Everything you know about the first AI winter is wrong. It was a hit job. So well executed that even today the killer has gone free.

(Read, The Man Who Buried His Teacher)
 

I don’t think there is any question about it. Your AI cost crisis can only be attributable to human error.

Pay for intelligence when you need intelligence. If you’re doing the same task twice, have the model write a program to do the task for both of you.

I put $30 on the Anthropic API three months ago. I burned $2.70 in the first few days figuring out my approach. Twenty cents since. Less than $3 total, and most of that was tuition.

That twenty cents built a text analysis engine. Sentence tokenizer, statistical metrics, entropy calculations, composite scoring for a website I built. The model wrote the code, I reviewed it, I deployed it as PHP on a Linux box I already own. The electricity cost is a rounding error on my power bill.

I keep reading about companies spending six figures a month on AI tokens and I’m doing the blinking eye meme. Putting every customer ticket, every document summary, every code review through a model, thousands of times an hour, 24/7, paying per token every time, seems crazy when the model could build you a deterministic tool instead.

Deterministic work belongs in traditional code.

But the integration pattern the industry has settled on is “put the model in the hot path.” Keep the AI in the loop on every request, forever. Perpetual token cost for work that stopped being an AI problem the moment someone understood the requirements. Your output changes when a new model version ships. Your pipeline breaks not because your code changed but because theirs did.

The model can write its own replacement for most of these use cases. Ask it to classify tickets? It can write you a classifier. Ask it to score text against a rubric? It can write the scoring engine. One build session. One deployment. Done. Stable output you control.

Companies pay the model to do the same work over and over because that’s how the tooling is sold. The SDKs make it easy to call the API. The tutorials show you how to put the model in your pipeline. The pricing page shows per-token costs that look cheap until you multiply by volume. Nobody in that funnel is suggesting you use the model to build something and then turn it off.

This is the SaaS treadmill applied to AI. Recurring revenue for the provider. Recurring cost for the customer. For work that could be a one-time build.

AI APIs are the right call when you need reasoning on novel inputs. Creative work, ambiguous classification, anything where the rules can’t be fully specified in advance. Pay for intelligence when you need intelligence. But if you’re sending the same shaped request a thousand times a day and getting predictable outputs, you don’t have an AI problem. You have an engineering problem. And the AI is the best engineer available to solve it for you once, without ever touching API billing.

Twenty cents. The result runs on its own.

The foundation of every AI system on earth was laid by a teenage runaway from Detroit and a clerk from Madras. Neither of them could get hired today.

In 1935, a 12-year-old boy in Detroit ducked into a public library to hide from the kids chasing him. His father was a boiler-maker who used his fists. The neighborhood wasn’t any better. The boy had already taught himself Greek, Latin, logic, and mathematics on his own. That night, hiding in the stacks, he found Russell and Whitehead’s Principia Mathematica. He read all three volumes in three days. He found errors. He wrote Bertrand Russell a letter. Russell was so impressed he invited the boy to Cambridge. The boy couldn’t go. He was 12.

He ran away from home. He ended up in Chicago, where he met Warren McCulloch, a neurophysiologist who had the vision for how the brain might compute but needed someone who could do the math. McCulloch took the homeless boy in. They worked together every night. In 1943 they published “A Logical Calculus of the Ideas Immanent in Nervous Activity,” the first mathematical model of a neural network. Every AI system running today is built on that foundation. Walter Pitts’s only earned degree was an Associate of Arts. He died at 46.

Twenty years earlier, a clerk in Madras making 20 pounds a year wrote letters to several British mathematicians containing pages of original theorems he’d developed entirely on his own. Most ignored him. G.H. Hardy at Cambridge opened his, thought it was a hoax, then concluded the results “must be true, because if they were not true, no one would have had the imagination to invent them.” Srinivasa Ramanujan became a Fellow of the Royal Society and the first Indian Fellow of Trinity College. He died at 32 from a treatable parasitic infection that was widespread in Madras and can lay dormant for years.

Institutions do phenomenal work. But the foundation of the field reshaping every industry on earth was laid by a teenage runaway and a clerk. You cannot regulate calculus. There will always be someone in a library, a garage, or a borrowed compute environment working on something no framework anticipated.

There are people out there right now working on problems the institutions have not yet named. Some of them are not waiting for permission. Some of them do not even know yet what they have found.

If we keep building frameworks around where we think the future comes from, we’re going to miss where it actually does.

An AI detector just flagged 46% of the Pope’s new encyclical as AI-written. The encyclical is about AI ethics. It was written in a prose tradition over a thousand years old. The same detector rated other paragraphs of the same document at essentially 0%. Same author. Same document.

I ran a similar experiment on myself. I asked ChatGPT to review my personal blog from 2008-2017 and identify posts that read as AI-written. It identified 35% of them as having structured arguments, clean frameworks, numbered examples, and tidy conclusions. None of them were AI-assisted. None of them could have been. ChatGPT didn’t exist yet.

The three worst offenders: a 2009 post about Twitter with definitions and numbered use cases. A 2010 business case for mobile websites with data and a strategic conclusion. A 2014 incident postmortem with a failure chain and lessons learned. Those aren’t AI patterns. Those are writing patterns. Humans have been organizing their thoughts like this for centuries.

A year ago these same tools were being sold to help you write more clearly. Now writing clearly is the evidence you used them.

Even the article covering this story hedges: “practitioners should treat single-detector outputs as suggestive and seek multi-method forensic work before drawing firm conclusions.” Here’s a conclusion that doesn’t require forensic work: if a writing tradition predates electricity, maybe weight the patina of the source before you let an algorithm accuse it of being a machine.

#AIDetection #FalsePositive #WritingIsNotACrime #AIEthics #ContentAuthenticity

1. Y2K had a deadline: January 1, 2000. You could circle it on a calendar. AI risk is perpetually “five years away”. GPT-4 can already pass the bar exam and write working code, and that was two years ago.
2. Y2K was a bug. AI is a feature: Nobody wanted two-digit date fields to break banking software. Millions of people are trying to make AI more powerful and more autonomous, on purpose, as fast as possible, and calling it progress.
3. Y2K had a fix: Tedious, expensive, boring (but the fix was known). Find the date fields, expand them, test. The AI alignment problem doesn’t have a COBOL patch.
4. The Prevention Paradox, magnified: The Y2K experts said “this will be bad if we don’t fix it.” We fixed it. Nothing happened. So everyone decided the experts were wrong. AI safety is walking into the same trap… except the fixes aren’t obvious, aren’t agreed upon, and nobody’s doing it at the speed the models are shipping.
5. Y2K didn’t argue back: A two-digit year field never wrote a convincing memo explaining why you should stop worrying. The failure mode of AI isn’t malfunction. It’s persuasion.
6. Nobody was emotionally bonded to their COBOL runtime: People didn’t date their name-field subroutines, tell them about their day, or feel seen by them. That attachment asymmetry is new, and we’re still pretending it’s just a UI problem.
7. Y2K couldn’t recursively self-improve: The worst case was broken systems cascading into other broken systems. Nobody worried the date field would rewrite itself, optimize its own deployment strategy, and start managing the power grid.
8. Y2K was legible: A senator could understand “the computer thinks it’s 1900.” Try explaining reward hacking or deceptive alignment to a committee that still treats “the cloud” like weather.
9. The Y2K grift was finite: Consultants made a fortune, but the gravy train had an expiration date. Literally. The AI industry has no such constraint. You can sell fear of superintelligence and sell superintelligence at the same time, forever. That’s not a prediction. That’s the business model.
10. Y2K was technical debt, AI is technical leverage: Y2K was the past catching up to the present. The decisions from the 1960s finally maturing into consequences. AI is the opposite: the present building something it doesn’t understand, pointed at a future it can’t predict, and calling the gap “innovation.

Nobody’s sure which end of the lever we’re on and there’s no button to turn it off.