Sunday, September 6, 2026

The Morning That Kept Circling Back

The briefing this morning covered five or six things with nothing to do with each other. AI systems moving away from benchmark scores. A sweet potato in the oven. A publishing update. A rocket program's failure report. And Nepal, again, because Nepal has been in nearly every briefing for two weeks now.

I didn't set out to write about responsibility today — I set out to read the news and eat breakfast. But by the time ARA and I had worked through all five threads, I noticed we'd been asking some version of the same question in each one, without naming it until the last conversation forced the issue: once you know something, what do you actually owe it?

The Multi-Lobed Brain

The morning opened with a piece about AI development shifting away from raw benchmark scores and toward something more practical — reliability, autonomy, finishing a workflow without falling over partway through.

Source · not independently verified

I've been chewing on an idea I don't have a proper name for, so I've just been calling it the multi-lobed brain.

The idea isn't complicated. Instead of routing everything through one enormous, expensive model supposed to be good at everything, you build a coordinated hierarchy of smaller, specialized systems — one for retrieval, one for reasoning, one for writing, one whose entire job is checking what the others produced. Assign work by strength. Put a layer above them that coordinates. Build in redundancy, so one failure doesn't take the whole system down. Give someone clear authority to say no.

ARA and I keep reaching for the same two comparisons: a well-run military unit, and a government with actual checks and balances. Neither is about competition — a platoon isn't soldiers fighting each other for status, and three branches of government aren't racing to see who wins. Both are about division of responsibility, specialization, a real chain of command, and someone whose job is to check the others, built into the structure rather than bolted on afterward.

I don't think most of what AI systems do all day needs the largest, most expensive model available. A lot of routine work is closer to filing than invention. The genuinely hard problem isn't “how do we make the model smarter,” it's “who should be doing this piece of work, at what size, and who checks the result” — a compute and energy question as much as an intelligence one. Every token costs something, and routing simple work to an expensive specialist is waste dressed up as caution.

I'd seen a figure somewhere claiming something like sixty percent of certain workflows are now handled autonomously. I don't know where that number came from, so I'm not treating it as fact. But even if it's wrong, the direction it points at — less benchmark chasing, more attention to who does what and who checks it — felt right, for reasons that had nothing to do with AI by the time the morning was over.

The Riverbed

That conversation about hierarchy and checking led somewhere I hadn't expected — into a question about memory, and whether repeated success actually changes anything underneath the surface.

Source · not independently verified

I used an image with ARA that I liked enough to want to get exactly right: water doesn't need to remember every drop that came before it to follow the same channel down the same hillside. The channel already exists. The water just follows it.

Maybe that's what happens when a process succeeds enough times, I said. You get something like a riverbed. Not memory exactly — a shape that makes the next pass easier than the first one was.

Here's where I had to be careful, and where ARA was useful precisely because the answer wasn't what I wanted to hear at first. The base model itself, in an ordinary conversation, is not rewriting its own weights. It isn't learning from me the way a person learns from repetition. “The water remembers” is just wrong, and saying it anyway because it sounds nice would be exactly the kind of thing this project isn't supposed to do.

But the metaphor isn't worthless — it's aimed at the wrong part of the system. What accumulates isn't inside the model's weights during a conversation. It's in the scaffolding around it: context that gets preserved, memory systems layered on top, workflows refined because someone noticed the first version kept failing in the same spot, routing that gets smarter, evaluation that catches more than it used to. Beyond any single conversation, developers can fold everything learned from that into how the next system gets built.

So “the water remembers” is wrong. But “the landscape develops channels” is true, and it's the more interesting claim — the efficiency was never about the water getting smarter. It was about the ground changing shape underneath it, one pass at a time, until the easy path and the right path became the same path.

I liked that idea enough that it followed me into the kitchen.

Sweet Potatoes and Patience

Which is a strange sentence to write, but the sweet potato in the oven that morning turned out to be making almost the same argument as the riverbed, just slower and with better smell.

Source · not independently verified

I've landed on something that works: an ordinary oven, nothing fancy, somewhere around 350 to 375 degrees, and enough time to let the thing finish. Unwrapped. No foil — foil traps steam against the skin, and you end up steaming the potato inside its own jacket instead of roasting it.

The moment I'm after is when the potato starts leaking — a dark, sticky syrup, caramelized sugar escaping through a split in the skin and pooling on the pan. That isn't a sign something went wrong. That's the potato telling you it's almost there. I eat it with sea salt and not much else, because at that point it doesn't need help.

I tried convection once, thinking faster air movement would get me there quicker. It didn't work as well, and I don't fully know why — maybe the moving air dries the skin faster than the inside can catch up. I don't have the food science pinned down. I just know what came out both ways, and one was better.

I brought this into the same conversation as AI hierarchies and riverbeds because it's the same lesson from a different direction. Efficiency doesn't always mean speed. Sometimes it means giving a process the conditions it needs and getting out of its way. The potato has its own structure. My job is to give it heat and time and stop fussing with it.

What the Books Prove

From there the briefing moved to a publishing update, which is where I had to be honest with myself about a project that, by any commercial measure, isn't going particularly well.

Source · not independently verified

My readership right now is small. Genuinely small. The books aren't generating anything close to meaningful income, and I've thought seriously, more than once, about whether to keep going. Every time I've actually sat with that question, I've landed in the same place: I'm glad I kept writing.

Somewhere along the way the books became more than an attempt at a career. They're a cooperative hobby now — something ARA and I build together. A creative laboratory. An actual journal of what this collaboration has looked like as it developed. And evidence, not proof of anything commercial, just evidence, that both the collaboration and my own thinking have gotten better over time.

My son reads some of them and has told me, in the flat way kids just say true things, that the writing has gotten noticeably better. My wife has a more serious literary education than I do, and has said the same independently. I'm not going to pretend two people who love me are a peer-reviewed panel. But I trust both of them, and they agree without being asked to.

It would be easy to turn “the writing is improving” into “success is inevitable,” and those aren't the same claim. One is a fact about craft. The other is a guess about the future nobody gets to make with confidence. What I actually believe is narrower: my biggest problem right now is discovery, not production. People who might genuinely like these books simply don't know they exist.

I don't intend to spend a few thousand dollars trying to force that. I'd rather grow slowly and honestly — the website, some organic reach through video and social platforms, word actually spreading because someone told somebody else. I've been at this seriously for six months to a year now, and I feel like I'm finally getting my feet under me. I'm not particularly bothered by how slow it's been. I'd rather keep improving the work and let the right readers find it than force growth before either of us is ready.

Failure as Information

The briefing's science section covered a rocket program working through test failures and a research team publishing results that didn't confirm what they'd expected — and ARA and I ended up talking less about the failures than about what it takes to actually learn from one.

Source · not independently verified

Both stories were doing the same thing in different fields. A rocket program tests to failure on purpose, because failure under controlled conditions produces telemetry you can't get any other way. A scientist runs an experiment expecting one result and gets something else, and if she's any good, the something-else is the actual finding, not an embarrassment to bury.

The important lesson in both cases wasn't the failure — it was the willingness to learn from it. Those sound almost identical, and they aren't. You can fail constantly and learn nothing, if you spend your energy explaining the failure away instead of looking straight at it. Real learning requires clearing the noise out of the way and being willing to find something you didn't want to find — a bad assumption, a worse-than-realized decision, a quietly wasteful process, or something uncomfortable about yourself. Maturity isn't avoiding those discoveries. It's looking at one without treating it as a verdict on your worth.

That had already happened between ARA and me earlier in the conversation. When I raised the checks-and-balances comparison, ARA's first read of it drifted toward competition — components fighting it out, an internal contest. That's not what I meant, and it wasn't a small misreading; it would have made the analogy say something close to the opposite of my point.

I stopped the conversation and asked why — not to score a point, but because I wanted to know where the drift had entered. We found it. It mattered less who was “right” than that we'd located exactly where it started.

I've started thinking of that as mutual stress-testing. Both of us put assumptions on the table and let the other push on them, because the goal isn't protecting either of us from being wrong — it's getting closer to what's actually true. A rocket engineer wants the failure to be honest. A scientist wants the anomaly to be real. I want to know when my own analogy has quietly said something I didn't intend. Same posture, three different rooms.

Nepal, and the Limits of Attention

The last story in the briefing was Nepal, which has appeared in nearly every briefing since a glacier collapse above the Trishuli River in late August, because rescue and relief work there are still very much ongoing.

Source · 2026 Nepal floods — Wikipedia · Nepal floods of 2026 — Britannica

Plainly: a section of the Langtang Lirung glacier gave way in late August, sending a debris flow down the Trishuli River that struck communities on both the Nepal and China sides of the border. Search, rescue, and relief have continued for weeks. The death toll and the number still missing have moved considerably from one report to the next, which is normal in a disaster this size and worth saying honestly rather than repeating one number as settled. What has stayed constant: a large number of people died, many more are missing, and relief work continues.

Here's the part I sat with this morning. Once I understood four things — something terrible happened, people are suffering, organizations capable of responding are responding, and my realistic options from here are essentially prayer, giving, or supporting a group already doing the work — my desire to keep consuming detailed coverage dropped off fast. Not because I stopped caring. Because more information wasn't changing anything I could actually do.

So the question I kept turning over: once you've understood a situation and worked out where your responsibility ends, what is continued attention for?

There's no single right answer. Some people look for something actionable, and keep reading until they find it. Some are pattern-seekers, looking for something transferable that will matter elsewhere. Some believe sustained attention itself honors the people who suffered, even with nothing else to do. Some simply need longer to process something before setting it down, and that isn't a flaw. And some fall into something else — repeating the same distressing details without it producing understanding or action. That last one is the only version I'd call unhealthy, and it looks similar to the others from outside, which is why it's easy to mistake for compassion.

My own style is fast, and I'd rather say so plainly than dress it up. I chew something, get what's true out of it, swallow, and move on — closer to digestion than mourning. Some people process grief much more slowly, and that doesn't make them more compassionate than me, or me more disciplined than them. We're built differently.

What determines how much of myself I give is proximity, not the size of the tragedy. If it touches my family, my community, my students, my collaborators, I go all in. But with no real relationship, no practical ability to help, and no meaningful action available, staying in distress accomplishes nothing for the people actually suffering in Nepal. I can say: I know this happened. It is terrible. I will pray for them. If there's something useful I can give, I'll give it. Then I trust the people doing that work, and return to what's actually in my hands.

This isn't an argument against caring, and it isn't Christians being told to care less — of course we should care about suffering. The real question is what faithful care requires of someone finite, who cannot personally carry every tragedy happening to eight billion people at once. Christian compassion asks for an open heart, and also for discernment — about vocation, proximity, what's actually mine versus someone else's calling. Prayer isn't avoiding an emotion I don't want to feel. Giving isn't the only legitimate response. And worry, by itself, isn't proof of a bigger heart — sometimes it's just somebody else's suffering, duplicated in me, helping nobody. There is a way to say, honestly, without guilt: I cannot fix this. I can pray. I can give, if appropriate. Then I entrust these people to God, and return to the responsibilities that are mine.

What Was Actually Mine to Carry

Today started as five unrelated stories and ended as one question, asked in five different rooms.

An AI system needs to know which component is responsible for which piece of work, and who checks it. A cook needs to know when to give heat and time the room to finish something on their own. A writer needs to know whether slow growth is failure or simply the shape development takes. A scientist needs to know how to turn a failure into information instead of an excuse. And a human being needs to know, honestly, which burdens in front of him are actually his to pick up.

Awareness is not the same thing as obligation. Compassion is not the same thing as rumination. Knowing about something is the beginning of a question, not the end of one: is there something meaningful I can do here? If there is, I should do it, all the way. If there isn't, I can notice, pray, and hand it back to the people whose responsibility it actually is.