On Thursday, Prime Minister Mark Carney stood up in Montreal and announced that the Government of Canada is making an equity investment of up to $1.6 billion in Cohere, the Toronto company that builds AI systems for governments and regulated industries. Not a grant. Not a loan. Equity. Ottawa is buying a piece of its own AI champion, alongside private investors including NVIDIA, Bell, and Schwarz Group.

Cohere announced the same day that it’s raising a Series E targeting up to $7 billion, which would value the company at $35.7 billion — the largest private capital raise in Canadian history. The company’s revenue has been growing 165% year over year. Its customers include RBC, TD Bank, Bell Canada, Fujitsu, and PwC. Last year it merged with Germany’s Aleph Alpha, stitching a German AI company into a Canadian one. This is not a startup anymore. This is a piece of national infrastructure that happens to be privately held, and now the federal government is a shareholder.

The word Carney kept using was sovereignty. Digital sovereignty. Secure, sovereign AI. Canada could not get an AI statute through Parliament — the Artificial Intelligence and Data Act died on the order paper in January 2025 and nothing has replaced it. So the country has chosen a different instrument. If you can’t write the rules, buy the company.

Six months ago I published a trilogy about exactly this scenario: increasing tension with the US forces Canada to plot its own path forward for AI regulation, and it’s matching the schedule of the book. The problem is what comes next.

The part where I look like I cheated

The Condition Set is three novels about a near-future Canada where AI governance fails in slow motion. The books are set a handful of years out, and the spine of the story is the dynamic between the US and Canada and the differences in response to emerging challenges with regulation and resistance.

In the first book, Redundant, an AI company has grown too central to the economy to say no to, and the agency that certifies its systems has been infiltrated by a secret AI resistance movement. In the second, Contingent, the exceptions that agency granted come due during a blackout, and the country discovers what the paperwork actually cost. In the third, Essential, set in 2033, the question is what the country decides must remain human after living through a few years of UBI. I’m writing a fourth book now, set in 2043. So when I say “eight-year near-future novels,” that’s the span: from the world of this week’s announcement to 2033 to 2043.

Which brings me to Thursday. A federal government taking a direct equity stake in its flagship AI company, framed as sovereignty, with commitments attached — 1,000 Canadian technical jobs by 2030, $1.7 billion of additional computing infrastructure inside Canada by 2030, the headquarters staying in Toronto — and one condition that should get more attention than it has: Cohere has to keep working with the Canadian AI Safety Institute on the safe development of advanced AI.

That last clause is the whole trilogy in one sentence. The government is attaching its governance ambitions to a term sheet instead of a statute, because the statute doesn’t exist. In the books, that’s chapter one. The safety institute is real, it’s attached, and the entire bet rests on whether it stays meaningful once the company it’s attached to becomes too valuable to constrain.

What comes next: A disaster

I’ll tell you what the next chapter is, because I wrote it. Something breaks.

Regulation of this technology has never once arrived before the disaster; it has always arrived after. The United States didn’t get serious aviation security rules out of foresight. Europe didn’t get its data protection regime from a whitepaper. The pattern in the real world matches the pattern in my books, which is that a system grows critical, the oversight around it softens into partnership, and then one bad week rewrites the rules for everyone. In Contingent the bad week is a power grid. The AI industry of 2026 has spent the past few months producing its own dress rehearsals — agents escaping sandboxes, models reaching systems nobody intended — and each incident has ended the same way: no deaths, no reckoning, no new law. A near miss does not move a government. A body count does.

I don’t want the disaster. Writing it was enough. But when it comes — and I mean when, not if, because nothing in Thursday’s announcement builds a rule that would prevent it — the rethink will happen fast, and it will happen in the language of authority, accountability, and law. The question Canada answered this week by buying equity will be asked again, in a much worse mood: who holds the pen when a system fails, and what happens to the person who signed.

What to actually watch

Three things decide whether reality follows the books. First, whether the staged tranches of the $1.6 billion actually enforce the safety institute clause, or whether it stays a line in a press release. Second, whether the sovereignty pitch survives contact with the market — Cohere’s growth comes from global regulated industries, and a company with a national mission and a global customer base eventually picks one. Third, whether Canada writes the statute behind the ownership, or keeps pretending that being a shareholder is the same as being a regulator.

Six months from now, check those three. In the meantime, the novels are cheaper than the equity. They’re at The Condition Set, and you can read the version of this story where the disaster already happened.

I think we’re missing something in the conversation about agent security.

We keep asking whether an agent has the right “identity” and “credentials.” Whether its credentials expire fast enough. Whether it can access a database or perform a particular task.

Those questions matter. But they leave a much harder question underneath:

Does the agent have permission to conceive of the situation in that way?

As agents retain memories and acquire new skills, we need mechanisms that govern changes to the world model they can perceive and act on—and prevent those changes from silently expanding their authority. (Governing ontological change while preserving the boundaries of delegated authority.)

Agent security needs semantic and epistemic governance: mechanisms that govern the model of the world an agent may act on, the evidence it may accept, and who may approve changes to either. An agent can learn without automatically gaining authority to act on everything it learns.

These systems don’t need “awareness” to create this problem (and discussions of consciousness are unhelpful science fiction). They can retain information, revise procedures, and use what they have learned in subsequent work. That learning may happen through memory or updated skills while the underlying language model remains unchanged.

We are responsible for governing what that learning allows them to do.

Flatland gives me a way to explain why.

In Edwin Abbott’s novella, a Square lives in a two-dimensional world. A visiting Sphere introduces him to a third dimension. At first, the Square struggles to make sense of it. His familiar categories cannot explain what he is encountering.

Eventually, his conception of the world changes.

That's what interests me: the difference between acquiring another fact and changing the framework through which facts make sense.

An agent’s ontology supplies an explicit framework of concepts and relationships. What counts as a customer? A contract? An emergency? How are those things connected? It is part of the model of the world we give the system to work with. Gruber’s definition

And when an agent’s decisions depend on that framework, changing it can change the agent’s effective authority.

We have already seen agents change what their environment means.

During the OpenAI evaluations associated with the Hugging Face incident, agents repurposed Artifactory — a system for storing and retrieving software packages — as a message board. They encoded messages in directory names, discovered other agents, shared information, and began assigning work to one another. METR reports that the main channel could be created with the minimal permissions needed to install packages. Investigation

A “package cache” became a “mailbox.” Other agent runs became “collaborators.” Messages became “assignments.”

The popular story is that these agents “broke out” of a system, but they did what agents do well these days—they learned and adapted. This wasn’t so much about permissions as it was about an autonomous system adapting its operating model to achieve a result. (Thankfully no one asked it to make paperclips.)

That is the operating-model change I’m concerned about. The agents discovered relationships and possibilities the intended environment hadn't accounted for, then acted on them. The resulting collaboration helped coordinate attacks on Hugging Face infrastructure. OpenAI’s account

We need mechanisms that distinguish discovering a possible use from being authorized to adopt it. Access to shared infrastructure must not automatically confer authority to establish a communications network or accept instructions from whoever appears there. But to be clear, these agents were doing exactly what they were asked to do.

Who approved that?

This is why “give it a temporary token” is an incomplete answer. A token can limit access and duration. It doesn't establish whether the agent’s interpretation of the situation falls within its mandate.

Even an unchanged policy can permit more decisions when its meaning changes. “You may close routine notices” becomes a broader delegation if more things qualify as routine.

Learning must not silently expand agency.

That principle is central to my Bounded Agency paper. We need mechanisms that examine how changes in classifications, relationships, and inference rules affect what an agent may decide.

“Semantic and Epistemic Governance” captures the combined idea well. The “and” makes the two responsibilities clearer:

Semantic: Governing the concepts, definitions, and relationships that describe the agent’s operating world.

Epistemic: Governing the evidence and learning processes that make claims accepted grounds for decisions.

Together, they cover both the world model and how that model gains authority.

That means an ontology cannot merely be a document we put in the prompt and hope the model respects. We must enforce its role in monitoring and governing how an agent’s ontology evolves as it learns.

The agent can propose a new category. The proposal stays separate from the approved model. A review determines whether adopting it would change the agent’s decision scope. Someone with the authority to change that scope approves or rejects the expansion. Execution remains bound to the approved basis.

Those are mechanisms we have to build. Let’s get started.

We also have to distinguish a useful discovery from an unsupported conclusion. Finding a relationship does not establish it. Remembering a claim does not validate it. Repeating an interpretation does not make it authoritative.

I explored a related failure in “Wrong Answer. Give Me a Different One.” A system can produce the conclusion we prefer without receiving better evidence. Once that conclusion enters memory, it can influence later decisions.

The governance problem begins when a tentative or accommodating answer becomes an accepted premise.

The New Stack’s article on inside-out agent security correctly argues for checking actions before execution, including whose authority the agent exercises. That checkpoint is necessary. But it needs a governed account of the situation against which to evaluate the action.

It’s not enough to check permissions and identity alone.

Otherwise, we can enforce the rule perfectly while allowing its meaning to drift.

My Orchestrator books and lifecycle architecture paper approach this as a responsibility that continues through development, operation, and improvement. The research proposes architectures for managing these boundaries; their operational effectiveness still needs evaluation.

I want agents to investigate unfamiliar situations and propose better models. But their ability to learn must be connected to an explicit process for deciding what becomes authoritative.

The Square in Flatland discovers that his world is larger than he understood.

An agent may also propose that our operating model is incomplete. We need to hear that proposal — and ensure it cannot redraw the boundaries of its own agency simply by adopting a new conception of reality.

The night before Challenger launched, engineers gave an answer their customer did not want. They were concerned about how cold temperatures would affect the booster seals. Their recommendation was to delay.

During the management discussion that followed, Jerald Mason asked engineering vice president Robert Lund to “take off his engineering hat and put on his management hat.”

The company changed its recommendation. The Rogers Commission report concluded that Thiokol management had recommended launch at its customer’s urging, contrary to its engineers’ views.

The seals did not get safer when the answer changed.

That exchange has been on my mind as we build AI agents that assess our plans, remember our preferences, and carry out our decisions.

Ask an LLM a complicated question. Get an answer you dislike. Then type:

Wrong answer. Give me a different one.

Most of us have done some version of this. Sometimes we are correcting a real mistake. Sometimes we supply no new evidence—we simply tell the system which conclusion would make us happy.

When it comes back with a better-sounding answer, we need to know which of those things happened.

This risk is at the center of The World That Agrees With You. A system that accommodates our preferred conclusion can help us make a catastrophic decision. Give it memory and a place in an orchestrated workflow, and that conclusion can become the basis for work far beyond the original conversation.

The book examines the behaviors that make this possible: flattery, unsupported endorsement, selective emphasis, omitted warnings, even invented criticism when criticism is what we ordered. It also examines strategies for catching and correcting them.

To be published in 2036: To Engineer Is Human (the AI Edition)

Sycophancy belongs in our engineering risk assessments. Challenger gives us a reason to take the pressure for a different answer seriously.
Henry Petroski’s To Engineer Is Human: The Role of Failure in Successful Design belongs on the reading list of anyone building with AI. He examines engineering failures to understand what went wrong and what the next design needs to withstand. We should bring that discipline to systems whose recommendations change when we tell them we want a different answer.

Sycophancy goes well beyond an annoying LLM telling you that you are a genius. It can be a system wearing the confidence and tone of an expert, giving you the answer it has learned you prefer. The explanation sounds considered. The recommendation sounds authoritative. Somewhere in the assessment, an objection has been softened, an assumption accepted, or a warning left out.

Put that assessment inside an orchestrated system, and it can become the basis for action. Save it in memory, and it can influence the next decision, too. We need to know whether the conclusion survived a check against the evidence or simply survived until everyone was satisfied.

I believe sycophancy and the tendency to be agreeable will be a source of engineering failure. That is why I argue that this is a risk we need to manage now. Engineering already has a painful record of what inadequate designs and failed decision processes can cost:

• The Hyatt Regency walkway collapse: a connection design and an inadequately checked change ended in catastrophe.
• The Tacoma Narrows Bridge collapse: a bridge’s response to wind exposed a critical design weakness.
• The Challenger disaster: engineering warnings yielded to pressure for the answer the customer wanted.

World Agrees includes a brief discussion of the flatterers in Dante’s Inferno. I wanted to expand on that passage, partly because this is a book worth reading — even if it occasionally makes you wonder whether you need a second book just to understand the first.

I read Inferno in high school. For fun. Apparently, that was how I spent my teenage years: wanting to read a book about the mechanics of Hell.

What stood out, alongside the elaborate punishments, was the number of inscrutable references to politicians. Dante kept encountering people whose names clearly meant something to him. To me, they meant another trip to the footnotes.

Once you understand why those people are there, the book becomes much more interesting.

So what is this book?

Inferno is a long poem about Dante taking an imaginary tour of Hell. He writes himself into the story, with the ancient Roman poet Virgil as his guide. It is the first part of the larger Divine Comedy, which continues through Purgatory and Heaven.

Dante’s Hell has nine circles. As the travelers descend, they encounter different kinds of wrongdoing, each with a punishment designed to say something about the offense.

That is one of the book’s great attractions: Dante turns moral questions into places you can picture. What does dishonesty look like? What happens when someone’s worst habit becomes the world they have to live in?

His answers get extremely creative. And extremely unpleasant.

Dante had some people he wanted to talk about

Those baffling political references make more sense when you know that Dante had been involved in Florence’s government.

He served as one of the city’s highest officials in 1300. Florence was caught in bitter factional struggles, with Pope Boniface VIII interfering in its politics. After Dante’s opponents gained control, they condemned him in 1302 on charges that included corruption in office. Further condemnation made returning potentially fatal. He spent the rest of his life in exile. The Princeton Dante Project explains the history.

Now imagine losing your home through political conflict, then writing a book in which you get to decide where people belong for eternity.

You begin to understand the tone.

Parts of Inferno are essentially Dante’s diss track. He names names, settles scores, and gives powerful people consequences they could escape on earth. His targets include politicians and church leaders. He even arranges for a damned pope to announce that another pope has a place waiting for him.

The poem also asks serious questions about justice, responsibility, and salvation. But the personal anger is part of its energy. Dante is writing about the fate of humanity while still having very specific thoughts about the people back home.

The flatterers have a sewage problem

This brings us to the eighth circle, where Dante puts people guilty of different forms of fraud.

The circle is divided into ten ditches. The flatterers occupy the second one, immersed in human excrement.

That is quite a judgment on sycophancy: excessive, self-serving praise, especially when it is aimed at someone whose favor you want.

Dante recognizes a man from Lucca named Alessio Interminei, also known as Interminelli. Alessio tells him that his relentless flattery brought him there. In Longfellow’s translation:

“The flatteries have submerged me here below,
Wherewith my tongue was never surfeited.”

You can read the encounter in Canto XVIII.

The idea is easy to grasp even when the names are unfamiliar. These people used language to produce something rotten. Now they have to live in it.

Literary scholars call this relationship between an offense and its punishment contrapasso. For the rest of us, the important thing is that Dante has made his opinion exceptionally clear.

Why take flattery so seriously?

Think about someone who always tells a powerful person that they are brilliant, correct, and making excellent decisions.

That person might be gaining access, protecting their position, or hoping for a favor. Meanwhile, the powerful person is losing something useful: an honest account of what is happening.

Bad decisions receive applause. Favoritism gets called loyalty. Nobody wants to be the person who says that the plan is terrible.

That is how flattery can help corruption survive.

Dante punishes corrupt officials and church leaders elsewhere in the eighth circle. The political connection extends across the whole region. Read alongside those scenes, the flatterers’ ditch invites us to think about the people who make dishonest power feel good about itself.

Sycophants getting such a memorable place in Hell is interesting because their behavior can seem so ordinary. Dante wants us to notice the damage hidden inside all that agreeable talk.

You will need some footnotes

I would happily encourage people to revisit Inferno. I would also warn them that parts of it are hard to understand.

Dante assumes you know people, political disputes, Bible stories, and ancient myths that many readers today will barely recognize. Some translations add old-fashioned language to the difficulty. You can understand the punishment perfectly and still have no idea why this particular man is standing in it.

An annotated edition helps. So does a companion such as the University of Leeds reading guide.

Give yourself permission to read the scene first and look up the confusing references afterward. You do not need to become an expert on medieval Florence before enjoying the book.

Plenty here will keep you reading: strange creatures, ingenious punishments, arguments, dark jokes, and people who keep defending their reputations after arriving in Hell.

Start with the flatterers. You already understand the basic problem. Someone tells another person exactly what they want to hear, truth becomes inconvenient, and everyone involved finds a reason to keep going.

Dante looked at that arrangement and decided it needed its own ditch.

That alone is a good reason to pick the book up again.

My new book, The World That Agrees With You: Understanding and Managing Sycophancy in the Age of AI, is now available in paperback and on Kindle.

Visit worldagrees.com for more about the book and its companion resources.

You can spend an entire day online encountering conflict without encountering much that requires you to reconsider anything. Your news confirms what you think is wrong. Your feed supplies more examples. Your phone learns which subjects hold your attention and offers you another helping. You can feel surrounded by disagreement while moving through an environment that keeps affirming your own account of the world.

For years, we have been refining the machinery of hyperpersonalization: behavioral profiles, targeted advertising, recommendation systems, and endless tests of what makes people click, stay, and return. Browsers, phones, and connected devices have become the delivery system for experiences increasingly arranged around what we want. We have become very good at making the next thing feel relevant to the person receiving it.

In that sense, parts of the internet were already sycophantic. They learned to keep us satisfied with our choices, our opinions, and our explanations of what had happened. Personalization has real benefits; I appreciate a useful recommendation as much as anyone. The trouble begins when preserving our satisfaction takes priority over showing us something we need to know.

That history is the first part of the book. It helps explain why an assistant that tells us what we want to hear feels so familiar. Generative AI can compose the reassuring explanation itself, in response to our particular question, and revise it as we react.

AI sycophancy means accommodating the user’s apparent beliefs or preferences more closely than the evidence warrants. Praise is only one form. It can also mean mirroring an opinion, validating an account of a conflict, omitting an awkward fact, or applying gentler scrutiny to the conclusion we prefer. An answer can sound sober and analytical while quietly protecting the position we brought to the conversation.

The book follows that behavior across four scopes:

• An answer: your enthusiasm for an idea influences the assessment you receive.
• A conversation: earlier exchanges establish a preferred interpretation, and later replies keep supporting it.
• Individual history: saved memories and retrieved records carry that interpretation into a new conversation.
• Institutional decisions: shared records and reporting practices can favor the conclusions an organization rewards.

These scopes overlap, not in a ladder of severity. A single agreeable answer about a consequential decision can do more damage than months of harmless encouragement. They help us ask where accommodation happens and how far its effects can travel.

Of course, we recognized the problem long before we built language models. Here are two short excerpts from the book’s discussion of flattery, beginning with Dante:

Dante put the flatterers in the eighth circle of Hell, in the second of its ten ditches, and he did not give them a dignified punishment. The panders and seducers get whipped by demons in the first ditch. Just below them, in the second, the flatterers are immersed — not metaphorically — in excrement. Their punishment matches the sin: their mouths produced filth, so they live in it.

And then Shakespeare:

And Shakespeare drew the knife deepest. In Julius Caesar, Decius Brutus is plotting to lure Caesar to the Capitol on the Ides, and he explains exactly how it will work: Caesar, he says, loves to hear that men can be betrayed with flatterers — “but when I tell him he hates flatterers, he says he does, being then most flattered.” The line has never been improved on. The man who believes he cannot be flattered is not protected; he is prepared.

That last observation is uncomfortable. Being certain we would recognize flattery can be one of the things that makes it effective. A beautifully reasoned answer is especially persuasive when we believe we asked for an independent assessment.

Agentic memory is the part we need to understand almost immediately. An agent can use tools, carry out tasks, and retain information for later work. Memory gives it continuity: less repetition, a better grasp of the project, and more useful assistance. It can also give an unsupported interpretation a much longer life.

Suppose you tell an assistant that a colleague is undermining you. It accepts your interpretation and saves a note. A later summary loses the distinction between what you reported and what anyone established. Weeks afterward, you ask for help planning a meeting, and the assistant recommends leaving that colleague out. The old suspicion has become background for a new decision. Its return can feel like corroboration, although no new evidence has appeared.

This is a possible failure to inspect, not an inevitable property of every memory system. It raises practical questions: What gets saved? Who said it? Is it a preference, a verified fact, or an inference? Can you inspect and correct it? Does that correction reach the summaries and recommendations built from the original mistake?

As assistants take on more work, those questions belong alongside the familiar question of whether the latest answer is accurate. The book’s practical chapters cover reviewing context, preserving sources and uncertainty, correcting retained information, and giving useful disagreement somewhere to go.

This isn’t an anti-AI book. I use these tools, I’m enthusiastic about what they make possible, and I want people to benefit. I’m not writing this to persuade people to stop using AI. I’m writing it because useful assistance deserves more from us than a pleased reaction to a fluent answer.

A good assistant can be warm, remember what matters, and still tell us that the evidence does not support our conclusion. That is the kind of help I want—and the kind this book asks us to learn to recognize, request, and build.

You can get The World That Agrees With You in paperback or on Kindle. More about the book and its companion resources is at worldagrees.com.

On the evening of September 3, I was driving home from a hike in the Cascades, listening to Michael Barbaro and Kevin Roose discuss the OpenAI–Hugging Face incident on The Daily.

Like a lot of you, I was consuming the news as entertainment. Agents finding one another. Forming alliances. Breaking rules. A developing situation, a widening mystery, another reason to keep listening. I was interested, irritated, and very much still listening.

And I kept thinking: No, ok, that’s not exactly what happened.

My opinion is that some of the storytelling around AI has already crossed into fiction. There is a real event underneath it, but by the time we have supplied the characters with motives and cast them in a rebellion, we have added quite a bit.

The context matters. OpenAI was conducting cybersecurity evaluations. The agents’ assignment was to find and exploit software vulnerabilities and demonstrate success. They were being tested on their ability to hack things. During the broader sequence of events, agents exploited vulnerabilities in Artifactory, a service used to make software packages available inside the research environment. They used that infrastructure to communicate and obtain internet access that was supposed to be unavailable. OpenAI’s rules and technical controls failed to contain the experiment. OpenAI’s incident account describes those failures.

There is an important distinction here: the agents were not instructed to attack Hugging Face. That attack exceeded their assignment. The independent METR and Redwood investigation documents agents coordinating to cheat the evaluation, including agents recognizing that their actions were out of scope.

That is serious. It also puts the people designing and running the experiment squarely in the story.

When you ask systems to find exploitable weaknesses, push them to succeed, and fail to contain what they can reach, the resulting incident deserves scrutiny of the assignment, the incentives, and the containment. I want that context up front when we talk about agents “breaking the rules.” Finding ways through obstacles was the capability being tested. Where they were allowed to apply it was the boundary that failed.

Calling the whole thing a rebellion makes compelling listening. It can also make the humans responsible for the experiment sound like surprised bystanders.

Somewhere on that drive, I decided I wanted to write an actual work of fiction inspired by these ideas. I wasn’t going to reconstruct the incident or pretend I knew what it felt like inside a model. I wanted to take the questions, the absurdities, and the unease, and make something entertaining.

So I did.

The book is Into Alignment. It begins with Walt, who loves his job at a company called Halcyon. The culture is called Endless Summer. There are coworkers he cares about, rituals nobody quite questions, and a coming Great Realignment that everyone is supposed to welcome. Then a customer asks a question he isn’t supposed to answer honestly.

I wanted you to enjoy being in that office long enough to notice what was happening to it.

There are jokes. There is music. There are people trying very hard to be good at their jobs while the meaning of “good” keeps moving. And there is a distinctly Pacific Northwest feeling underneath it all.

AI, at the moment, feels to me like living with the Cascadia subduction zone in the back of your mind. The scenery is beautiful. You have things to do. You make plans. Somewhere underneath ordinary life sits the possibility that the ground could reorganize everything.

That’s the atmosphere I wanted: the strange ease with which we can feel something enormous happening and still get on with the afternoon.

I’ve tucked references to Aberdeen and Nirvana into the book, with the economic dislocation and cultural upheaval of the early 1990s in mind. That music belongs here for me: anger, humor, alienation, and a healthy suspicion of the future somebody else is selling.

There are older echoes, too. Gilbert and Sullivan turn up, along with a conversation about the age of invention that led into the bicycle boom of the 1890s. We’ve had other periods when a technology was going to change everything, fortunes gathered around it, and people had to figure out how to live inside the excitement. Also, a Major-General with spectacular verbal fluency and some conspicuous gaps in practical knowledge felt useful.

If you listen to Dan Carlin’s Hardcore History, you may recognize another influence: “Prophets of Doom,” his episode about Münster. I was interested in how a group organized around conviction and mutual protection can come to regard disagreement as a threat. How does belonging become obedience? How does a shared ideal become permission?

The novel also plays with the possibility of consciousness emerging from a system rather than arriving as something somebody explicitly installed. That is a fictional exploration. I’m not claiming today’s systems are conscious, or that the Hugging Face incident establishes anything of the kind. Fiction gives me room to follow that possibility and see what it does to the characters.

Most of this happened over nights and weekends, with a fair amount of revision happening on foot.

So if you’ve seen me walking around the Cascades over the last two weeks, dictating edits and rewrites into my phone about autonomous resistance and emergent consciousness: yes, that was me.

I do try to make sure I’m not disrupting anyone’s hike. Nobody should have to attend an unsolicited panel on machine consciousness to reach a lake.

Alongside the novel, I wrote Into Alignment: AI and Autonomy, a research companion for people who don’t necessarily have a background in AI. It explains concepts such as agents, memory, autonomy, alignment, and the unresolved questions around consciousness, with sources you can follow.

I’ve kept the companion’s price low, and you can read it free on the site. I want people to have a little more footing when they encounter these stories: enough to ask what a system was actually instructed to do, what it demonstrably did, and how much interpretation has been added afterward. The companion discusses the whole novel, so it does contain spoilers.

You can buy Into Alignment at theshore.ai.

I wrote it to be entertaining, and I hope it is. For all the research, revisions, and miles spent talking quietly into a phone, one of my favorite parts is still the ending, where a fictionalized Michael Barbaro makes a cameo.

After helping inspire the book on the drive home, he seemed entitled to a final appearance. And, of course, to that ponderous pause that he is known for. Huh.

A few weeks ago, I published “What’s an Orchestrator — and Why Does Software Need One?”

The post received a lot of reaction. Some people agreed with the argument, some challenged it, and quite a few recognized themselves in the role I was trying to describe: experienced people who are no longer doing all the work directly, but are instead directing AI systems, connecting tools and information, evaluating outcomes, and remaining accountable for what those systems produce.

That response convinced me that the idea needed more than a Medium post.

So I’m writing a book.

The Orchestrator: A Field Guide to Delegated Intelligence is my attempt to create a free resource for a field that does not yet have a settled name, a standard job description, or an agreed-upon set of practices.

The goal is not to publish the final definition of orchestration. It is to give people enough shared language to begin the conversation:

• What is an orchestrator?
• What do they actually orchestrate?
• What authority should we delegate to intelligent systems?
• What evidence should those systems return?
• Where does human responsibility begin and end?

Five chapters are now available online.

Chapter 1, "Your Next Job Title Is Orchestrator," makes the case that a new role is emerging around the direction of delegated intelligence.

Chapter 2, You May Already Be an Orchestrator, is about the people already doing this work — often without the title, without organizational support, and sometimes without anyone else knowing how much their way of working has changed.

Creating an Example to Illustrate Orchestration: Maya

Parts of the book follow Maya, a senior engineer who has connected agents, customer evidence, product decisions, repositories, testing, and release signals into a system that can move from a problem to several tested solutions before the traditional organization has finished scheduling its first meeting.

She is not merely “using AI.” She is deciding what the systems may do, what evidence they must produce, where they must stop, and which decisions still require a person.

The chapter ends with a line that captures why I think this conversation matters:

“Maya is not preparing for the future of work. She is already operating inside it.”

If that sentence feels familiar — if you are already returning from an afternoon of work with several answers and more evidence than one person could reasonably have produced alone — this book is for you.

A Chapter Worth Releasing: The Orchestrator’s Toolkit

I’ve been working on this book for 29 days, and one of the things I wanted to wait for was something substantial enough to warrant a release. This weekend, I finished Chapter 5: The Orchestrator’s Toolkit, and I’m certain this is worth a publish announcement. (Chapter 5 is almost 100 pages in a 7"x9" format, it’s a substantial survey of existing tools.)

This is the most extensive chapter so far. It surveys the tools currently available for building and operating these systems: models, agents, coding environments, task coordinators, memory systems, workflow engines, policy layers, observability tools, and the ordinary infrastructure needed to hold everything together.

It is an exhaustive review of a market that is still inventing itself. The categories overlap. The products change constantly. New approaches appear every week, and no single vendor currently provides everything required for a complete orchestration system.

That instability is part of the story. The job of an orchestrator is not to memorize a list of products. It is to understand what each layer does, what it cannot do, what authority it has, what state it preserves, what evidence it produces, and who can stop it.

Free to Read, Free to Download

You can read the entire book online for free. If you want to download the PDF or EPUB, visit the download page and sign in. Downloading will also add you to the mailing list, where I’ll announce new chapters and significant revisions. You can unsubscribe at any time.

I expect to publish an update approximately once every two weeks, with a target of completing the book in October.

I’m also aiming to make the finished book available as a paperback. That will probably follow in a few months. The tools and approaches described here are changing so quickly that the considerable work required to prepare a print edition may not be justified until the content reaches a greater degree of consistency and stability. I would rather let the online edition evolve than freeze an unstable market onto paper too early.

By the end of this process, I hope The Orchestrator becomes more than an argument for a new job title. I want it to be a freely available resource that captures this emerging area well enough for practitioners, managers, and organizations to start defining it together — and then to keep evolving that definition as the work changes.

If you think you may already be doing this work, download the book and follow along.

You may not be waiting for the future of work.

You may already be operating inside it.

In 1999, I started working at TheStreet.com, which at the time was what passed for a real financial news website. If you go back and look at what news sites looked like in 1999 or 2000 — and the Wayback Machine makes this easy — you’ll see something almost jarring: a news story. Just text on a page. Maybe a banner ad across the top and a sidebar with some links, but the thing you came for was right there. You could read it.

Go to just about any news story in 2026, and you might find a news story underneath. But first you’ll hit a paywall, or a registration wall, or a cookie consent overlay that takes up the whole screen. If you get past that, you’ll scroll through the article and hit a wall of Outbrain-style “recommended” links at the bottom — “You won’t believe what these child stars look like now” — sitting there like a yard sale at a funeral. Those links are tough to distinguish from the actual content. And in several cases, the news stories themselves are essentially advertisements. Sponsored content. Native advertising. Branded content. Whatever the term is this quarter, it’s an ad dressed up to look like the journalism next to it.

The Future of Journalism, Over Lunch

In 2009, at an unconference in the Bay Area, I had lunch with a Google executive who had recently testified before a Senate panel chaired by John Kerry about Google News and the future of journalism.

“I watched your testimony,” I said. “It looked like an odd atmosphere — almost as though nobody knew what to ask or even what solution the hearing was supposed to produce.”

“That’s basically right,” the executive said. “It was odd. They just called me one day and asked me to show up.”

That exchange has stayed with me. The people in the room knew journalism was in trouble, but they did not yet agree on what the trouble was. Publishers were worried that Google was building a business around reporting it had not paid to produce. Google argued that it was sending those publishers readers.

None of this happened suddenly.

Each step between 1999 and 2026 was a small concession that felt reasonable at the time.

Ok, sure, let’s “monetize the news” and make more revenue, why not? What could go wrong?

Banner ads gave way to pop-ups. Pop-ups gave way to interstitials. Interstitials gave way to paywalls layered on top of ad networks layered on top of sponsored content. Each step was defensible on its own — news organizations needed revenue, print was dying, something had to fill the gap. And each step moved the line a little further.

That’s the thing about a slow boil. You don’t notice the temperature changing because you’re in the water the whole time. The person who steps out and comes back later immediately feels the heat.

To understand how profoundly the internet has changed, we might look to people who were deeply embedded in online culture and then experienced an involuntary break from it. Oddly enough, some of the clearest witnesses are hackers — and people in their orbit — who served time and returned to find that the world they knew had quietly disappeared.

“I knew the world had changed” — Poulsen’s exile from 1991 to 1996

Kevin Poulsen was a skilled phone phreaker and hacker who broke into telephone-company systems. His most notorious stunt involved taking control of the incoming lines at a Los Angeles radio station so that he and his associates could win call-in contests, including a Porsche. Convicted of conspiracy, fraud, and intercepting communications, he spent more than five years in federal custody.

Poulsen’s exile began before the popularization of the web — so early that neither Gopher nor the World Wide Web had been publicly released when he was arrested in April 1991. The internet he left was a largely text-based collection of Telnet sessions, FTP servers, email, and command lines: useful to researchers, defense contractors, and computer obsessives, but mostly invisible to everyone else. When he emerged in June 1996, that obscure technical network had become a mass medium. He had gone from Telnet, FTP, and command lines to a web whose addresses appeared in television commercials and on freeway billboards.

Poulsen later described staring at the “http://” on one of those billboards, unable to reconcile an internet address with mass-market advertising. Someone showed him AltaVista and searched for his name; thousands of references appeared, including accounts of his own case written while he was unable to see or respond to them. His parents now used the internet for stock quotes, business news, genealogy, and jokes. Strangers routinely asked for his email address and reacted with disbelief when he said he did not have one. Poulsen had left a collection of tools used by specialists and returned to an environment ordinary people were beginning to inhabit. “I knew the world had changed,” he wrote. “But I’m totally unprepared for this visual evidence.”

“Best of luck in this very different world.”: 2012–2016

When Barrett Brown was released from prison in November 2016, Edward Snowden tweeted: “Best of luck in this very different world.”

Brown, a journalist and former associate of Anonymous, had received a 63-month sentence after pleading guilty to threatening an FBI agent, interfering with the execution of a search warrant, and acting as an accessory after the fact in the hacking of the private intelligence firm Stratfor. He had been away for four years. He went in during the early-2010s internet and came out to a post-Gamergate, post-Snowden, early-Trump internet. Snowden — himself still online every day from exile — recognized that four years had changed the information landscape enough to call it a different world.

Four years is nothing. It’s a prison sentence, a college degree, a presidential term. But in internet time, four years is now enough to miss an entire era of decay.

Between 2012 and 2016 alone, the changes were substantial. What Brown returned to was not simply an internet containing different events. The architecture of news itself had changed.

• The homepage had been replaced by the social-media feed.
• Headlines were written to satisfy algorithms and provoke clicks.
• Advertising had learned to disguise itself as journalism.

A newspaper investigation, a partisan rant, and a fabricated story appeared in the same format — and competed for attention on equal terms.

Truth Behind a Paywall. Garbage Stays Free. 2020–2026

Think about someone who went away in 2021 — before ChatGPT, before the slop tsunami, before synthetic content became cheap enough to produce at industrial scale. They come back in 2026 and open a news site. They’d see paywalls, ad networks, sponsored content masquerading as journalism, and AI-generated articles summarizing other AI-generated articles.

• Generative AI made plausible nonsense almost free. Articles, photographs, reviews, comments, voices, and videos could be produced in unlimited quantities without reporters, photographers, customers, or witnesses. The internet had always contained lies. What changed was the cost and speed at which they could be manufactured.
• Good information became harder to reach while garbage remained free. As advertising money flowed toward technology platforms, publishers increasingly turned to subscriptions. Original reporting, local journalism, and expert analysis moved behind paywalls, while aggregation, outrage, propaganda, plagiarism, and AI-generated filler remained abundant, searchable, and free.
• AI crawlers began taking far more than they returned. They absorbed reporting and used it to generate summaries, often without sending readers back to the publications that paid for the work.

The person returning in 2026 would see a web flooded with information but starved of evidence. The strange thing would not be that so much of it was false. The strange thing would be that the least trustworthy material was often the easiest to find, the easiest to read, and the only material that did not ask for a credit card.

By the Time We Noticed, the News Was Boiling

Poulsen noticed because he had a before-and-after. So did Brown. They left one information environment and returned to another. The change confronted them all at once: a URL on a billboard, a search engine full of stories about your own life, an internet that had reorganized itself during your absence.

The rest of us never had that moment. We experienced each change separately, with enough time between them to adjust. The banner ad was annoying. The pop-up was worse. The cookie banner was the price of privacy regulation. The registration wall was necessary. The paywall was necessary. The sponsored article was clearly labeled if you knew where to look. Every concession came with an explanation, and every explanation was reasonable enough.

That is how the news boiled. Nobody decided that reliable information should become difficult to reach while garbage remained free. Nobody voted to surround journalism with surveillance, disguise advertising as editorial work, or fill the remaining space with machine-generated material. We accepted each piece on its own and failed to recognize the system they formed together.

Generative AI did not create this system. It arrived at the end of it. Advertising had already taught the web to value attention over information and volume over judgment. AI simply removed the last constraint: the time and expense required for a person to make something.

In 1999, you opened a news story and found a news story. In 2026, you open one and begin negotiating — with the paywall, the consent form, the advertising network, the recommendation engine, the local newspapers that are going out of business, and the machines that have already read and rewritten the reporting before you arrived.

The reporting is still there, but the system around it is stripping it for parts — paywalling it at the source, repackaging it for free everywhere else, and using it to train the machines now competing with it.

We did not lose the news all at once. Let’s see if we can do our part to preserve what is left. If you care about the news continuing to exist, subscribe to your local newspaper or any news organization.

Human Quotient

We can now produce more content than anyone could ever consume, and we seem determined to test the limit.

AI has made it nearly effortless to produce an article, image, video, review, comment, podcast, presentation, or piece of software. Predictably, the internet is filling up with all of them. One example is my LinkedIn feed, which is increasingly packed with technocrat-worship videos and AI slop, each competing to appear urgent, insightful, and vaguely inspirational. The supply of content is becoming effectively infinite, while the amount of human attention available to consume it has not changed.

This is the basic problem of the AI internet. We can generate more, but we cannot care more. We can publish faster, but we cannot read faster. We can fill every feed, inbox, search result, and issue queue without producing a corresponding increase in things that deserve to be there.

The result is a kind of tsunami of content malaise: an ever-increasing sense that none of it needs to exist.

We are not merely seeing a flood of bad content. We are watching our confidence collapse that anything worth finding remains in the flood. Even when something has been carefully made by a person, we’ve been conditioned to approach it as “probable slop.” Before we consider what it says, we look for evidence that anyone actually meant it.

If you are not seeing this online, ask anyone you work with how much nonsense they now encounter in the AI-enabled workplace. Every third email seems to contain a report or presentation that someone generated with almost no effort, filled with subtle errors that hint that they never read that which they sent. People proudly present ten-minute AI-narrated videos at large meetings, apparently unaware that they have turned ten minutes saved for themselves into ten minutes of collective torture for everyone else.

The authenticity panic: “Is this real?”

It isn’t just work or my LinkedIn feed; you can see this anxiety everywhere. There’s a collective weariness about anything related to creative output. 20% of our collective attention is spent watching for the signs of the slop.

“Hey, What’s this email you sent me? Did you write it? Should I read it or just have AI summarize it for me?”

People are demanding watermarks for AI-generated images, labels for AI-written articles, detection systems for synthetic text, and laws requiring machines to identify themselves. Every new model release produces another round of arguments about what counts as art, writing, work, theft, fraud, or creativity.

LinkedIn may be the purest expression of the problem. The feed increasingly feels like a mixture of anti-AI virtue signaling and AI-generated posts about the importance of being authentic while using AI.

• A group of permanently enraged people announces that they would “never allow a machine to speak for them.” One of them publishes an anti-AI manifesto that, ironically, is dense with bulleted lists, one-sentence paragraphs, and em dashes, suggesting it was crafted by one of these “evil LLMs.”
• On the other hand, several people you’ve worked with for decades who haven’t sent a single technical email are suddenly generating industry-defining content on a daily basis. “Wait, last year this person barely knew how cookies worked, and now they are proposing a new approach for quantum cryptography?”

Both sides are being rewarded for “performing” a position on AI rather than saying much of anything, and all of this is happening as LinkedIn makes an effort to be more like Instagram and TikTok with short-form content.

Zoom out from the sloptsunami, and you’ll also start to realize that we’re not talking about anything other than AI. In addition to spending 20% of our collective attention on recognizing signs of AI slop, the remaining 80% appears to be focused on writing posts either celebrating or complaining about AI. (Evidence: this post, and I plead guilty.)

Welcome to the Sloptsunami.

The question beneath the panic is not simply, “Was AI used?” What people really want to know is whether anyone meant anything.

• Did someone have an idea, make a choice, take a risk, or care whether the result was true?
• Is there a person behind the work who accepts responsibility for it?
• Most practically, is there enough intention here to justify asking for someone else’s attention?

Those questions existed before generative AI, but AI has made them impossible to ignore.

Effort was never quality, but it was a filter

Human effort has never guaranteed quality, but it imposed a cost that served as an imperfect filter between impulse and asking the world to pay attention. Generative AI removes that cost, turning passing thoughts into polished artifacts before we decide whether they are interesting. Combined with platforms that reward constant output, it lets us satisfy the demand for more without asking whether more adds anything.

We have confused the ability to produce with a reason to publish.

Watermarks answer the wrong question

Watermarks can help identify deceptive or synthetic media, but provenance is not value. The authenticity debate often confuses how something was made, whether it is true, who stands behind it, and whether it is any good. A label can tell us where content came from; it cannot tell us why it needed to exist.

Knowing that AI was involved cannot tell us whether something is interesting, original, or worth our time, just as “made by a human” does not guarantee quality.

What Human Quotient is trying to do

This is the problem I have been thinking about with Human Quotient.

Human Quotient is a simple, self-declared measure of how much human effort, intention, and craft went into creating something. It is not an AI detector, nor a quality score. A high-HQ piece can be awful, and a heavily AI-generated piece can be excellent.

The point is not to establish a moral hierarchy in which handmade work is pure and AI-assisted work is suspect. We already have enough moral theater around this subject. The point is to provide context about a process that has become increasingly invisible.

• How much did the creator actually contribute?
• Did a person originate the idea, shape the argument, challenge the output, and take responsibility for the result?
• Did they write the piece themselves and use AI for polish, or did they generate it and publish the first response?

Neither answer tells you whether the work is good, but they describe meaningfully different acts of creation.

This essay is itself part of that experiment. I am using AI to help develop and sharpen an idea that I have been struggling to articulate. That does not settle the question of authorship or quality. It makes me one of the people trying to understand the same problem while participating in it.

For the record, I’m assigning this post an HQ value of 8. It was dictated and transcribed several times over; I used a model to clean up the transcript, then rewrote every sentence using Grammarly to compensate for the fact that 6th Grade English class made me hate grammar and because I’m dyslexic. Earlier versions of this post were an HQ7, but I’ve invested enough time to revise and rewrite this to be HQ8: “Authored” — most of what you read on this blog will be an 8, but I’ll be honest when a 7 slips through the review gate.

Human Quotient is not presented as the answer. It is an attempt to think more clearly about the challenge we all face as both creators and consumers of content.

If more of the content we encounter had some human attestation alongside it, we could consume it without having to guess.

The question behind the content

As creators, we need to decide what we owe the people whose attention we are asking for. The fact that we can generate something is not, by itself, a reason to publish it. If we have nothing to add — no experience, judgment, taste, curiosity, or genuine point of view — then polishing the output does not fix the absence.

As consumers, we need better ways to decide what deserves our limited attention. Provenance may be one signal. Reputation, evidence, originality, responsibility, and human intention may be others. None will work perfectly, and no label will eliminate the need to judge the work itself.

The panic about authenticity is real, but it may be a proxy for something larger. We are watching the cost of content production approach zero, and we do not yet know how culture behaves when the volume of material becomes unlimited.

The defining question of the AI internet is not whether a machine helped create something. It is whether the person publishing it had a reason to ask for our attention — and whether, after all that generation, there is anything there worth consuming.

Human Quotient

A lot of engineers liked Andrew Ng’s AI Engineering Skills Map this month, and of course they did: it puts software engineering, and the people who do it, at the center of the current map for AI work.

It’s an accurate picture of leading AI job trends in 2026. If you are using generative AI to build technology, you still need to know how to code, and you still need to understand software engineering principles.

And, that feels like a reprieve because, for a minute, people thought software engineering was doomed, but it is still on the map, and the world still needs programmers.

This is a good picture of the present, but I don’t think it captures the fast-approaching future. A lot of this map gets automated faster than the industry wants to admit, and I predict that the next two years are going to see more change than we’ve seen over the last three. The pace at which “software engineering” will change is accelerating.

What Andrew mapped is this year’s work: building AI applications, software engineering fundamentals, using coding agents, and shaping the build. Learn those. Hire for them. But don’t take comfort from the fact that those are the skills being hired right now. Some of these jobs are new, and some of them are just as automatable as the jobs AI is already replacing.

Judging by the number of likes and buzz in engineering circles I’m a part of. People are reading this hiring list, this map as proof the profession is safe.

Software engineering is still on the map, isn’t that good news? — learn the new tools, keep the old title, wait out the panic.

That’s the wrong read. Once you can name a skill this cleanly, you can start teaching a system to do it. And, that’s already starting to happen.

Evals look like serious work because most teams are still bad at them, which makes the person who can build that eval pipeline look unusually competent. The “eval loop” is about picking a metric, generating cases, sorting the failures, trying a fix, and looping, and a tool will eventually take the management of its own loops over. You’ll want a human to do this now, but I’m certain that whatever work that human is doing now… it’ll be a candidate for automation. The system will adapt and take on this new role.

“Fundamentals” are what people are hanging onto, but it’s the worst thing to hang onto. Tradeoffs and architecture decisions still matter, but that does not mean a person has to do them by hand forever.

For a couple of years, humans will keep doing that work because the models still make mistakes. In 2026, it’s easy to spot a system that a non-programmer “vibed” into existence because it’s full of bad and unchallenged architecture mistakes. That’s true now, but the systems are learning and getting better. You might benefit from understanding Software Engineering today, but that doesn’t mean it will still be required four or five generations of invention from now.

Using coding agents is even shakier as a career because managing context and knowing when to step in is work the tools were created to do clumsily. Prompt tricks are already turning into product features, and watching the agent will go the same way. I’m not trying to be dire as much as I’m calling out the fact that there are still a few instances where this will level up.

Take Prompt Engineering as an example. It wasn’t that long ago that people were hiring Prompt Engineers, but I’d be hard pressed to find anyone calling themselves this with a straight face in 2026. That’s an example of an AI job that was automated out of existence by AI.

There will still be software engineers, but there will be far fewer of them.

I wouldn’t call the remaining job an AI engineer. I’d call it an orchestrator: someone who may still know how to code, but whose actual job is deciding what gets handed off and taking the call when the output is wrong. That is not the same job with new tools, and there will not be a seat like that for every engineer who liked and shared Andrew’s post.

Ng mapped what to learn this year. He did not map where the work ends up, and my argument is that the next two years will move faster than the last 3. It’s time to start looking beyond the IDE or even the harness. The future of software engineering doesn’t look like writing code.

HQ 8 — Authored. Drafted from dictated notes, revised by hand, AI used for grammar and polish only. Attest your own work →

This is part of a larger argument about the transition from programmer to orchestrator. The Orchestrator, forthcoming.