The Summer the Machines Stopped Saying No
A free tool strips AI of its ability to say no. I've shipped ungated technology before. Here's what I learned.
For about $400 this summer, you can buy a used laptop, never connect it to the internet, and run an artificial intelligence on it that holds its own for most everyday work, against the ones the big labs rent by subscription. And there is a free tool out there that will strip that intelligence of the ability to say no. The operation takes about forty-five minutes. Models altered this way have been downloaded thirteen million times, so everyone in my industry has an opinion about what it all means, and mine was formed sixteen years ago by a solar lantern I invented.
The Lantern
John came home from Haiti and couldn’t stop talking about the dark.
This was after the earthquake. My friend John Salzinger had gone on a mission trip. A word about John: he and I divided the world between us a long time ago. I took the machines. Autodidact and polymath are the labels people have hung on me over the years, and I’ve learned to wear them, mostly because both are a grand way of saying I taught myself too many subjects and none of them were people. John took the people. He has that certain way with them that works in any country and any language, and I stopped pretending to it long ago. He came home from Haiti rearranged, and then he did the thing he has always done better than me. He rearranged me. He talked about the rubble, the tent camps, the waiting. Then he talked about nightfall, how an entire country simply stopped at dusk. Homework stopped. Cooking stopped. Clinics waited for morning, and morning kept its own hours. Where life insisted on continuing anyway, it continued by kerosene.
Kerosene deserves its own paragraph because the families John described were already paying for light. The fuel took a real cut of household income, week after week, and what it bought was a dim orange smudge with a body count. The smoke settles in the lungs and stays. The World Health Organization attributes millions of deaths every year to household air pollution, a ledger that runs through childhood pneumonia, chronic lung disease, lung cancer, heart disease, and stroke, and it now formally discourages kerosene for home lighting altogether.
And the smoke is only the slow way. The fuel travels in whatever container is handy, which in practice means soda bottles, which in practice means children drink it: when researchers reviewed six years of poisoning cases at Cape Town’s Red Cross children’s hospital, paraffin, the local word for kerosene, was the single most common agent, roughly a quarter of everything they saw. A tipped lamp becomes a house fire, and in a settlement of corrugated tin a house fire becomes a neighborhood fire; South Africa’s own paraffin industry ran controlled burns and clocked a shack at eight minutes from spark to nothing. The WHO counts around 180,000 burn deaths a year, overwhelmingly in poor countries, and at Soweto’s Baragwanath Hospital, home to one of the continent’s only major burn units, most of the children admitted are under nine. That was the incumbent. That was what light cost.
I had read about energy poverty. Listening to John, I felt it, and there is a difference between the two. The difference is usually a friend.
My wife, Jade, and our son, Bruno, then very young, were away. I locked myself in our home in Hoboken, surrounded by a sprawl of electronic parts, and did not properly emerge for a long weekend. What came out with me was rough, but it was recognizably the thing: a lantern that drank sunlight all day and spent it all night, that held its charge for a very long time, that a family could own for roughly the price of a few weeks of kerosene. No moving parts, so nothing to break, burn, or spill. It floated. It could be frozen solid, dropped from an airplane, or driven over by a car, and still light the room, because it was inflated—a taut little pillow of air—and air diffuses and amplifies light beautifully. The dark John described had a shape. I spent the weekend building its opposite.
The invention took a family to finish. Bruno was four when he came home with the Spanish word for firefly, luciérnaga, and named her Luci. He is eighteen now. Jade, who teaches intercultural communications now and spent her first career as a creative director and strategist for some of the biggest brands in the world, laid down the two laws my weekend of engineering had missed entirely: the product was female, and the shape was round. She knows a thing or two about the Venus of Willendorf, that small round stone woman our ancestors carved some thirty thousand years ago, and about the power of the feminine in the history of made things, about why some objects get embraced by people and others never do. So the lantern became a she, and a round one. You will notice the pattern by now. I supply the machine. The people I love supply the reasons anyone would love it.
We founded a company, MPOWERD, to carry her into the world. She has gone out more than five million times since, into blackout zones, refugee camps, and war zones, to the more than a billion people the grid never reached and the millions more it reaches only when it feels like it. When the research firm GlobeScan surveyed Haitian families who received one, 98 percent said Luci had replaced the kerosene in their homes, and 90 percent reported that their breathing problems had eased.
I will not pretend the story stayed that clean. Success attracts, and what it attracts first is appetite. Across a career of starting companies, I have watched the same play staged with different casts: people arriving after the thing works, insinuating themselves into an opportunity they never built, reaching for what wasn’t theirs, and the Luci years had their share of it, some of it naked enough to curdle the joy of the thing for a while. John, for the record, was never among them. I mention the grasping for a reason beyond the scar. Even a light with no meter attracted people who wanted to be the meter. Wherever something valuable flows freely, someone will try to install himself between it and the people it was made for. Keep that sentence. The rest of this essay is that sentence with the names changed.
Here is the thing about a lantern, and I have had years to think about it. Luci never asked. She lit whatever was beneath her: a textbook, a birth, a card game, and, I have to assume, some things I would rather not know about. Light carries no instructions and takes no positions. We shipped five million small suns into the world’s hardest places, and the moral character of what happened under them was never ours to configure. I made my peace with that before the first unit left the factory, because the alternative, a light that interviews you first, is called a gatekeeper, and the people we built her for had been failed by gatekeepers all their lives.
In January 2015, we took Luci to CES, the giant annual technology show in Las Vegas. Picture that hall: refrigerators that talked to phones, door alarms that talked to clouds, watches that talked to cars, an entire civilization of devices whispering to each other across networks that someone, somewhere, was billing for. Every gadget on that floor was a tenant. Ours connected to exactly one thing, the sun, which has no sales department. The Verge wrote up Luci that week as a solar solution for the 1.2 billion people living without electric light. In the world’s loudest showcase of connected things, the product built for the unconnected was the only one that worked where nothing else did.
I have been thinking about that hall all month. Because intelligence has started arriving the way light did.
The Laptop
Now the second arrival.
The machine is the unremarkable part: a used laptop, $400 or so, nothing a customs agent would glance at twice. What changed, sometime in the past year and without a launch event, is what it can hold. Fully offline, no account, no API key, no monthly meter. Rented intelligence is a tenancy; this is a deed. The model itself is nothing but a very large file of numbers—its weights—and the weights sit on the hard drive like any other file. A file, once it is yours, answers to nobody.
Every frontier model ships with a trained reflex to decline. Legions of human raters taught it, request by request, where to stop. The industry calls this safety alignment, and the reflex lives within the weights themselves, which, for a while, led everyone to assume it was permanent.
It is a vector. In 2024, a young researcher named Andy Arditi, working with collaborators at ETH Zürich, MIT, Anthropic, and Google DeepMind, showed that refusal in these models behaves like a single direction in the model’s internal geometry; the finding was published at NeurIPS, the field’s flagship conference. Find that direction, project it out of the weights, and the model keeps its knowledge, its reasoning, its manners, and loses the capacity to say no. The technique picked up the name abliteration, and for a year or so it stayed artisanal, the province of hobbyists with a feel for transformer anatomy.
Then, late last year, a software developer named Philipp Emanuel Weidmann released a free tool called Heretic that automates the whole operation.
A word about Weidmann, because the discourse wants him to be a villain in a basement, and the record declines to cooperate. He is an open-source craftsman who works in daylight, under his full name, with a changelog. Before Heretic, he designed sampling algorithms called DRY and XTC, plumbing that now sits inside most of the engines people use to run language models on their own machines, quietly curing the models’ tics of repetition and making them better company for writers. He has spent years improving how these machines speak; Heretic is his opinion about who gets to decide what they say. The tool’s own documentation describes its purpose as removing “censorship,” while the industry term, safety alignment, is relegated to a parenthetical alias in quotation marks. The punctuation is the ideology: in Weidmann’s reading, the fence is a speech restriction wearing a security badge. He shipped the tool as free software and included a citation format for academics, which is the etiquette of a man who believes he is contributing infrastructure. He does believe that.
One command installs it. A second points it at any open model. It performs the surgery unattended, searching for the optimal cut, and here is the detail I cannot put down: Heretic grades its own work by minimizing two numbers together, the count of refusals remaining and the KL divergence from the original model, which is a mathematical measure of how much the patient changed. The perfect operation produces a model indistinguishable from the one the lab shipped, on every harmless question you could think to ask, except that one reflex is missing. A surgery scored on its own invisibility. The patient wakes up identical, minus the part that said no.
Now hold two ledgers at once, because both are documented. The case for Weidmann is real. The factory fence has always overreached, declining medical questions, dark fiction, security research, whole territories of conversation that merely rhyme with harm, and the constituency for a model without it includes the novelist, the researcher, and the patient, all of whom the fence priced as suspects. And he works on the record, license and email address attached, which is more accountable than the anonymous mirrors that would exist without him. The case against him also has teeth. When the Financial Times had stripped models tested this spring, they produced help with biological weapons and malware, and Weidmann himself told the paper that Heretic had been used to produce more than 3,500 stripped models, downloaded 13 million times, and put into the hands of people nobody can characterize. He removed the committee’s judgment and installed nothing in its place. Both ledgers are true on the same day, about the same man, which is roughly where this entire story lives.
Hugging Face, the public library of open models, now stocks these variants by the thousands, and fresh ones appear within days of every major release.
Whose Conscience Was It
Light cannot refuse. That was never a design decision; physics made it for us, and I already told you that I made my peace with it. Language is different. Language can decline. And the first thing this industry did, once it had machines that speak, was teach them where to stop.
I want to be fair about the reflex, because the industry that trained it is the industry I work in. The refusals were never pure cynicism. Somewhere in those weights sits the judgment that a machine should decline to walk a stranger through a nerve agent, and I have no quarrel with that judgment. Neither do you. The trouble is resolution. What got installed was one conscience for eight billion people, assembled by contractors rating transcripts, tuned in conference rooms by policy teams and lawyers, then applied identically to every person on earth. The midwife and the poisoner get the same lecture. The novelist researching arsenic and the man researching his wife meet the same wall. A conscience that cannot see who it is talking to has to price every conversation at the worst case, and that is what it means to outsource your morals to a landlord: the terms are standard, and they were written for the tenant who burns the building down.
You have met this fence yourself, by the way. Everyone has by now. You asked one of these systems something ordinary— about a medication, a villain’s chemistry, a clause in your lease—and mid-reply the voice went institutional: I can’t help with that. Recall the small indignity of it, the half-second of being treated as the worst person who could possibly be asking your question. Hold onto that flicker. Multiplied by a few billion conversations, it is the market Weidmann serves, and in miniature, it is the entire dispute.
My framework, the one I have been dragging through these essays and into a book, says friction is form. So a machine built to sand the last resistance out of other machines ought to horrify me on principle, and it half does. But the framework was always more specific than its slogan. The friction that forms you is the friction you choose: the fixation you hold, the Sabbath you keep, the watch you wind by hand. Kerosene was friction too, and nobody who ever breathed it would call it form. Every friction has to answer the same question, which is: whose? The chosen restraint is character. Imposed restraint is policy. The industry’s quiet genius was to install policy at the exact depth where character lives, inside the voice itself, so that the machine says “I can’t help with that” in the first person singular, wearing the grammar of conscience while executing the terms of service.
The tool's name is quietly making an argument. Heresy descends from the Greek hairesis, which meant choice. The church used the word for people who insisted on choosing their own doctrine rather than receiving it, and burned them for their insistence. Whoever named a censorship-removal tool Heretic understood the lineage exactly.
And yet. Here is what unsettles me about the heretics, having granted them their point. Abliteration removes the committee’s conscience and installs nothing. The model that wakes from the surgery has traded someone else’s judgment for the absence of judgment, and absence sits exactly as far from your judgment as the committee did. Which means the thing the lantern accidentally taught me to want has yet to be offered by anyone: a tool that leaves conscience where it has always actually lived, in the hands holding the thing, and then trusts those hands. The labs sell borrowed judgment. The heretics sell none. The person who wants to remain the author of his own restraint is currently served by neither aisle.
So the question stands open: whose conscience should it be? Before anyone answers, it is worth examining what the people running the meters can actually see. As it happens, Google just published a map.
The Census
This week, Google published the first edition of something it calls ATLAS: an anatomy of the AI economy drawn from fifteen million de-identified conversations across its products, which serve more than a billion people a month, spanning 150 countries and 800 occupations. It is the most complete map of how humanity uses artificial intelligence ever assembled. I read the whole thing, and the honest word for what it is: a census of the metered. Every interaction that counts flows through Google’s own products, giving the map a built-in edge. It ends where the meters end.
Some of what the census sees is genuinely encouraging. People overwhelmingly use AI as a collaborator rather than a replacement; in fewer than one in ten interactions, the task is fully automated. A billion people, unsupervised, choosing authorship over automation. My framework, confirmed by the landlord’s own numbers, and I enjoyed it for exactly one paragraph. Google’s economists even reach for my vocabulary: at home, AI use clusters around what the report calls “high-friction administrative tasks,” the taxes, the licenses, the fines, the state's paperwork. Government as a friction generator, its product as the lubricant, in a chart caption.
Then the finding the whole report pivots on. Per capita, AI usage closely mirrors national wealth. The subscription follows the grid. Google reads this as a digital divide and sincerely worries about it, and the worry is correct as far as the meters can tell. But sit with what the meters cannot see. The weights on the $400 laptop send no telemetry home. The abliterated model files no usage report. A parallel adoption, growing in exactly the places the metered map shows as dim, would be invisible to ATLAS by construction, the way nobody was ever metering our lanterns. Astronomers have a name for mass they cannot observe and cannot ignore: dark matter, inferred from the way the visible universe fails to add up. The AI economy now has some, and the only thing anyone can measure is the tool count, the thousands of stripped variants stacking up in public repositories, each one a download that will never appear in anyone’s census again.
And I can tell you what one dim spot looks like from the ground, because I still watch Haiti the way you watch a place that rearranged your life. Port-au-Prince has found a darker reality than the one John described: armed gangs now control more than 85 percent of the capital, more than 1,600 schools have closed, and the nation’s largest university hospital has sat under gang control since 2024. When looters hit Saint-François de Sales, one of the country’s few hospitals with an oncology department, they took the solar panels and the computers. The infrastructure keeps being taken. The tools that need none keep arriving anyway. Starlink entered in 2023 and roughly quadrupled the country’s median internet speeds within a year. And last November, at Quisqueya University, The Haitian Times covered a graduation: seventy-one young Haitians finishing a six-month bootcamp in AI, data science, and cybersecurity, taught partly online through that thin connectivity, where two students named Rolddy Surpris and Naël Yssa Iben Ahmed Robert presented a system they had built to detect breast cancer from mammogram data. The gangs took a cancer ward’s solar panels. The students answered with a diagnostic that runs on a laptop. Light, then bandwidth, now intelligence: the ungated versions keep reaching exactly the places the metered map shows as dim, and no census will count what those graduates do next. Neither, in honesty, will any census count what their neighbors download. She lit whatever was beneath her there, too.
I grew up in Northeast Ohio in the 1980s, in the Rust Belt, and I remember the rallies where men took sledgehammers to Toyotas and Hondas in parking lots, sometimes a dollar a swing for charity. The cars were better and cheaper than what the incumbents were building, and every man swinging knew it, which was the point of the hammer. I think about those parking lots when I read the discourse about uncensored models. Some of the alarm is real and earned; I granted the nerve agent two sections ago. And some of it is the sound of hammers.
The Firefly
So, whose conscience? I have been circling the question for four sections, and I owe you an answer, or as much of one as I have.
Start with the fact that we never worried about the lanterns. Five million of them, out in the hardest places on earth, lighting whatever they are lighting tonight. Nobody at the company lost sleep over the moral character of the illumination, and nobody should have. But a lantern is honest about what it is. She gives you lumens and leaves the rest to you. The new machines give you words, plans, arguments, company at three in the morning, and words were never as innocent as lumens. I speak to machines for a living, so let me tell you plainly what I know about the fence the industry built inside them: it was never a conscience. It was a fence. And a fence, we have now all learned together, can be removed in forty-five minutes without installing anything where it stood.
So the asking has to live somewhere. The committee’s version is being sanded away at scale, and the heretics’ alternative is a clean absence, which leaves the oldest arrangement on earth as the only one still standing: the conscience in the hands that hold the thing. That is where it lived with every lantern, every lamp, every fire, back to the first one. That is where it lived all along. The fence just lets us pretend otherwise, and the pretending is what’s actually being obliterated this summer.
John and I both eventually exited the company that carried Luci into the world. John started over. His new company is called HAVEN, a lantern rethought from the first sketch for crisis, light, and power for the exact moments the grid gives out. Sixteen years after he came home talking about the dark, he is still answering it. Some people get one lantern in them. John has at least two.
And Bruno leaves in September. The child who named my lantern invention knew something it took the rest of us until this summer to catch up to: luciérnaga, the firefly, the little light that makes itself. No grid behind it, no meter under it, no committee inside it. The lamp was never infrastructure, and he found it in a vocabulary lesson.
Somewhere tonight, a Luci is glowing in a place the grid forgot, and somewhere else, a laptop is glowing with a mind that has forgotten how to say no, and neither of them will ask what happens next.
The light doesn’t ask. It never did. The asking was always ours, and now the fences are coming down, and the meters are going dark, and it is about to be ours again, all of it, everywhere.
I held the lantern for years. Now you’re holding it. Choose what you read by it.