Empty Calories

In the fall of 2022 I co-founded a software startup called Rezelut. The product was an energy planning platform for enterprise executives: an assessment engine, a scenario planner, a performance dashboard. Building the prototype took a company and a broad team of skilled contributors. Two co-founders contributed capital and their firm’s development teams; around them, designers, product managers, and data scientists. Months of billable development hours went to people who weren’t founders, across six months of concentrated build, inside nearly two years of my own effort. We shelved it at the prototype stage, caught in the oldest trap in startups: too early to show revenue, too late to keep building without it.

This spring I built the Crux Resilience Snapshot, a ten-domain household assessment with a weighted recommendation engine, in a few days. Alone. In my home, for the price of a software subscription.

They are not the same product. Rezelut was far more ambitious. But they are the same discipline: a data-driven assessment engine, a recommendation layer, a front end. One took a company. The other took less than a week. Or, to use an analogy, one was cooked from scratch by a full kitchen. The other arrived like Uber Eats: fast, cheap, no cooking required. With an important caveat: I know how to cook (more or less). That detail is central to this article.

That ‘Uber Eats’ development speed wasn’t me becoming a brilliant coder. The tool I used didn’t conjure the ability to build software out of nothing. It was distilled from millions of code repositories, from the prose of working writers, from research papers and manuals and a quarter century of forum answers, from the accumulated output of everyone who ever published how to do anything. The collective creative work of a civilization, harvested, compressed, and sold back to us in a form that is fast, cheap, and effortless to consume. Processed food. And as with processed food, the question isn’t whether it’s convenient. It’s what got stripped out in the processing, and what happens to a population that eats nothing else.

How the pantry was stocked

We know how the harvest happened, because it was documented while it was happening. In August 2023, The Atlantic revealed that a dataset called Books3, more than 183,000 pirated books, had been used to train AI models at Meta, Bloomberg, and elsewhere, and published a tool that let authors search for their own titles. Meta’s own research paper had described the pirated collection as a publicly available dataset for training. The lawsuits started within months of the tools reaching the public: visual artists in January 2023, over a training set of five billion scraped images; Getty that February, over twelve million photographs; novelists that summer; The New York Times that December, arguing the models were being built into a market substitute for the very journalism they were trained on. Hollywood’s writers spent 148 days on strike in 2023 and won the first collective-bargaining agreement anywhere with guardrails on generative AI. The people whose work was being absorbed said so, loudly, in real time. It didn’t matter.

Making the meal-replacement smoothie

Then the system did what systems do: it converted an existential objection into a transaction. In June 2025, a federal judge ruled that training on legally purchased books was “quintessentially transformative”, fair use, no license required. The pirated portion was another matter, and Anthropic (full disclosure: the maker of the tool I used to build the Snapshot app) settled that portion for $1.5 billion; the largest copyright settlement in American history was granted final approval in July of 2026. It worked out to roughly $3,000 per book. Once. For past conduct only. Publishers, meanwhile, converted their lawsuits into licensing agreements. Within three years, the question “can you just take it?” had been replaced with “what does it cost after you already took it?” The answer: about $3,000 a book, set against the trillions in market value the corpus helped create. We didn’t negotiate the price. It was set after the value was already gone.

Supplements aren’t food

That was the one-time charge. The recurring one is quietly insidious.

I recently wrote about an economy that has run off the cliff and hangs mid-air, suspended by momentum. The same physics apply here, to a different asset. AI-assisted work looks great right now, and it looks great because it is being produced by people whose expertise was formed before the tools existed. The experienced engineer prompting a coding model knows what good software looks like because they spent years writing it ‘by hand’. That experience enables the tools to amplify and accelerate formed judgment. They are running on legacy human capital. What will replace it?

The formation pipeline is where it shows first. Stanford researchers analyzing payroll records for millions of workers found that employment for software developers aged 22 to 25 fell nearly 20 percent from its late-2022 peak, while employment for older workers in the same occupations kept growing. The same split repeats across the most AI-exposed occupations: experienced workers stable or rising, entry-level falling. Expertise has always been formed by doing the low-stakes work first. The low-stakes work is disappearing.

The consumption side now has a name: cognitive debt. That’s the term MIT Media Lab researchers used after wiring essay writers with EEG sensors. The group writing with AI assistance showed the lowest neural engagement of any condition, and when the tool was taken away, they performed worse than people who had never used it at all. A Microsoft and Carnegie Mellon study of 319 knowledge workers found the same pattern from another angle: the more people relied on AI output, the less critically they evaluated it. Looks great. Tastes great. Could we still make it ourselves? Measurably: fewer of us, and less each year.

You don’t have to guess where this goes. Recursive self-improvement, AI building the next AI, is the stated objective of the labs building it, and they report that their models already write a growing share of their own code. The handoff isn’t a dystopian projection. It’s the roadmap we are already traveling.

Start with quality ingredients

The Snapshot build worked because of twenty-five years of doing it from scratch, in energy and in product, not instead of them. I knew what a product was, what an assessment engine had to do, and what ‘the goal’ looked like when it appeared on the screen. The tool was a multiplier of what already existed. There isn’t a substitute for quality ingredients.

Keep the kitchen skills

That’s the same test as everything else in these field notes. The generator, the second water source, the volunteer airlift: what do you still have when the primary system is replaced, throttled, or gone? I use these tools, both gratefully and skeptically. But I also do the work: draft the concept, critically review the output, invest the time to rewrite, review, and refine. Read the code before you ship it. Work the problem before you ask. AI slop is sloppy. And lazy. We don’t need to keep mindlessly consuming it—or making it. The enjoyment of a well-prepared, well-sourced meal shared with others provides more than a ratio of carbs to fats to proteins. It is both nutrients and nurturing. The capability you build into yourself is the one asset that doesn’t evaporate; the current cheap, abundant version of these tools is a subsidy, not an ongoing promise. Who pays when that bill arrives is for another article, coming soon.

How to think about risk, redundancy, and building capability before disruption forces the issue is the subject of the book, CRUX: Risk & Resilience. See us at cruxresilience.com for a comprehensive resilience assessment and mitigation plan.