If you’re a brand called Tide, and you sell laundry detergent, your job is to get a bottle of it onto the shelf at the Walgreens on the corner of 13th and U, and then to get another one there next month, and the month after that, in roughly the same condition, at roughly the same price, with roughly the same brand experience, from now until the end of time.

Almost everything about how you operate flows from those assumptions. You want consistent processes, durable roles, learning that compounds across cycles, an org chart that doesn’t crater when one person leaves, a finance team that can forecast the next quarter, and a brand team that protects a vibe you’ve spent decades building. You measure your performance against sales growth, market share, and net margin. Your whole shebang — your strategic plan, your OKRs, your annual budget, your performance management system — is built around the fact that what you want to do is deliver something, repeatedly, consistently, well.

If you run a soup kitchen, the same sort of thing is true. A very large percentage of people who come in on Wednesday will be hungry on Thursday. To solve for that, you want predictable supply, predictable staffing, a kitchen that runs the same way regardless of who’s volunteering, and a donor base that mails you the same set of checks every fall. You measure your performance against the meals you serve, week over week. The shape of your work and the shape of your organization match each other, and they should.

Now consider what happens if your job is to eliminate a disease.

Almost none of the previous logic pertains. You actively don’t want consistency; you want the willingness to abandon a strategy that isn’t working at month eighteen, and the agility to try something else by month twenty-one. You don’t want durable roles; you want different combinations of people feeding into different dimensions of the work at different stages — surveillance, then strategy, then regulatory coordination, then deployment, then mop-up. You don’t want learning that compounds inside one organization; you want learning that flows out across an entire field of actors (most of whom don’t report to you). You don’t want sales growth as your metric, or for that matter anything rooted in what you yourself are doing; what you want is the absence of cases — a metric that points at a state of the outside world.

The cascade of assumptions built into the commercial firm-form — and into nearly every organization the social sector inherited from it — is correct for delivery work, including a great deal of valuable social-impact delivery work. But it is not correct for outcome work. Running an outcome-accountable organization with delivery-accountable habits keeps you married to the wrong design.

Which is one reason a recent piece by Anastasia Gamick is so exciting.

Share

Anastasia is a co-founder of Convergent Research, the studio that incubates Focused Research Organizations — purpose-built, time-bound nonprofit research outfits designed to attack scientific bottlenecks that fall through the cracks between universities and startups. Nearly a dozen of them now exist. She has been building this kind of organization for years, and the piece reflects that; it reads like someone reporting from inside the work, not theorizing about it from outside the fishbowl.

Her piece, Working on vs. Solving, draws a binary that names the design problem directly. Working on a problem is what delivery-shaped organizations do — funding research, publishing roadmaps, building fields, running programs, coordinating grantees, year over year. Solving is what outcome-shaped organizations do. It means picking a finish line that is a state-of-the-world rather than an activity, and owning the imperative to reach it. The Carter Center’s Guinea Worm Eradication Program reduced Guinea Worm cases from 3.5 million annually to fewer than 20 because Donald Hopkins, with Jimmy Carter, picked a finish line of zero and held himself accountable to it for decades. The Montreal Protocol banned CFCs in 1987 and the ozone layer is on track to recover by the second half of this century because the treaty bound its signatories to a phase-out timeline rather than to a research agenda.

What both examples share is a person or a structure that owns the outcome. Anastasia’s piece proposes a piece of vocabulary for the kind of leader who could play that role — a general manager — and lays out the five tools such a person would need to wield: grants, investments, new organizations, advance market commitments, and policy work, all simultaneously, with the flexibility and capital and time to actually finish.

The point is that — far from simply representing a clever way of doing things — this ‘general manager’ approach is a design solution to a structural problem that makes adopting an outcomes orientation difficult under typical circumstances. The piece is downstream of an earlier essay by Nan Ransohoff, which makes the case that there’s a category of important problems for which no specific person feels accountable for the whole thing, and proposes the general manager as a solution to this state of affairs.

They’re not the only ones to have noticed.

Around the social sector, an entire class of organization has been quietly forming over the past two decades. Their daily operations share very little in common. What they share is a posture. They each reject the central assumption that an organization exists to deliver a defined activity, and instead organize themselves around what it would take to deliver an outcome.

Adam Marblestone — Anastasia’s collaborator at Convergent — has been arguing for and building Focused Research Organizations, time-bound nonprofit research outfits structurally committed to dissolving when the work is done. Michael Kremer’s advance market commitment mechanism — a way of getting suppliers who don’t yet exist to come into being — has been deployed by the Gates Foundation against pneumococcal disease, by the U.S. government during Operation Warp Speed, and by Nan herself at Stripe through Frontier’s $1 billion commitment for durable carbon removal. DARPA’s program-manager model — tenure-limited managers, expiration dates on the badges, three-to-five-year programs that end when they end — has been refining a version of this for sixty-five years. Funder collaboratives like Blue Meridian Partners and the Audacious Project pool capital across philanthropies precisely so no single funder’s institutional constraints set the ceiling.

The institutional shape civil society inherited is mismatched to much of the work civil society is now doing. As this structural problem becomes increasingly explicit, design responses like these keep emerging across contexts. But they haven’t yet fully coalesced as a formal class.

My collaborator Ben Marshall and I wrote about another potential instance of this class in the Stanford Social Innovation Review this past February. We tried to make the case that a ‘creative paradigm’ for social value — organized around time-bound projects, auteurial vision, and discernment-based evaluation — carried proven success in other outcome-oriented contexts, and offered an alternative to an ‘institutional paradigm’ optimized for delivery.

This argument generalizes. What we see over and over again across the sector is that the traditional firm-form serves as a kind of gravity well, pulling organizations into a delivery posture even when their stated mission is something a delivery posture categorically cannot achieve.

—

How does this ‘gravity well’ work?

The tax code is part of it. Treasury Regulation 1.501(c)(3)-1(b) — the “Organizational Test” — requires a nonprofit’s articles of incorporation to “limit the organization’s purposes to one or more exempt purposes” and “not authorize any activity that doesn’t further those exempt purposes except as an insubstantial part of operations.” Note activity as the unit of analysis, and to a lesser extent purpose (rather than e.g. goal or outcome) as the organizational imperative. This then allows IRS Form 990, filed annually, to go and audit actual programming against that stated purpose — creating a scenario wherein substantial mission drift is a paperwork ordeal at minimum and a tax-status risk at worst.

Such legal architecture locks each organization into a purpose-specific ‘groove’ from its first day of existence.

But the much bigger part is everything that follows from the tax code’s premise: that such an organization is the kind of thing that does a defined activity, repeatedly, in a measurable way. Foundation program officers are reviewed against discrete attributable bets. Donor reporting expects activity-level outputs. Strategic plans cover three-to-five years and typically assume stable program areas. Job descriptions describe ongoing functions. Annual reports tell stories about what the organization did, not about what changed in the world. The year-over-year budgeting process assumes steady-state operations. Even the language people use about their own work — we run a program, we deliver services, we maintain a presence in the field — encodes assumptions that fit a delivery operation and don’t fit anything else.

Again, none of this is per se bad. Lots of organizations need to consistently deliver programs. I am glad legal aid clinics deliver legal aid services. The key is just that when you’re trying to scope a new organization or shift an existing organization to mobilize around some kind of outcome, it’s difficult to shrug off all of this inheritance. You can push back against all of it, but you have to do so actively, every day, against every default that surrounds you.

How? There are options.

Lewis Bollard’s farm-animal welfare work at Coefficient Giving has directed over $130 million toward a single outcome — animals affected by improved welfare standards — by treating that number as a portfolio problem rather than a programmatic one, funding corporate cage-free campaigns and welfare science and alt-protein companies and movement-building in over 20 countries, with the weight shifting across these moves as the binding constraint shifts. Frontier, mentioned earlier, doesn’t fund carbon-removal companies the way climate philanthropy ordinarily would; it signs binding offtake agreements that let traditional capital flow into companies that wouldn’t otherwise be fundable, then publishes its contracts open-source so the rest of the market can use them. Cultivarium, a Focused Research Organization at Convergent Research, is a five-year nonprofit startup making 170,000 microorganisms accessible to synthetic biology — recruiting staff for the duration, releasing tools open-source to the field, and dissolving when the work is done. Convergent has launched many others on the same template, including E11 Bio, which is building an imaging system to map brain wiring at a hundredth of current cost.

In each case, the move that mattered was a rejection of the default cascade of assumptions — about what consistency means, what good operations look like, what success is measurable as, what makes a leader effective, what the relationship to time should be, and what the organization is built for — coupled with the construction of new, non-obvious mechanics fitted to whatever is necessary to actually enact the outcome. These aren’t just nice, one-off stories. They’re templates for solving entire classes of problems.

—

So, how to refer to an organization built this way? The phrase that’s been useful in my own thinking is focused mission organization (FMO). An FMO, like an FRO, is a nonprofit organization with a defined finish line — but where an FRO drives toward a research imperative, an FMO drives toward a state-of-the-world outcome. The FMO inherits from the FRO the time-bound posture and the willingness to plan for one’s own obsolescence. It generalizes the FRO beyond science bottlenecks to anything outcome-shaped — policy actuation, market shaping, disease elimination, civic infrastructure. Freedom to Marry, which won federal marriage equality and then dissolved on schedule, is a canonical American example of an outcome-driven project built this way.

But the particular phrase isn’t important. What’s important is that the category exists, and that recognizing the category gives practitioners a place to locate their work and gives funders a class of bet to recognize. Right now, an organization built like Frontier or Convergent or Freedom to Marry finds itself in a bind. It says it’s a nonprofit, and the wrong cascade of assumptions kicks in. It says it’s a “campaign” or a “fund” or an “initiative” and a different, distinct stack of wrong assumptions kicks in. The absence of a category leaves practitioners reinventing their self-conception every time they try to explain their work.

Nan and Anastasia have given the field the general manager as a piece of conceptual infrastructure — a phrase practitioners can pick up, that funders can recognize, that lets a shared vocabulary form. Marblestone gave it the focused research organization. Kremer gave it the advance market commitment. Each idea is a foothold against the default cascade of assumptions, a little flag planted in territory that mostly doesn’t yet exist.

There should be more flags.

—

The thing I most want a reader to take from this cascade of reasoning is not that any particular one of these architectural innovations is the answer, and we should all drop what we’re doing and sprint over to our funders shouting from the hilltops that we are now building one of them. It’s simply that there exists a meaningful new category, and the category is not yet well-populated. As both funders and operators, that gives us an opportunity.

If you work in civil society, at least some of the assumptions built into how your organization runs — or what your organization supports — could be misaligning your work with the outcomes you say you want. The strategic plan, the job descriptions, the performance reviews, the way you talk about what you do at the all-hands — any one of these could have fundamental misalignment baked in. Many of these norms feel necessary because they look like how a serious organization operates. And they feel that way for a reason: they are how a serious delivery organization operates. The sector has had one primary model in front of it for so long that pursuing an alternative can feel more unprecedented than it actually is — almost reckless, an indulgence for the sake of change.

But building this kind of organization isn’t reckless. It can be risky, but that’s different — most real work is risky. The question is just where we embed the risk. The risk of novel strategies and structures is that they’re unproven and may not work. The risk of the default setting is that we know it cannot work for outcome-shaped problems.

Nan and Anastasia are building organizations whose shape matches the outcomes they’re accountable for. So are Marblestone and Kremer and Bollard and the people behind every advance market commitment and every FRO and every patient funder collaborative. The work of the next decade in civil society is to build many more of them, and to build the resourcing infrastructure capable of recognizing them as a class.

The institutional shape we inherited is not the only shape available. It is one shape among many — the one we built defaults around, the one our tax code anticipates, the one our boards know how to govern. Others have to be made. They will not arrive on their own. And the more they integrate themselves into the existing ecosystem and architecture, the less stridently that existing architecture will push back against and ultimately reject them.

So, let’s talk about them. Evangelize them. Fund them. Sit on their boards. Write the contracts that let them work. Build the metrics that let them be evaluated honestly. Treat them, in other words, not as curiosities or novelties or even valuable one-off solutions, but as institutional innovations allowing us to deal with a poison pill baked into the heart of our social problem-solving architecture.

The category exists already. The question is how quickly we open our eyes to it.

A homie of mine, as a side project, is conducting a serious industrial-base diagnostic on what it would take for the United States and its allies to build a thirty-foot bipedal combat robot.

His current draft cites, by document number, GAO reports and Defense Science Board studies and a 1983 National Academy of Sciences report on the collapse of American machine-tool manufacturing. It excludes the People’s Republic of China per the Defense Production Act, NSIBR, and DFARS regulations — but it includes Korean industrial capital inside the U.S. industrial base via the Hanwha Philly Shipyard, which Hanwha bought in 2024. All of this is rendered in a crisp, visually-distinctive typographic register that looks like 1972 IBM improbably bred with 1942 trade-journal masthead. The project also generates its own reading syllabus, sourced from the citations, in case you want to verify any of it.

He did all of this himself. In his spare time. While getting, like, a good night’s sleep.

Said homie’s name is Emmet Penney. He’s a Senior Fellow at the Foundation for American Innovation, where he focuses on infrastructure and energy policy, and hosts a popular podcast called Nuclear Barbarians. Other résumé-padding? Former editor-in-chief of Grid Brief. Contributing editor at Compact Magazine. Multiple Emergent Ventures grants, a Robert Novak journalism fellowship, work that’s appeared in American Affairs and Claremont Review of Books.

He is, in other words, a working professional in a serious policy domain who — not to get too online too fast — codes as much more of a ‘wordcel’ than a ‘shape rotator.’

But he has also, in roughly the last year, become one of the most concretely fluent individual users of frontier AI tools I know. The Gundam project is the most extreme example. There are many others.

I got him on the phone for an hour this past Monday to nerd out about all this, mainly because it’s cool as hell, but also because one big question couldn’t stop nagging at me ever since I saw the first outlines of his first jawbone-unhinging dashboard prototype.

That question was: am I actually interested in becoming — with real, but manageable effort — way better at my job?

Should I be?

Should you?

Share

I’m going to invite you to navigate your browser to this here URL. Then I’m going to ask you to bear with me for a second.

What you’re looking at is a navigable map of Illinois. Eleven operating nuclear reactors — more than any other state. The only commercial away-from-reactor wet spent-fuel storage facility in the country. Class I railroads, navigable waterways, the Strategic Highway Network corridors that the Department of Defense certifies for heavy industrial transport. Two national labs. Three university nuclear programs. The brownfield sites at Zion and Dresden 1 — decommissioned, but permitted, sitting there waiting.

Each layer toggles. Each pin clicks. The sidebar walks you, criterion by criterion, through the Department of Energy’s January 2026 Request for Information for a Nuclear Lifecycle Innovation Campus and explains how Illinois meets each one. Every claim is sourced. Every dataset is cited. The flat typography and decisive visual hierarchy convey, before you’ve read a sentence, that someone with a point of view made it.

What this tool does, functionally, is make the question of whether Illinois is a serious candidate for a multi-billion-dollar federal nuclear campus answerable in five minutes by a person who is not a nuclear engineer.

How do I know? Well, among other reasons, the state response team submitted Emmet’s GitHub link alongside their formal RFI submission to the DOE.

Now, this is all very inside baseball. Nobody really cares what an RFI is. But that’s sort of on us, because that RFI might be the reason a whole-ass nuclear campus gets built on Site A rather than Site B — the mobilization of hundreds of millions of dollars and century-scale infrastructure powering homes and offices and factories and the laptops upon which we read Substack posts like this one.

The mechanic that lit the spark of all of this — what got the entire process started — was, until January, a simple PDF file. That was the state of the art.

Until, pardon the expression, Emmet went nuclear.

A year ago, this kind of artifact would have involved a six-figure consulting engagement with a six-month timeline and a bunch of developers you’d need to herd like mewlings into product meetings. Now it’s something a guy can whip up for fun between anime episodes.

Out the other end pop about eight billion kilowatt hours of energy annually. Stick that in an annual report.

On the one hand, this is all just straightforwardly impressive. But on the other, it feels revolutionary — in both the ‘changes everything’ and ‘dudes with pitchforks’ sense. Because it definitely feels like this might be doing things to the structure of professional life that we have not yet come to terms with.

I’m one of the people who hasn’t come to terms with it. Because I do not have the juice that Emmet has.

We’ll come back to that.

Emmet also has something else. He calls it a stack. He could just as easily call it scaffolding, or a rig, or — to use the word that I think is closer to right — an apparatus. A latticed thing he built around himself that gives him altitude.

He’s chosen a deliberate vocabulary for it. There are tools named after famous Chicago opinion writers (Royko, after Mike Royko, for tweet-thread chunking); fictional librarians (Ultan, after the librarian in Gene Wolfe’s Book of the New Sun); and authors of stories about fictional librarians (Borges, for historical research). His after-action protocol is Ignatius, presumably for the Jesuit founder. His adversarial multi-agent review pipeline is called Inquisitor, which depending on your perspective codes either markedly less or markedly more Jesuit.

This stack is sort of like a cast of characters in a play, and sort of like an org chart of employees at a company. They’ve all got jobs to do.

Ignatius sits at the end of every project phase and interviews him about what worked and what didn’t. Then it writes a best-practices markdown the next project will read before doing anything else. Inquisitor runs an adversarial review pass against work product, surfaces errors with severity tags, applies the must-fixes, queues the rest for his judgment. ADRs — Architecture Design Records — version-control the evolving assumptions of every project, timestamped so the entire decision history is legible at a glance. When a Claude session inevitably approaches its context limit, the system writes its own handoff document with an ideal opening prompt for the next session.

Emmet pastes the prompt in. The new session reads the handoff. Work resumes. Previously, on Lost…

When he conducts research, the system pulls 200-page PDFs from naval archives and returns them as cross-referenced markdown files with confidence vectors on individual quotes — a number, attached to each piece of evidence, indicating how firmly the source supports the claim. It checks author bios for biases (“this person has this CV, and so we infer their biases to be...”) and flags adversarial voices missing from the literature — telling him, unprompted, you don’t have enough people pressing against the fundamental assumptions of this part of the project. All of this is catalogued in note cards, which are generated and sorted automatically.

When he builds a new tool, his own best practices require him to first read all of his existing tools and borrow architecture from them where it fits. The note-card cataloguing skill, for example, borrowed its interview architecture from the Ignatius after-action skill.

The system, in other words, is composing itself.

“Yeah, I’ve basically solved the memory across projects problem,” he told me, with the casual air of someone who has done something he understands is significant but hasn’t yet been instructed to act humble about.

But so, okay, let’s just take a look at all the stuff we walked through, one more time.

A quality-control pipeline. An adversarial review function. A document-citation system with confidence vectors. An author-bias scanner. A devil’s-advocate function that hunts for missing critique. A version-controlled architecture for tracking strategic decisions across the lifetime of a project. A library catalog. An archive. A retrospective process. A best-practices document that compounds across projects. An onboarding memory. A handoff protocol. A budget officer.

These are scripts, sure. But really they’re departments.

More specifically, they are the actual departments of an actual institution that produces actual research, of the kind that lives on the eighth floor of a building in a city you’ve heard of, with a director who flies to conferences and an annual budget your local school district would envy.

Emmet has, in other words, built in his spare time the operational equivalent of a small policy shop. No employees. No board. No budget. No bitter break-room coffee staining the floor of any building anywhere. You can imagine the whole thing fitting in a snowglobe on his desk. The organization’s one employee is also its CEO.

This is the part of the conversation where I started sweating a little bit and compulsively squeezing the armrests of my chair.

Emmet has built himself an apparatus. That apparatus can be described as a sort of latticed self that has evolved a working person into something like an institution.

He hasn’t gotten better at his job by working harder. He’s gotten better at his job by building something taller, grasping at the sky.

Share

One standard objection to all of this is that AI flattens everything it touches into the same featureless register. Emmet’s work refutes this so completely it becomes its own argument.

He calls his system the Penney Design System, and it began with intuition. It’d be sick if I had my own style I could use for Google Slides. So he had the model study Sears Roebuck’s full design history, then Electricity World (a defunct trade journal from the early electric age), then the first ten years of RAND Corporation reports. Three completely distinct sources and traditions. Then he asked the model to compare them, and made decisions: 1910 Sears Roebuck for clarity, 1940s utility trade journal for layout, early RAND for voice of authority. Weighted the poles. Told the model to emphasize the trade-journal axis. Let the system construct a design language out of the result.

Now everything he ships looks like it came from somewhere. Somewhere specific. Somewhere chosen. Somewhere with values.

“People always say, like, ‘oh, taste is the final moat,’” Emmet tells me. “I think people just mean that, like, being cool is still important. But that’s not a meaningful thing to say. Maybe this is going to sound pretentious, but I think one of the things that’s important that AI can do for you, by being stylistically assertive, is to help cultivate yourself. Which is really the wellspring of taste.”

The model isn’t generating the taste. Emmet is. Then the apparatus executes the taste, consistently, across hundreds of artifacts he would never have had the time to produce by hand.

All of which is to say the bottleneck on creative work in 2026 isn’t the labor of execution. Nor is it some fundamentally, irreducibly human verve. It’s the willingness to do the work of figuring out what you actually want.

But so, I write professionally in roughly four registers depending on the audience, and produce probably three messaging artifacts per week for various projects, and I tweet entirely too often about all of it. I do not have a design system for any of that, and at no point until this conversation had I ever considered that I should.

Is that authenticity? Laziness? Or is it cope?

On the internet, a popular meme involves the process of carcinization — how nature and evolution wind up independently turning all kinds of different animals into crabs.

Emmet has caused me to realize a similar process is going on for work.

He was telling me about his friend Sterling, a graphic designer who has worked for A+ brands you’ve heard of — and has also, in the last year, become unreasonably good at vibe coding. The two of them had been comparing notes.

“The thing that we noticed,” Emmet said, “was that more and more of our lives were about making editorial decisions.”

What Emmet and Sterling were noticing — across two completely different professional domains — was the same shape converging on both of them. The work didn’t go away. It shifted. Execution became direction. Labor became judgment. The bottleneck on knowledge work moved up the stack, from can you make this? to can you decide what should be made?

Crab editors all. Red ink and claw marks.

Which explains why Emmet’s personal applications can also be so casual: he can whip them up as fast as he can do the task itself. He built his wife a meal-planning app that lets her drag recipes onto a calendar, scrapes ingredients, deduplicates, and emails her the shopping list. That took an afternoon. He built himself a Juggernaut Method workout app — a powerlifting program he’d never been willing to manually write down in a journal — that lives on his phone for his basement lifting session. That took him thirty minutes — barely enough time to scurry from the sand into the waves.

But where does this stop? Does it stop at all?

Most professional life, as currently practiced, runs on a quiet assumption: there’s a ceiling on how good you can get at your job in the middle of doing it. You can read a few books a year. You can attend conferences. You can develop, over a decade or two, the small set of practices that distinguish you from your peers. But the actual work is the work, and the work has its own gravitational pull, and the pull is hard enough that the question of whether you could be substantially more effective than you currently are doesn’t really come up. The ceiling is somewhere up there. You’ll get to it eventually, maybe. You’re doing fine.

This assumption is shattering.

What Emmet has demonstrated is that the ceiling has moved by an enormous distance, and the cost of climbing toward it has fallen, and the work of climbing is no longer fundamentally different from the work of having a hobby. The endless ladder is not totalizing. It is not even particularly hard. It is, in fact, very close to the cognitive register of reading nonfiction — rewarding, ongoing, accumulative, optional in any given week, fairly easy to integrate around the rest of your life.

And if it isn’t that costly, and if it’s actually kind of enjoyable, and if there’s a dimension of personal in addition to professional self-cultivation…isn’t it our duty to be doing this?

If you, like me, are working on something that you believe genuinely matters — whether the United States has enough electricity to power itself, whether a parent in a news desert knows what her school district is doing with its budget, whether a person facing eviction can navigate a courthouse in her own language, whether the policy idea you’ve been working on for three years can actually move — then the math has changed.

The cost of operating with much higher leverage has fallen substantially. Continuing to operate at the same level is a choice.

The people who depend on the work we do — the constituents, the readers, the parents, the patients, the citizens — are bearing the cost of that choice. That seems important to act on.

But also, I used to play poker semi-professionally. I wasn’t good enough to spike the tournament circuit, but what I could do was play eight or nine tables simultaneously online. And while that was a great way to help pay for college, it was a horrible way to live. Because you could always, always, always be making money. Lots of it. Every meal out, every pick-up ballgame, every trip to the gym or catch-up with a friend or party at the creole place downtown had a price of several hundred hypothetical dollars in expected value.

So the choices were to swim through a constant state of once-removedness from all life, or hemorrhage thousands of dollars in value to fully participate in literally anything.

Then again, the money was great, and I’d do it again in a heartbeat.

Which brings us back to the Gundam.

A Gundam is a fictional Japanese mecha — a bipedal armored combat platform piloted from inside the chest cavity, the centerpiece of a long-running anime franchise that has spanned forty years and several generations of weeb culture. It is, by design, a serious-looking thing dressed in maximally unserious circumstances.

Emmet’s project takes the construction of one — specifically, the Heavyarms variant, a 20-to-40-ton Gundam after Trowa Barton’s Heavyarms Custom — as a benchmark. The premise: what would the United States and its allies need to be able to do in order for the construction of a Heavyarms-class platform to be possible, and what does the gap between that capability and current capability tell us about the actual industrial base?

“It started as a joke,” Emmet told me. “I was like, what would it take to build a Gundam? Like, procurement-wise? I don’t know. But it’d be very funny to find out.”

What started as a joke kept turning up things that were not actually funny. The aforementioned 1983 National Academy report on the collapse of American machine tooling. The German robotics firm whose acquisition by a Chinese company forced a six-month pause in F-35 production while the Pentagon vetted a single robotic arm.

Forget building a mecha. We’re having trouble building anything.

How do we know? Emmet’s current MVP renders a tech tree across seven research tracks: propulsion, structural materials, mobility and balance, sensing and autonomy, armament integration, industrial base, doctrinal utility. Twenty-one gating research nodes. Six are researched at source-card depth. Thirteen are not yet started. The source-cards, where they exist, are exhaustive.

Taking the joke seriously generated exactly the kind of structured industrial-base inquiry that ordinarily costs a defense consultancy six figures and twelve months.

“It became a serious project,” Emmet said, “because I was like, oh, I’m starting to realize what’s going on with American munitions, and the defense industrial base. This could be a really interesting way to bring people into this problem.”

I asked him how close we are to building a Gundam.

“I would assume far. But the reason you set Gundam Level Zero is because it would stress test every single aspect of what we try to do. And it would also be very beneficial for, like, every other thing we want to build.”

As far as sci-fi inspirations go, Gundams have fallen out of fashion. Much more readily-available when it comes to product comparisons are things like Neal Stephenson’s “Young Lady’s Illustrated Primer” from The Diamond Age, or how Samantha from Her led to an awkward public kerfluffle between ScarJo and OpenAI.

But fundamentally what we’re doing is sitting here talking seriously and casually about building Gundams like that’s a totally normal conversation to be having, and — far from being an indulgent waste of time amongst hyperactive nerds — it’s surfacing real material constraints on our capacity as a nation and a nation’s military to build anything.

It’s only going to get weirder from here. But more than that, if we’re not getting weird, we’re probably not flying high enough.

The first time I saw one of Emmet’s dashboards, I thought: cool. The second time I thought: how? The third time I thought: oh.

Should I be working at his level? I think so, maybe. Will I? Possibly. Should you? I don’t know. That depends on what you’re working on, and what the people who depend on you have a right to expect, and how honest you are willing to be with yourself about whether the gap between what you are producing and what you could be producing is one you can keep tolerating.

Assuming that’s the right paradigm at all.

I am generally averse to maximizer culture. This tends to be for the straightforward reason that what’s most legible is typically not what’s most valuable, and a great deal of maximization simply involves solving for what’s easiest to see. Poker taught me what a life lived at constant expected-value optimization actually feels like, and it isn’t a life. But there’s also no romantic repose in willfully being worse at a job grounded in a moral imperative to improve the world. The question is whether the waxwings we build for ourselves are up to the task of soaring through the air.

The point of the metaphor, after all, isn’t that it’s dangerous to fly. It’s that, after a certain point, the sun can get too hot.

The wings Emmet has built are not a set of obligations to do more, in the maximizer sense, until you get burned. His apparatus is constructed out of a set of capabilities for spending less of his life on work that matters less, so that the work that matters more — the work that calls him to do it in the first place — can get the leverage it deserves.

That’s true of work that matters to him, and that’s true of work that matters to each of us. There shall be cathedrals. Ubiquitous AI infrastructure has arrived, and it’s here to stay. The ceiling is much higher than we think. The horizon is much weirder.

But at the root of all of it is that simple question of what matters — a question that, previously, has been easy to brush under the rug.

So tell me. What matters to you?

Does it really?

You walk into the meeting having done all the prep. Folder of charts, memo with three options pre-mortemed against every objection a normal human could plausibly raise, the elevator pitch and the dinner-party version and the under-five-minutes version. You deliver beautifully. The skeptic’s objection comes — you saw it from a mile out — and you parry with the kind of grace that briefly makes you wonder whether you’ve been undervaluing yourself in the comp band. The skeptic concedes. The boss nods.

Then the boss goes and does what she was already going to do.

You have been in this meeting. So have I. The folder didn’t matter. Or rather: it mattered a great deal to you — you wouldn’t have walked into the meeting without it — but the relationship between the existence of the folder and the actual outcome of the meeting turns out to be considerably more complicated than one might suppose.

I want to suggest that this scenario contains parallels relevant to the current AI debate.

Share

Leave aside process automation for a second (we’ll get there). In a lot of organizations, the role of an AI system is to literally or metaphorically populate something like that folder of charts. Even with access to actual omniscience — a Magic 8 Ball that, against every prior expectation of Magic 8 Balls, is actually right about everything — the rest of the meeting still happens. The skeptic still objects. The poise is still required. The boss can still go and do what the boss was going to do anyway.

The standard counter to this is something like: well, if everyone in the room had access to the Magic 8 Ball, and they all agreed it was right, then of course the boss would do what the Magic 8 Ball said. There’s something to this. But many of us have lived long enough to notice that, even in conditions of near-universal information access, sometimes people just do what they want to do because they want to do it.

A friend of mine reports to the CTO of a very large multinational. The data scientists in his organization have produced careful, well-reviewed analyses of strategies that would meaningfully improve the business. But the CEO just doesn’t like several of those strategies, and so they don’t get done. Another friend is an executive at a major sports betting company, where rote arithmetic regularly demonstrates that various promotional offers are actively unprofitable. They get run anyway, not because of some galaxy-brained four-dimensional chess but because a regional VP has an intuition about something and is sufficiently senior that nobody is interested in fighting that fight.

But so okay, those are anecdotes. What happens in real life?

Well, take basketball.

The basic insight that drove the change between the top and bottom legendary Kirk Goldsberry charts has arguably been available in some form since the 1996-1997 season, when the NBA began collecting ‘zone-tracking’ jump-shot data: a three-pointer is worth 1.5x as much as a two-pointer, but it is not 1.5x as difficult a shot1.

The math is not subtle. You don’t need a PhD in statistics to see it. But in 2005 — almost ten years after the first smattering of data was sitting there for anyone to notice — look at that top chart. The midrange is a galaxy. The three-point line is a few sparse satellites.

It took a second decade for change to actually happen at scale. The 2014–15 Golden State Warriors — the team that famously ushered in the modern game by winning a title on the strength of their long-range shooting — averaged 27 three-point attempts per game.

But even that number, the cutting-edge revolutionary number that won them a championship and turned Stephen Curry into a global icon, would be dead last in the NBA today. By a meaningful margin. The implementation revolution keeps-a-chugging, long after the information revolution rendered the problem (nominally) solved.

In contrast to Steph, let’s look at KG.

Kevin Garnett is a first-ballot Hall of Famer in one of the GOATED Hall of Fame classes2. He is among the greatest players ever to step on a court. Now consider this hilarious compilation of his utter dominance at the heinously long two-pointer — a shot that, by the math we just discussed, is the worst shot in basketball.

Garnett wasn’t dumb. His coaches weren’t dumb. The Celtics’ analytics staff did not need to extend a lucrative consulting gig to Count Von Count to hold forth at length about the difference between two- and three-point shots. Rather, he took those shots because in the context of an actual NBA game, against actual NBA defenders, while running actual NBA plays, the long two was the shot the system generated and the shot he had put in an entire career’s worth of reps to perfect.

My friend Kyle Boddy, who runs Driveline Baseball — a player development company that uses biomechanics data to help baseball players improve specific mechanical skills — talks about this all the time. Getting a professional athlete to consciously focus, in the heat of competition, on changing a deeply ingrained habit will almost always make them worse, at least in the short run. The information that the habit abstractly ought to change is just one input into a tangle of muscle memory, in-game pattern recognition, and ten thousand small contextual decisions per game.

You can nudge the system at the margins. You cannot reason it into a different state.

The lesson here involves how long it takes actual humans — even highly compensated, ferociously competitive, professional humans, running organizations whose entire existence is dedicated to winning — to convert known information into actual behavior.

Now consider the kind of system that involves more than one person doing more than one thing, in an information environment more complex than the arithmetic difference between “2” and “3”.

This, I am suggesting, is the situation most organizations are in, most of the time.

To be clear, I am not arguing that this analogy proves AI won’t change everything. I am extremely sympathetic to the view that AI will change a great deal — I’ve devoted the next phase of my professional life to a bet that it will.

What I am arguing is: the question of what AI will change, and how fast, is best answered by looking at the conditions under which better information actually moves systems. We have a lot of evidence about those conditions, and the evidence is consistently more interesting than the discourse.

There are situations where new information propagates through a system almost instantaneously. When Apple ships a software update, a hundred million devices change behavior overnight. When Tesla pushes new firmware, every car in the fleet drives slightly differently the next morning. And when Tom Kalil — famous to roughly seven giga-nerdy DC policy people including yours truly — wrote save the world one document at a time on his EEOB3 whiteboard, his point was that the actual text of a forwardable, no-edits-required, just-press-send email is itself an enormous lubricant for moving an idea up a chain of command.

These are situations in which the friction between knowing the right thing and doing the right thing is close to zero, because the system was designed to translate one into the other. It is much easier for me to compare a thousand flight prices on Google Flights than it is for me to do the same thing by calling airlines, and the result is that I now compare a thousand flight prices on Google Flights.

In other systems, this is not the case. Jim Cramer’s net worth is a useful sociological data point on the question of whether people act on better-quality investment information when they have access to it. While we’re on the subject, most American retail investors continue to put their money in actively managed funds despite fifty years of consistent data showing that low-cost index funds outperform after fees4. The information is publicly available, regularly cited, taught in introductory finance classes — but the behavior doesn’t follow.

Outside of finance, let’s turn to pharma. The active ingredient in Tylenol is acetaminophen. The active ingredient in the store-brand bottle next to it is also acetaminophen, at less than a third of the price. But Tylenol still outsells it by a comfortable margin. In sports, our beloved Cleveland Browns could, at any time over the last thirty years, have called any one of America’s 15,800 high school football programs and asked them how to run a football team. They evidently have not, and their record shows it. And anyone who has worked in social impact has had the experience of a brainstorm producing Yet Another Dashboard And/Or Database, which is dutifully built, never used, and then referred to apologetically in subsequent grant reports.

Now, importantly, AI is not just an information-quality intervention. It also automates large chunks of work that formerly required people, which moves our analysis in a different direction — when the AI is the agent rather than the agent’s advisor, a different set of frictions and constraints apply.

But the basic mechanics of how change actually propagates through systems remain in force, and will for a substantial amount of time. For now, the AI that drafts the memo still has to navigate a meeting, or delegate someone to navigate it, or concede the work of interpersonal persuasion. The AI that runs the spreadsheet might auto-execute. But somebody chose to aim it at this spreadsheet rather than another, with success criteria reflecting someone's incentives, on a budget that won out over seven other projects.

The action can automate more easily than the authority and the incentives and the resources. Automation moves the friction; it doesn’t eliminate it.

So if we want to think clearly about which parts of the world AI will reshape, and how fast, a few specific things seem to matter more than others:

• Lift. How much work is required to translate new information or capability into action? If it’s a button on a phone, very little. If it’s a behavior change requiring sign-off from three layers of management, considerable.

• Path dependency. How locked in is the existing way of doing things? An organization that has spent twenty years building muscle memory around a particular approach will not pivot for the mere reason that the alternative is now demonstrably better.

• Marginal value of quality. Even literally perfect bleach is not going to topple Tide. A great many products and services are bought on brand and habit, and the difference between “good” and “perfect” is not going to change behavior.

• Legibility of correctness. In some domains it is easy to verify which answer is right. In others — medicine, foreign policy, hiring, fine art — it is difficult or time consuming even when objectively possible, and “correctness” can itself become fraught.

• Principal-agent dynamics. The person making the decision is often optimizing for something other than what the decision is nominally about. Being right on the merits is, in a lot of cases, the wrong axis on which to win.

• Variability of value. Some systems care a great deal about getting closer to optimum. Others operate happily at 70% of the way there.

A reasonable next step is to draw a two-by-two — you almost can’t help it — with some multi-correlate aggregation representing ‘friction’ on one axis and another representing ‘the value of being correct’ on the other.

TBD on that terminology, but you see what I’m saying:

• Low friction plus high value-of-correctness is where AI changes things fastest (Google Flights, the forwardable email, the software update).

• High friction plus low value-of-correctness is where AI mostly produces dashboards nobody opens.

These two cells are relatively straightforward. Far more interesting, by a wide margin, are the other two — the ones that look like the NBA in 2005, where the ‘right answer’ is in some sense already plainly available from the 8-ball, and the system grinds slowly and inscrutably toward it while the rest of us watch in wonder or frustration from the stands.

Share

On Monday I got a beer with the founders of an AI company, and over dinner that evening I told my wife about it. What’s unusual enough to write about is: the conversation actually made sense.

In my day job, I’m the co-founder and president of an organization perhaps best described as a “venture studio.” If that combination of words means anything to you at all, please join me in my decades-long quest to explain my career to my Mom — you might have better luck than I’ve had. Point being, I am overwhelmingly used to working on things near-zero people find interesting.

So it was a wild experience to grab a beer at Crooked Run with Shannon and Travis — co-founders of an awesome new company called GroundVue — and then describe the conversation to Maria over dinner and watch her nod.

Here’s how GroundVue works. There are roughly 90,000 local governments across the United States, and most of them hold public meetings where people argue about things that matter to them: zoning, Medicaid, school budgets, data centers, SNAP benefits — all the unglamorous infrastructure of actual civic life. Those meetings get recorded, but it’s nobody’s job to process the recordings. They just sit there. GroundVue’s idea was: how about we go and make it possible to engage constructively with all of that?

Being a 40-year-old dude in a major city with a Twitter account, I talk Various AI Things at my wife a lot. As a consequence, she has learned to deploy the gracious but unmistakable expression of a person who has been asked to be excited about something she does not care to understand. This time was different, though. Having relayed the conversation with the kind of full-on golden retriever energy exclusive to a card-carrying pick-me like yours truly, I got a response I wasn't expecting: “They're making sure we know what happened at government meetings? That's great. Seems like we need that.”

We need that. Not “that sounds impressive,” which is what people say when they don’t care about something but want to be supportive. Not “wow, scary,” which is what people say when they care just enough to be alarmed. We need that — the sound of a use case that requires no translation.

And it hit me how rare that reaction is when describing an AI product.

Compare GroundVue’s pitch — “we help you know what happened at public meetings that affect you” — with what the people who run the AI industry actually say about their technology.

Sam Altman told Axios earlier this month that a world-shaking cyberattack enabled by AI is “totally possible” this year. He expects “significant threats we have to mitigate from cyber.” He has separately described the prospect of AI-engineered pathogens as “no longer a theoretical thing.” OpenAI’s corporate mission defines “AGI” as “a system that outperform[s] humans at most economically valuable work” — which is to say, outperforms me and you at our jobs — a formulation that bakes ‘this will replace you’ into the grammar of the goal itself. Dario Amodei, Anthropic’s CEO, has written at length about a future in which AI compresses a century of biological research into a decade.

These are not fringe characters. They run the industry. And they are telling us, with increasing specificity, that the technology they are building may reshape civilization — and therefore reshape all our lives — in a matter of years.

This is a lot to hear over a bowl of cereal.

The natural response is to find a frame that lets us stop worrying about it — even if these frames result in realities that are worrying on a different level. Bloomberg Opinion offered one such claim up on Wednesday: when tech CEOs acknowledge AI risk, the argument went, they’re actually selling you a product. Fear, after all, is the ultimate sales pitch, right?

And honestly? After twenty years of Silicon Valley overpromising — Google’s “don’t be evil,” Facebook ‘connecting the world’, a whole paradigm of the internet that was supposed to topple dictators but instead built them a surveillance apparatus — suspecting the sales pitch is a reflex that makes sense.

It just isn’t what’s actually going on.

The people who think and talk about AI risk do actually believe it. They don’t believe it incidentally; it’s why they’ve devoted their lives to this project. Many were thinking, and talking, and posting about it before the labs even existed. I’ve spent a good part of my adult life adjacent to this world, and the thing I keep having to explain to friends is: they mean it. Not as a pitch. Not as a positioning strategy. They got into this field because they believed they were building something with the power to reshape the world, and they are now telling you — on the record, without metaphor — that it’s happening.

This should be more alarming than the marketing theory, not less. And yet the marketing theory is more popular, because it’s the one that lets us all disengage.

Share

The reason the labs keep defaulting to the sci-fi register — curing cancer, living forever, capturing God in a bottle, whatever else — isn’t some kind of galaxy-brain tactic. It’s that they find the sci-fi register genuinely more interesting than the alternative.

I’ve talked to enough people inside these companies to know: a meaningful share of the people building frontier AI just authentically aren’t very interested in quotidian applications. A volunteer firefighter in South Dakota who replaced a paper map with an AI-powered evacuation tool? Cool, sure. But small potatoes next to nuclear fusion, or autonomous drones, or — and people do say this with straight faces — birthing a new form of life into the universe. Those ambitions are real, and those who hold them are not faking.

But here is the problem: when you lead with “we might cure cancer but also possibly kill everyone,” a normal person doesn’t hear “there is a lot of upside, but also some important downside.” What they hear is “I am fixated on grand ideas to which you are incidental” — which sounds a lot like “I am messing with you.”

And then the companies try to close the gap with advertising, and make it worse. Google’s “Dear Sydney“ Olympics ad showed a father using AI to write his kid’s fan letter to her favorite athlete; Google pulled it after the internet asked why you’d outsource the one thing a child should be learning to do. Apple’s “Crush“ ad flattened pianos, cameras, and paint cans into an iPad; Apple apologized within 48 hours. These keep happening because the people making the ads seem not to understand what the ads communicate: that technology is here to do your thing for you, or to you, but definitely not with you — and you should be impressed by this. Because you are the sidekick, and the technology is the star.

Meanwhile, the most widespread felt experience of AI for millions of Americans is their kid using ChatGPT to bullshit a homework assignment — which means the technology’s dominant cultural association is not “marvel of engineering” but “cheating.” Another prominent use case is as our favorite unlicensed therapist that’s been engineered to tell us what we want to hear.

When you combine a messaging strategy of “maybe end the world or take your job, but also might do something cool eventually” with a personal experience of “my teenager uses this to fake a book report, and also maybe confides in it about things he won’t tell me,” there are only so many viable alternatives. Acceptance isn’t one of them — that feels ridiculous, since after all none of this seems real. Skepticism might be an option, but that’s a kind of half-measure. What comes most naturally is dismissal: the reaction, whether out of fear or insecurity or sheer implausibility, that this is not a thing serious people need to pay attention to.

That dismissal is what worries me. To be clear, I don’t think everyone should just be categorically excited about AI. I think there’s plenty to be worried about, and plenty more we simply don’t understand. But the technology is real, and it’s advancing whether you or I are excited or not, or worried or not, or whether we understand it or not.

And a society that can’t develop an informed relationship with it will not govern or steward it well.

The way to build that relationship is not a better keynote speech, or a more reassuring ad campaign, or for the labs to start pretending they care about school-board analytics when what actually gets them out of bed is the robot alien shoggoth. That kind of pretending is what we’ve come to expect from Big Tech over the last two decades. It would be transparent, and cringe, and would only compound the trust problem.

The way to build it is fluency — not excitement, not conversion, but the kind of operational familiarity that comes from using these tools on problems you can see, discovering where the technology works and where it hallucinates, and developing a felt sense of when to trust the output and when to check it twice. The way we’re all fluent in spreadsheets: imperfectly, but functionally.

That fluency won’t come from the labs, because these sorts of day-in, day-out, mundane applications aren’t what draws most of the people in the building into the room. That’s great, that’s chill, and that’s fine — that’s why they are creating the most important technology since the internet, while I am sitting here writing this Substack. But it could come from the GroundVues — from the former federal technologists and the concerned PTA members and the legal aid clinics that are already doing this work, without begging for anybody’s permission, because they saw a problem and the tool was there.

That work is already happening — mostly without fanfare, mostly without funding, mostly because someone saw a problem and messed around until they cobbled together a tool to try and solve it. What’s striking is how little attention this whole world of use cases gets relative to so much discourse about whether AI will save or destroy us.

The debate generates enormous energy. The GroundVues of the world generate evidence. I know which one I’d rather bet on.

On Tuesday, the AI company Anthropic announced that they had built a model so capable of finding and exploiting software vulnerabilities that they couldn’t release it.

A model called Claude Mythos — an instance of nominative determinism up there with “Altman” and “Amodei,” some dental-hygienist uncle taking pot shots at shaping destiny — had autonomously discovered thousands of zero-day flaws in every major operating system and web browser. Some of these had survived twenty-seven years and millions of automated security tests previously undetected. In one case, the model chained together four separate vulnerabilities to escape a browser sandbox entirely on its own. No human involved. No weird guardrail caveats. The capability just emerged, the way a child who learns to write his name eventually learns to forge a signature.

Anthropic’s response was to launch something called Project Glasswing: give the model to Apple, Microsoft, Google, AWS, and about forty other organizations and say, essentially, patch everything you can before someone else builds something like this and doesn’t tell you about it. They committed a hojillion dollars in credits. They called it “urgent.” They weren’t playing.1

I read all this while eating a bowl of cereal. The Axios headline had little sirens in it: 🚨🚨🚨

“Huh,” I said, and sent a Tweet or something.

Judging by the relative lack of headlines beyond Axios, our collective reaction as a nation was similar. It was like we heard an asteroid landed in our neighborhood, but decided not to look out the window to see for sure because we needed to go clip our toenails.

This is not a column about cybersecurity or frontier-model regulatory policy. I don’t know enough to write one, and people who do — Dean Ball, Seb Krier, Mikko Hypponen, literally anyone at Glasswing — should be the ones you listen to. I defer to them with great enthusiasm and zero shame.

What I want to talk about is the cereal and the “huh.”

Your average Senator didn’t understand the internet in 1996. By 2006, Ted Stevens was describing it as “a series of tubes” and Congress was still trying to figure out what a hyperlink was. By 2016, Mark Zuckerberg was explaining to a roomful of legislators what a cookie is and why it didn’t involve any Monsters. This was all cute and cringe and funny. It generated a lot of memes. It was also, in retrospect, quite bad — bad enough that a generation of teenagers got psychologically rewired by algorithms nobody in power understood well enough to regulate.

But the internet was never going to delete anyone’s operating system overnight. We had time to be stupid, and we used all of it — fifteen years of getting blindsided by things a subreddit mod could have explained to us, while a lot of people got hurt by our sluggishness. It was bad, but it wasn’t existential.

With Mythos-class capabilities, appropriately enough, the stakes have become…mythic. A similar distance exists between what the technology can do and what the people who govern it understand — but compressed from a decade into months. It’s like our dental hygienist uncle just figured out how to shut down every computer on earth. And a society that can’t even locate its own reaction to that news — that reads about it over cereal and shrugs — is a society that is not equipped to navigate whatever comes next.

So how do we close that kind of distance? Not by pushing our glasses up our noses and ranting about Mythos over dinner, or giving talks in conference rooms about how transformative AI is going to be, or convening blue-ribbon commissions.2 That’s not going to cut it.

The way you close any distance between an abstraction and a reality is the same: you make the abstraction concrete, put it in someone’s hands, and let the understanding grow from the inside out.

Almost exactly two years ago, my wife Maria decided to open a nail salon. She had never run a business. She speaks English as a second language. She is not a person you would describe as “extremely online” (that would be me). What she had was a vision, a laptop, and a ChatGPT subscription.

So, over three weeks, she used it to write her business plan, draft her lease emails, design her price list, figure out what kind of LLC to form in D.C., write her website copy, compare ventilation systems — everything. As she buzzed along she would sing a little song — “G-P-T, GPT,” to the tune of that Rosé/Bruno Mars number.

She did not do any of this because she was abstractly excited about AI. She was not trying to prove some kind of concept. She did it because she had a problem, and the tool solved it, and so she kept using it. By the end, she didn’t just have a nail salon (which — please drop by!). She had what I would conservatively describe as a better working understanding of these models’ strengths and weaknesses than most members of Congress. Because she’d used the thing, and using it taught her something.

Matt Yglesias recently ran an experiment along similar lines, but pointed at something debatably more ambitious than “elevated beach shack chic” waterless nail hygiene. When the Washington Post was imploding in February, he got interested in whether AI plus open data could produce cost-effective local journalism. He launched a Substack called Ten Miles Square, fed local D.C. datasets to Claude, and had it generate analytical stories. He did a whole debrief on the ups and downs.

The results weren’t perfect. But a single person with no coding background, working with publicly available municipal data, generated in a few weeks the kind of unglamorous, civically vital, base-hit local reporting that used to require a whole newsroom. Test scores broken down by demographic shift. Police overtime driven by recruiting shortfalls. Metro ridership patterns that showed commuting hadn’t recovered from Covid even though weekend traffic had. Stories that matter to actual communities — but that weren’t being written, because dozens of person-hours would otherwise be spent on each one, and those kinds of business models aren’t viable anymore.

Now imagine that experiment run not by a Substack celebrity but by a community organization in Jackson, Tennessee. Or a legal aid clinic in East St. Louis. Or a volunteer fire department trying to figure out evacuation routes from a county with a paper map and a phone tree.3 Imagine it in two hundred cities across America. Imagine all that being built in a year.

The technology exists right now. It’s happening in pockets all across the country. But it isn’t happening everywhere, at scale, because nobody has built the connective tissue between the people who make these models and the communities that need them — and funded the small, scrappy, technically curious teams that would actually do the work.

This kind of work is a far cry from Mythos. But that’s exactly the point: it doesn’t feel like sci-fi. We can start to wrap our heads — and our hands — around it right now.

The people who need to develop a felt, operational understanding of these tools aren’t just the legislators who’ll vote on regulating them — though God knows they need it too. It’s the whole ecosystem that shapes how a society relates to a technology: the county administrators who decide what software to buy, the school superintendents who set policy for ten thousand kids, the philanthropists whose funding decisions determine which problems get worked on, the nonprofit directors and journalists and thought leaders whose framing determines whether the public sees AI as a toy, a threat, or a tool.

These communities learn from each other. A funder who has watched a team process a decade of benefits data in an afternoon thinks differently about what’s possible than one who’s merely read an op-ed about it. And right now, across all of these categories, a staggeringly small number of people have reps — the kind of repeated, load-bearing, this-is-part-of-my-actual-job engagement that teaches you not just what AI can do but how it fails, where it hallucinates, when to trust it and when to check it twice.

People had computers in their offices for well over a decade before it was normal to have them at home. And it was the office — not the living room — that taught them not to click on the email from the Nigerian prince.

Now, that scammy Prince is a model that can crack the entire internet.

I’ve talked to enough friends in enough industries to know that corporate America is struggling with this too. I know people whose organizations have spent more on AI strategy decks than on AI itself. The distance exists there also, and it’s not closing on its own. But I don’t know enough about enterprise deployment to fight that particular fight in corporate America.

What I know here in civil society, on my home turf, our lack of reps is kneecapping us.

My contention is that our collective cereal-bowl reaction happened because this stuff still doesn’t seem real. It feels like sci-fi even as it’s happening, and so it outpaces our collective ability to process it.

That won’t change until a vastly larger circle of people possess operational knowledge — the kind that comes from deploying the technology, repeatedly, in contexts where the stakes are real and going through the process teaches us something. Whether a court filing summary needs a human to double-check it before it goes to a judge, or whether it’s fine to flag the edge cases and let the rest through. What “good enough” means for a benefits enrollment form versus a zoning analysis, because when the consequences are different, the tolerances should be too.

This kind of operational knowledge is what will build the institutional muscle memory that lets organizations use these tools the way they use spreadsheets: imperfectly, but competently, with an earned sense of when the output is solid and when it’s going to get someone in trouble.

We do not have a decade to figure this out. The internet took roughly that long to go from Yahoo’s homepage to our moms’ Facebook accounts, and the basic infrastructure of society was never at risk of being compromised overnight. With Mythos-class capabilities, every month we fail to build this knowledge is a month where the task in front of us gets even harder. We weren’t prepared this past Tuesday. We can’t afford not to be prepared tomorrow.

I keep coming back to the name. Claude Mythos. This “aw-shucks” everyman grafted onto Aristotle’s term for the structure of a tragedy. The incomprehensibly powerful dressed in the absurdly ordinary. Milk and cereal. Cereal and milk.4

But it’s also where the opportunity lives — in the ordinary. To deal with a challenge on the scale of “Mythos”, we need deployment on the scale of “Claude”. My wife didn’t develop her understanding of AI by reading Works in Progress (alas). She developed it by opening a business and stumbling through.

To be clear, the answer to the many Cereal Moments we’ll have over the course of our lives — because this is not the last time we’ll read about something cataclysmic over breakfast and shrug — is not to feel worse about them. It’s not to generate more anxiety, or more white papers, or more takes.

It is to get in the reps. All of us. To build things. To fund and support the work. To talk about it. To learn, by doing, how to do it well — so that when the next Mythos arrives (and it will) we greet it with a felt relationship with the technology rather than a blank stare into a bowl of cereal.

The asteroid didn’t miss us. It landed. But we can go outside and check it out.