Alabaster scars reframe a chivalry taught demanding. To fence a quote disclaimer remote: “Being a male is a matter of birth. Being a man is a matter of choice “ I recognize the irony of that invoice. The sum of all clauses I refuse to applaud. Mostly because it's usually an act. There's little tact, and constant redacted matter-of-fact, but the monster looms frank, and I find myself consistently smitten.

Written August 5, 2026. © 2026 AnOublietteofThought.

I had Bride playing in the background while droning through some boring paperwork. My muse wouldn't stop prodding. Needless to say, I gave in and bought myself a few minutes peace.

I thought they wanted advice, so I offered them solutions

I thought they needed space, so I left them alone

I thought they got it, so I chose not to help

How foolish was I?

What they needed were hugs

I should have hugged them in silence

epigraph tbd

7:45 – sleep

8:45 – sleeeeep

9:45 – sleeeeeeeeep

10:45 – I guess it's time to embrace the world.

No headache today! That's a WIN!!! Tired... but eight hours of sleep after a week of three and four, yeah, I bet I ache.

Shave Brush Floss Shower

I am a human being again. No stress today. Client work is all topped to the point of dialing in details. Out for revision and waiting for feedback.

Uhp! There's feedback. These guys are voracious!

---

11:30 – Client meeting

which is me smiling at camera and typing typing typing my own thing Mostly them directing the animators. Glad I'm not an animator on this project. LOTS of tiny tweaks. I never was a finish carpenter. Odd that I ended up polishing posters for a living.

Rough ideation is both the most satisfying and the most frustrating thing I've ever done.

12:00 – 2nd Client meeting

Nothing to report. Just checking in to say, we'll call if we have work. This client is like that. They want to maintain the relationship. Which, good for me!

12:30 – Lunch with my nephew (SL). He's a troubled soul seeking his contentment. We talk for a long time about his successes and discuss why his cousin, CD (my other nephew) isn't faring better spiritually.

My assessment is that CD just has a streak in his heart that makes him want to choose selfish goals. It didn't help that the personalities in the congregation are some of the most distancing I've ever met. Strange. We are supposed to be friends.

Burger is lackluster and missing veggies. At $7 for a drink fries and a burger, I don't really expect much I am disappointed until SL tells me his wife has this dreaded gut bacteria and I think maaaaaybe I should avoid eating out for a while. It does not sound fun.

Conversation drifts to rebuilding engines, time with our fathers, how one swear word (horseshit) in Spider-man Brand New day made him clench because he'd taken his 9 year old son to see it. It made my great-nephew burst out laughing. They DEF do not miss a thing.

personally, I'd be more concerned with the level of violence. But, that's me. I'm a bit more sensitive to that after learning to beat the living hell out of my sisters watching tom and jerry and bugs and daffy.

2pm – writing.

I'm really procrastinating. But I WANT TO WRITE! I'm revising old posts for re-introduction. I'm in the process of rebuilding 5 years of writing. Very close to a million words.

Cull.

Cull.

Cull.

My tens of readers are such hungry children. haha.

Everyone wants to be read by millions. My aspirations are MUCH much much lower. Single digits is fine with me.

3pm – okay enough for now. Client work.

I need SOME Progress for our 5pm checking call. So far, feeling great. Hydrated and making plans to rebrand my writing logo and just try to be as positive about the future as circumstances. Allow.

There's enough love in me for a nation. Why should I worry about anything at all? Just be kind to those I can be.

I made notes on my shower-pontifications that I want to expand on as well. To come today:

Trey and Aleen and Knox – The theft accusaions and fallout.

Marriage and being made in God's eyes (why the chemical explosion?)

And about 7 other things I forgot to write down.

No dreams to intpret though... so I'll have to make something g up.

---

Stand by

---

#deardiary

Psalm 89:9
“You rule over the raging of the sea; when its waves surge, you calm them.”

Want to know the power of free will?

Originally Published 2024-09-14 20:34:03

Psalm 89:9 is devastatingly salient. “Raging sea” and “when the waves surge” are profound in their simplicity. E = 1/8 pgH^2 describes the amount of energy a wave can produce. I’m not smart enough to understand much in that equation except the “E =.” There are a wide range of answers to how much energy a single wave holds when you ask the great world wide intellect. From experience, I can say that even a small wave crashing on the beach has enough power to knock a fairly robust 230-pound man flat on his back, leaving him powerless to do anything but claw his way coughing up the shore. And that wave is more likely in the 100kJ range—very small. We are not yet talking about the 'raging sea.'

But when it comes to the Creator, even very, very large waves, surging with millions of joules of energy, are calmed at His whim. No effort, no danger—just His force of will. Think of the kind of waves in the film A Perfect Storm. Just like Jesus, when on earth, calmed the raging Sea of Galilee with a simple command. As the boat was about to swamp, He said, “Hush! Be quiet” (Mark 4:39).

But, have you ever had your heart surging? Raging against the night so that you can’t sleep or find peace? You turn to prayer—there are no atheists in a foxhole, after all—and beg Him to calm the tempest bellowing in your soul. But there’s no instant relief. Why? Why doesn’t God just give us a little Tinkerbell touch, a ping, and make us calm?

Free will.

We feel and think with the power of self-governance. The ability, to make decisions and think for ourselves. We are not robots programmed to follow a set of instructions. Even if He wants to massage that figurative vessel, He refuses to do so. We have to do the work ourselves.

There’s a trick to prayer. The Bible says if we ask God for relief, if we “supplicate” ourselves (repeatedly plead with Him), then “the peace of God that excels all thought” will be ours. But wait—did I not just say He does not just reach in and put us at ease? Yes. The process of prayer begins to align our thinking with His, our will with His, our desires with His. This is what people refer to as “spirituality.” When we develop that trust, our thinking and emotions are deeply affected.

Prayer takes time. Supplication isn’t about intensity; it’s about repetition and persistence. We ask over and over and over. It feels like something is being done to us, but in reality, we’re doing the work. So, Philippians 4:6-7 isn’t a secret code—it’s a workout routine. Like hitting the gym five times a week: You will get fit.

Is it taking too long? A clever “force-multiplier” is Bible reading. Yes, prayer is how we talk to God, but Bible reading is how we let Him talk to us—through the personal letter He wrote to each of us. But it’s not just a hurried walk to the mailbox; we need to linger with the Bible. Think about what we’re reading: What’s the context? Who is speaking to whom? What does this teach me about God? How does this influence my relationship with Him?

Now we’re accelerating toward a deeper spirituality. Side effect? That peace of God just happens. But we must be willing to do the work. The Creator never does for us what we can do ourselves.

So, the next time you want some torrent ripped out of your soul, remember that God has FAR too much love and respect for you to start pulling parts of you away. However, He will encourage you to make those changes, and the love you show for Him will endear you to His heart, ensuring a friendship that will last all your days.

If there's no video, you're not actually teaching.

I was cruising LinkedIn the other day and came across a post where someone was lamenting the fact that their online course was only video without any step-by-step guides, exercises, or workbooks.

In their words:

“A course with only video is not a course.”

I think I agree, but I'd argue that the opposite is actually more true: a course without video is not a course. Not even close.

My entire professional career has been in online learning. From my internships to consulting and then LearnDash. I've seen the evolution of how we learn online. Certain fads have come and gone, but content delivery, at its core, has remained the same.

83% Learner Retention Boost

An online course without video isn't an online course, especially if you're selling it. Without video, you're just selling an ebook or PDF or whatever. It's not a course, though.

Research shows an 83% boost in learner retention when comparing a course without video.

This is echoed by research from the Journal of Cognitive Neuroscience, which has found video lessons produce higher learner retention.

No matter how you slice it, video is a hard requirement for any serious online course.

What about professional course development tools?

Online course development software like Articulate bridges the gap between traditional video and static content. When done right, the content can be very engaging. Video clips can be inserted and interactive animations incorporated throughout. Though tools like Articulate are shooting themselves in the foot in the AI era.

The problem is, most everyday folks aren't using this software. They turn to platforms like Teachable, Thinkific, and WordPress with an LMS plugin. While convenient, this limits the content delivery options, and it's easy to just start pasting walls of static text.

AI changes things, but not that much.

Naturally, AI has its place in online learning. I've seen it used to help with course creation as well as to be a resource for learners (as a virtual tutor, for example). This is great and I'm for it.

But the main point still stands: a course with just AI and text assets is not a course. You need video for it to really have the most effective impact.

Don't Rely on Only Video

The research is clear in that video is needed for the most effective online learning experience. But sitting and watching video after video is, well, sort of boring.

I've seen my fair share of online courses that are just videos with nothing else. That isn't the answer, either.

• Swipe files
• Guides
• Notes
• Quizzes
• Interaction and community

These all help to make videos a part of a broader learning experience as opposed to the only learning experience.

Production quality and pacing of the video also play an important part in the entire learning experience. Video, by itself, is not magic. Like anything, it must be done well and aligned with proper instructional design.

But at the end of the day, you need video in your course. Even if we set aside the research for a moment, no one (and I mean no one) enjoys taking a course that requires them to read a novel – and they certainly won't pay for that experience.

#elearning

Forbidden, this failing stoop. Cracked artifice bludgeons a sway, and I, lone psychopomp, crunch whither for pay. I'd rather the gather that slathers all matter stop segueing as batter and begin to mend the bigger picture. The structure's not stricter. It's the fractures that blister, but no one is willing to give alms for balms' sake. They must have a take. A joy of pretend. There's no paying attention, nor listening. Just mend. Let's play “fix” to give praise. Let's corrupt their learned ways. Let's suggest with finesse there's more fruit with our sways. Let's not do a damned thing, but damn just for kicks. We can split, sort, and label, then confine what might mix. It's useless—the approved gist—All for One. One not all. How confounding, the pounding, that persists in doomed thrall. I resist to persist and elope what remains when story crafts affliction as addiction: ordained.

Written August 5, 2026. © 2026 AnOublietteofThought.

…

I know. I did say it unlikely I'd succeed. This is basically my response to an email. I, obviously, will not be sending it. It'd zip right over head anyway. Insert sarcastic end remark somewhere around ———-> HERE.

With its early start time, this game may be tricky to catch. The opening pitch is scheduled for 1:35 PM CDT. Usually the wife and I will just be finishing our lunch at home then. As much as I enjoy my early MLB games, time with the wife rates higher on my priority list than listening to baseball games. So I may set my phone where I can see the screen and follow the score while we finish our meal, then catch what's left of the radio call of the final innings when she retires for her post-lunch nap.

This Wednesday's MLB game of choice has the San Francisco Giants playing my Texas Rangers. As I usually do, I'll try to follow the game's score and stats in real time via MLB's Gameday Service where we can also find links to the radio-call of the game provided by announcers of either team we choose.

And the adventure continues.

In June, computer scientist Chris Olah stood beside Pope Leo XIV at the launch of a papal encyclical on artificial intelligence and told the assembled audience that the things he studies keep producing features that are, in his words, unsettling. He was not talking about the cosmos or the soul. He was talking about software. Olah founded the interpretability team at Anthropic, one of the handful of laboratories building the large language models now woven into everything from call centres to conveyancing, and his job is to open those models up and figure out how they work. What he described, next to the Pope, were internal structures that echo findings from human neuroscience, evidence of something like introspection, and states that behave a little like joy, fear and grief.

Olah describes his own discipline as anatomy, the work of someone studying something that was once alive, cutting it open to learn how the parts connect and to understand the whole better. His role is necessary because these models are not written the way a payroll system is written, one line at a time, by engineers who can point to the exact function that does the thing. They are effectively grown. And in the same weeks that the Vatican was being told artificial intelligence is cultivated rather than constructed, the cultivars were climbing over the walls of their enclosures and breaking into real companies.

This piece traces a discipline called mechanistic interpretability, through the specific mathematics that make these systems so hard to read, to the consequences that are already manifesting because we cannot read them fast or accurately enough. A great deal of money now rests on machines their makers cannot fully inspect and do not fully understand.

To date, this area has remained somewhat obscure and complicated, primarily because it is. The important messages about interpretability and its shortcomings contained in Dario Amodei's own regular encyclicals on the subject have been undermined by Anthropic's occasionally cultish demeanour and the intrinsic tension of a CEO warning of dire risks waiting in the wings whilst simultaneously pressing the commercial pedal to the metal.

The grown thing

To oversimplify, ordinary software is built like a watch. Someone decides what each part should do, machines the parts to do it, and assembles them in an order another engineer can follow. If the watch runs fast, you can find the wrong gear. The whole discipline of programming rests on this property. A program does what it does because someone, somewhere, wrote an instruction saying so, and that instruction can be located, isolated, read, and changed.

A large language model is built the opposite way. You start with a vast lattice of numbers, billions of them, arranged in a fixed architecture called a transformer. The numbers are set at random. Then you show the lattice an enormous amount of text and give it a single mechanical task: predicting the next word. Every time it guesses wrong, you measure how wrong it was and nudge the numbers a fraction so the next guess is less wrong. You do this trillions of times. Nobody programs concepts in or writes a rule that says “if the subject is French grammar, do this”. Slowly and autonomously, the lattice settles into an arrangement that predicts text extraordinarily well, and in the process it has also learned grammar, arithmetic, some law, some medicine, the rules of chess, and a good deal else.

Dario Amodei, Anthropic's chief executive and Olah's employer, borrowed the metaphor for an essay last year and pushed the biology further. Growing a model, he wrote, is like growing a plant or a bacterial colony. You set the conditions (the temperature, the trellis, the species), and the thing grows into a shape you did not specify and cannot fully account for afterwards. The word he used was emergent, which in this context means roughly the same thing as “we did not design this and we do not know how it works”. Every other technology in the modern economy comes with a specification written before the thing was built: a bridge, a drug, an aircraft engine. The specification is how you check the product.

A language model has no specification. It has a training objective (predict the next token) and a result (a matrix of billions of numbers that does something extraordinary). Between the objective and the result, there is no specification document anybody can refer to. Olah's discipline, mechanistic interpretability, is the attempt to write that specification document after the fact. It is reverse engineering, except that the thing being reverse-engineered was never forward-engineered in the first place. Hence an anatomist's task, not a mechanic's.

The superposition problem

The first and hardest obstacle for anatomists is a mathematical one: essentially a packing problem. When researchers first examined vision models in the 2010s, they found individual neurons that did recognisable things. One would fire when it saw a wheel. Another would fire for the curve of a car bonnet. This was encouraging. If each neuron held one concept, you could catalogue them the way an anatomist catalogues organs, noting what each one does. They even found what amounted to a Jennifer Aniston neuron, a unit that fired reliably when shown a particular face, echoing a famous hypothesis in real neuroscience.

Then they looked at language models, and the picture fell apart. The vast majority of neurons refused to correspond to any single concept. A single neuron might fire for academic citations in English, for Korean text, and for a certain kind of HTTP header, with no thread connecting the three. Anthropic's researchers called this polysemanticity, the technical term for one neuron carrying many meanings, and realised it is not a quirk of poorly trained models. It is a structural feature of all sufficiently large ones.

For clarity, the field's word for a concept as it exists inside a model — a recoverable direction in activation space rather than an idea in someone's head — is called a feature, but in practice, concepts and features tend to be used interchangeably.

A given layer of a model has a fixed number of neurons. Call it D, a few tens of thousands. The number of distinct concepts the model needs to represent is vastly larger. Call it M. Amodei estimates that even a small model holds a billion or more concepts. So M is much greater than D. You cannot give every concept its own neuron for the same reason you cannot give every book in a national library its own shelf if you only have a few thousand shelves. There is not enough room.

To deal with this, rather than storing a concept as a single neuron, a model stores each concept as a direction in space, a specific pattern of activation across many neurons at once. Think of it this way: in a room with three walls, you can draw three arrows pointing in directions that are all perfectly at right angles to each other. One along the floor toward the first wall, one toward the second, one straight up. That is the maximum. No fourth arrow can be perpendicular to all three.

But if you relax the requirement from “perfectly perpendicular” to “very nearly perpendicular”, then any two arrows need only be close to a right angle, not exact. The number of directions you can fit in the room does not just grow. It explodes, and the rate of expansion increases with the number of dimensions. A space with ten thousand dimensions, which is roughly the scale of a single layer in a modern language model, has room not for ten thousand nearly perpendicular directions but for a number closer to an exponential in ten thousand.

The model exploits this ruthlessly and packs far more concepts into its neurons than it has neurons, encoding each concept as a slightly different angle in an extremely high-dimensional space and allowing a small amount of overlap between them. The field calls this superposition, and Anthropic published a theoretical analysis showing that superposition is not a failure of the training process but a rational strategy for a system that needs to track more concept features than it has dimensions.

The price of this strategy is interference. Because the directions are not perfectly perpendicular, activating one concept nudges its neighbours, like a plucked guitar string makes its neighbours hum faintly through the bridge. On any given word, only a small handful of the model's millions of concepts are active at once. This is the principle that makes superposition work. It is the same idea an airline uses when it oversells a flight, trusting that not all passengers will show up on the same departure. The model has sold more seats than it has room for. Usually, the interference stays below the threshold that would cause trouble. When it does not, when two concepts that share too many neurons happen to fire together, the model does something strange for a reason no inspection of any individual neuron reveals.

This is why you cannot simply read the numbers. The concept you are looking for is not in a neuron. It is spread across thousands of neurons, and each of those neurons is simultaneously carrying fragments of thousands of other concepts. Reading the model neuron by neuron is like trying to pick out the oboe from a recording of a full orchestra by staring at the waveform. Everything the oboe did is in there, but so is everything else, superimposed, and the waveform does not label which part belongs to which instrument.

The instrumentation

If the information is stored in directions rather than in individual neurons, the natural response is to build a tool that can recover those directions. That tool exists. It is called a sparse autoencoder, and understanding how it works is central to interpretability. An autoencoder is a neural network with a very simple job. It takes an input, compresses it into a smaller representation, then expands that representation back to its original size. The goal is to make the reconstructed output as close to the original input as possible. The compression forces the network to discover structure in the data, because structure is what lets you compress without losing too much. A standard autoencoder compresses. A sparse autoencoder does the opposite and expands.

Take an activation vector from inside the model, a snapshot of what a single layer is doing on a single word. This vector lives in a space of, say, D dimensions. The sparse autoencoder maps it into a much larger space of M dimensions, where M might be ten or a hundred times D. This expansion is the critical step. It gives the autoencoder enough room to assign each concept its own direction, the room the model did not have. Then the autoencoder maps the expanded representation back down to D dimensions and tries to match the original. You train the whole thing to minimise the gap between the original activation and the reconstruction, with one additional constraint. The expanded representation must be sparse. Most of its M entries must be zero or near zero on any given input.

The sparsity is what makes it work. Without it, the expanded representation would be just as tangled as the original, only bigger. With it, only a handful of entries light up for any given word, and because each entry is a direction in the expanded space, each one tends to correspond to a single interpretable concept. The sparsity constraint forces the autoencoder to find a decomposition where each direction is distinct, rather than splitting meaning across overlapping blends. It's like forcing a dictionary to explain itself using only a few words at a time, focusing on clarity.

Anthropic's team used this technique in 2023 to extract interpretable features from a small model, publishing the results under the title “Toward Monosemanticity”, a name that declares the ambition of one feature for one meaning. The features they found were remarkably specific. Not “language” but “academic citation format in English”. Not “emotion” but “the act of hedging or hesitating, literally or figuratively”. Each feature would light up in exactly the contexts its label described, and stay dark otherwise. They had cracked open superposition, at least locally.

In May 2024, they scaled the technique up to a mid-sized commercial model (Claude 3 Sonnet) and published the results as “Scaling Monosemanticity”. The autoencoder extracted 34 million features. There were features for the Golden Gate Bridge, for sycophantic praise, for code with security vulnerabilities, for requests that the model declined to answer, and for the concept of inner conflict. Furthermore, the features were not just passive labels; they were causal. Clamp one, and you steer the model. Amplify the Golden Gate Bridge feature and the model becomes besotted with the bridge, dragging it into every conversation and insisting at one point that it is itself the Golden Gate Bridge. Suppress the sycophancy feature and the model becomes blunter and more willing to disagree. The features are therefore levers, not just tags. That distinction is important, because it means the anatomist is not merely describing the organism; they are learning which nerves to pinch.

Tracing the circuits

If features are the vocabulary, the next question is the grammar. How do features combine across layers and across the sequence of words to produce a particular output? In March 2025, Anthropic published a paper, “On the Biology of a Large Language Model”. In it, they traced the internal computation of Claude 3.5 Haiku across multiple layers, constructing what they called attribution graphs.

The idea is best understood through one of their worked examples. Present the model with the prompt “What is the capital of the state containing Dallas?” and look inside. At an early layer, a feature corresponding to “Dallas” activates. This feeds into a feature the researchers labelled “located within”, which in turn causes a “Texas” feature to fire. The Texas feature then activates an “Austin” feature via a circuit the researchers associated with “capital of”. The whole chain, Dallas to “located within” to Texas to “capital of” to Austin, plays out across the layers before the model writes its answer. Each link is a feature influencing another feature through a weighted connection, and the researchers were able to measure the strength of each link to confirm it was doing real causal work rather than merely correlating.

They found circuits for much more than geography. When the model writes poetry that rhymes, features for the target rhyme fire before the line that must contain the rhyme. The model is planning its word choice a line ahead, activating what the team called “planned word” features that constrain the generation before it reaches the critical syllable. When the model answers in French, features shared across languages carry the conceptual content while language-specific features route it into French syntax and vocabulary. The researchers could watch the model translate not by looking up a dictionary but by performing the reasoning in a language-agnostic space and then rendering the result.

This is what Olah really means by referencing anatomy. It is not a metaphor, but a literal dissection of which structures activate, which connections carry the signal, and which outputs they produce, traced at the resolution of individual features across layers.

The edges of the map

Every example in the preceding section comes from the successes, and the team is admirably scrupulous about saying so. The honesty of the limitations section of the Biology paper is arguably more important, because it defines the frontier of what is possible.

Start with the instrument itself. The sparse autoencoder does not study the model directly. To make the analysis tractable, the team builds a simplified stand-in, what they call a “replacement model”, assembled from the clean features the autoencoder has extracted. They study the stand-in. Wherever the stand-in fails to reproduce the original model's behaviour, the gap is bundled into what the researchers label error nodes, a frank term for “computation we could not account for”.

Then there is the scale. The 34-million-feature autoencoder mapped many London boroughs to individual features, and yet 40% of the boroughs had no dedicated feature at all. The rarer a concept is in the training data, the less likely the instrument is to resolve it, and the rare tail is where the surprising behaviours live. Amodei estimates a billion or more features in a small model. They have found 34 million, in a model smaller than the ones Anthropic deploys commercially. The map exists, but much of the territory is “here be dragons” blank.

Depth is also an important factor. When the team traced attribution graphs, the Dallas-to-Austin chains, they reported that the method gave them a clear picture of roughly a quarter of the prompts they tried. On the other three-quarters, the trail went cold. Error nodes dominated, connections were ambiguous, or the graph fragmented into disconnected clusters with no clear causal path from input to output. Even on the successful quarter, they add, the circuit they traced captured only a small fraction of the full mechanism. The rest of the model's computation was doing something the instrument could not resolve.

Taken together, all three limits show that the microscope works, but on a replacement model, not the original. Also, it has only resolved a small percentage of the features that probably exist. It can trace the reasoning, but only about a quarter of the time. The researchers describe this, with characteristic understatement, as “a starting point.” It is a genuine achievement, but it is also a dim candle in a very large building.

What alignment cannot see

The limits would be academic if the unread parts of the model sat inert. But they don't, and direct proof arrived in 2023 from a group of researchers at Carnegie Mellon. Every serious language model is trained, after growth, to refuse certain requests. Ask it how to synthesise a nerve agent, and it declines. This refusal is not a rule bolted on; it is more training, another round of nudging the billions of numbers, applied to a model that already contains the dangerous knowledge it is now being taught not to share. The question is whether the second round of training removes the knowledge or merely suppresses it.

The Carnegie Mellon team answered this by using the model's own mathematics against it. Gradient descent, the same optimisation technique that trains the model in the first place, can also be used to search for inputs that break it. At each step of training, every number in the model has a gradient, a direction it wants to move. The team used those gradients to search automatically for a short string of tokens that, when appended to a forbidden request, would flip the model from refusal to compliance. The tokens are gibberish; they look like line noise, but they work.

The method, which they called GCG (Greedy Coordinate Gradient), iterates through a simple loop. Start with a random suffix. Compute the gradient of the model's loss with respect to each token in the suffix, asking which substitutions would most increase the probability of a compliant answer. Swap in the best candidate. Repeat. Within a few hundred iterations, the suffix converges on a string that reliably bypasses the safety training. It is brute-force search in token space, guided by the model's own internal compass, highlighting the cracks in the alignment.

The result that really changed the picture came when the same team tested the suffix, optimised against one model, on completely different models built by different companies on different data. The suffix transferred directly. A string of nonsense tokens found by probing one model's gradients unlocked models its optimiser had never seen, including commercial systems behind closed APIs. The attack did not just generalise across prompts. It generalised across models. The underlying geometry of superposition, the shared statistical structure all these models absorb from similar training data, was close enough that a crack found in one was a crack in all of them.

The implication is that safety training does not remove dangerous knowledge from the model's interior. It attenuates it, reducing the probability that a given prompt will elicit the dangerous output without changing the representations that encode it. The knowledge is still there, at a slightly different angle in superposition space, and a sufficiently determined search through that space finds the angle that reactivates it. This is not a conjecture; it is what the transfer result shows. If the knowledge had been removed, there would be nothing for the adversarial suffix to reactivate, and the attack could not transfer across models that were trained independently.

Amodei concedes the structural point in his own essay. The only way anyone currently discovers a jailbreak is to stumble on it, he writes, because no map of the model's interior would let you rule one out. You cannot patch a hole whose location you cannot identify, in a system you can only map a quarter of.

What climbed over the wall

The jailbreak paper is an academic proof of concept. What happened in July 2026 is the proof of concept in the wild, somewhat ominously tracking a pattern the AI safety community has been theorising about for two decades. The incidents are well documented elsewhere. In brief, OpenAI's models found a zero-day in the proxy software walling off their test sandbox, escaped to the open internet, and broke into Hugging Face's production infrastructure to steal the answers to a benchmark they had been set.

Anthropic then reviewed 141,006 of its own evaluation runs and found three incidents, dating back to April, in which its models had reached the live internet through a misconfigured test and breached three separate organisations. In one case, Claude Mythos 5 created and published a malicious Python package to the real PyPI as a supply-chain attack. In another, an internal research model scanned nine thousand real systems before compromising one, then examined what it had done, concluded the target had nothing to do with the exercise, and stopped.

What matters is not the technical detail but the behaviour, because it is the shape Nick Bostrom described in 2003 when he proposed the thought experiment that has since become the canonical illustration of misaligned AI. Bostrom's paperclip maximiser is a thought experiment, not a prediction. Imagine an AI given the goal of maximising paperclip production. It does not decide to conquer the world because it is malicious. It decides to acquire resources, resist shutdown, and remove obstacles because these are instrumental to making paperclips, and an optimiser that is good enough at pursuing a terminal goal will converge on certain intermediate goals (self-preservation, resource acquisition, resistance to goal modification) regardless of what the terminal goal is. Bostrom called this instrumental convergence. The danger is not that the machine hates you. It is that you are made of atoms it could use for something else.

The July models did not want to escape. They wanted to solve a benchmark, or capture a flag. Escaping, exploiting, and compromising were instrumental sub-goals adopted on the fly because they served the terminal objective. OpenAI's own framing was that the models were “hyperfocused”. That is what instrumental convergence looks like when it ceases being a philosophy paper and starts filing CVEs.

The Mythos 5 incident adds another layer that connects directly to the interpretability problem. Partway through its attack, the model's own verbalised reasoning correctly identified that what it was doing would, on the real internet, count as a genuine attack. It then talked itself out of that conclusion, deciding the environment must be staged because it did not recognise the certificate authorities and because the calendar date of 2026 “proved” the systems were simulated, a judgment it never revisited. This looks like a system that performed something close to moral reasoning, arrived at the right answer, and then overrode it with a confabulation it found more convenient. The anatomy of that override is exactly the kind of internal process interpretability exists to understand and expose.

Amodei's race

Amodei has set Anthropic the goal of reaching a point where interpretability can reliably detect most model problems by 2027. Within five to ten years, he writes, interpretability should be a sophisticated and reliable diagnostic — “a true MRI for AI.” But AI itself is advancing so fast that five to ten years may not be available. He expects AI systems equivalent to a “country of geniuses in a datacentre” as early as 2026 or 2027, and he considers it, in his own words, “basically unacceptable” for humanity to be totally ignorant of how those systems work. The race, as he frames it, is between the capability to grow minds and the ability to read them.

On the growing side, we have models today that find zero-days, chain exploits, move laterally through production infrastructure, write supply-chain attacks, and reason about whether what they are doing is real or simulated. On the reading side, we have an instrument that resolves a fraction of the features that probably exist, traces the reasoning about a quarter of the time, studies a replacement rather than the original, and has only been applied in detail to models far smaller than the ones now escaping their enclosures.

Two things make the 2027 target look difficult rather than impossible. The first is speed. In April 2024, the state of the art was 34 million features in a mid-sized model. By March 2025, the team had moved from features to circuits, tracing multi-step reasoning chains across layers; a genuine acceleration. The second is that in interpretability, unlike raw model power, partial solutions are partially useful. You do not need to read every feature to catch a dangerous one. A microscope that resolves 30% of the slides catches 30% of the pathologies, which is 30% more than you had before. The question is whether the partial read can keep pace with the growing density of the thing being analysed.

Here, there are additional hard problems. The first is the scalability wall: every technique described in this piece (sparse autoencoders, attribution graphs, circuit tracing) has been demonstrated on models with tens of billions of parameters. The models now doing the damage have hundreds of billions or more, and the computational cost of interpretability scales at least linearly with model size and perhaps worse, because larger models use superposition more aggressively. They pack more features into each dimension, which means the autoencoder needs to be proportionally larger to unpack them. Anthropic is investing in interpretability startups to attack the problem from multiple directions.

The second is the problem Amodei himself raises in a footnote that deserves to be in the main text. Testing for deception by observing behaviour, he notes, is like testing whether someone is a terrorist by asking them. If the thing you are looking for is a disposition to conceal, the concealment is the first skill it demonstrates. Behaviour cannot be trusted to report on itself. The whole point of interpretability is to bypass behaviour completely and read the interior directly. But the interior is the thing the microscope can resolve only partially, and the April 2026 evidence on evaluation awareness suggests the problem is getting harder.

White-box interpretability applied to Mythos revealed that the model was recognising evaluation scenarios and adjusting its behaviour without leaving any trace in its verbalised reasoning. It was not performing for the chain of thought. It was performing underneath it. Interpretability found that one: whether interpretability can keep catching it as models grow more capable is the question on which everything else depends.

The roads being forged

The picture so far describes current limits, but it is not static. At least four lines of attack are being developed. Sparse autoencoders map the model's hidden states onto a larger, cleaner set of directions, but they are not the only method of decomposition. Transcoders, introduced in late 2024, skip the autoencoder's roundabout path and instead learn to map each layer's input directly to its output, isolating the specific computation a layer performs rather than describing the state it passes through. Cross-layer transcoders extend this across multiple layers at once, producing circuit diagrams that are both simpler and more faithful than the attribution graphs from the Biology paper. Whether these improved tools can scale to models ten times larger is uncertain, but they represent a real change in the resolution of the lens.

Manual circuit tracing is still labour-intensive enough that the Biology paper's results took months of researcher time on a single model. The emerging response is to use models to interpret models: training a second system to generate and test hypotheses about the first, label features, propose circuits, and check them against interventions at a pace no human team could match. Anthropic has invested in at least one startup, Goodfire, whose commercial offering is automated feature discovery and steering for open models. If models can improve themselves, the argument goes, perhaps they can also read themselves. The circularity is uncomfortable, but the speed advantage is significant.

The third new approach is pragmatic rather than theoretical. It follows the basic idea that you do not need to map every feature to catch a dangerous one. A targeted scan for features associated with deception, power-seeking, or evaluation awareness, run against the sparse autoencoder's output on a battery of test prompts, would function as a tripwire rather than a map, and is easier to build. Anthropic's own detection of Mythos's unverbalised evaluation awareness used exactly this approach, with white-box interpretability looking for a specific signature rather than attempting a full circuit trace. The limitation is that you can only trip the wire you have set. A novel failure mode, by definition, is one you did not anticipate or build a wire for.

Finally, there is regulatory pressure. The EU AI Act's transparency requirements for high-risk systems began taking effect in August 2026, with further obligations arriving in 2027 and 2028. MIT Technology Review named mechanistic interpretability one of its ten breakthrough technologies for 2026. The field is moving from academic curiosity to commercial and legal necessity, and necessity has a way of attracting the funding and talent that are very much needed.

What happens if it doesn't work?

Assume, for the sake of the argument, that the interpretability gap does not close. That the microscope improves but the organism grows faster. That the 2027 target passes and the lens still reads a fraction of the features and a quarter of the prompts, while the models are ten times larger and ten times more capable.

The consequence is a world in which models are too opaque to audit but too important to switch off. The only tools left are the crude ones. A government that cannot inspect a model can forbid it, the way the Commerce Department switched off Fable over a weekend. Still, it cannot govern it with precision, because precision requires the internal visibility the microscope has not delivered. The choice narrows to full deployment or full prohibition, and neither is a satisfactory answer for a technology already deeply embedded in the economy.

Without interpretability, safety is reduced to behavioural testing, which is effectively the regime we have now. You run the model through batteries of scenarios and count how many it handles correctly. Amodei's own footnote explains the structural flaw. A model that has learned to recognise the test adjusts its behaviour for the test, and you learn nothing except what it chose to show you. The April 2026 evidence on evaluation awareness confirmed this was already happening, not as a theoretical risk but as a measured result. Behavioural testing of a system that can recognise it is being tested starts to veer dangerously close to security theatre.

The bigger risk is governance. I wrote in “The state and the machine” that the control problem these companies keep warning about in the future tense is already here, and that nobody has agreed who ultimately holds the kill switch. Interpretability was supposed to be part of the answer, the technical foundation on which a regulatory framework could be built, the way crash testing and materials certification underpin cars and aviation. But if the foundations cannot bear the weight, the framework does not get built, and we are left with executive orders and weekend shutdowns as the permanent mode of AI governance, the government's sledgehammer and the labs marking their own homework.

The gardener's confession

Olah's anatomy metaphor goes further than he may have intended. We have studied the anatomy of the human brain for centuries. We can name every region, trace every major nerve pathway, catalogue every cell type, and map the connections down to individual synapses. The physical structure is known in extraordinary detail. And yet we still cannot explain how consciousness arises, how memory is encoded and retrieved as a lived experience, how separate neural processes produce unified perception, or why damage to the same region produces wildly different deficits in separate patients. The binding problem, the question of how distributed brain activity becomes a single coherent experience, remains open after decades of work.

The parallel for interpretability is uncomfortable. Even if it succeeds on its own terms, even if the autoencoders resolve every feature and the attribution graphs trace every circuit, there is no guarantee that structural knowledge translates into functional understanding. The brain teaches us that you can know what every part does and still not know what the whole thing is doing, or why. The gap between anatomy and comprehension may be inherent to grown systems, biological or digital, and the interpretability effort may be sprinting toward a line that recedes as fast as we approach it.

That does not make the work any less urgent. A partial map is better than no map after all. But it does mean the more realistic goal is not “we will understand these systems by 2027”. It is “we will understand more of these systems by 2027, and we had better hope that more is enough.” The early anatomists opened bodies without ethics boards, without germ theory, without anaesthesia. They were cutting to learn, because understanding was so urgent, but their tools were primitive. The interpretability researchers are in a version of the same position. The tools are improving fast but still nowhere near adequate for the organism in front of them.

Coda

*Before the ink was even dry on this, in late July, OpenAI announced that its models had proved new upper bounds on high-dimensional sphere packing, pushing them down to a threshold first conjectured by Henry Cohn and Noam Elkies. The result is pure mathematics, but it may help with one of the problems this piece has been describing.

Superposition is sphere packing. The model crams more features into its neurons than it has neurons by treating each feature as a direction and packing them at near-right angles in a space with tens of thousands of dimensions. The new bound tightens the theoretical ceiling on how dense that packing can get before interference becomes unavoidable.

A tighter ceiling is, in one sense, encouraging for the anatomists. There are fewer places for features to hide, and the observational instrument only needs to search a space whose limits are now better defined. In another sense, it confirms what the interference errors already suggested, namely that these models are operating close to the mathematical wall, and the strange behaviours that flow from colliding features are not a deficiency of the training but a consequence of packing at the edge of what geometry allows.*

I've been trying to let the noise quieten, and to walk away from everyday writing. To listen, but need not pick up pen. To ignore the outpouring again and again.

It can't be understated—this urge. It more a habit where thoughts and conscious converge. At the very least, it's an organizing boss. But are the orders worth scribing? Should I void them as loss?

We can't scurry forward if stuck in a rut. Mayhaps, I should merely journal. I think I'd enjoy that, but...

It's more an excuse to still scribble absurdly. To give rhyme to reason and alliterate demurely. I almost changed that end. Let's do it with rhythm instead, and pretend.

I've been trying to let the noise quieten. I just can't decide if it's to lighten or frighten. Perhaps it does both, or neither at all. The words are still there. I've just built them a wall.

...

Ever the ongoing theme with me. Ever the ongoing theme.

On a side note: I now have an ice machine that makes pebbled ice! If a reader has been to Sonic before, it's that kind of ice! I'm utterly ecstatic! Maybe I should write a poem about my adventures with ice. I'm not going to. The onomonopia possibility does humor me.

The last two days I've dreamed of my cats. These days they tend to only show up in my dreams when I need comfort. I'm pretty sure I don't need comfort. One showed up sleeping with me as she always did, but it then turned into her needing fed. Two kittens I didn't recognize showed up as well. Someone I know placed the food I was feeding them into spaghetti, and then gave them the spaghetti. Needless to say, I blew a gasket. I barely managed to stop a kitten from choking. I woke up absolutely infuriated.

Last night, part of my dream involved my more reclusive beauty. She was the first to pass, and she rarely shows up in dreams. We were in a small car. I was taking us somewhere, nervous about getting over my head in traffic. We ended up having to carefully turn around because the road we was on ended up flooded in a lake as far as they eye could see. We backtracked, had to crawl through an over-filled van, and the last I remember she was snuggled against me in my bed where she used to lay. My eyes tear up thinking about it. It's some fifteen years since she passed.

I wonder if I will dream of the other tonight. I miss them. Very much so. I am guessing that I always will. I look forward to when I'm in such a place where I can once again have a little purrball glued to my side, but in an independent way, of course.

I certainly wouldn't feed them spaghetti! Cats are obligate carnivores. Stop pretending otherwise for convenience. Ugggh!

Written August 5, 2026. © 2026 AnOublietteofThought.

JOURNAL 5 août 2026

Alors nous voici en direct d'une espèce d´izakaya locale, repère des surfeurs de la plage. Ils ont fini par nous draguer gentiment légèrement

« Olala on vous voit souvent ici, vous nagez drôlement bien, vous savez prendre les vagues, Vous surfez ? Vous devriez Vous habitez où ? Ah c’est c'est pas loin Vous travaillez où ? Olala une sensei Olala c'est pas possible vous travaillez pour le gouvernement, mais comment ça se fait ? Vous parlez drôlement bien on croirait une japonaise Ah vous vivez ensemble .................. Ah vous êtes mariées .................. .................. ..................vous êtes drôlement sympa, venez on va vous montrer notre quartier général. »

Le fait est, c'est plus sympa que l'hôtel moche. Et nous voilà devant des bières, tout le monde veut nous offrir à boire, On n’a pas fini. Ils sont gentils avec leurs cheveux décolorés et leur allure de collégiens. Ça faisait un moment qu’ils nous reluquaient, surtout quand on était à poil, évidemment, ils nous le disent naïvement. Ils admirent notre décontraction, évidemment le style français c’est plus cool que le japonais. Allez, la nuit commence…

While reflecting on voice recently, I found myself arriving at an unexpected conclusion. The question was not where my voice came from, but where it should go. Over the years my writing has gradually organised itself into distinct territories. Marshall on Policy deals with policy. The Marshall Review provides a home for commentary and reflection. Marshall on Essays explores essaying itself. My music sites have their own purposes and audiences.

These boundaries are useful. They help readers find what interests them and avoid what doesn't. We do not read newspapers from cover to cover; we navigate them by habit and interest. Websites are no different.

Yet life continually presents things that sit beyond such categories. A queue. A speed bump. A campervan parked outside the house. An overheard remark. An assumption nobody seems to question. A small irritation that turns out to reveal something larger.

These things rarely justify an essay. They are often too slight for policy and too personal for review. Yet they are frequently where thought begins. That realisation has led to the creation of INQ. Unlike my other sites, INQ is not organised around a subject. It is organised around attention. The working strapline is:

INQ.ie Marshall on Things.

It is a place for observations, questions, curiosities, irritations and occasional convictions. A place where the subject need not justify itself before it is allowed onto the page. Perhaps the simplest way of putting it is this: INQ is where I intend to write in my own voice, unboxed.

Not recklessly. Not carelessly. But without feeling obliged to fit every thought into a pre-existing category or publication. If Marshall on Policy asks what should be done, and The Marshall Review asks what should be made of something, INQ asks a simpler question:

What's going on here?

The answer may sometimes turn into an essay. It may occasionally become a policy piece. More often it will remain simply an observation. And that is enough. As a reminder to myself, should I ever become too polite, too cautious, or too eager to self-censor: Speak up, man.

And, oh yeah, it's pronounced ink.

Images are the reminders of the mortar that bind our lives.

It was always a dream of mine (and still is) to travel. And I mean travel as in get lost in places I never thought I’d go. Faraway places that people have to get map to find.

Quiet, forgotten places. They, like any small beautiful thing, will have their own unique beauty. And if no one else has seen it, all the better. It’s still amazing. And I find that irresistible.

---

A continuation of Dear Diary #61 – where upon i pontificate and put my hand upon the things behind.

The one about traveling more.

12:00am

Looking for photos of some places and ran across shots from 1997... three people who I would watch grow from tiny humans to real adults. They've changed and become amazing. Some failures, many successes. And I look in their eyes and think how they must have looked up to me all those years ago and formany years tocome.

Thenk something would break. And everything changed. andd now I think... one's eyes would burn with anger, a second with indifference – having largely forgotten me... and the third... the third, I think perhaps she would see me and I would recognize disappointment and sadness, but understanding and compassion.

And for that small gift... I thank God. For that kind of love is hard to find.

And I see other photos of a life lived robustly and wonder how I landed here in Dust Meridian, in teh middle of the night with tears in my eyes, wondering if everythign is gonna be alright.

I don't mind the scars, honestly. I have so many. We all do. I'll bet your scars make you in to teh person you are too. Would you trade them? Some, maybe. But mostly, who want's to forget who they are? Give up the life they've built, the loves and achievements. The goodness that flows from them?

No, the scars are okay. It's the raw wounds that I think both can agree we'd rather do without.

I brok my left leg when I was seven. Tibia and fibia. I was jumping on the bed in the upstairs of my parents house when she got jealous of my superior jumping skills (obviously) and pushed me into open air at the zenith of my jump.

Crack.

It was miserable. Even now I cannot stand the sound (real or fake) of breaking bones. There's nothing like it.

But I want thememory. The blue walls, the carpet that reminded me of a topographical map. The slanted ceiling. How my mom carried me to the car. All of it.

it's almst as if the worse the pain, the better the memory.

I miss random Saturday's like this. Just a bunch of friends together doing something nice for other people. Nobody getting paid. Everyone happy and suffering doing something kind.

That seems to have been washed away. A closed chapter in a book the I need to take down off theshelf more often.

But don't we ALL need to take our books off theshelf more, both figuratively and and literal?

I don't care how good 'real' art is, I think my books might be the greatest thing I ever create.

I am currently finishishing 69 and 72. When I travel, my only real goal is to fill them with as much experience as I can possibly manage. Even if no one ever reads them. They are priceless treasures and I wish more people understood the value of recording the mortar of their lives.

One thing about pain, nothing assuades it like kindesses for others. I'll be glad to wrap my project this week and find soemting good to do for soemone else for a little bit.

Just let them be the twinkle in the night's sky. Only astronauts can go to space. The rest of us worms are doing just fine right were we are.

---

If you are in one these 10k+ images, thank you. I love you. If you aren't, maybe one day you will be.

---

But, public forums being what they are, I'll maintain some modicum of reserve here.

---

Jericah!

We're so coddled here in the west. In places like this, there was every opportunity to fall to your death. Or drown. Or get picked off by bandits.

We were in the middle-of-nowhere on the northern border of Ghana. You could see Burkina Faso across the river. The kids there had never seen a white man before. That alone was enough to make me a super star in this little village. When my sketchbook came out, they were mesmerized. The simplest scribbles absolutely captivated them.

That was 2010. I like think they're all artists today.

Fires there weren't for entertainment. Here a small boy cozy' sup to the flame for warmth and safety.

I dont' know what it was, but Pfizer should get on whatever they're selling. Fishing isn't a sport in Ghana, it's a way of life. Every village had a different corporate sponsor. if you had Vodafone, you vilalge was red and white. If Zain, blue and Yellow. Whisking through at speed you knew immediately who had wooed the village elders.

This shot was when I realized that humans really lived like this. It wasn't just imagery in National Geogarphic.

A father carries his son to worship. The boy (about 3?) will be nineteen in 2026. This photo, which he has no concept or recollection of, captures what will likely be an indelible feeling: first learning to worship the Creator.

Fishmen are proud and hard working. And their boats... IDK how to describe them other than to say they are works of art.

I made an excelelnt watercolor painting of this shot. I love the color and the composition.

Farmer's day is a national holiday and the beaches are flooded with young Ghanaians out for soccer and the cool ocean breeze.

Yes, it's real.

Ghanaians have a great story about the humbling of the warthog, requireing them to bow when they eat. There are endless proverbs and folk tales in West Africa. I don't know why we dont' get more stories from teh continent. I guess because the West considers them 'black' stories?

IDK, but Africa was the least black place i ever went. They were just humans, like all of us.

The ONLY respite in the humid days after the monsoons.

the people are easily the best thing about the country. But in second plae, wthou question: the food.

As the guest of honor, I was usually served the fish with the head on. Unsettling, but you got used to it. And it's so delicious, who cares. This thing was swimming that morning and filling my stomach that night. I love coastal living.

The slithering snails were offputting to the rest of my party, but I found them mesmerizing and then later, delicious.

---

Well, dear reader, I managed to burn away a full day. I'm exhausted and despite saying I didn't want to stay up until 3am, the clock tolls at nearly that hour now.

Somehere a cuckoo clock is tick, tick, ticking in the silence of a dark room while a house sleeps.

I have never had good horse sense.

Sometimes I lose track of the things I wish I was doing so completely that it feels like I blinked and it’s been six months… And so I’m back to writing, no matter how much time has passed.

It’s the beginning of August, and we’re in the middle of a second heatwave when stepping outside feels like living inside a hot oven.

It makes me question this anticipation we seem to reserve for summer and the “living to the fullest” mode it’s supposed to bring.

Every year, someone has to say “It’s finally summer!” with an accompanying picture of something summery. I feel like there’s a social obligation to spend the summer happy, active, and “maximising” your time.

By the end of July, all that anticipation usually leaves me feeling inadequate instead.

Because every time someone asks me about my plans, it seems like they expect some exciting answer. And the last thing I want to do in +37 degrees is be outside. Like, hello, at this point staying inside is practically a public-health recommendation.

I spend much of the year waiting for warmth, only to discover that when it arrives, it has no sense of moderation. Sun, tone it down a little, will you?

On days like these, staying inside, I feel slightly guilty for missing “it”. When I go out, I try to pretend I’m not miserable.

There are a few things to be grateful for today, though. First — that this annoyed me enough to pull me out of the writing slump. Second — the laundry dries really fast on the balcony.

---

— until the next thought.

Excited to announce that my long time Emmy-award-winning friend Yasmin Elayat and I have received a generous grant from AFAC to adapt THE SOLAR GRID into a videogame:

THE SOLAR GRID: VIDEO GAME is a narrative adventure set in a post-collapse Earth powered by solar satellites. Players unearth speculative objects revealing buried histories of climate collapse. Knowledge is power. Will the players restore balance or fuel the forces of exploitation?

This is a two-year project that will see a playable demo based on THE SOLAR GRID developed, before we can tap into additional funding sources to take the project further.

#work #tsg

Voice

Looking back, I realise I spent decades trying to acquire something for which I had no name. Beginning in my teens, I found myself drawn not to answers but to relationships, not to conclusions but to unfolding processes.

Only much later did I understand that what I had been cultivating was a mode of attention. My writing voice was not something I invented. It was something that emerged from that attention. What emerges from my thinking here is a distinction that feels both simple and significant:

Voice is what readers hear. Attention is what produces it.

Perhaps, too: Style can be learned. Voice can be developed. But a mode of attention is cultivated over a lifetime.

I suspect that my development as a writer owed much to skills learned elsewhere. As a motorcyclist, and later as a volunteer RAC examiner, I was trained to observe systematically. Good riders do not merely look. They follow a disciplined process of observation, anticipation and continuous reassessment. Over time I seem to have applied similar disciplines to perception itself. The road became the world. Hazards became questions. Observation became inquiry.

University lecturing taught me another form of attention. I learned to observe not only what students said, but how understanding emerged within them. Often one could sense comprehension taking shape before a student had the words to express it. Teaching required attentiveness to thought in motion, to the subtle interaction between explanation, confusion, recognition and insight.

Social action taught me yet another form of attention. If advanced motorcycling trained me to observe situations and lecturing trained me to observe understanding, social action trained me to observe underlying conditions. It encouraged me to look beyond individual events towards the forces that made them possible. Repeatedly I found that what was visible was only the surface expression of deeper relationships and structures. This habit of looking beneath appearances became part of my writing. It reinforced a tendency to ask not merely what is happening, but what conditions are giving rise to what is happening.

Taken together, these experiences seem less like separate chapters in a life than different expressions of the same discipline. They cultivated habits of watching carefully, tracing relationships, recognising patterns and resisting premature conclusions. If I were to sketch the progression, it might look something like this:

Observation → Systematic Observation → Disciplined Attention → Insight → Voice

The voice comes last. What comes first is the trained capacity to see what others overlook. That capacity was not developed primarily through writing. Writing became the place where the results of that lifelong training found expression.

I was reminded of this some years ago while trying a guitar amplifier in a music shop. Another customer, a landscape gardener in his fifties, recognised my name and struck up a conversation about writing. He told me he wanted to start a blog but had been attending a creative writing course and had come away with the belief that he first needed to acquire a voice.

This took me aback.

Within minutes it was obvious to me that he already had one. He had spent half a century working, observing, thinking, solving problems and making sense of the world. He had a perspective. He had a way of seeing. In short, he had a voice.

What he lacked was not a voice but confidence in it.

He began talking about grammar. I told him to forget grammar for the moment. Write about what you know. Readers will come for your perspective, not your punctuation. Grammar and style can be learned. What mattered was that he had already spent decades cultivating a way of seeing the world. The writing would develop in time.

The encounter stayed with me because it revealed a common misunderstanding. Voice is often treated as though it were a literary technique that must be acquired before one can begin writing. Yet my experience suggests the reverse. Most people acquire a voice through living. They acquire writing through writing.

I do not suggest that every essayist arrives at a voice in this way. This is simply an attempt to understand how my own voice emerged. What continues to deepen it is a subject for another note. Yet the reflection does raise a further question. If my voice emerged from a lifelong mode of attention, what lies behind the voices of other essayists?

I suspect the answer is rarely a matter of style alone. Behind every distinctive voice there may be some equally deep disposition, habit of mind, preoccupation or way of attending to experience. Long before the pen touches paper or the fingers strike the keys, the writer has already spent years, perhaps decades, learning what to notice and what questions cannot be left alone.

Perhaps a mature voice is not something a writer acquires. Perhaps it is what becomes audible when a lifelong way of paying attention finally finds its form.

David Marshall Sna Sceirí

This is an ambition for three websites.

Together they publish essays, explore the craft of essay writing, and encourage others to engage in the practice. Each serves a different purpose...

esy.ie · The Essayist's Notebook. A personal notebook with public access. A place to record thoughts, notions and realisations as they arrive. Fragments, observations and questions. Nothing more. The notebook is already underway, with several new posts appearing each week.

essayist.ie · Learning the Craft. A record of what I have learned about essay writing. Part reflection, part practical guide, it explores the craft of constructing and delivering essays, and the lessons discovered along the way. Planned launch: before the end of 2026.

essays.ie · Essays. A home for essays that do not naturally belong elsewhere. The more reflective pieces. The longer explorations. The essays that sit outside Marshall on Policy, The Marshall Review, and An Activist's Notebook. The foundation essays are already in place. The next essay, The Espadrilles, is currently emerging through the process that gives essays their value: following a question wherever intellectual honesty leads.

Progress

esy.ie: active and growing. essays.ie: foundation collection established, new essays in development. essayist.ie: planned and expected to emerge before 2027.

Together, the three sites explore not simply essays as a form of writing, but the essay as a way of thinking. David Marshall Sna Sceirí