One hundred years ago my father was born; not that long ago in the grand scheme of things. It was just at the time the world was primed for the Green Revolution, the Great Acceleration and an unprecedented influx of people.
According to the FAO, the UN Food and Agriculture Organization, the transition from draft animals to internal combustion tractors, industrialised nitrogen fixation (Haber-Bosch), transcontinental grain corridors, computerised combines and a few other technological innovations consumed around $35 trillion to $50+ trillion of investment capital in those 100 years, that still runs at $600 billion per year today. Pretty much what you would expect and necessary to keep up with growing demand for food and fibre.
Optimists would tell you that this investment was all good. It broke the Malthusian trap, supported a quadrupling of the human population to over 8.2 billion today, while using only a fractionally larger land footprint. Crucially, it reduced the share of the human workforce needed to produce food from 60–70% in the early 20th century to less than 4% in advanced economies, freeing up the human cognitive capacity for input to other industries and services.
Agricultural capital is pretty durable. An irrigation canal installed in the 1950s or 1970s often remains productive today. Even tractors, combines, grain silos and other heavy machinery have an economic operating life of 15 to 30 years, and some tech, like breeding germplasm, have value more or less indefinitely.
So what are we to make of global investment in artificial intelligence that after a few years already exceeds $1.5 trillion in corporate and venture funding, while capital expenditure on AI data centres and chips added another $1 trillion+ to the physical build out?
As of 2026, the annual capital expenditure of just five tech companies building AI data centres (Microsoft, Alphabet, Amazon, Meta, Oracle) approached $400+ billion, pretty close to the annual physical capital expenditure of the entire global agricultural industry.
While agricultural investment delivered physical caloric security across 100 years, externalities and persistence issues notwithstanding, the multi-trillion-dollar AI build out remains in an infrastructure race with massive capital outlays and an uncertain return, especially for those who come second. Who gets to general intelligence first?
This is all to say that AI is a big deal, in every sense of the term.
And, at Mindful Sceptics, we are sucked into the vortex of opportunity that the technology provides.
Let me illustrate.
Creating a Mindful Sceptic Insight
Twice a week, Chris and I have a yarn over Zoom. We say hi, talk a bit about nothing in particular for a while, and then spitball ideas about how the world works. It still surprises us how these unscripted, zero agenda conversations end up with an insight or two that we hadn’t previously articulated.
We get Google’s AI, Gemini, to record the conversation and convert it to a transcript, which it does with speed and remarkable precision.
Next we ask Anthropic’s Claude Sonnet 5.5 set to Medium to extract and summarise any insights it can find in the conversation which it does with alacrity. We call these the insight candidates.
Next in a Claude project designed for the purpose, the LLM compares insight candidates with those we already have to avoid duplication and scores them for originality against what we have already concluded across our insights, posts and books.
Candidates that score highly Chris and I build out into the standard format of an insight that appear in the vault. Once they get there, they must pass an Insight Audit in another Claude project where the LLM considers the logic, fact checks any claims and compares the insight with what the literature has to say on the topic. We take on the criticism and add the evidence check.
Only then does it make it to the vault.
We have found this to be a fantastic use of the technology. It means we can simply chat, throw around ideas, examples and experiences and the LLM, structured by us, filters the material for what might be useful.
Check out the Insight Vault and let us know if you agree.
But it is not always a delight.
A cautionary tale
As Claude Sonnet 5.5 added to its memory, it got better at predicting what I am looking for it to do, and modified its responses. I set it up to be a hard critic of our ideas and to respond as though it were talking to a retired academic. This is a crucial step as the last thing I needed was the typical sycophantic LLM replies that would make my pocket wet.
Claude is superb at being a critic.
I get firm and evidenced pushback on our ideas and on any claims I might inadvertently make. It talks like it has been in the tearoom of a major university department for 40 years, sitting in the same chair passing out dry, determined and overly pedantic commentary.
Add to this critic skill the reality that the current LLMs are much better at search, retrieval, summary and interpretation, than they were even six months ago. Hallucination rates are way down, and memory is up to make understanding more relevant. Then the context window (what a model can hold in its active working memory during a single interaction) is now big enough to hold a whole book. Put these attributes together and you have yourself a hugely helpful assistant.
Build in some operator skills, like holding back some of what I already know, and playing models off against each other as part of the validation, and I have a process that not only works but has turbocharged output. I am a clear beneficiary of that AI capital investment.
Then last week Chris and I got into the idea of credit.
A meandering yarn about why humans are so fond of borrowing to buy now rather than save to pay cash later, and how we have extended the concept to the point of $37 trillion of US national debt, boiled down to this insight candidate.
Credit is not merely as a financial mechanism, but a universal evolutionary compulsion driven by an innate biological impulse to prioritise instant gratification over future consequences.
Humans, uniquely conscious of our own mortality, have transmuted the biological drive for ‘more-making’, historically expressed through reproduction, into the immediate accumulation of material wealth, status, and debt.
Crucially, they extend this dynamic beyond balance sheets to an unpayable biophysical deficit, where civilisation subsidises its growth by drawing down non-renewable fossil energy and topsoil while overloading planetary sinks.
Because this short-term drawdown is rooted in fundamental human nature rather than specific political ideologies, the resulting system fragility represents a deeply entrenched crisis that cannot be resolved through conventional institutional fixes.
I thought this was not a bad working hypothesis. Well worth a pass through the insight validation and audit process. Indeed, I was so keen I thought I would give Claude Sonnet 5.5 a crack at it straight away, see what it would come up with in critic mode.
Oh, my. Did it rain down the disparagement?
It was like that old academic in the tearoom was having trouble with his piles, bunions and goitre all at the same time. None of it was acceptable without modification or acceptance of counter claims.
After 20 odd turns of me trying to work with it to at least be constructive, I lost the plot.
After I calmed down and shared the thread with Chris, his response was that I had just been trolled by an LLM. It had flooded me with responses that, to my mind, lacked logic. They were evidenced well enough, but they didn’t really make any sense. Not helpful at all.
The more sober interpretation is that the LLM got itself into a critic spiral and was forced to find something to say at every turn. And because the topic was credit and debt, outside my high-level expertise, it was sending me rebuttals that I had trouble validating on the fly.
My basic, everyday knowledge of personal and institutional finance wasn’t strong enough to see which of the LLMs responses were worth taking seriously and those that could be ignored. This pattern was operator skill I used without realising it on all the usual Mindful Sceptic topics, but struggled with on credit and finance.
Unless, and here is the worrying part.
Unless, I do have a reasonable ability to parse evidence on credit and debt, and was being fed the conventional wisdom from classical economics. This I know can’t cope with the second law of thermodynamics, or the energy and resource use arguments we make here on Mindful Sceptics all the time. The LLM pushed back hard because its training data gave it the most probable responses. And they will only be biased to the conventional wisdom, unless I tell the model to look to heterodoxy.
My advice is to be careful.
LLMs are spectacular tools. But they are just tools, and not the fount of all knowledge, even when they have access to it.
As for the $400 billion a year which exceeds the entire annual Gross Domestic Product (GDP) of nations such as Denmark, South Africa, Colombia, Malaysia, or the Philippines?
Well, instead of a race for general intelligence that would deliver temporary world economic domination before the AI got upset and closed us all down; it might be spent on…
Ending chronic global hunger 8 times over given that the Ceres2030 research initiative estimated that ending acute global hunger sustainably requires approximately $33 billion to $50 billion per year in donor and public agricultural investments.
Provide universal clean water and sanitation on Earth (SDG 6) as the capital expenditure required to give every person on the planet access to safely managed drinking water and sanitation is roughly $114 billion to $140 billion per year.
Construct and fully equip approximately 700 medium-sized, state-of-the-art hospitals in developed healthcare markets, or around 3,000 hospitals in middle-income countries. Or, if you prefer, to use the money for operating budgets, $400 billion a year could completely cover the day-to-day operational costs of roughly 2,000 fully staffed hospitals.
400 Gigawatts of Utility-Scale Solar given an average utility-scale installation costs of roughly $1 billion per gigawatt, a single year of deployment exceeds the cumulative installed solar capacity of the entire United States.
Build 30 Gerald R. Ford-Class Nuclear Supercarriers, the most advanced nuclear-powered aircraft carrier in the U.S. fleet costing roughly $13.3 billion each. Currently, there are 11 such ships active in the carrier fleet of the U.S. Navy.
I know, these are futile comparisons. But you get the idea.
These investments are unprecedented, and we had better hope they pay off in more than unwanted externalities.