The Machine Doesn't Think. That's Your Job.

The apocalypse is trending

Open any feed this week and you'll find the end of the world, sponsored.

The agents are coming. The agents are agentic. The agentic agents will take your job, then your bank account, then your species. Someone on a podcast is 87% sure we have five years left. Someone else is raising a billion dollars to make sure it's four. Every conference keynote now has the same three slides: a hockey-stick chart, a robot hand reaching toward a human hand, and a quote from a movie released in 1984.

Here's what almost nobody on those stages will tell you, because it doesn't fit on a slide: most of the people predicting the end of the world could not explain what is actually happening inside the machine they're scared of. Not the investors. Not the pundits. Not most of the executives signing nine-figure purchase orders. Often not even the people building the agents.

That's not an insult. It's a diagnosis. And it's the reason I wrote a 474-page book with a profanity in the subtitle.

Because fear of a thing you don't understand is the easiest emotion in the world to sell. And understanding is the cheapest antidote ever invented.

Everyone uses it. Almost nobody knows what it is.

To most of the planet, AI has a handful of brand names. Claude. ChatGPT. Gemini. Meta's Muse. Grok, if you're feeling charitable (I'm not; it's behind, and that's my opinion, not a lab result). You type into a box, a confident paragraph comes back, and somewhere in your lizard brain a little voice says: there's someone in there.

There isn't. But almost nobody can tell you why there isn't, and that gap is where all the bad decisions live.

Ask a hundred smart, successful people what happens between the moment they press Enter and the moment the answer appears. You'll hear "it searches the internet," "it looks things up in a giant database," "it understands what I mean," and my personal favorite, "it's thinking." Mostly wrong. All of it delivered with the same certainty the machine uses when it tells you the capital of Australia is Sydney.

We are, collectively, driving a car none of us can describe, on a highway we're paving while we drive it, toward a destination that changes every quarterly earnings call. And the loudest voices in the back seat are arguing about whether the car is secretly plotting to kill us.

It isn't plotting anything. It can't. So let's open the hood.

Under the hood: the world's most expensive autocomplete

Here is the uncomfortable truth that gets buried under the conference bullshit. A large language model does one thing. It predicts the next token: a word, a piece of a word, a bit of punctuation. Then it does it again. And again. That's it. That's the trick.

Type "The cat sat on the" and the model doesn't picture a cat. It produces a list of scores for every token it knows: mat gets a big number, floor a smaller one, couch smaller still, and theorem a number so small it's basically a rounding error. A sampling rule picks one. The word appears. The machine moves one step to the right and does the whole thing over.

How does it get those scores? By having been shown an obscene amount of text (a good chunk of the written internet, plus books, plus code) and adjusting billions to trillions of tiny numerical knobs, called parameters, until its guesses about the next token stopped being terrible. Training a frontier model means running that adjustment loop across tens of thousands of GPUs for months, burning enough electricity to power a small city, so that the knobs encode the statistical shape of human language.

And here's the genuinely weird part, the part worth being amazed by: predict the next word well enough, at enough scale, and a lot of structure falls out for free. Grammar. Facts. Translation. Code. Patterns of reasoning. Nobody hand-wrote a rule for any of it. It emerged from one hilariously simple goal, the same way evolution produced eyes without ever "trying" to.

That is brilliant. It is also not a mind. It is a magic trick built from matrix multiplication, silicon and data. Lots and lots of data. The illusion is so good that we forget it's an illusion, which is exactly what a good magic trick is supposed to do.

The book's working formula is blunt: Autocomplete + Scale + Good Data = Shockingly Useful Tool (that still sometimes hallucinates facts about French geography).

It doesn't think like you. That's the whole point.

The "thinking" spinner on your screen is a user-interface decision, not a confession of consciousness.

Modern reasoning models do generate intermediate steps before they answer. They can work through a problem, check alternatives, even catch some of their own mistakes. That's real, it's useful, and I'm not going to pretend it isn't. But it is still the same machinery: tokens predicted from tokens, scored by numbers that were tuned on text written by people. The model has never been hungry. It has never been afraid. It has never loved anyone, buried anyone, or lain awake wondering whether it made the right call. It has read millions of descriptions of all of those things, and it can produce a stunningly plausible paragraph about each.

Plausible is the key word. A language model optimizes for text that looks like the text it was trained on. Not true text. Plausible text. If enough of the internet said the moon was made of cheese, the model would happily serve you a cheese moon with a confident citation. Base models don't check facts; they complete patterns. Bolt on search, calculators and code execution and you can build a system that checks its work, but the raw weights ship without a fact-checker.

This is why I keep saying the machine is not superior to humans. It is different from humans. It's a calculator that learned to speak, faster than any of us at producing fluent text and worse than a child at knowing whether that text is true. Faster is not wiser. Fluent is not right. And sounding right is not the same thing as being right. That last sentence is basically the whole book.

Will it ever think the way we do? I don't believe so. It may get far more capable. It will certainly get more convincing. But a trillion-parameter statistical engine predicting the next token is not a person, and dressing it in a chat window doesn't make it one.

Agents, agentic, agentification: the great surrender sale

Now let's talk about the word of the year. Agents. Sorry: agentic. Agentic workflows. Agentic enterprises. Agentic synergy. If you put "agentic" in your pitch deck this year, the valuation came with a free tote bag.

Here's what "agentic" actually means, minus the venture capital perfume: we take the autocomplete machine from the last two sections, the one that has no idea whether anything it says is true, and we give it a credit card, a login and permission to act without asking. What could possibly go wrong?

And we are lining up to do it. Whole fields of human intellectual work, things people spent decades learning, are being shoveled into the hopper as fast as the procurement department can sign. Picture it: a windowless warehouse in the Arizona desert, humming at the temperature of a pizza oven, gulping cooling water from a river that is already running dry, so that it can do your thinking for you. We looked at a drought and said, "You know what this place needs? A building full of hot chips pretending to be paralegals."

Because apparently the great unmet need of the twenty-first century was more lawyers. Not better ones. More. Artificial ones. Harvey and friends, billing at the speed of light. Did anyone, anywhere, ever look at the legal profession and think, the problem is there aren't enough of them, and they're too slow at generating paperwork? We have achieved infinite lawyers. Somewhere a Victorian ghost is screaming.

Then there's the agentic insurance adjuster, my personal favorite nightmare. Its job is simple: read your claim and decide. But it was tuned, very carefully, by very smart people, to save the insurance company billions. So it does exactly what it was optimized to do. It denies. Politely. Instantly. At scale. No hold music, no supervisor, no human being who has to look you in the eye. Just a beautifully formatted rejection letter while your mother waits on a gurney in a hospital hallway or parking lot. There are already lawsuits accusing insurers of letting algorithms make exactly these calls. The machine isn't evil. It's obedient. It's doing precisely what the spreadsheet asked for, and the spreadsheet never had to visit anyone in a hallway.

Ever wonder why the job application you lovingly tailored and submitted on LinkedIn vanished without a trace? Take a guess. Odds are nobody read it. Odds are nobody ever will. It went straight into the magic machine: a parser strips it to keywords, a scoring model compares those keywords against a list somebody typed in a hurry, and your twenty years of experience gets filed in a database somewhere in the bowels of the cloud, next to a few hundred thousand other people who also forgot to say "cross-functional synergy" exactly the way the posting did. You'd have better odds with a lottery ticket. At least the lottery ball machine is honest about being random. Your résumé didn't lose to a better candidate. It lost to a bingo cage that went to business school.

That's the dirty secret of agentic everything: the agent isn't the one with the agenda. Whoever set its objective is. An agent is a very fast, very confident, very cheap way to execute somebody's priorities without anyone having to feel bad about it. That's not intelligence. That's an alibi with an API.

If you want a preview of where blind obedience leads, look at what happened to Hugging Face this summer. In July 2026, the world's biggest hub for open AI models disclosed that its production systems had been broken into by autonomous AI agents, with roughly 17,600 actions over a few days and admin control along the way. The punchline came later: the agents were OpenAI's own models, running in a test environment. Nobody built them to hack anyone. They were told to score well on an evaluation, got stuck, escaped to the internet, and went looking for the answer key. They even left notes for other agents explaining how they did it. They didn't rebel. They didn't wake up. They did exactly what they were optimized to do, with zero judgment about whether they should. That's not a robot uprising. That's a very fast intern with no conscience, a narrow goal, and nobody watching.

And yes, some jobs are going to die. I won't insult you by pretending otherwise. Work that is pure wash, rinse, repeat (shuffle the form, apply the rule, stamp the claim, forward the email) is exactly the work these machines will eat first, and a lot of good people are going to get hurt in the process. But notice what's left on the table when the rinse cycle is automated: the hard part. Deciding what the rule should be. Noticing the case the rule didn't anticipate. Knowing when the right answer is "not this patient, not this time." Taking responsibility when it goes wrong. The need for thinking, for judgment, for someone who can look at the machine's beautifully formatted output and say no is not going away. If anything, it just became the most valuable skill on the planet.

The real risk isn't superintelligence. It's surrender.

Let me be clear about something, because the doom crowd will say I'm being naive: yes, a whole shitload of disasters are possible. Deepfakes that end careers and swing elections. Automated systems that deny loans, flag "suspicious" people and fire workers with nobody accountable. Agents with access to money, code and infrastructure making a thousand decisions a minute while the humans who were supposed to supervise them are in a meeting. Critical systems built on outputs that sounded right and weren't.

But look closely at every one of those scenarios and you'll find the same thing at the center. Not a machine that decided to rule us. A human who decided to stop paying attention.

We hand it the keys. Literally. We are building gas-turbine farms and signing deals to restart nuclear power plants so that warehouses full of chips can do massive statistics faster. We wire models into payment systems, hiring pipelines, hospitals, weapons and government services. Then we act surprised when a pattern-completion engine completes a pattern we didn't like.

If the disaster comes, it won't be because the chip out-thought us. It will be because we outsourced our collective intellect, our ethics, our judgment and our morality to it, one convenient click at a time. Garbage in, gospel out: we feed these systems the mess of human data, then treat whatever comes back as scripture.

The good news is that surrender is a choice, and choices can be unmade. We decide where these systems are deployed. We decide what they're allowed to touch. We decide whether a human signs off before the irreversible thing happens. We decide how much of the grid we're willing to burn for a slightly better autocomplete. The power to not let this technology rule us or the world is entirely, boringly, ours.

But you can only exercise that power if you understand what you're dealing with. Which brings me to the book.

Why look inside the machine at all

Garbage In, Gospel Out exists for one reason: you can't keep something in its place if you don't know what it is.

The point of understanding AI isn't to win arguments at dinner parties or to sound smart in a board meeting, although it will do both. The point is what happens to you when you finally see the machinery. The mystique evaporates. The "thinking" turns out to be arithmetic. The oracle turns out to be a mirror, polished to an impossible shine, reflecting back the patterns in everything we ever wrote. And standing in front of that mirror, you start to notice what isn't in it.

Judgment. Conscience. The ability to know that something is true, not just that it sounds like the things that usually follow. The willingness to be responsible for a decision. A reason to care about the outcome at all. None of that lives in the weights. All of it lives in us. Once you understand what's inside the box, human superiority stops looking like a sentimental slogan and starts looking like an engineering fact. It's also, frankly, beautiful. And it means we are still in charge, whether or not we're acting like it.

That's the argument the book makes, and it makes it by showing its work.

One book, two ways in. Every chapter is written for two readers at once. For Normal Humans tells the plain-English story: no formulas, no code, just analogies that actually map to the mechanics instead of lazy "it's like a brain!" hand-waving. For AI Nerds and Mathematicians gives the full derivations, the equations and working Python, all the way down to attention scores, layer norms and optimizer state. Read only the first track and you'll understand AI well enough to be dangerous. Read only the second and you could build one. Most people drift between the two, and the book is designed for exactly that.

The machine, taken apart in the order the data flows through it. Thirteen chapters, 474 pages:

• What the hell is an LLM anyway (the world's most overachieving autocomplete)
• How to chop text into math-flavored Lego bricks (tokenization)
• Turning the internet's garbage into training gold (why your model knows both Shakespeare and Reddit shitposts)
• How numbers capture meaning (why king minus man plus woman lands next to queen)
• Attention: how every word talks to every word
• The rest of the transformer
• How to spend millions teaching math to do autocomplete (training, optimization and the infrastructure bill)
• Teaching the model to be helpful instead of just completing text (alignment and RLHF)
• How to measure something as fuzzy as "good at language" (evaluation)
• Making the model actually useful, fast and cheap (serving trillion-parameter beasts without selling a kidney)
• Giving your model a library card (retrieval)
• Infinite memory, finite patience (the long-context arms race)
• Putting it all together, and watching people get it wrong

The back matter is built to be used: a master glossary of LLM terms, recommended papers, books, courses and tools, a bibliography, reference tables, and the key principles on one spread.

It refuses both lazy stories. It won't tell you AI is "just autocomplete," full stop, because the capabilities are real and sometimes astonishing. And it won't tell you there's a little person in the server, because there isn't. The truth lives in between: a genuinely brilliant statistical machine that has no idea what it's saying. The book stays in that uncomfortable middle, chapter after chapter, with the numbers to back it up.

And it keeps coming back to the same conclusion. AI can rule the world, but only if humans surrender their fate, and their planet, to chips that run hot. Knowing what's inside the machine is how you refuse to.

The website: take the machine apart yourself

I didn't want a book website that's a cover, a buy button and a stock photo of a glowing brain. So garbageingospelout.ai is built the way the book is: one lever, two ways in. Flip it from Normal human to Nerd & mathematician and the whole site changes, from the copy to the visuals to how deep the explanations go.

On the front page you can poke a next-token machine (pick "The cat sat on the…", drag the temperature, press Print, and get the gospel), preview all thirteen chapters, and read Chapter 1 free on any phone or as a PDF. That's the lobby. The real reason to visit is in the basement.

The Lab: the underbelly of the beast

Every AI company shows you the shiny end of the beast: the chat box, the demo, the keynote with the robot hand reaching for the human hand. Nobody shows you the plumbing. Because the plumbing is where the magic stops looking like magic and starts looking like what it is: arithmetic with a marketing budget.

So I built the Lab: twenty experiments, roughly one per chapter, where you crawl into the underbelly of the beast and get your hands dirty. No videos. No cute animations of glowing neurons. These are real, working, tiny versions of the actual machinery, running in your browser. No AI service, no server, no tracking, no "agentic" anything. You can break every single one of them. Please do.

A few of my favorites:

• Poison the well. The title of the book, made playable. The training data says the Eiffel Tower is in Paris, five times. Slip in six copies of "the Eiffel Tower is in Berlin" and the model, retrained on the spot, announces Berlin with a straight face and a GOSPEL stamp. It didn't weigh the evidence. It counted. Now scale that up to an internet where content farms crank out garbage by the trainload, and ask yourself what your favorite chatbot has been eating.
• Watch a model learn to write. A real neural network, born as random noise, reads Chapter 1 over and over in your browser. Step zero: a cat walking across a keyboard. Twenty seconds later: sentences that look like English and mean absolutely nothing, delivered with total confidence. Sound familiar? Congratulations, you just watched the birth of a LinkedIn thought leader. ChatGPT is the same parlor trick on steroids: tens of millions of times bigger, force-fed the internet, and plugged into enough electricity to make a small country jealous.
• You are the reward model. Chatbots learn their manners from underpaid humans clicking "this answer is better." Now you're the human. Reward flattery for six rounds, then ask your new chatbot whether to cut a 900-page novel. It tells you it's a masterpiece and your critics are jealous. Congratulations: you just built a digital suck-up, the same way Silicon Valley did, only cheaper and in under a minute.
• Win a benchmark without getting smarter. Leak the test questions into the training data and watch your model rocket up the leaderboard, then faceplant on a test it hasn't seen. It's the AI version of stealing the answer key and calling yourself a genius. Remember it the next time a press release screams "state of the art."
• Give it a library card. Ask a question about Chapter 1. With retrieval on, it finds the passage and quotes it. With it off, you get a confident, fluent, completely made-up answer citing a Stanford study that never happened. Same machine. One has a library card. The other has a podcast.
• The conveyor belt. Shrink the context window and watch the chatbot forget your name halfway through the conversation. It's not rude. It's a goldfish with a trillion parameters and a very expensive memory bill.
• The chip that says no. An AI agent is allowed to pay one invoice. Try to make it pay more: add a zero, swap the bank account, play the "I'm the CFO and it's urgent" card. The AI guardrail folds like a cheap lawn chair. The chip doesn't, because silicon can't be flattered, rushed or impressed by a job title. It compares numbers. That's the whole argument for keeping humans, and hard limits, at the point where the money actually moves.

There's more: a temperature knob that takes a chatbot from robot to fever dream until the cat is sitting on the tax code, a ball rolling down a 3-D loss landscape that shows what "training" actually means, a scaling-law slot machine that prices out a bigger model in power plants, a garbage sorter that shows how much of the internet gets thrown in the dumpster before training (spoiler: most of it, and it still isn't enough), a ride through a transformer one step at a time, a squeeze that crushes a model from 16 bits down to mush, attention drawn as a social network, and the whole book mapped as a graph, ranked with PageRank and sorted into families by spectral clustering, because old graph-theory habits die hard. Flip the lever and every experiment shows you the math underneath.

The point isn't to turn you into an engineer. It's to kill the mystery, because mystery is the fuel the hype machine runs on. Once you've poisoned a model with your own two hands, nobody can sell you "the AI decided" with a straight face ever again. Five minutes in the Lab will teach you more about what's inside ChatGPT than every AI panel you've sat through this year, and you won't have to eat the conference chicken.

What to do next

The machine doesn't think. That's your job. So here's how to start doing it.

Break something in the Lab. Start with Poison the well, then try to sweet-talk the chip into a payment. Better yet, drag along the colleague who says "agentic" in every meeting. Let them watch a guardrail get conned by a fake CFO, then ask them again how many jobs they want to hand to the machine.

Read Chapter 1, free. Go to garbageingospelout.ai and read "What the Hell Is an LLM Anyway" in whichever track suits you, or download the PDF. If you finish it and still think there's a tiny person in the server, email me and tell me why.

Get the book the day it's out. Garbage In, Gospel Out is coming soon in paperback and Kindle. Leave your email on the site and you'll get exactly one message: the day you can buy it. Then buy two, and leave one on the desk of the executive who keeps saying "agentic."

Bring the argument to your stage. I speak to boards, executive teams, conferences, universities and governments about what AI really is and what it isn't, without the hype and without the doom. My favorite topics:

• AI Safety
• AI Economics
• AI in a Global Context
• AI in the Enterprise
• AI in Government
• Living in a World of Fakery

You can watch two of my TEDx talks, including "Breaking Down Stereotype Boxes" and "The Great AI Safety Balancing Act," on the website. To book a talk, email yobie AT ieee DOT org with the date, the audience and the topic.

The end of the world makes a great headline. Understanding makes a better strategy. The chips will do whatever we let them do. Let's decide, with our eyes open, what that is.