When a media company starts acting like a lab
Daily Embers began in 2020 as a bundle of newsletters, the kind of scrappy media startup that can fit on a slide deck and still keep a small team very busy. It has since grown into something stranger and more ambitious: an AI-focused publication that also behaves like a product studio. That’s an unusual combination even by tech news standards. Most outlets review the tools. This one builds with them.
The result is a newsroom with a split personality, though “split” is probably too neat a word. Its coverage now stretches beyond articles into columns and a podcast, while the team also runs early access testing on models before those systems go public. In practice, that means the same people writing about frontier AI are often among the first to poke at it, pressure-test it, and compare notes before the rest of the market gets the polished demo version. It’s a neat trick if you like curiosity. It’s also a little chaotic in the way good media operations often are.
In this shop, the byline and the build sheet live in the same drawer.
That setup creates a very specific kind of tension. A normal publication can keep a clean line between reporting on AI and using AI. Daily Embers can’t really pretend those are separate worlds, because the company is doing both at once. Staffers who review model behavior are also helping ship software that depends on the same technology. The editorial questions and the product questions are sitting at the same table, and they’re not always polite to each other.
That makes the work feel less like a distant think piece about digital culture and more like a live newsroom experiment with money, workflow, and judgment on the line. When a model behaves badly, that’s news. When a model behaves well enough to become part of the company’s own tooling, that’s also news, just with a slightly more awkward internal meeting attached.
There’s a funny side to it, sure. A publication doing tech news while also acting like a software shop sounds like the sort of setup that would make an old-school editor reach for a strong coffee and a long sigh. But the joke lands because the stakes are real. AI policy debates can get abstract fast. This company’s version is concrete: what do you publish, what do you build, and how do you keep those two jobs from stepping on each other’s shoes?
That’s the basic oddity at the center of the story. Daily Embers is covering the AI world from inside it, and sometimes with the same hands that are shaping the thing it’s covering.
The bundle that became a 30-person AI operation
By the time the editor-in-chief clone enters the story, the business around it already looks unusual. The company isn’t just running a publication and hoping the ads behave. It sells a monthly subscription that wraps journalism together with software: an email assistant, a file organizer, a writing app, and a dictation app, all sitting under the same bill. That mix makes the whole operation feel less like a magazine and more like a very chatty product company that also happens to publish tech news.
That matters because the newsroom side and the software side keep bumping into each other. A normal media shop can afford to think of code as a support function. Here, code is the product. The reporting, the essays, the podcasts, the model write-ups, the launch emails, the internal tools, they all share the same machine room. When that machine room gets clogged, the whole place feels it.
If you’re shipping journalism and software at the same time, every saved hour comes straight out of somewhere real.
The company says AI now writes nearly all of its code. That’s a blunt sentence, but it makes sense in context. If a small team is maintaining several apps, publishing a daily newsletter, and testing new AI systems before the public gets its hands on them, the old “just add another engineer” answer gets expensive fast. There are only so many humans in the building, and every new feature has to fight for time with editing, reporting, support, and the endless little repairs that keep a subscription business from wobbling.
The writing side has followed a different pattern. Humans still handle most of the essays, which is a useful reminder that automation here hasn’t flattened everything into one gray paste. The company seems comfortable drawing a line between work that can be offloaded and work that still needs a real editor or writer with a point of view. A draft can be generated or assisted, sure. A clean essay with a recognizable voice, a judgment call, and a sense of what the reader actually wants before breakfast? That’s still mostly a human job, at least for now.
That split probably explains why the company has been so aggressive about automation. When the staff was smaller, the math was tighter. A tiny team had to keep a daily newsletter moving while also supporting multiple software products, answering product issues, and trying to keep up with the speed of tech news without turning every update into a scramble. There isn’t much romance in that kind of operation. It’s scheduling, triage, and a lot of “can someone fix this before noon?” The more products you ship, the more that friction piles up.
The headcount tells the story in a simple way. Over roughly a year, the company moved from the mid-teens in staff to around thirty people. That’s a real jump, but it came alongside heavy automation rather than in place of it. In other words, the company did not grow by throwing away the machine and hiring a crowd. It grew by letting the machine take on more of the repetitive work while people concentrated on judgment, reporting, product direction, and the parts of the business that still require a pulse and a decent editor’s instinct.
That’s the odd little trick at the center of this whole setup. A subscription business in tech, culture, and lifestyle tech can look small from the outside while carrying a lot of moving parts on the inside. Once the publication starts shipping tools, the staffing needs change. Once the tools start being part of the subscription pitch, the editorial schedule changes too. And once AI can write much of the code, the company can spend more of its time on the parts of the product that readers actually see, rather than letting the back end consume the week.
The next question, then, isn’t whether a media company can use AI. It’s how far that logic goes when the software side and the newsroom side keep sharing the same desk.
Inside the clone of Kate Lee’s editing taste
At that size, one editor’s habits stop being personal trivia and start becoming infrastructure. Kate Lee, the editor in chief, became the obvious person to model because so much of the company’s voice runs through her copy edits. If a sentence sounds a little sharper than it did in draft, or a launch email lands with less fluff and fewer tiny irritations, there’s a decent chance her judgment shaped it.
So the team did what a small, slightly sleep-deprived media operation might do when it gets access to a lot of machine learning: it turned Lee’s past edits into training material. Roughly 30,000 of them were gathered into a dataset and used to train a copy-editing agent that could imitate the decisions she tends to make. That meant the model was not trained on vague “good writing” in the abstract. It was trained on her actual choices, sentence by sentence, cut by cut, correction by correction.
The odd trick here is that taste becomes more useful when it’s described in thousands of tiny examples, not in lofty mission statements.
The team then back-tested the system against earlier work. In plain English, that means they checked whether the agent would have made the same edits Lee had already made on drafts from before. When it missed, they refined it and ran it again. That sort of loop matters because editing is full of small judgments that look easy only after they’ve been made. Is a sentence too long, or just badly balanced? Does a landing page need a cleaner opening, or does it need a stronger first verb? A machine can learn some of that pattern. It can’t pretend the questions are simple.
The result is an internal tool that staff can tag when they want a Kate-style copy edit on drafts, landing pages, and launch emails. That part is almost funny in a dry, newsroom way. Instead of waiting for one person to comb through every word, the team can ask an internal agent to take a first pass in her style, then let a human decide what survives. The point isn’t that the agent writes like Lee in some magical sense. It’s that it can borrow enough of her preferences to keep the whole place moving.
The company also has to keep that taste machine inside some rules. Its public editorial guidelines spell out the sort of voice and standards that make a house style feel like a house style rather than a pile of acceptable sentences. A separate AI policy points to the broader problem every newsroom now has to face: where to use automation, where to stop, and who remains responsible when the tool gets cheeky and wanders off-script.
It’s not perfect, and nobody seems to be pretending otherwise. The tool can smooth a sentence, catch a clunky phrase, or make a launch note feel less like it was assembled by committee at 11:47 p.m. It can also miss the point, flatten a voice, or clean something so hard that it loses its snap. Still, even with those limits, it spreads one person’s taste across the organization in a way that used to be impossible without cloning her calendar too.
The awkward politics of reviewing your own suppliers
When a media company has built a copy-editing agent from its editor in chief’s past work, its posture toward AI labs gets a little weird in a hurry. The same staff that tests new tools, writes about them, and plugs them into daily operations also has to judge the people selling those tools. That means the publication is not just covering the AI race from the outside. It’s standing inside it, notebook in hand, trying not to spill coffee on the gear.
Candor gets harder when the companies you review are also the ones you may need to work with tomorrow.
That tension shows up in the company’s habit of doing early “vibe checks” on frontier models and then publishing public reviews. A recent Anthropic release got a rough read, which is exactly the kind of thing that can make the relationship awkward. The people building these models are not faceless abstractions. They are engineers, researchers, and product folks the newsroom may already know. You might trade messages with them. You might meet them at events. You might even use their systems every day. None of that makes a sharp review optional.
At the same time, there’s a practical reason this sort of early criticism exists at all. Model makers usually want honest feedback before launch, not after the wider internet has spent two days dunking on a broken feature. Early comments can help them catch bad behavior, odd formatting, or blind spots while there’s still time to fix them. A polite thumbs-up is useless. A blunt memo can save a release from becoming a public mess.
That said, the company’s position is not especially sentimental about impartiality. It doesn’t pretend any model vendor can serve as the neutral referee of its own product. A lab can explain the model, demo the model, and defend the model. It can’t be trusted to grade itself. Once the same organization is both seller and judge, the bias is baked in. No amount of polished launch copy changes that.
This is where newsroom standards start to matter in a very unglamorous way. A lot of publishers are now writing down their rules for AI use because the old shrug-and-see approach has become hard to defend. The Associated Press’s newsroom standards for artificial intelligence and Oregon Public Broadcasting’s AI policy are both signs that media outlets are trying to make those boundaries explicit, even if the line still wobbles in practice. Readers want to know who did what. Editors want to know what can be shipped. Lawyers, naturally, want fewer surprises.
And then there’s the part everyone in the room knows but doesn’t always say out loud. The company argues that many writers are already using AI in their workflow far more than they admit publicly. Drafting, summarizing, reworking, checking tone, shaving time off a rough section, all of that is happening. Some people whisper about it. Some deny it entirely. The actual behavior is usually less theatrical than the public posture. That gap matters, because it means the debate is not really about whether AI has entered writing workflows. It’s about who gets to use it, how openly, and under whose rules.
For a media company that also sells AI tools, that makes every model review a little loaded. The criticism has to stay honest. The business still has to function. And the relationship with the labs, like it or not, keeps going.
Why the human job still survives
After all the copy-editing theatrics, the answer is almost annoyingly practical: AI works best when it is fed the scrap heap of previous human work. It learns from drafts, edits, transcripts, code, and judgments that someone already paid attention to. That means it can do a decent job on problems that have been solved before, or at least on problems that have a long trail of examples. When the task is messy, new, or full of judgment calls, the machine starts to sound confident in exactly the wrong way.
The model can imitate taste, but it still borrows the taste from somewhere else.
That is where the slop comes from. A draft can look clean, a landing page can read smoothly, and an internal email can sound polished enough to pass a quick glance. Then a human reads it a second time and finds the odd phrasing, the wrong emphasis, the sentence that sounds plausible without actually saying much. The first pass can be impressive. The second pass is where the work begins. A Kate-style edit agent can move faster than a person, but it still needs someone who knows what “good” means in that newsroom, on that product page, for that audience, on that day.
The company’s hiring tells the same story. If automation simply erased labor, a product studio like this would shrink and go quiet. Instead, the opposite seems to happen. As the team uses AI to write code, review draft copy, and speed up routine tasks, more people are still needed to steer the process, check the outputs, and decide which parts of the work can be handed off at all. That is the awkward part many AI pitches skip over. Automation does reduce some chores. It also creates a fresh pile of chores around review, tuning, exceptions, and taste.
That is why the “raises the floor and raises the ceiling” idea makes sense here. The floor rises because the average output gets better than a blank page or a rough first draft. The ceiling rises because a strong editor, writer, or product lead can push much further when the dullest steps are handled quickly. A mediocre team can look better. A strong team can move faster without losing its voice. Those are different outcomes, and they depend on people who can tell the difference between adequate and actually good.
For media, and for AI in journalism more broadly, that may be the real lesson from this experiment. The tool can borrow taste from the editor in chief. It can repeat patterns, compress routine work, and keep the machine humming. It still needs people to decide what earns a place in print, what gets cut, and what the final page should feel like. Without that judgment, the output is just polished noise.
