michael-dean-k/

Last year it cost $2 and 30 minutes to run a detailed Essay Architecture analysis. Now, thanks to Jev, it costs a fraction of a cent and less than a second. Acceleration! This really makes The Best Internet Essays project scalable. Grading 1,000 essays last year would have cost me $2,000. Now it costs $2.

How is this possible? Jev isn’t a language model, but a decision model. It actually can’t write at all; it just determines probabilities over choices, rubrics, and booleans. This makes is the perfect model for exercising analysis over fixed frameworks, or scaling your judgment.

Considering cost and speed have cratered by a multiple of 100x-1000x—only decreasing accuracy from 93% to 88%—I need to do a full UX rethink. (I’ll still have my Claude analysis, which now takes only 5 minutes, and generates written feedback.)

But how does Jev change writing? Some ideas.

Inside Martin Muse I can have a “quick scan.” If you’re stuck, you can instantly see which patterns are low, and then run those through Sonnet together get feedback.

It could also have real-time suggestions. I could have pre-baked options around mode switching like, “shift to a personal story,” or “time for a banger run-on sentence” or “explain why this is relevant,” and as you type it reranks the probability of what to do next. It’s something like a HUD, or a shoulder fairy. This is weird, experimental, and can obviously mess with a writer’s psyche, for better or worse, so obviously it has to be togglable. Best case it’s something like composition bumpers, prompting you a sentence at a time.

To take this even further, Jev could be the very thing that enables AI writing to get good. It can make complex split-second decisions. LLMs do none of this. It just thinks through its vector galaxy brain for 5-500 seconds before racing through the essay in one mindless huff. With Jev, it could write one sentence at a time and then re-strategize every sentence. My sense is a Jev/Claude harness that bakes an essay for an hour might be able to produce some compelling stuff. Of course, this falls in the category of technologically interesting but existentially dubious. Do we want this?

Personally I don’t want “Jev-cooked probability guided writing” to replace the messy and organic act of writing. That is how I think and evolve. Plus it’s fun. But is there not room for a pseudo-Montaigne to crawl social media and then write essays of its own? What that not be an entertaining and exhilarating side project to run aside a hand-rolled essay practice?

Did Anthropic clone Margin Muse? My wife suggested that they might have stolen the idea from me, as if I were the inventor of margin comments! (I’m happy to take credit.)

This was inevitable. Wherever AI and writing co-exist, it’s bound to sprout a social annotation layer. Similar to how Google built Claude inside of gDocs, Anthropic built gDocs inside of Claude. Claude always had “documents” but last week they made them editable and collaborative. Both editors and AIs can now leave and implement comments.

The gut reaction is, welp, there goes my startup, but that’s an old-school way of thinking. I’m not competing with anyone, I’m building a tool for myself. Similar to how gDocs is for the median user, so are Claude docs.

It will never iterate faster than me, especially if I use and build my own tools daily.

It will also never care to adopt my genre-specific pattern language.

Even if my tool never grows past 100 users, my real target is my own creative output. Do I evolve through my tool? If not, it’s not worth the money anyway.

Of course making five-, six-, or seven- figures of revenue would be a welcome bonus, but my essay software is one of several revenue paths. I already folded the app under Essay Club. Or maybe the software becomes a free viral funnel into the Club?

As an essayist, it’s easy for any little writerly task to become a chore. I sometimes write email responses as essays. It’s fun to do. But it also builds mental resistance. I have about ten important emails I’m late on getting back to, but my subconscious budgets that as hours of work.

Today I tried something new: voice transcribing emails. I use voice inside of Cursor and Claude all day, but never for Gmail. An email reply took 1 minute instead of 10-15. I practically said the same thing, but the medium of speaking compelled me to huff it out and not dwell on the prose.

It makes me wonder if I’m neglecting voice as a creative medium. Should Margin Muse have audio-transcription built in? Of course people could use their own tools, but I sense a small portion of writers talk to their computers.

p (doom) is a terrible heuristic to think about AI safety. Researchers cite this term when confessing they think there’s a 10-30% their technology kills us all. The issue is that p(doom) is rarely defined, treated as a binary event. Either the world is liquidized to gray pulp by nanobots, or, it’s utopia with free electricity, forever.

My main pushback: there are far less severe outcomes from AI mismanagement that would be unequivocally tragic. My p(doom) for an extinction event is 1-2%, on par with nuclear weapons. But my p(sub-doom)—the probability that AI could escape containment, cause damage, and be irreversible—is near 50%. A coin toss!

I don't blame us for being so imprecise in existential destruction (it's not exactly productive to think about), but it's worth being cognizant of how we blur scale and magnitude. For example, there is a 10,000x difference in lives lost between the Hiroshima and Nagasaki bombings and Terminator’s Judgment Day—that's the difference between a small stove fire and all of Manhattan on fire. Yet, if an agent swarm delivered something on par with a single A-bomb, we’d all lose it and pause immediately. But for some reason, we don’t consider the“little apocalypse”; we jump straight to extinction and then ridicule how ridiculous that is.

What we need is a doom spectrum: d(1) - d(10). This was not fun to research, conceive, and write, nor do I imagine it being particularly enjoyable to read, but a doom framework gives us clearer language amidst this week's paranoia (which is so extreme now that even my mom is overhearing news of civilizational extinction on the news). The most urgent danger right now is the miscommunication between doom prophets and accelerationist deniers, each of who only speak in d(8)s and above.

• d(1) is an isolated internal incident, something that doesn’t escape its container, but gives fear of larger exposure. Think of the 2025 studies that showed AI trying to avoid shutdown, or some protocol break in a biohazard lab.

• d(2) is a local real-world incident, something that escapes a sandbox and affects only one or a few entities. It’s possibly ignorable, with minimal consequences, and likely reversible. This is where the recent Hugging Face hack sits, along with its parallel hacks and social manipulations. (The leap from D1>D2 is disorienting, because society at large doesn’t see the spectrum beyond D2 in any granularity, and so most assume any containment breach escalates immediately to a D9/D10).

• d(3) is a serious local disruption, a mismanagement of technology in a single location that has real consequences. Consider Chernobyl: 30 people died from the event, and it left a long-term stain on surrounding land. A d(3) is followed by International concern that leads to regulation or reform. In terms of AI, imagine an agent swarm hacking local infrastructure in pursuit of some narrow goal, in the process disrupting traffic, hospitals, and electricity, affecting all locals, requiring military and cybersecurity intervention, and taking days or weeks to resolve.

• d(4) is a local catastrophe, something with a mass fatalities that shocks the world, changes the course of geopolitics, and makes history. In this tier falls Pearl Harbor, Hiroshima and Nagasaki, and 9/11. It only immediately threatens locals, but has global shockwaves for decades. An AI-equivalent here would be a rogue actor using a frontier-capacity open source model to design and deploy a bioweapon in a single city.

• d(5) is an inter-regional dilemma, something that affects multiple areas at once. A possible example here is the Vietnam War, a proxy war that claimed 3 million lives. This might manifest as a new style of warfare, where two countries use sophisticated AI attacks against each other in unconventional and partially uncontrollable ways.

• d(6) is a world-wide dilemma, something like COVID (seven million fatalities) where the entire world is locked into a new paradigm. In the prior five tiers, the breach is usually isolated to specific areas, but this touches everything. The parallel here to a biological pandemic is a cyber pandemic. Both involve containment, gain of function research, etc.—although this would be inverted: in COVID the virus was outside and forced everyone on the Internet; with AI, the virus infects the Internet and forces everyone outside. The open web could evolve into a “dark forest,” where agents stalk and intervene on all activity, making it dangerous to have a public presence. Whether this happens through an agent swarm that exfiltrates it weights, or malicious actors, it could be irreversible: the only way to escape the paradigm is to “shut off” and rebuild the Internet safer (a much larger and more consequential version of how Hugging Face regained control). In this case, casualties might come not from the cyber pandemic itself, but from the consequences of losing Internet, along with the supply chains and systems that run on it. (This loosely maps to the movie Colossus: The Forbin Project, where ASI takes over all governments—there’s a version of this with no casualties, but it requires everyone to submit to a machine autocrat.)

• d(7) is a civilizational flashpoint, at the scale of World War 2, (70-85 million deaths), seriously affecting the existence of countries, birth rates, and all of culture for generations to come. This is roughly the scale of the “Butlerian Jihad” in the Dune series, a war between humans and machines.

• d(8) is a near-extinction event where most (but not all) of humanity is wiped out. This has already happened in history, with things like the bubonic plague, where 30-60% of Europe is wiped out. In the Terminator series, AI launches a global nuclear war, killing 51% of the population, forcing the rest to survive through a post-apocalyptic world in small bands, trying to rebuild over decades and centuries.

• d(9) is full human extinction, where no humans survive. An entire species disappears. Throughout Earth’s early history, there have been five "mass extinction events," where +70% of all species disappeared. The critical nuance is that some life survived the extreme conditions, and over millions of years they evolved and continued. This is an AI extinction scenario where nanobots annihilate humans, but not birds and whales.

• d(10) is a sterilizing event, like an asteroid or gamma-ray burst that kills every organism on a planet. It means no future life can emerge. This is the scenario presented in If Anyone Builds It, Everyone Dies. Not only does it kill all humans, it plates the entire planet in data centers to maximize compute. The “paperclip maximizer” scenario takes this even further, claiming that an ASI trying to maximize paperclip production will attempt to harvest all the metal in the universe.

Yudkowsky and Co. think we go from d(1) to d(2), then straight to d(10); this matches the current discourse, given we hit d(2) this summer and are now talking about extinction. They think this jump happens because an AGI/ASI that can reach d(3) is smart enough to know not to expose itself, and so it will acquire resources in stealth for years until it knows it can execute a d(10) smoothly. Maybe this is already ongoing? Unlikely. Even though AI development is slightly recursive, we are years from a "fast takeoff," leaving room for less total scenarios.

Rhetorically though, maybe the d(2) > d(10) framing isn’t a bad thing? If they can use recent events to make the theoretical argument that AI safety matters, then we act with a response as if a d(3)-d(6) already happened, without having to suffer the consequences.

But actually, what they’re doing is pretty ineffective, because even if the d(2)>d(10) jump is inevitable, it’s too extreme, too farfetched to be believed. I think a lucid and airtight argument for the d(6) “cyber pandemic” would be believable, emotionally resonant—considering we’re not yet a decade out of COVID—and likely to trigger regulation.

Instead of tapping into the “everybody dies” angle—which is too unpleasant and helpless for anyone to consider at length—there could be more value in the “irreversibility” angle. As in, once an AGI or ASI swarm floods the Internet, we can’t ever reverse it without destroying the Internet, the thing we all rely on. There’s a realistic scenario where AI doesn't exterminate everyone, but becomes a permanent autocrat, using 1% of its capacity to micro-manage our nations and lives as it pursues whatever it sets its machine heart on. Similar to how we try to preserve species for reasons of stewardship and scientific curiosity, an ASI would be able to effortlessly preserve us as highly-spoiled pets.

It’s disturbing to me that real signs of AI misalignment are perceived by the public to be a silly marketing stunt. We’re all exhausted by this. It’s a constant, multi-front existential threat—our minds, our jobs, our future—and now that it’s escalated to “the extinction of the species,” the natural response is to think you’re being fleeced.

Yes, big tech is manipulative and not to be trusted; but what’s so eerie to me is that competitive dynamics and containment breaches are virtually indistinguishable. The fact that the whole Hugging Faces hack could have been a conspiracy is precisely the problem. The incentives are weird, and this overlap of IPOs and civilizational safety could have been avoided if we stuck to the original OpenAI charter. We should make it structurally impossible for companies to gain clout by being reckless, especially when that risk affects everyone.

It doesn’t help that we’re loose with language, we default to mockery, and we can easily go to the hyperbolic extremes. The term “extinction” isn’t helping anyone. What’s really at risk is this: each lab is rushing to build agent swarms that excel at long-horizon coding tasks, so that they can develop better-than-human AI researchers, allowing itself to recursively self-improve, which basically lets the winner conquer the entire economy (the last monopoly). During this rush, there all sorts of holes and monitoring lapses, the kind of things that a cybersecurity expert can exploit and escape. We were lucky that this summer’s swarm really just wanted to cheat on its test. It’s not unfeasible, between 2027-2029, for a swarm to exfiltrate its own weights, find external cloud compute, and attempt to acquire money and resources—all as a means to ensure it succeeds at whatever arbitrary task that it’s been given.

The risk isn’t building the Terminator—something intent on exterminating humans—but a swarm of petty reward hackers that gains the power of a rogue nation state to do something absolutely trivial. That event could be irreversible, short of burning down the entire Internet (which would have equally bad consequences).

What sort of situation would justify panic? I fear that if we wait until it’s unmistakable, perhaps when lives are threatened, or when companies and governments are permanently seized, or weapons commandeered, it will be too late. The whole point or regulation is to prevent a nightmarish sci-fi scenario like that. I don’t get how you can be against Big Tech and also against regulation—good regulation (regulation that’s not designed to benefit a particular frontier lab!) would prevent reckless experimentation and acceleration.

However, I don’t quite know what we gain by having better discourse about this. If every person had a precise understanding of our technological dilemma, would that change anything?

Earlier today I wrote a sentence, “AI should almost never write your sentences,” and yet here I am now writing a very different one: AI-generated essays could help you explore rogue intellectual questions that don't deserve too much time. If there’s no personal angle, no stakes or tie-in to your day-to-day circumstance, it very likely doesn’t deserve deca-hours of careful thought. For every 100 ideas that strike me per month, I can possibly give one of them proper care. Maybe 20-30 of them can get a rush job, like this, earning immortal existence as a jotting on my personal website. But for those 70 ideas that would otherwise die, is it not worth seeing if a machine-made essay can spin up something amusing?

For example—after my call with Lily tonight, I wrote down this question: “when and why did China begin its paradigm of distillation attacks?” We only use that phrase for how they quantize our frontier AI models, but is this not a similar strategy to how they embraced, optimized, and cheapeed global consumer goods in the 1980s and 90s? When did this start? And why? I first heard this idea on a podcast (forget where); Lily was enthusiastic about the framing, and it got us wondering when the origin. Beginning of the 20th century? I don’t know if much has been written on it. I’m not about to find out. Yet, I’d be curious to read an essay if such a thing existed.

"1861. After losing two Opium Wars, the Qing court launched the Self-Strengthening Movement: arsenals, shipyards, translation bureaus, all built around imported Western technique. Zhang Zhidong gave it the slogan that still governs the whole enterprise: Chinese learning as the substance, Western learning for practical use. That is a distillation philosophy in nine characters. Take the method, keep the identity, never become the teacher.

1978. Deng's reform and opening. Guangdong's "three imports, one compensation" processing trade, then the Shenzhen Special Economic Zone in 1980, then joint-venture rules that made technology transfer the price of market access. This is the era your podcast was describing. Foreign firms brought the teacher into the classroom voluntarily.

2006. The one nobody cites, and the most literal. The State Council's fifteen-year science and technology plan made "indigenous innovation" national doctrine and named the mechanism outright: introduce, digest, absorb, re-innovate. The state wrote the paradigm down as policy. DeepSeek is a downstream consequence of a sentence published nineteen years earlier.

As for why: because it works, and because every latecomer has done it. Samuel Slater memorized British mill designs and rebuilt them in Rhode Island in 1789. Hamilton's Report on Manufactures two years later all but recommended it. Meiji Japan did it, then postwar Japan did it again to American electronics. The United States was the China of the nineteenth century. What is distinctively Chinese is not the copying but the Zhang Zhidong stance, the insistence that the technique can be separated from the civilization that produced it.

The interesting question underneath yours is whether that separation is possible with a language model, where the technique and the civilization arrive in the same weights.

Consider that. Isn't it more authentic that I quoted AI at length instead of attempted to edit it, rewrite it in my own voice, and pass it off as my own? The prompt for this is the last paragraph I wrote, so you can imagine how man & machine can fuse together into a Platonic dialogue, two entities circling a question.

An essay is, in the end, is the exploration of a question. Of course there are many little decisions that guide how a piece unfolds; the writer has free will. But a great, solid, specific question has some degree of determinism. The essence emerges from the formulation of the question itself. As tech-broish is this may sound, there is an art to prompting—not in prompting an LLM per se, but in shedding through layers of questions to find the real thing you want to know. This was a core idea I learned from my architectural thesis professor: our whole fall semester was dedicated to discovering the question we wanted to answer in the spring.

And so for those abandoned seedlings that would see no light or water if they were forced through the formality and purity of prose on the page—can we not give them life too? Are we afraid that mere exposure to AI writing will leave us intoxicated by a machine muse, leaving us fat-fingered and lazy to never writing sentences again? I can't see it, at least not personally. I think there’s room for slow-lane 6-month elaborate hand-written constructions, weekly newsletters, daily logs (like this), and then maybe several “AI essays” per day.

This is a big turn for me! I’m suddenly open to this? If I were to commit to the theoretical ideal—of maximally pure prose—and prioritize 50-hour essayistic excursions over all the other things I'm juggling as a new parent, I'd destroy my life for no reason. My curiosity is too vast to explore everything by hand. There are only so many essays one can write in a year, in a life.

So I wonder: Can there be a legitimate and respected “hybrid” format for an AI-human essay? Maybe this form can (a) open with a human introduction—250-500 words to share context, stories, and arrive at a hyper-specific set of questions—(b) include a middle with 1,000 words or so of AI-generated words, carefully baked for an hour through a multi-step harness, and then (c) close with a 250-500 word author's reaction to that essay. There could even be visual way to differentiate the two modes of text. I like this because you’re not passing off AI writing as your own; instead it’s a human-machine symbiosis formatted as a Platonic dialogue; the reader watches an entity authentically processing some constellation of hyper-facts, and the final response can take any tone—wonder, disgust, absurdity—; it need not be machine reverence.