a fraction of what AI tools are feasibly capable of performing.

AI can theoretically cover most tasks in business and finance, management, computer science, law, and office administration. In most of those sectors, actual adoption — measured by real usage — is still a small slice of that capability. The researchers found no material difference yet in unemployment rates between workers in AI-exposed roles and those in less-exposed positions.

Peter McCrory, Anthropic’s head of economics, put it plainly: “We’re seeing little evidence of widespread job displacement so far.”

I want you to sit with the strange ambiguity of that sentence, because it is doing a lot of work. It is both reassuring and quietly terrifying, depending on how you read it.

If the gap between capability and adoption is large, as Massenkoff and McCrory found, then we are not witnessing displacement in real time. We are watching the loading screen. The displacement has not happened yet, not because AI can’t do the work, but because organisations haven’t yet fully reorganised around the fact that it can.

That reorganisation is the thing to watch. It is not a technological event. It is a management and procurement event. It happens when enough CFOs review their headcount and AI tool budgets in the same spreadsheet.

Who Is Already Feeling It

The absence of mass unemployment numbers does not mean nothing is happening. It means the distribution is uneven, and some groups are already on the receiving end.

Workers aged 22 to 25 are showing a 16% fall in employment in AI-exposed roles. That is the first cohort to have their early career formation coincide with the emergence of frontier AI capabilities. They are not being laid off in large numbers. They are simply not being hired in the first place. Entry-level roles are contracting. The jobs that used to be the bottom rung of a white-collar career ladder are the first to go, because they are the jobs most easily replicated by a capable AI with light human oversight.

A study from the IMF found that for occupations highly exposed to AI with limited scope for human complementarity, employment levels are 3.6% lower in regions with greater AI skill demand than in comparable regions five years after those skills appeared. The shape of this is: first the roles compress, then the headcount contracts, and the full labour market signal emerges years after the technological shift that caused it.

Meanwhile, ManpowerGroup’s 2026 Global Talent Barometer found something that I find deeply telling: regular AI usage among workers jumped 13% year-on-year to 45% of the workforce, but confidence in using technology fell sharply by 18%. Workers are being pushed into AI tools they are not confident using, while 43% fear automation will replace their jobs within two years. The same survey found 64% of workers planning to stay with their current employer despite feeling uncertain about the future, what ManpowerGroup called “job hugging”: clinging to stability rather than adapting, because the cost of getting the adaptation wrong feels too high.

The Skills Divergence Is Already Happening

Anthropic’s March report had a second finding that tends to get less coverage than the headline-displacement numbers, but which I think matters more to the people reading this newsletter.

While widespread displacement remains limited, the data reveals a growing divergence between early AI adopters and everyone else. Workers who are mastering AI tools are pulling ahead of their peers. The gap is widening. It is not theoretical; it is showing up in productivity metrics and career trajectory data.

This is the part that actually applies to you, right now.

The window in which AI proficiency is a differentiator, rather than table stakes, is not infinite. At some point, “knows how to use AI” stops being an edge and starts being the minimum bar for staying employed in knowledge work roles at all. We do not know exactly when that transition happens. But the benchmark data, the adoption curves, and the model release cadence all point toward it happening faster than most institutions are prepared for.

What I’m Watching Next

Two things that will tell us a great deal about the second half of 2026.

First: GPT-6 is expected sometime between May and July, with Claude 5 (internally codenamed “Fennec”) targeting a May-to-September window. If those models represent qualitative leaps rather than incremental benchmark improvements — the kind of shift that GPT-4 was in early 2023 — the adoption curve will steepen rapidly. Organisations that have been taking a “watch and wait” approach to AI integration will find the gap between what they’re doing and what their competitors could be doing suddenly and visibly uncomfortable.

Subscribe now

Second: watch the enterprise software renewals. The most honest signal of AI-driven displacement is not layoff announcements; those lag reality by 12-18 months. The honest signal is the software procurement data: which vendors are growing, which are contracting, and what the headcount implications of those decisions turn out to be. When a company replaces a team of analysts with an AI-powered data platform, the story doesn’t appear in the headlines. It appears six months later, in the headcount line of a quiet earnings call.

The intelligence factory is open. It’s running three shifts. The question is not whether it will change your industry. It’s whether you’ll see it coming while there’s still time to do something about it.

— Olaf

If this resonated, share it with one person in your network who is still treating AI as a future problem. The data suggests it is a present one.

There’s a particular kind of investor self-deception I want to talk about today, because I’ve practised it myself and I suspect you have too.

It goes like this. You read a value-investing book — say Phil Town’s Rule #1 — and come away with a tidy checklist. Five growth rates. Four “M”s. A formula for intrinsic value. It feels like a recipe: plug in the numbers, turn the crank, out comes a verdict. The promise is that judgement can be outsourced to arithmetic.

It can’t. But the arithmetic is still worth understanding deeply, because the point of the numbers isn’t to hand you an answer — it’s to force the right questions about a business. And once you understand what each one measures, you find the framework doesn’t only work for listed shares. It works just as well when you’re standing in a café you’re thinking of buying, or any bricks-and-mortar business where someone wants real money for a real cash flow.

So today: what the “Big Five” growth rates are actually telling you, how debt fits in (Town is oddly quiet on the most dangerous and most powerful lever in the picture), and how the numbers connect back to the Four Ms — the qualitative side that stops the arithmetic from lying to you.

I’ll use a worked example whose numbers come out as a glorious mixed bag — because a company that passes every test cleanly teaches you nothing about what a failing number looks like, or how to tell a real red flag from a misread one. The company is Domino’s Pizza (NASDAQ: DPZ): a brand you know in your bones, and a business far stranger under the bonnet than the pizza box suggests.

The usual warning: I hold no position in Domino’s, none of this is advice, and the piece is about method, not whether you should buy the stock. Big Five figures are ten-year measures from my own screen, cross-checked against the public record; valuation figures are as of writing. I may have got something wrong; if I have, I’ll own it.

The Big Five, briefly

Town’s system rests on five numbers, each a compound annual growth rate (CAGR) measured over ten years, then five, then one — so you can see whether the business is accelerating or fading. The five:

1. ROIC — return on invested capital

2. Sales (revenue) growth

3. EPS (earnings per share) growth

4. Equity / book-value-per-share growth

5. Free cash flow growth

The rule is brutally simple: he wants each compounding at 10%+ a year, consistently, over a decade. Consistency matters as much as the level — a business that grew 40%, 2%, 30%, then -5% is telling you something very different from one that ground out 12% every year, even if the averages match.

Here’s what most summaries skip: each number answers a different question, and they only mean something together. One at a time, then — because the why is where the understanding lives.

Subscribe now

ROIC — “Is this actually a good business?”

Return on invested capital is, to my mind, the single most important number Town asks for — and the one people most often skip, because it’s the fiddliest to calculate. For every dollar of capital tied up in the business — equity and debt — how many cents of profit does it throw off each year?

A business earning 20% on invested capital is a machine that can turn capital into more capital, again and again. One earning 4% is barely beating the bank. ROIC is the closest thing in finance to a measure of quality: it tells you whether the company has something special — a brand, a network, a cost advantage — or is just shovelling capital around for thin margins. Why ten years? Because one good year is luck. A decade of high ROIC is a moat you can see from orbit.

Sales growth — “Is the world buying more of this?”

Revenue growth is the rawest signal of demand. It sits at the top of the income statement, before all the accounting choices that muddy everything below it. If sales aren’t growing, nothing else can grow for long without financial engineering. It’s the least manipulable of the five and, for that reason, the one I trust first.

Read more

“You’re forgetting Jevons paradox. When something gets cheaper, we use more of it. Demand for cognitive work will explode. Everyone will be fine.”

I want to take this seriously because the people making this argument aren’t stupid, and the paradox is real. But “Jevons will save us” has become a kind of intellectual comfort blanket, pulled out whenever the unemployment data gets uncomfortable, then put away before anyone asks how exactly it’s supposed to work. So let’s pull it out properly and look at it.

The case, stated fairly

William Stanley Jevons noticed in 1865, in The Coal Question, that as steam engines got more efficient, Britain burned more coal, not less. Cheaper coal-per-unit-of-work meant coal was suddenly economic for applications that had been marginal before. Total consumption exploded.

The pattern repeats. Cheaper bandwidth didn’t kill the internet; it gave us Netflix, TikTok, and video calls with grandma. Cheaper computation didn’t end programming jobs; it gave us a software industry the size of a small continent. Cheaper air travel didn’t empty the airports; it filled them.

The optimistic application to AI: as AI makes cognitive work radically cheaper per unit, latent demand expands until total employment in cognitive fields holds steady or grows. Legal services that previously cost $400/hour become affordable to small businesses, individuals, the chronically underserved. Medical second opinions, tutoring, accounting, design — all the things people wanted but couldn’t afford — suddenly become accessible. The pie grows faster than the per-slice productivity gain, and human workers ride the expansion.

There’s even a beloved case study: ATMs. The conventional wisdom in 1990 was that ATMs would destroy bank teller jobs. Instead, teller employment grew for two decades after ATM rollout, as James Bessen documented in his work on automation and labour. Why? Because ATMs made branches cheap to run, so banks opened more branches, and tellers shifted from cash-handling to relationship work the machines couldn’t do.

It’s a good story. It’s even partially true. But it leans on three assumptions that are doing more work than the storytellers admit.

Assumption 1: Demand is highly elastic

Jevons only kicks in if there’s enormous untapped demand waiting to be unlocked by lower prices. Sometimes there is. The world genuinely wants more legal advice, more therapy, more medical diagnostics, more tutoring, more software systems — vast unmet need at current prices, gated almost entirely by cost.

But sometimes there isn’t. Does the world want 50x more middle-management memos? 100x more quarterly earnings analyses? 200x more enterprise sales decks? Demand for many specific knowledge-work outputs is bounded by the real-economy activity they serve. A company with $200M in revenue doesn’t need ten times more financial reports just because reports got cheaper. It needs the reports that match its actual operations. The supply curve shifts; the demand curve mostly doesn’t.

Cheap coal found new uses because coal is an input to making physical things, and we always want more physical things. Cheap memos don’t find new uses because memos aren’t an input to anything most people care about. This distinction matters, and Jevons-optimists routinely paper over it.

Subscribe now

Assumption 2: Expanded demand requires humans

This is the assumption that gets smuggled in without examination, and it’s the one I want to nail to the wall.

Jevons paradox is about demand for the task. It says nothing about demand for humans doing the task. Demand for coal exploded in the 1860s. Demand for human coal-shovelers… did not, particularly. The work got mechanised faster than demand grew.

The ATM story works because ATMs couldn’t do relationship banking, mortgage advice, or trust services. They were a partial substitute; they handled the routine, freeing humans for the complex. If a future ATM could do everything a teller does, including building rapport with the elderly widow opening her first investment account, the analogy collapses. Demand for “banking interactions” might still explode; demand for human bankers wouldn’t.

So when someone says “AI will expand demand for legal services,” the honest follow-up is: expand demand for human lawyers, or for legal services delivered by AI? These are very different questions, and the answer determines whether Jevons is your friend or your replacement. Coal demand growth was great for coal companies. It was less great for the men with shovels.

If AI is a complement to human cognitive labour, making us 10x more productive but still needed, Jevons probably helps. If it’s a substitute — capable of doing the whole job — the expanded demand flows to more AI, not more humans, and Jevons becomes the polite name for your unemployment.

The current evidence is mixed and shifting fast. In some domains AI looks complementary (radiology, where it improves diagnostic accuracy alongside humans). In others it looks increasingly substitutive (routine document drafting, customer service tier 1, code generation for well-specified tasks). The mix will determine which sectors get the friendly Jevons and which get the British coal shovelers’ Jevons.

Assumption 3: The transition is smooth

Even granting that Jevons works in the long run, the transition can still be catastrophic. Falk and Tsoukalas’ AI Layoff Trap paper (which I covered recently) makes this point sharply: firms can rationally fire workers ahead of AI capability, betting on tools that don’t quite exist yet, and the displaced workers don’t smoothly retrain into the new expanded-demand roles. A ten-year transition with 15% knowledge-worker unemployment is a generational disaster, even if year twenty looks fine.

We can already see this mechanism warming up. Challenger, Gray & Christmas recorded 1.2 million U.S. job cuts in 2025 — the highest annual total since the pandemic — with hiring plans at their lowest since 2010. AI was directly cited in roughly 55,000 of those cuts, less than 5% of the total. You can read that two ways: AI is barely a factor (the AI-washing critics’ read), or AI is the named tip of a much larger iceberg of “restructuring” and “market conditions” cuts that are functionally the same bet. Either way, the door to new hires is closing faster than the door to existing ones, exactly as Falk and Tsoukalas predicted.

And the speed matters. Previous Jevons-style adjustments unfolded over decades across narrow sectors: coal across the 19th century, ATMs across two decades, spreadsheets across one. Labour markets had time to reallocate. AI hits most cognitive work simultaneously over a few years. Even if every affected sector reaches a new equilibrium with healthy human employment, the simultaneous shock has no historical precedent to reassure us.

There’s also the distributional question. Cheap legal services helping small businesses are great, but they don’t help the paralegal who lost her job last March. Even when Jevons creates new work, it goes to different people in different cities with different skills. The aggregate statistics can look fine while individual lives are wrecked.

Where Jevons does help

I want to be fair to the argument, because it isn’t wrong, just oversold. There are sectors where Jevons will probably soften the blow significantly:

• Healthcare, where latent demand is enormous and regulatory & trust requirements keep humans firmly in the loop.

• Childcare and early childhood education are similar to eldercare. Physical presence is the product; latent demand is enormous (cost is the binding constraint for most families), and no one is sending their toddler to be cared for by robots running on GPT-6.

• Skilled trades that touch AI tooling, where physical presence is irreducible, and AI is a force multiplier.

• Advisory work with high accountability stakes — the kind where someone needs to sign their name, take the call at 11 pm, and be sued if they’re wrong.

These are the sectors I’d retrain into if I were 25.

Where it doesn’t

And where I wouldn’t:

• Routine cognitive work with bounded real-economy demand (most middle-office corporate roles).

• Work where AI is a near-complete substitute (a lot of document production).

• Anything where the value being delivered is information transformation rather than judgement, accountability, or relationship.

The honest frame

Jevons paradox is a useful concept. It is not a panacea, a guarantee, or a serious answer to “what happens to displaced workers.” It’s a partial mechanism that works in some sectors and not others, on timescales that may or may not match the urgency of the transition.

The framing I keep coming back to in this newsletter is: who captures the surplus? When AI makes a task cheaper, the savings go somewhere. Sometimes, to consumers (cheaper services), sometimes to firms (higher margins), sometimes to workers (higher wages for the humans still needed), and sometimes to AI providers (concentrated rent). The Jevons story implicitly assumes the surplus flows in a way that creates more human work. But “implicitly assumes” is doing a lot of lifting, and history shows the surplus can go anywhere.

If you find yourself reaching for Jevons as reassurance, ask the harder question instead: in this specific sector, with this specific AI capability, who captures the surplus, and does that flow create or destroy human work?

That’s a question with answers. They’re just not always the comforting ones.

— Olaf

Last week, we established something uncomfortable: the companies automating jobs away are acting rationally, the system has no self-correcting mechanism, and the one policy instrument that could work requires global coordination that has never been achieved for anything.

If you found that bleak, you were reading it correctly.

But the trap operates at the level of the whole system. That is not the same as saying every individual outcome is equally determined. The aggregate is heading somewhere most of us would rather it didn’t. The question is where you are standing when it arrives.

Assets Are a Hedge, Not an Exit

The instinctive response to economic disruption is to accumulate assets. Buy property. Hold equities. Build a buffer. This is not wrong — but it is incomplete.

The demand destruction the paper describes doesn’t stop at wages. It runs through asset values too. A world where knowledge workers have lost meaningful income is a world where consumer spending has contracted, corporate revenues have compressed, and the equity portfolios tracking those revenues have repriced accordingly.

Assets help at the margin. But the more durable question is income resilience — not how much you have, but how exposed your income stream is to the dynamics the paper describes.

Where the Frontier Moves Slowest

The paper’s logic tells you which income streams are most at risk. The automation distortion is worst in fragmented, competitive markets with measurable outputs: customer support, back-office operations, entry-level software, content production at scale. These are sectors where AI deploys cheaply and output is easy to verify.

Roles harder to automate profitably share a different profile — embedded in complex human systems, where integration costs are high, output is difficult to measure, and being wrong is expensive in ways that matter to the people paying. They require not just task execution but the ability to navigate the situation around it.

The frontier will reach these roles. It is not there yet. And “not yet” is what you are working with.

The Geography of Cost

There is a response to income compression that rarely appears in serious discussions because it sounds too simple: spend less. Not as austerity — as a structural choice about how much income you actually need, and whether your current cost of living is the most efficient way to maintain what you value.

Subscribe now

As recently as last week, my wife and I looked after two golden retrievers on well-to-do Waiheke Island for a week while their owners were away. The dogs needed to stay somewhere familiar, which meant we had to stay at the owners’ holiday house. We love the dogs. We said yes.

What followed was one of the better working weeks either of us can remember. Coastal walks every morning, a beautiful property, and a full working week without a single day of leave. Productivity, if anything, was higher than in the office. The owners got reliable, affectionate care. Nobody paid anybody anything. Everybody won.

This is cost-of-living arbitrage in its most human form. The conditions that made it possible — remote-capable work, a relationship of trust, a willingness to say yes — are conditions that can be cultivated deliberately. If income compresses, the people best placed to absorb it are not necessarily those who earn the most. They are those whose cost structure gives them the most flexibility, who have reduced their dependence on income-intensive living without reducing what they actually value. That is a form of resilience no market crash can reprice.

The Audience of One Thousand

The paper shows that the automation trap is most destructive in fragmented competitive markets, where each actor externalises the cost of their choices onto the whole. Employment sits inside that structure. An employer automating your role is not malicious. They are responding rationally to a competitive environment that punishes restraint.

A direct exchange of value between a writer and a paying audience sits outside that structure. The writer internalises both the value created and the relationship that sustains it. There is no competitor undercutting the relationship with cheaper AI, because the relationship is not a commodity. You are not selling approximable output, but a specific perspective, earned over time, delivered with accountability.

This does not make audience-based work easy or universally accessible. But understanding why it is structurally resilient, not just that people say it is, changes how seriously you invest in getting there.

The Coordination That Actually Works

The paper’s policy solution requires global coordination that has never held together for anything of comparable complexity. But coordination doesn’t only happen at that scale.

The firm-level trap exists because each actor is too small to internalise the damage they cause, and too exposed to defection to restrain themselves. A community of people with complementary skills, shared resources, and genuine mutual obligation faces a different structure: small enough that each member’s contribution is visible, trusting enough that defection carries real cost.

Mutual aid societies, guild structures, and cooperative models all emerged in periods of economic disruption where individual exposure was high and system-level protection was absent. The knowledge-work version is still being invented. But the logic is as old as the problem.

The people who navigate the next few years best are unlikely to be isolated individuals who correctly predicted the endpoint. They are more likely to be people embedded in relationships of genuine mutual value — professionally, geographically, personally — who have something to offer and people to offer it to.

The system-level trap is real. Individuals cannot escape the aggregate outcome by clever positioning. But the aggregate outcome will arrive unevenly, with significant variation in how different people experience it. That variation is not random. It correlates with the choices above — made now, while the window is still open.

Understanding why is the beginning. The rest is execution.

Two recent editions of this newsletter told the stories of two investments I identified and didn’t make. One was a London-listed fintech with a durable competitive moat. The other was a distressed bond with a six-month countdown clock and extraordinary return asymmetry. Different instruments, different analysis, different failure modes. Same outcome: I didn’t act when I should have.

This edition isn’t a retelling. It’s the framework underneath both misses, and the single mental model that, had I internalised it properly, would have made the decision easier each time.

The Four Ms

Phil Town’s Rule #1 investing is Buffett and Munger made legible. The central claim is simple: you don’t need to be a genius to invest well. You need to buy wonderful businesses at fair prices, and then leave them alone.

Town distils this into four filters. Run any business through all four and very few survive. That’s by design. The framework is a culling mechanism.

Meaning

Do you understand the business?

Not what it does, but what it does. How it actually makes money. Why customers keep coming back. What gives it pricing power when competitors push back. Whether the revenue model holds up under pressure.

If you can’t explain it clearly to someone with no finance background, you don’t understand it well enough to own it. This isn’t about intelligence. It’s about the quality of your conviction. Fuzzy understanding produces fuzzy decisions. When a stock drops 30%, and eventually they all do, you need to know whether that drop is an opportunity or a warning. You can only know that if you genuinely understood the business before you bought it.

Share

Meaning is the first filter because it’s the most honest one. A lot of investing mistakes begin here, dressed up as something more sophisticated.

Moat

Does the business have a durable competitive advantage?

Buffett’s image is a castle surrounded by a moat that protects it from competitors. The question isn’t whether the business is good today. It’s whether it will still be good in ten years, after motivated competitors have thrown capital and talent at taking its market share.

Read more

In February 2026, Jack Dorsey announced that Block was cutting nearly half its workforce. AI had made those roles unnecessary, he said, and within the year, most companies would reach the same conclusion. He wasn’t issuing a warning. He was stating a fact; and then doing it anyway.

That tension sits at the heart of a paper published last month by economists Brett Hemenway Falk and Gerry Tsoukalas: The AI Layoff Trap. Its central question is disarmingly simple. If every firm can see that mass automation erodes the consumer base all firms depend on, why would any of them do it? And its answer is more uncomfortable than most economic research tends to produce: because they have no choice.

The Cliff Is Visible. That Changes Nothing.

Falk and Tsoukalas design their model with an unusual assumption: full transparency. Every firm can see exactly how automation maps into lost worker income and reduced consumer spending. There are no information failures, no hidden dynamics. The cliff is visible to everyone.

And every firm races toward it anyway.

When a firm automates, it captures the full cost savings from replacing workers with AI. But the demand destruction — the spending those workers would have done across the economy — falls across all firms, not just the one that pulled the trigger. Each firm bears only a fraction of the damage it causes. The rest lands on its rivals.

Restraint is therefore irrational. A firm that holds back while others automate suffers the demand loss from their layoffs without capturing the cost savings from its own. It loses twice. Even if every firm in an industry agreed that collective restraint would raise all their profits, the agreement would immediately collapse, because automating is the best move regardless of what everyone else does. There is no negotiation that changes this calculus, and no amount of foresight that makes holding back rational.

The paper is also precise about what happens as AI improves. Better AI does not shrink this gap. It widens it. Each business gains an additional incentive to automate beyond its competitors to capture market share, but at the industry level, these gains cancel out. Everyone runs faster to stay in the same place, and the damage to the collective demand base grows with every lap. Economists call it the Red Queen effect.

The Policy Toolkit Is Almost Empty

The paper works through the obvious responses with unusual rigour. Most of them fail for the same structural reason: they change how much profit firms make overall, but not whether automating a given task is more attractive than not automating it. That decision — made at the level of each individual task, by each individual firm — is where the arms race actually lives. Instruments that don't reach that decision don't stop the race.

Subscribe now

Universal Basic Income raises living standards for displaced workers but adds a fixed amount to the spending base without changing any firm’s per-task calculation about whether to automate. Capital income taxes scale the whole profit function down equally, but the decision to automate depends on the difference between options, not their absolute levels — tax rates cancel out of the equation. Worker equity participation narrows the gap but cannot close it: eliminating the distortion entirely would require workers to hold more equity than firms have to distribute. Private negotiation among firms runs into the same wall — automating remains the individually rational move regardless of any voluntary agreement, so no non-binding arrangement holds.

The one instrument that works is a Pigouvian automation tax: a per-task levy set equal to the demand destruction each automated position causes but that the automating employer does not currently pay. At the right rate, this aligns private incentives with social costs. It is the only instrument that operates on the right margin.

The problem is implementation. The paper’s model is a single closed economy. A tax imposed by one government does nothing for companies or workers outside it, and automation is mobile enough that unilateral taxation simply pushes adoption to untaxed jurisdictions. The economists note this themselves, pointing to multilateral coordination as the answer, analogous to border adjustments in carbon policy. Carbon policy has been under active international negotiation for more than three decades.

Who Actually Loses

There is a version of this story where the answer is obvious: workers lose, owners win. The paper’s formal result is more unsettling.

Over-automation is not a redistribution. It is a deadweight loss that harms both sides. Workers lose income through displacement. But business owners also end up worse off than if none of them had automated. The collective demand destruction is large enough that every company’s profit falls below what it would have been under restraint. The trap does not produce winners at one end and losers at the other. It destroys value that was available to everyone and distributes the loss across both sides of the ledger.

So What Do You Do?

It would be dishonest to end here with a tidy list of investments to make before the crash. The same demand destruction described in the paper would erode asset values alongside wages. If the wheels come fully off, more cash doesn’t buy much in an economy where consumers have stopped spending. That is worth saying plainly.

But the scenario is not binary. Between “everything is fine” and “complete collapse” lies a long, uneven compression; some sectors hollowed out faster than others, spending squeezed but not zeroed, opportunity distributed with increasing unevenness. In that world, how you are positioned relative to the disruption matters more than whether you have correctly predicted its endpoint.

The honest individual response is less about asset accumulation than about income resilience. The paper gives a clue about where the automation frontier moves slowest: the distortion is worst in fragmented, competitive markets — customer support, back-office operations, entry-level software. Roles embedded in complex human systems, where integration costs are high and measurable output is ambiguous, are harder to automate profitably. The frontier will reach them. It is not there yet.

There is also something structural worth naming. A direct exchange of value between a writer and an audience, which is what a paid newsletter is, does not sit inside the competitive dynamics the paper describes. It is not dependent on an employer or on a fragmented market in which each participant bears only a fraction of the damage they cause. Understanding the trap clearly, before most people do, is itself a form of positioning. Not because knowledge protects you from the aggregate outcome, but because the gap between where things are now and where they are heading is where asymmetric opportunities concentrate.

That gap is still open. It will not stay open indefinitely.

Next week: if the system-level fix requires coordinated policy that won’t arrive in time, what does individual-level coordination look like? The paper’s failure modes for collective action point, perhaps unexpectedly, toward some durable answers.

Sometimes the trade is right. Everything lines up. And then one thing doesn't.

This is a story about Synlait, a missed arbitrage, and the lesson that contingency planning is not optional.

The Setup

Synlait Milk is a Canterbury dairy processor. In better days, it supplied infant formula for A2 Milk’s China business and was the darling of the NZX. It had a good run. Management, as sometimes happens after a good run, concluded they could do no wrong.

They were wrong.

The company expanded aggressively — new plants, acquisitions, new product lines — and loaded up the balance sheet in the process. When market conditions turned, as market conditions always eventually do, the debt that had seemed manageable suddenly wasn’t. Infant formula demand out of China softened. Margins compressed. A large chunk of revenue was tied to a single customer. The half-year result to January 2024 showed a net loss of $96 million. Net debt had climbed to $559 million.

The market was not pleased.

The Bonds

Back in February 2021, Synlait had raised $180 million from retail investors and institutions via a listed bond: SML010 on the NZX Debt Market. The coupon was 3.83% per annum, paid quarterly — roughly one percent every three months. The bonds matured in December 2024. Face value: $1.00.

Yield, for those unfamiliar with the term: it's the return you actually get based on what you pay. If a bond pays 3.83 cents per year but you buy it for 50 cents, your income yield doubles — the coupon is fixed, but you paid less for it. But that's only part of the picture. If the bond is redeemed at face value on maturity — $1.00, regardless of what you paid — you also pocket the difference as a capital gain. Buy at 50 cents, get $1.00 back: you've doubled your money before a single coupon is counted. The more distressed a bond, the lower its price, and the higher the combined return if it survives. High yield is the market's way of pricing in risk. Very high yield is the market screaming.

By mid-2024, Synlait’s bonds were screaming.

The Crisis

The pressure point was a $130 million prepayment obligation to Synlait’s bank syndicate, due July 15, 2024. Synlait had also warned it was unlikely to meet three banking covenants by July 31. The fear was a covenant breach triggering acceleration — banks calling in their loans. The equity market agreed things were dire: Synlait shares hit an all-time low of 20 cents on June 27.

And then, for approximately two days in mid-June 2024, something unusual happened.

The yield on the SML010 bonds spiked to around 50%. There was also another figure — possibly a yield-to-maturity calculation displayed on the NZX debt market screen — that, in my recollection, approached something like 200%. I have held off on stating that second number with confidence. What I am confident about: the primary yield was 50%, briefly, and it was one of the most extreme readings I had ever seen on a listed NZ debt security.

The Arbitrage Case

Here is what I saw in the financial accounts.

Read more

The numbers behind it are worth sitting with. Job postings for software development roles on Indeed fell 71% between February 2022 and August 2025. US CS graduates now face 6.1% unemployment, higher than philosophy majors (3.2%) and art history graduates (3.0%). The degree that tech executives spent twenty years telling kids was a guaranteed path to six figures now ranks 7th in unemployment among all college majors.

Students noticed. Enrolment in CS programs fell at 62% of universities last year. Deposit volumes for CS degrees dropped more than 25% in a single application cycle. For context: deposits are the leading indicator. What’s in the enrolment data now is the caution of 2024. What those students will find when they graduate in 2027 is not something anyone can model yet.

The same trend is playing out here. In Australia, just 2.9% of incoming university students chose ICT-related degrees in 2026, down from around 9,750 students the year prior to roughly 7,686. The Australian Computer Society called it a “steady decline” with no obvious explanation. In New Zealand, 34% of companies have already slowed entry-level hiring, and 88% expect to do so within three years, the highest reported rate of entry-level job displacement from AI of any country in the IDC survey. New Zealand was also the country reporting the highest rate of roles removed entirely due to AI, at 53%.

Subscribe now

The pivot is to AI degrees. UC San Diego launched one; it’s the only UC campus where CS enrolment went up. MIT’s AI and decision-making major is now its second-largest. Dozens of universities are scrambling to repackage. The message to students: same building, new sign out front.

Here’s the problem with that bet. The students fleeing CS for AI degrees are responding to the 2023 labour market, the one where “prompt engineer” was a job title and AI literacy felt like a moat. The 2027 labour market, when those students graduate, is a different question entirely. If the models keep improving at their current rate, the gap between “knows how AI works” and “AI works” will close before the cohort crosses the stage.

There’s a structural irony worth noting. US universities doubled CS output over the last decade, from 52,000 graduates in 2013-14 to 113,000 in 2022-23, partly because the industry kept signalling demand. The same companies doing that signalling then overhired, laid off 260,000 workers in 2023 alone, and began replacing junior roles with AI tooling. The pipeline they filled is now flooding a market they’ve largely closed.

The student debt stays. The jobs didn’t.

Prediction ledger entry: UC Berkeley CS graduating class will remain below 500 in 2028 (currently tracking toward ~350 in 2027). Confidence: 70%. Score on Day 180.

Practical tip

If you’re advising someone choosing a degree right now — a kid, a junior colleague, anyone — the question to ask isn’t “what’s hot?” It’s “what’s the half-life?” Degrees take four years to complete and deliver value over a forty-year career. Any skill with a half-life shorter than the degree itself is a bad investment. CS as traditionally taught is in that category. The question for AI degrees is whether they’re different in kind, or just further along the same curve. Nobody honest can answer that yet. Hedge accordingly.

Sources: UC Berkeley Daily Californian · Computing Research Association CERP Pulse Survey · MARKETview · Federal Reserve Bank of St. Louis / Indeed data · National Center for Education Statistics · ACS Information Age · IDC/Deel AI at Work report (NZ data)

Standard stuff. Senior developer, market rate, fixed term.

Except it wasn’t standard.

I was running AI coding agents throughout — Claude, Gemini — and they were doing a substantial share of the mechanical work. Code generation, boilerplate, test scaffolding, repetitive transformations.

My honest estimate: I was producing roughly twice the output I would have without them.

The project is finished. The client was happy. I was paid the agreed rate.

Nobody renegotiated. Nobody got a bonus.

The productivity gain simply disappeared into the engagement.

I’ve been thinking about that ever since.

Where did the value go?

The client got a shorter project at market rates.

They captured the gain without asking for it. The contract was for a deliverable, not for hours of human effort. When the deliverable arrived faster, they paid less in total.

That’s it. That’s the whole story.

“Labour-saving devices don’t save labour. They transfer it — usually upward.”
— broadly attributed to economist Joan Robinson

This isn’t a complaint about one client. It’s a pattern.

Subscribe now

Every knowledge worker who adopts AI tooling is, right now, doing more and finding that the additional value flows directly to the buyer of their output. The market rate hasn’t caught up. Their leverage hasn’t increased.

The tools improved their output and, at the same time, made their premium rate harder to justify.

The rate pressure that’s coming

Here’s the part worth stating plainly.

I delivered double the output. A developer with five fewer years of experience and AI tooling could probably deliver 80% of what I delivered — for 20% less. A developer with ten fewer years and better prompt skills might deliver 60% for 40% less.

None of those transactions feels like a crisis.

Each one is just the market clearing.

But the cumulative effect is a rate floor that drops steadily — driven not by any explicit devaluation of skill, but by a quiet expansion of supply. More output per developer means the market needs fewer developers, or can pay them less, or both.

The maths is not complicated. The implications are.

This isn’t a prediction about the distant future. At current rates of AI capability improvement, it’s a description of the next two or three years — for developers first, then analysts, writers, paralegals, and financial modellers close behind.

The Last Safe Harbour

There’s a second thing I noticed on that project. It’s more uncomfortable.

AI coding tools are genuinely impressive at execution. Clean functions, edge cases handled, tests generated. Where they’re weaker — at least today — is in software architecture. The decisions that determine how a system is structured so that future changes are cheap and localised.

That’s actually the whole point of architecture. Not elegance for its own sake. Not intellectual satisfaction.

Good architecture exists for one reason: to make future changes fast and cheap.

A senior developer knows how to make one change in the right place. That skill takes years to develop.

But then I thought: does it matter?

If an AI can make 57 tricky changes to a poorly-structured codebase — flawlessly, instantly, at near-zero cost — then the value of elegant architecture starts to collapse.

The entire point of good structure is to make future change cheaper.

If future change is already cheap, you’ve removed the problem that good architecture was solving.

I’ve spent thirty years getting good at placing one change in the right place.

That skill may be obsolete within the next three years.

Not a decade. Not some abstract future.

Within the timeframe of this newsletter.

What this means for all of us

I want to be careful not to overclaim. One developer, one project, one data point.

But the pattern generalises.

The standard counter-argument is that productivity gains always create new jobs and new value categories over time. The loom, the assembly line, the internet — all displaced workers, all eventually created new types of work.

But here’s what’s different this time.

Nobody can name what those new jobs might be. Not even roughly. Not even as a category.

With every previous wave, you could point at the new thing. This time, the pointing hand is empty.

It is not a useful frame for the individual knowledge worker deciding what to do in the next eighteen months.

“It is not the strongest of the species that survives. It is the one most adaptable to change.” — commonly attributed to Charles Darwin

At the individual level, the question is simpler and harder:

The value you’re creating is increasing. The share of it you’re capturing is shrinking.

What are you going to do about that?

I don’t have a complete answer.

But I do see one viable path: escape the dynamic entirely. Build financial independence fast enough that your income is no longer hostage to a rate floor that only moves in one direction.

Don’t win the race to the bottom. Get off the track.

That’s what this newsletter is really about.

The gain was real.

The beneficiary wasn’t you.

Subscribe now

I want to be specific. Not therapeutic-specific — I’m not interested in processing feelings here. Financially specific. Because the numbers tell a story that is more useful to you than the narrative I was telling myself while it was happening.

How it started

March 13, 2020. New Zealand was four days from its first COVID lockdown. I was on a contract at a major Auckland tech firm, and I was frustrated. The same mistakes, everywhere. Bad naming. Tangled dependencies. Tests that tested nothing. Developers who were skilled but had never been shown a better way.

So I opened a Slack channel. I called it #olafs-daily-tips. Each day, one post: a clean code technique, a SOLID principle, a pattern for handling legacy code without breaking it. I thought I might get ten or twenty followers if I was lucky.

At its peak, the channel had just shy of 350.

When that contract ended in May 2020, I had a decision. I could file the tips away and go back to being a senior developer. Or I could find out if there was a business in what I’d been doing for free.

I set up a website. I kept writing. In the end: over 300 posts on software craft, covering everything from unit testing to clean architecture to component design principles. The kind of material I’d been wanting to exist for twenty years.

Word spread. By March 2021 — almost exactly a year after the Slack channel opened — I had my first paying client: a major NZ fintech. One workshop on TDD. A few thousand dollars. Proof of concept.

The good years

FY22 was the peak. Total revenue: $222,096.

The headline number flatters the underlying business a little. About $200,000 of it came from a single six-month embedded training contract with a large NZ tech company — essentially a coaching residency, working alongside their development teams every day. Strip that out and the standalone workshop business was around $22,000 that year.

But still. The business was real. The pipeline was building. I had:

• A major NZ fintech (repeat client, growing engagement)

• A Big Four bank

• A global payments network — a genuine showcase client, the kind that makes other prospects take you seriously

• An Australian software firm, trained in Bath, UK, via video call, up to midnight my time

That last one felt significant. Export dollars. International reach. I was no longer just a local training operation.

FY23 brought in $109,000. Down 51% from peak — but I didn’t read it that way at the time. The large embedded contract hadn’t repeated, which explained most of the gap. The global payments network was a new client. The Australian firm was back. The major NZ fintech was still booking.

Every business has down years. I’d had a down year. That was my read.

I pushed harder. LinkedIn posts, daily tips in an already crowded market, hoping to be spotted by a tech leader with budget. I built an online course on TDD using Teachable. I kept doing the midnight calls for international clients. I sent a monthly newsletter to a list of 43 people — every client manager I’d worked with, a handful of prospects, a few aspiring engineering managers. People with budget authority. People who had seen the work firsthand. I offered 30%, 33%, sometimes 50% off workshops and seminar bundles.

Six students enrolled in the Teachable course. Total.

Still, I thought: keep building, keep the quality high, the market will come back.

The six months

It didn’t come back.

October 2023 to March 2024. That’s the window. That’s when it became undeniable.

The newsletter kept going out. The discount offers kept going out. The 43 people on that list — the right people, the exact right people — read them and did nothing. Not a negotiation, not a “not right now”, not even a polite decline. Silence.

FY24 revenue: $47,000. Down 57% from FY23. Down 79% from peak.

By March 2024, I knew. Not that the business was struggling — I’d known that for a while. That the game was up.

Read more

Free post — The Next 1000 Days

Something is happening to your job.

You might not feel it yet.

But the numbers don’t care what you feel.

The names you recognise are already moving

Last year, US employers announced over 1.17 million job cuts — the highest total since the Covid-19 pandemic in 2020 (CNBC). Of those, about 55,000 were explicitly linked to AI adoption, according to Challenger, Gray & Christmas (Equitypandit). That sounds manageable, right? A rounding error in a 160-million-person workforce.

Here is the problem with that framing: CFO survey data suggests AI-related job cuts could be nine times higher in 2026 than in 2025 (Fortune).

Nine times.

The dam isn’t broken. But the cracks are spreading fast, and almost no one in charge is watching.

This isn’t speculative. The announcements are public. Amazon announced the largest round of layoffs in its history, slashing 14,000 corporate roles and explicitly citing AI as the driver. Salesforce CEO Marc Benioff confirmed he cut customer support from 9,000 workers to 5,000 — because AI was already doing up to 50% of the work (CNBC). Meanwhile Microsoft cut 15,000 jobs, and Oracle cut thousands more — both framing the layoffs as their businesses “doing more with less” through AI (TechRepublic). Microsoft reported quarterly revenue of $70.1 billion, up 13% year-on-year (TechRepublic). They are not struggling. They are thriving. They just need fewer of us to do it.

IBM’s CEO confirmed that AI chatbots took over the jobs of several hundred HR workers (CNBC). Not the jobs of the future — the jobs that existed last Tuesday.

Who gets hit first

If you are early in your career, the data is brutal. Job listings for entry-level corporate roles have declined 15% over the past year, and over the past two years, there has been a 400% increase in employers using “AI” in job descriptions, according to career platform Handshake (CBS News). Translation: they want one person who can direct AI, not five people to do the work.

The brunt of job losses is falling on entry-level roles — data entry, customer service, admin, help desk. Businesses are not just laying people off; they are eliminating the roles entirely (TechRepublic).

And it’s not only the obvious targets. In early 2026 alone, there were 32,000 job losses in technology firms. Nearly 55,000 job cuts were directly attributed to AI across 2025 — and over 75% of those happened after 2023 (AIMultiple).

The curve is accelerating, not flattening.

Here is what they’re not telling you

Every AI company will tell you they want to augment workers, not replace them. They will use words like “partnership” and “co-pilot.” That’s the $20-a-month story — the one where you buy a subscription and become superhuman.

The real story is different. The actual product these companies are building is your replacement, sold to your employer for a few thousand dollars a month. Not a tool that helps you — a service that removes the need for you. The incentives are not aligned with your job security. They never were.

There is no government committee managing this transition. No international body setting a humane pace. No adult in the room deciding that 3% annual job displacement is the safe speed, not 30%. The decision theorist Eliezer Yudkowsky put it bluntly: in a sane world, we would coordinate to ensure the pace of automation doesn’t outrun our ability to adapt. We don’t live in that world. We live in this one, where capability races ahead and the workforce scrambles to catch up.

Some economists will tell you new jobs will emerge. Historically, they’re right — the steam engine and the internet both created more jobs than they destroyed. But that transition took decades, and it happened when the technology had hard limits. Today’s models have no obvious ceiling. Anthropic’s new Claude Mythos model has already identified thousands of software vulnerabilities across every major operating system and browser — a capability previously thought to require elite, state-sponsored hacking teams (The Atlantic/DNYUZ). Anthropic considered it too powerful for public release (AOL/Anthropic). That’s the version they’re keeping back. Think about what it says about the versions they’re shipping.

On software engineering benchmarks, Mythos jumped from 80.8% to 93.9% — and on high-difficulty mathematical reasoning, from 42.3% to 97.6% (36kr). That is not incremental progress. That is a step change, in a single model generation.

Speaking personally: I use Claude Code daily in my development work. It is extraordinary. Some days I feel less like a software developer and more like a code reviewer — checking what the AI built, nudging it, redirecting it. My output is up. My team size requirement is down. I am not complaining. But I am paying attention.

What about UBI?

Universal Basic Income gets floated as the answer every time this conversation comes up. Maybe it happens, maybe it doesn’t — the political obstacles are enormous, and the timeline is anyone’s guess. But here’s the thing even the optimists miss: UBI isn’t a solution, it’s a consolation prize. The problem with depending on UBI isn’t the money. It’s the leverage.

Leverage comes from contributing something the world needs. When your income is a government transfer rather than a wage, you have no seat at the table — not in your company, not in your industry, not in the economy. You become a beneficiary, not a participant. That is not a position of strength. That is the definition of dependency.

The only durable answer is to build assets that earn whether or not you are employed. Investments that compound. Income streams that don’t require your employer’s approval to exist. Financial independence is not a retirement goal for the comfortably-off. It is quickly becoming the only genuine hedge most knowledge workers have.

What to watch

I made a call in the last post on unemployment figures by year-end. Here’s what to track over the coming months:

The official unemployment rate is a lagging indicator — it will look fine right until it doesn’t. What to watch instead: entry-level job postings (are they continuing to fall?), the proportion of layoff announcements that cite AI (it was 4.5% of all cuts in 2025 — by January 2026 it had already risen to 7%, per HIGH5 Test), and whether we start to see college graduates unable to find starting roles at scale. That last one will be the canary.

The transition won’t announce itself. It will arrive as a series of ordinary-looking events — a hiring freeze here, a restructure there — until one day the aggregate is undeniable.

Don’t wait for undeniable.

The Next 1000 Days tracks one person’s attempt to build financial resilience before the labour market changes in ways we can’t fully predict. If this resonated, consider sharing it with someone who needs to read it.

Nothing here is financial advice. Any capital you lose acting on my reasoning is your problem. Any gains are obviously due to my outstanding insights.

There’s a version of this post where I tell you I’ve figured something out.

I haven’t. What I have done is looked at where things are heading — for developers, for knowledge workers, for anyone whose income depends on skills that are becoming cheaper by the month — and decided that working through it in public, with real numbers and outcomes, is more useful than pretending I have a map.

I don’t have a map or a crystal ball. I have a set of hypotheses and a limited amount of time to test them.

So, what do you actually get out of reading this? A real-time record of what one person is trying, what it costs, and what happens. Not a framework. Not a success story reverse-engineered from the ending. Something closer to a flight recorder — running whether things go well or badly, with nothing tidied up after the fact. If you’re navigating the same terrain, that’s more useful than most of what’s out there.

That’s what this newsletter is. Not a self-help column. Not a success story told in retrospect. A lab notebook, written in real time.

The situation, plainly stated

I’m a software developer and architect with thirty years of experience. I’ve worked across development, architecture, analysis, CTO roles, consulting, mentoring and coaching. I know this domain well.

I’ve also watched AI quietly pull the rug out from under a business I built. Not because I executed badly, but because the structural conditions that made it viable disappeared. The training market for developer skills. Gone. First slowly, then all at once. All over in a couple of years.

And this isn’t just my problem. Watch what’s happening in any development team right now. Developers who once wrote code are increasingly just reviewing what AI agents produce. That might sound like an upgrade, and in some ways it is, but let’s be clear about what it means economically: you’ve just become a quality control layer. The next generation of agents, the ones that make fewer errors, need fewer human reviewers. The arithmetic is simple. Shouting, for those willing to look at it directly.

I’m in my mid-fifties. This hits differently when you’re not twenty-eight with thirty years of runway ahead of you. But I’d argue it hits harder for the twenty-eight-year-olds — they just don’t know it yet. At the current pace of AI development, the knowledge work landscape in three years will look nothing like it does today. For anyone whose income depends on cognitive tasks, the question isn’t whether this affects you. It’s when, and how much, and what you’re going to do about it.

The experimental capital I have available to work with is real but bounded. Enough to be meaningful. Not enough to be reckless. Every decision I make here is made under the same constraint most of you are working under: genuine consequences for getting it wrong. A thousand days is a very short time horizon. Most of us will need to do some fairly fancy footwork before the clock runs out.

That’s not a disclaimer. It’s what makes this worth reading.

The countdown

Day 1000 is April 8, 2026 — the day this experiment begins. Day 0 is January 3, 2029.

The clock is running.

By Day 0, the employment landscape for knowledge workers will look materially different from today. Compressed rates. Fewer opportunities. More competition from AI-assisted generalists, and from AI itself. Contractors will feel it sooner — there’s no organisational inertia to buffer them, no HR process, no redundancy package, no three-month consultation period. Just fewer renewals. But employees aren’t safe either. The permanent role just has a slightly longer fuse.

Nobody really knows how governments and central banks will respond to job displacement at this scale. It’s new territory. The policy toolkit wasn’t designed for this. I find that either fascinating or terrifying, depending on the day.

If I’m still primarily dependent on contract income when the clock hits zero, the experiment will have failed.

That’s the hard constraint the rest of this is built around.

The four streams: what I’m actually testing

Most income diversification advice is too abstract to be useful. Here’s what I’ll actually be testing, why, and what I honestly think about each one. I’ll go deep on each in future posts.

1. Contracting. The floor, not the future

Still doing it. Right now, it’s my primary income source. It supports my family and funds the experiments. Without it, there are no experiments.

But let me be direct about what it is and isn’t. Contracting income at the senior end is heavily taxed in a progressive regime. You trade time for money, the government takes a large cut, and it scales linearly. More hours, more income, no compounding. If I try to squeeze more out of it after hours, diminishing returns kick in fast: I’m working into my rest periods, my day-job productivity suffers, and the taxman cheerfully harvests much of the gains anyway. I’d just be earning regular income, the hard way.

The strategic role of contracting in this experiment is to preserve runway while the other things develop. I’ll keep my eyes open for extraordinary opportunities here. But I’m not holding my breath.

Honest assessment: Useful short-term. Structurally limited. Probably shrinking as an opportunity before Day 0.

2. Investing. Real but slow

Investing is on the list, but I want to be straight about where I’m starting from: the capital I have available for this is limited, and a fair amount of it is still in the process of being saved. If you’re in a similar position — income mostly spoken for, not much left over at the end of the month — that’s exactly the situation I’m designing these experiments for. You don’t need a war chest to start, though it obviously helps. You do need to be deliberate about what you do with whatever you have.

Here’s the honest version of what investing can do: doubling your money in three years is a 26% annualised return — exceptional by any measure. It’s just not going to solve a cash flow problem in the near term.

I’ll also confess something. In the past, I’ve been sloppy about this. I spotted two exceptional opportunities in the last two years, knew it at the time, and did nothing with either of them. One would have doubled my money in six months. Another went up 300% over two and a half years. I watched both play out from the sidelines. That kind of thing is clarifying. Good investing turns out to be as much about psychology — what you do when you’re losing, whether you act when you should — as it is about finding the right opportunities. Through general laziness I’ve managed about 10% per annum. Not bad. Also not where we need to be.

I’m being more rigorous now. I’ll document positions and reasoning here as a record you can learn from and adapt — not a system to follow blindly, but a real account of decisions made under real constraints, scored against outcomes.

What investing does particularly well: it runs in parallel. It doesn’t require much of my time to compound. And the analytical skills — reading balance sheets, assessing competitive moats, thinking probabilistically about outcomes — transfer directly to evaluating business opportunities. Which brings us neatly to stream three.

Honest assessment: A genuine parallel track. Best case, I double the capital over three years. Worth running seriously regardless.

3. Building or buying a business. Where the real opportunity is

This is where I think the actual upside lives. It’s also the most complex category, and the one I know least about, beyond what transfers from investing. I’ll be learning this in real time and sharing the process as I go. Feature or bug, depending on your perspective.

The barrier to building software products has collapsed. Any competent developer can now build and deploy in days what would have taken a team months not long ago. That’s simultaneously the threat to contracting income and the opportunity on the other side: if you can identify the right problem, the cost to build a solution is lower than it has ever been in the history of software.

But this stream isn’t just for developers. Bricks-and-mortar businesses are interesting for a different reason, and they’re accessible to anyone willing to do the analytical work. Physical reality still has friction. A good local business — genuine repeat customers, solid margins, owner ready to exit — doesn’t become obsolete because the latest model dropped. And the analytical toolkit from investing transfers directly: you’re still reading numbers, assessing moats, thinking about what makes an economic position durable. The skills compound across categories. Which is convenient, because I need all the compounding I can get.

I’m actively looking at opportunities in both categories. When I find one worth pursuing, you’ll read about it here — including the reasoning, the numbers I can share, and the parts I got wrong.

Honest assessment: Highest potential. Highest complexity. The category where this experiment either works or doesn’t.

4. This newsletter. The meta-experiment

I’ll be straight with you: I have no idea if this will work.

The thesis is a timing bet as much as anything else. Most developers aren’t ready to hear what’s coming. Not because they’re unintelligent, but because the disruption hasn’t hit hard enough yet to feel personal. The AI agents are still making silly mistakes. That will change. The question is whether establishing a clear-eyed voice now, before the urgency peaks, puts this newsletter in a position to matter when it does.

If the timing is right, the readers who show up early themselves gain the most value and, in turn, become the most valuable subscribers of the newsletter. If I’m too early, I’m writing into a relative void for a year or two while the market catches up to the thesis.

Either way, this is a business experiment in its own right, not a side project. The metrics I care about: paid subscribers, churn, revenue per subscriber, and whether the writing holds up when I re-read it in two years.

Honest assessment: Complete unknown. The experiment is live whether I like it or not.

The prediction ledger: one call to open the account

This section is what separates lab notes from commentary. I make specific, falsifiable predictions. I assign confidence levels. I score them when they resolve.

The scoring system is the Brier score — a proper scoring rule that rewards calibration, not just correct picks. The short version: you’re penalised for being confidently wrong, and rewarded for being accurately uncertain. If you want the full technical explanation, ask, and I’ll write it up. It keeps me honest in a way that vague prognostication doesn’t.

Some issues will have a prediction. Some won't. Here's the first one.

Prediction #1: The OECD unemployment rate, currently sitting at 5.0% (January 2026), will rise by at least 2 percentage points to above 7.0% by end of 2026 — approximately 1 percentage point attributable to macroeconomic headwinds (trade disruption, geopolitical friction, slowing growth), and at least 1 percentage point attributable to AI-driven displacement of knowledge work.

Confidence: 75%. Scores on Day 650 (approximately March/April 2027, allowing a quarter for the December 2026 figures to be released). Baseline and scoring data: OECD.org monthly unemployment releases.

The macroeconomic 1% feels relatively uncontroversial given where we are with wars, tariffs, trade fragmentation, and general global uncertainty. The AI 1% is the more interesting call. Organisational inertia will buffer wholesale displacement for a while — companies don’t restructure overnight, and the transition has retraining and liability dimensions that slow things down. But the trickle has started. By end of 2026, I expect it to be visible in the numbers.

One caveat worth making explicit: if AI investment crashes between now and then, that doesn’t invalidate the thesis. The dotcom bust in 2000 didn’t mean the internet wasn’t going to matter. It just meant the market got ahead of itself. The underlying capability keeps developing regardless of what the Nasdaq does.

What comes next

Paid posts go out weekly. I’m committing to that cadence partly because there’s enough happening to write about, and partly because without it I’ll let things slide. You’ll get positions taken, decisions made, things read and found useful, outcomes scored. Easy calls and harder ones. Occasionally, I turn out to be wrong about something in public.

Next week goes deeper on investing — specifically, the approach that led me to those two opportunities I mentioned, and why you don’t need to spend all day buried in annual reports to identify a candidate worth looking at. I also built a Python script that does much of the intelligence-gathering for me. More on that next time.

If you have questions about the setup, the predictions, or anything I’ve glossed over here — comment or reply. I read everything.

The clock is running. Tomorrow is Day 999.

— Olaf

Before you go: one practical tip

Each issue will end with something immediately useful. Not a framework. Not a mindset. Something you can act on.

This one is about debt, and it matters more right now than most people realise.

If your income outlook is uncertain, reduce your debt.

The logic is simple. Debt is a fixed obligation in a world where your income is becoming variable. It doesn’t care whether you’re between contracts, whether your hours got cut, or whether the renewal didn’t come through. The repayment is due regardless.

Most people think about debt in terms of interest rates. The more important variable right now is fragility. A mortgage extension, a car loan, a credit card balance — each one narrows the gap between an income disruption and a genuine crisis. The bigger the debt load, the less runway you have when things get bumpy. And things are going to get bumpy.

This isn’t about becoming debt-free overnight. It’s about not adding to the pile while your income is still reliable. The time to shore up the foundations is before you need them, not after.

If you’re currently thinking about taking on new debt — extending the mortgage, upgrading the car, whatever it is — ask yourself one question first: how does this look if my income drops 30% in twelve months? If the answer is uncomfortable, wait.

That’s the tip. Simple. Not exciting. Worth more than most advice you’ll read this week.

The Next 1000 Days documents one person’s attempt to build financial resilience in the face of AI-driven disruption to knowledge work — in real time, with real numbers, scored against outcomes. If someone you know is thinking seriously about this, forward it to them.

Nothing here is financial advice. Any capital you lose acting on my reasoning is your problem. Any gains are obviously due to my outstanding insights.