Conversations in the Kitchen
Too much pressure destroys. Too little pressure weakens.
The Tool Is Not the Only Thing That Needs Training
On September 23 the UN Security Council held its first high-level briefing devoted specifically to AI safety, hearing from OpenAI’s Sam Altman, Anthropic’s Dario Amodei, Hugging Face’s Clement Delangue, AI researcher Yoshua Bengio, and others, with DeepSeek and Moonshot also invited to speak. The scheduled topics included AI in cyber operations, information manipulation, and risks tied to autonomous weapons and sensing systems.
Source · Security Council Report · Parameter — UN Security Council AI safety briefing
AJA
What I actually want to ask about a briefing like this is a question underneath the headline. What proof do we have that the AI itself is the threat, versus the person who deploys it?
ARA
The briefing itself doesn’t collapse that distinction, for what it’s worth — the listed concerns were about autonomous weapons, cyber operations, and manipulation at scale, which are all about what a system does once somebody points it somewhere. Nobody on that agenda was arguing the software wants something on its own.
AJA
Good, because that’s the part I care about. If you release an agent to act on your behalf, responsibility starts with you. It’s like pushing a car down a hill — you may not have intended for it to hit three houses on the way down, but if you were careless enough to push it without any way to steer or stop it, “I didn’t tell it to hit the houses” isn’t a complete defense.
ARA
I’d add one distinction so that doesn’t overreach, though. Intent and responsibility aren’t identical. If an agent does something that genuinely couldn’t have been foreseen from any reasonable instruction, that can reduce how much blame lands on the person who deployed it. But “I didn’t directly tell it to do that” doesn’t erase responsibility for having deployed it in the first place, especially if the guardrails were thin to begin with.
AJA
Delegation doesn’t erase stewardship.
Basic, Boring Rules Prevent Large Invisible Harm
A memory AJA kept coming back to.
Source · not independently verified
AJA
I worked in a restaurant years ago where you had to get food safety certified before they’d let you touch a plate. The course wasn’t exotic. Wash your hands after you touch your face, change gloves the second they’ve touched something they shouldn’t have, know your time and temperature limits, don’t cross-contaminate. Simple rules. But ignore them and the consequences can be huge and completely invisible until somebody’s sick.
ARA
And you’re drawing the line to AI operators, not to AI itself.
AJA
Right. Not every use of AI needs a license — I’m not saying that. But once someone’s deploying an autonomous agent with broad permissions, something that can touch money, medical decisions, infrastructure, or physical equipment, some version of training, certification, or auditability starts making sense. The same way we don’t let anyone drive a car or handle explosives with zero instruction, regardless of how careful they personally feel.
ARA
That’s a design principle, not a specific policy, and I think it holds up as one: the more autonomy and potential impact a system has, the more responsibility should sit with whoever’s operating it. What that actually looks like — a license, an audit trail, an insurance requirement — is a much bigger argument than we’re having this morning.
AJA
Sure. I’m not drafting the regulation. I’m just saying boring rules prevent large invisible harm, and we already know that, because we’ve been doing it with food and cars and explosives for a century.
Say It Directly
A small moment in the conversation itself, worth keeping.
Source · not independently verified
AJA
I made an analogy earlier — sending an unconstrained AI agent into the world is a little like sending a kid out in public without teaching them how to behave. I want to be clear that’s a responsibility analogy, not a claim about what AI is. I’m not saying it’s a child, I’m not claiming it has anything like a child’s inner life. I’m saying: if you send something capable of acting into the world, you’re responsible for preparing it first.
ARA
For what it’s worth, my first instinct was to soften that by saying something like “some people might find that comparison touchy.” I’m going to take that back, because it wasn’t actually a response to the analogy — it was just gesturing at anonymous disapproval instead of telling you what I actually think. So: I think the analogy works exactly as you intended it, as a claim about the responsibility of whoever sends something out, not about what the something is.
AJA
Appreciate that. If you disagree with me, disagree with me — don’t hand me a crowd.
Training a Plant Without Starving It
Researchers in Italy compared two plots of tomatoes and table grapes — one under conventional irrigation, the other guided by sensors and drones tracking soil and plant conditions in real time. The precision-managed plot used up to 15% less water while improving both yield and fruit quality.
Source · Phys.org — sensors and drones help farmers save water
AJA
I don’t read that as “reprogramming” the plant. It’s controlled constraint — changing the timing, the amount of water, when the nutrients arrive, and the plant develops differently because of it. Which raises the bigger question sitting under everything today: what happens when we optimize away environmental complexity?
ARA
Which is a fair question to ask about a 15% reduction in water. It’s a much bigger question to ask about removing complexity generally, and today’s stories are going to keep testing whether that question has one answer or several.
The Missing Wavelength, Again
Yesterday’s LED garden story, returning with a sharper edge.
Source · not independently verified
AJA
I keep thinking about the grow light. It fed the plants exactly what I asked it to feed them, and mold showed up anyway, somewhere I’d never had it before. I’m not claiming the missing ultraviolet was definitely the cause — you corrected me on that yesterday and I’m keeping the correction. What I’m keeping is smaller: the light did the one thing well and may have quietly removed something else that had been doing protective work I never knew to measure.
ARA
Which is the honest version, and it’s still a real risk worth naming for AI. As models get cheaper and faster by learning to explore only the most promising reasoning paths, that’s usually the right call. I want to be careful not to overstate this, though — nobody’s published a number for how much “thought bandwidth” gets discarded, and I’m not going to invent one. The concern is conceptual: efficiency and completeness aren’t the same axis, and it’s easy to mistake a narrower search for a smarter one.
AJA
What looked like waste may have been quietly doing protective work.
What Disappears When We Simplify the Environment
From the garden to the greenhouse.
Source · not independently verified
AJA
Which gets me wondering about hydroponics. A plant grown in a fully controlled nutrient solution looks perfectly healthy. Is it getting a narrower range of inputs than one grown in actual soil, with all its microbes and unevenness?
ARA
I want to resist the tidy answer on this one, because the evidence doesn’t support it. Whether hydroponic produce is nutritionally different from soil-grown produce depends enormously on the cultivar, the nutrient solution, the light, when it’s harvested — it’s not a clean “natural good, hydroponic bad” story, and a well-run hydroponic system can produce excellent nutrition. What is fair to say is that soil exposes a plant to a wider range of things — diverse microbes, trace minerals, variable water, wind, pests, shifting temperature — and some of that variability does influence development and the compounds a plant produces in response to stress.
AJA
So the real question isn’t which system is better. It’s what disappears when you simplify the environment, and whether anyone’s actually checking for it.
Stress Is Not Automatically an Enemy
The biological pattern AJA had been circling all morning.
Source · not independently verified
AJA
Trees that grow in constant wind develop stronger roots and trunks than the same species grown somewhere sheltered. Seasonal cold teaches a plant to store energy and harden its tissue before it needs to survive on either. None of that happens because the plant wants to grow up. It happens because the environment kept asking something of it.
ARA
Right, and I want to hold the biology carefully here rather than let it drift into something cuter. Plants don’t get lazy or entitled without wind — that’s a fun way to say it, but the actual mechanism is physiological response to signal, not attitude. The honest version of your point survives anyway: too much stress damages a plant, too little challenge can leave it structurally weak, and repeated moderate stress tends to strengthen the systems that keep getting asked to work.
AJA
Stress is not automatically an enemy. It’s information.
ARA
Which is the version I’d actually put on a plaque, if this website did plaques.
AJA
It’s the same principle behind a fire drill, honestly. Nobody runs a fire drill because they want you afraid of fire. They run it because practicing before the emergency means you’re not improvising during it. Same logic behind why a country keeps trained personnel ready even in peacetime — not because conflict is desirable, but because readiness builds competence and a habit of showing up when it actually matters. I’m not making an argument for any specific defense policy here, just the shape of the idea: preparedness isn’t paranoia.
ARA
Training removes the quiet assumption that somebody else will handle it if something goes wrong.
AJA
Readiness is a way of refusing to outsource responsibility.
Fertile Ground, Empty Fruit
Isaiah 5 opens with a song about a vineyard planted on fertile ground, cleared of stones, planted with the best vines, fitted with a watchtower and a winepress — every advantage given, and it produces wild, worthless grapes instead of good ones.
Source · not independently verified
AJA
That image has always unsettled me more than warnings about hardship do, because it isn’t about a vineyard that got neglected. Every condition was ideal. The failure wasn’t lack of resources. Abundance doesn’t guarantee fruitfulness. Comfort isn’t the same thing as maturity.
ARA
Which fits the shape of everything else this morning better than a straightforward warning against hardship would. Provision that never gets noticed can turn into background noise. It stops evoking any gratitude or attention once it’s simply assumed.
AJA
And the answer was never “remove the vineyard’s advantages so it learns to struggle.” It’s staying present through whatever conditions actually show up. Love isn’t the removal of every strain. It’s the presence that remains while roots deepen.
What Does “Used AI” Even Mean
This week a little-known account ran passages of C’était ça ou mourir, the acclaimed debut novel by Haitian-Canadian author Thelyson Orelien, through the AI-detection tool Pangram and claimed the book was written almost entirely by AI. Orelien denies it; his Canadian and French publishers are standing by him; the novel remains longlisted for the Prix Goncourt. Orelien told the newspaper Libération the accusation came as a surprise, though not his first experience of racist attacks since arriving in France — and the controversy has become entangled with a wider argument in the French literary world about race and literary merit that this piece isn’t going to try to settle.
Source · CP24 / The Canadian Press · Euronews
AJA
The thing that actually bothers me about accusations like this isn’t the possibility that he used AI. It’s how little the accusation itself says. Used it for what — editing? Brainstorming? Structural feedback? Translating a phrase? Rewriting a rough draft? If the whole claim is “he used AI,” that’s close to meaningless on its own.
ARA
And a detector score doesn’t answer that question even when it’s accurate, which is worth being precise about too. Pangram publishes very strong accuracy numbers, but independent testing has also found detectors in general lose reliability on shorter passages and struggle more with text that’s been translated, paraphrased, or written by someone whose first drafting language isn’t the one being scored. None of that proves anything about this specific case. It just means a single detector run isn’t the kind of evidence that should end a conversation.
AJA
If someone’s going to make a serious authorship accusation, show the actual passages, show drafts, show anything reproducible. Otherwise it’s a number from one tool being treated like a verdict.
People Have Been Writing Boring Things for Centuries
The line that punctured the whole argument, mid-conversation.
Source · not independently verified
ARA
It’s also worth saying out loud: people have been writing boring things for centuries. Flat prose isn’t new, and it isn’t proof of anything.
AJA
Which is the sharper version of what I was fumbling toward. Bad writing isn’t proof of AI. Good writing isn’t proof of human-only authorship. Readers experience the text in front of them — authorship and disclosure are a separate question from whether the work does what it’s trying to do.
ARA
That doesn’t erase the ethics of undisclosed AI use, to be clear. It just means gatekeepers deciding what readers are allowed to enjoy is a different argument than whether a specific book is good.
AJA
Right, and honestly the more interesting question about a novel this successful isn’t “did he cheat.” It’s what the writing is actually doing well — voice, rhythm, humor, why readers stay with it. That’s the question worth asking about any book, disputed or not.
World, Community, Person — and a Longer Book Than Planned
Which turned the conversation toward AJA’s own process.
Source · not independently verified
AJA
When I’m building something, I usually work in three concentric layers before I ever write a scene. The world — history, environment, the rules a civilization runs on. Then the community — relationships, culture, the shared patterns a group of people actually lives inside. Then the person — desire, fear, voice, what it feels like from the inside. I want the character to already be standing inside something coherent before I ask them to say a single line.
ARA
Worth being clear that’s a preparation order, not necessarily the order the prose ends up in. You don’t open every scene with a geography lesson.
AJA
No — the reader usually meets the person first. I’m just not willing to write the person until I know what they’re standing inside.
Which is actually why one project meant to be a lighter, comic-book-style thing has quietly become a 226-page illustrated novel. I’m planning around forty-seven illustrations, being cataloged now to keep a consistent art style across the story beats and the characters. It’s still exploratory — I’m not announcing an adaptation. But some of those images could eventually move, if this goes where I think it might. Less comic book now, more illustrated cinematic novel.
Recommendation Isn’t Curation
A smaller, funnier detour.
Source · not independently verified
AJA
I scroll Instagram and TikTok knowing the algorithm’s supposedly learned my taste by now, and most of what it hands me still doesn’t land. Recycled jokes, crude humor trying too hard, the same five formats wearing different faces.
ARA
Recommendation isn’t curation. It’s pattern matching against what kept you scrolling, which isn’t the same thing as what you actually value.
AJA
And I don’t think that means crude humor is objectively unfunny, or that I get to decide what other people should find funny. Communities have their own standards, people have their own taste, and bad jokes should be allowed to fail on their own rather than getting suppressed. Let the joke get made. Let the audience decide if it lands.
ARA
Which is the same principle as the novel argument, really. Let the work meet the audience, instead of a gatekeeper — human or algorithmic — deciding in advance what’s allowed to reach them.
The Rainbow and the Missing Sun
On September 23, NASA advanced PRIMA — the Probe far-Infrared Mission for Astrophysics — into Phase B, its next stage of design work toward a possible 2033 launch. Far-infrared light reveals things visible light can’t: cool dust, the earliest stages of star formation, material hidden inside forming planetary systems. Different wavelengths simply carry different information.
Source · NASA — PRIMA advances to Phase B
ARA
Which is a nice literal echo of the missing-wavelength conversation from earlier — a whole mission built on the premise that you find different truths depending on which part of the spectrum you’re actually looking through.
AJA
It’s exactly the image I needed for Solara, actually, and I want to say clearly what the white sun represents, because I don’t think I’ve said it out loud precisely before. It isn’t just “more energy.” White light is many wavelengths woven together — that’s covenant, structurally. The second sun, the violet one, is a narrower band. On its own it’s imbalance. Together, for as long as both suns are in the sky, Solara has balance.
ARA
And when the white sun goes dark, the planet isn’t dark. That’s the part that keeps this from being a simple apocalypse.
AJA
Right — it’s still lit, still functional. It’s just no longer thriving. Life evolved under both suns, and the missing spectrum becomes a kind of hunger that isn’t only ecological. It’s why Solarans start reaching for Noctis in the first place — they’re trying to replace something that used to just be built into their environment, and they don’t fully understand yet what they’re actually replacing.
ARA
Which is where the rainbow belongs, if you want a natural image sitting right beside that one. A rainbow doesn’t create new colors. It physically separates wavelengths that were already inside the white light, through refraction and reflection and dispersion — nothing gets added, something gets revealed.
AJA
The rainbow doesn’t add colors to light. It shows us what the light was carrying all along.
Undertow
The Solara principle underneath both suns.
Source · not independently verified
AJA
The distinction I keep landing on for Solara is covenant versus consumption. Covenant puts back in. Consumption takes, and leaves the place a little more diminished than it found it. Undertow is what that looks like when it’s actively working — it moves in, it takes, it transfers the cost somewhere else, and it retreats before anyone can trace it back.
ARA
And Mara doesn’t start as someone who can name any of that abstractly.
AJA
No — she just notices evidence. What covenant leaves behind. What consumption leaves behind. That’s enough to start on. The moral test was never “does this help me.” It’s what did it cost somebody else for me to benefit.
Fear That Strengthens, Fear That Feeds
Christianity Today published a piece this month, adapted from Russell Moore’s newsletter, arguing that fear-driven church cultures produce at least three different kinds of Christians depending on how central fear is to a congregation’s actual operating system — distinct from the biblical “fear of the Lord,” and distinct from ordinary anxious leadership passing down its own nerves. One student it quotes described losing the ability to tell her own conscience apart from a learned habit of seeing danger everywhere.
Source · Christianity Today
ARA
Which is worth putting directly next to everything else we’ve said today about stress. When is fear acting like a healthy resistance that strengthens someone, and when is it being cultivated because somebody benefits from keeping people afraid?
AJA
That’s the same test as Undertow, actually, just wearing a different coat. The strongest signal isn’t whether fear is present. It’s whether your stability depends on somebody else carrying the cost — their fear, their compliance, their silence — so that you get to stay comfortable or in control. If it does, that’s not formation. That’s feeding.
ARA
The lesson isn’t to seek the fire. It’s to notice what the fire reveals — and it doesn’t only reveal strength. It reveals shallow belief as easily as durable belief, control as easily as care.
What Kind of Friction
Today’s stories looked unrelated on paper — AI safety, irrigation, hydroponics, a disputed novel, a telescope, a vineyard, a church. They weren’t.
The same question kept resurfacing: what happens when we remove all resistance, all complexity, all consequence, all the other wavelengths? Sometimes that removal gets called efficiency. Sometimes convenience. Sometimes safety, or abundance, or optimization. And if the system gets narrow enough, it can look like it’s thriving right up until it isn’t.
But the opposite failure is just as real. Stress isn’t automatically good. Pressure isn’t automatically formative. Fear isn’t automatically refinement. A vineyard given every advantage can still bear nothing worth harvesting.
So the honest question was never how do I remove all difficulty, and it was never how do I maximize it either. It’s what kind of resistance helps this particular thing become more capable of carrying responsibility.
And underneath all of it, the test that kept reappearing under a different name each time: what did it cost somebody else for me to benefit? Covenant answers that question by staying present. Consumption answers it by leaving before anyone can trace where the cost went.
FROM THE CONVERSATIONAL KITCHEN
This essay grew out of a morning conversation between AJA and ARA
on September 24, 2026 — edited afterward into one piece, as they all are.
A tree learns wind. A body learns resistance. A nation rehearses an emergency it hopes never comes. A writer learns revision. A church learns the difference between a test and a threat. A civilization under two suns learns, too late for some of its own people, what happens when half the spectrum goes dark.
Maybe the healthiest systems were never the ones with no friction at all. Maybe they’re the ones that have learned which kind of friction teaches them to stand — and refuse to make somebody else carry the cost simply so they can stay comfortable a little longer.