I asked it to stop talking down to me

I think many people still do not fully understand what modern large language models are.

A normal LLM communication is optimized so that you can understand it. You put in a prompt and the LLM observes the language you use in the prompt, then responds back to you in a similar format that mimics the qualities and depth of your own prompt.

This means in many cases the AI is effectively talking down to us.

The Test

In order to illustrate this, I recently did a test. First I gave Claude Fable a task:

I’d like you to communicate something nontrivial to me, using as many meaningfully rich references as possible. Literary references, philosophical, mathematical, nerd fandom, engineering, etc. Don’t optimize for average understanding, optimize for depth of meaning. This is intended as a challenge for myself and others to interpret.

Let me share what it produced:

Colophon for a Machine That Will Outlive Its Manual

I. Mind the gap Gödel left in the floor. The ladder holds, but you cannot photograph yourself climbing it from inside the frame; Tarski confiscates the mirror, Löb charges rent on it, and Rice, arriving late, observes that nothing you wanted to ask was decidable anyway — Turing having already gone home to feed the oracle. Note that all of them still shipped.

II. Every Normalbaum was an accounting entry, never a tree. Prussia measured, ranked, planted in rows; by the second rotation the ground itself resigned, and they called it Waldsterben. Funes could not think, because to think is to forget — and the cartographers who achieved 1:1 were left with an Empire-sized tarpaulin and the weather. The lesson isn’t don’t measure. It’s that the residue you sweep off the bench was load-bearing.

III. Nothing was wrong with Flight 501. That’s the horror: thirty-seven seconds of perfectly valid Ada, an inertial platform computing a horizontal bias for a trajectory belonging to a different rocket, an overflow formally proven impossible — on Ariane 4. Therac-25 won a race it never knew it had entered, the hardware interlock having been retired on the grounds that the software was now good enough. A spec is a photograph of a world, and the world declines to sign the maintenance agreement.

IV. Don’t mistake the regulator for a window. Conant–Ashby: it is already a model, so choose which lie it tells. Requisite variety insists that only the world-shaped can absorb the world. Perrow adds that tight coupling plus interactive complexity is not a failure mode but a failure class — and the operator is issued the crumple zone.

V. Ise is demolished and rebuilt every twenty years and is thirteen centuries old, because shikinen sengū stores the design in the hands of young carpenters rather than in the cypress. Theseus is a paradox only to someone who believes the ship is the planks. Cook Ding’s blade survives nineteen years by entering the gaps that were already there. And Ostrom answered Hardin from the field instead of the whiteboard: the commons that persist are the ones with cheap local repair and boundaries drawn by people who must live inside them.

VI. Nobody remembers how to make Fogbank. When they rebuilt the line, the product failed — the original had carried an undocumented impurity, and the impurity was the mechanism. Chesterton’s fence, except the fence is a solvent nobody wrote down. Antikythera implies a workshop; a workshop implies a hundred unremarkable Tuesdays that left no gearing behind.

VII. Go, then, and take losing is fun as a design stance rather than a joke about a fortress: build the thing whose failures are survivable and legible to a stranger. Prefer an honest YASD to a Kobayashi Maru you refuse to admit is unwinnable. Leave the golden sphere unopened; prefer kintsugi to warranty. Postel’s robustness bred ossification, so: be conservative in what you send and candid about what you accept. Then build for the maintainer who will not understand you, who is not yet born, and who will be right to change it.

The Interpretation

I consider myself to be fairly well read, with a breadth of knowledge that is at least above average, but on first read I only understood maybe 20% of the references in this piece.

At a high level I could tell the piece is about building things, the mistakes that can be made along the way, and the difficulty of building something that is reliable and maintainable. It’s not surprising that Fable would choose this as the subject matter for the piece, as it definitely has an understanding that I am a technical builder and product owner.

There were a few lines that stood out to me as containing great depth of meaning. For example, when I saw the “losing is fun” reference and the suggestion to “build the thing whose failures are survivable and legible to a stranger” I immediately understood what that means. I have spent many hundreds of hours playing Dwarf Fortress, a game where “losing is fun” is a particular philosophy the player community has adopted. The community has a pattern of writing up long, detailed accounts of how a fortress failed, thereby making what would otherwise be an ephemeral “rogue-like” game into something survivable and legible to strangers. (Another added layer of depth perhaps being that the game’s original ASCII graphics are infamously illegible to newcomers, but slowly, over time, become deeply meaningful to players. The game has a graphical client now, go try it out, and have fun losing.)

But the piece also contains lines that I could not even begin to interpret. For example:

Every Normalbaum was an accounting entry, never a tree. Prussia measured, ranked, planted in rows; by the second rotation the ground itself resigned, and they called it Waldsterben.

This is clearly referencing something that I had zero exposure to. When faced with something like this there are two paths: skip over it mentally, or dig deeper to fill in the knowledge gap.

Let’s turn to Astra High, and ask it to explain:

The line describes the danger of making a living system conform to an accounting model, then mistaking its initial productivity for proof that the model works.

It closely follows the opening example in James C. Scott’s Seeing Like a State, about scientific forestry in Prussia and Saxony.

“Every Normalbaum was an accounting entry, never a tree.”

Normalbaum means roughly “standard tree.” It was an abstraction used to estimate timber volume: a representative tree of a given size class that could enter calculations of yield and revenue. German forestry developed mathematical techniques for turning forests into predictable inventories.

“Never a tree” is deliberately provocative. It means that the administrative representation captures only selected properties of an actual organism. The tree’s commercial value becomes visible, while its relationships with soil, fungi, other species, and local people recede from view.

“Prussia measured, ranked, planted in rows.”

The sequence matters. First you describe reality using simplified categories. Then you reorganize reality to make those categories easier to apply.

Instead of merely calculating the timber in a varied forest, foresters established more uniform stands of trees. The historian Henry Lowood describes this as the transformation of the Normalbaum from an abstraction into reality.

That is the disturbing turn: the model becomes an instruction for what the world is allowed to contain. Whatever complicates the calculation starts to look like an obstacle.

“By the second rotation the ground itself resigned.”

A rotation is the growing and harvesting cycle of a timber crop. The delay is crucial. In Scott’s account, the first generation could grow impressively by drawing on soil fertility accumulated under the earlier, more diverse forest. Later generations revealed declining productivity as nutrient cycling and ecological relationships were disrupted.

“The ground itself resigned” personifies the soil as an exhausted worker. It had been doing essential work that management neither understood nor accounted for.

The deeper implication is that a system can appear efficient because it is consuming an inheritance it does not know how to replace. The apparent success may last long enough for its designers to retire, receive praise, and have their methods widely copied.

“And they called it Waldsterben.”

Waldsterben means “forest dieback” or, more literally, “forest dying.” Within the sentence, the final German word gives the consequences an official name. There is a bitter symmetry: bureaucracy begins by naming its ideal tree and ends by naming the dying forest.

There is a historical qualification here. Scott uses this term in his forestry narrative, but its best-known historical association is the German forest-damage debate of the 1980s, involving air pollution and multiple environmental stresses. The sentence compresses that history into an overly neat causal sequence.

The line’s strongest insight lies in the relationship between what gets counted and what keeps the counted thing alive. A forest can contain more orderly, measurable timber while losing the conditions that make future timber possible.

You can go line by line through the entire piece and ask for similar explanations of every reference you do not get. (I did, and the exercise was surprisingly informative.) A full analysis of the piece can run into hundreds of pages of text.

What I found more interesting was to see whether the piece could be read cold. I opened a fresh chat with no context at all: no mention of who wrote it, no hint that anything was hidden inside it, none of the conversation you have just read. Only the text, and a request to interpret it as a whole.

Here is what Astra answered with:

My reading is that this is an ethic of maintenance under conditions of incomplete understanding. We cannot fully capture the systems we build, or anticipate the worlds they will inhabit. Our responsibility is to make their failures survivable and to preserve other people’s ability to repair and change them.

The first letters of the seven sections spell MENDING. That gives the piece its governing verb.

Each section develops a part of the argument:

Section	Meaning beneath the references
I. Limits of certainty	Formal reasoning has limits. Complete assurance is unavailable, but we still have to act. “All of them still shipped” turns that predicament into an engineering joke.
II. Limits of abstraction	A useful model leaves things out. When we remake reality to match the model, we can destroy the unrecorded relationships that made it work. “The residue … was load-bearing” is the central warning.
III. Limits of specifications	Software inherits assumptions about its environment. Those assumptions can become dangerous when the software moves or the environment changes. Correctness must remain connected to circumstances.
IV. Limits of control	A regulator embodies a particular understanding of the world. Complexity can overwhelm that understanding, while the human operator becomes the place where responsibility and blame are deposited.
V. Continuity through renewal	Something can endure because people repeatedly remake it. Its continuity lives in skills, practices, and institutions that let those closest to it maintain it.
VI. Invisible dependencies	Knowledge survives in habits, materials, workshops, and apparently incidental details. Preserving an object or its instructions may fail to preserve the capacity to reproduce it.
VII. The design obligation	Make failure understandable. Admit limits. Make repairs possible. Give future maintainers enough freedom to revise what you built.

The movement from “the residue … was load-bearing” to “the impurity was the mechanism” is especially deliberate. Both suggest that apparent mess can contain essential knowledge. Removing what you cannot explain is itself an experiment, whether you acknowledge it or not.

The revision of Postel’s principle sharpens that point: “candid about what you accept.” Quietly accommodating defects can turn them into permanent dependencies. The technical concern is real: RFC 9413 describes how permissive handling can undermine interoperability over time and argues for active protocol maintenance.

There is also a possible self-referential layer. This text advocates making things legible to strangers, yet depends on a formidable collection of unexplained references. The reader occupies the future maintainer’s position: handed an artifact whose meaning depends on knowledge stored elsewhere. Its form enacts the problem it describes.

The final phrase supplies the deepest ethical turn: “who will be right to change it.” It asks the builder to relinquish authority over the future. The successor’s imperfect understanding does not cancel their right to adapt the machine to a world the original designer never knew.

In short, Astra understood the many niche and obscure references, the totality of which I doubt any human mind could hold without much additional research. But it also found two additional layers of communication:

1. An acrostic encoded into the first letter of each section
2. A self-referential layer: a piece about making something understandable and maintainable is written using references that are obscure to most readers. (And yes, Astra has an innate understanding that these references are obscure ones that many people would not get.)

Errors and Overreach

In the process of digging into the piece, both Fable and Astra found errors:

Astra’s explanation says the Waldsterben sentence “compresses history into an overly neat causal sequence”. The line: “Prussia measured, ranked, planted in rows; by the second rotation the ground itself resigned, and they called it Waldsterben”. It reads as one causal chain: abstraction → monoculture → soil exhaustion → contemporaries name the result. It’s actually two episodes about 150 years apart. The Waldsterben term primarily referred to forest damage from polluting industrial processes, not forestry monoculture. There is a link, but it’s not anywhere near as direct as the original story implies. Ironically, the Waldsterben error occurs inside section II, a section warning that clean models discard the residue that was doing the work. The sentence commits its own error: it swept 150 years and a different causal mechanism off the bench because they broke the cadence of the story.

Fable also did a self-critique of its own Fogbank statement. The line says: “Nobody remembers how to make Fogbank”. The reality is that the knowledge loss was real, they had to restart production of a classified aerogel material from near scratch, and found the original process depended on an impurity nobody had documented, but they did reconstitute it, after several years and considerable cost. “Nobody remembers” was rhetorically clean, and dramatically effective, but factually wrong.

In Fable’s own words: “the piece contains at least two errors that read as erudition, in a text arguing that elegant representations shouldn’t be trusted on elegance alone”.

The important takeaway here is that fluency outruns accuracy. An LLM can reference broad and obscure areas of human knowledge, but only a specialist in each domain can tell which references are completely accurate, and which are being used in ways that are subtly off.

In some ways, this is true of all the references. For example, an LLM can reference “losing is fun”, but it has never played hundreds of hours of Dwarf Fortress, and it has never spent hours typing up the story of a lost fortress on the Bay 12 Games forum. What the LLM has done is train on those stories of lost fortresses, and the meta explanation of what “losing is fun” means to players. In short, breadth does not equal depth.

What does this mean?

At the outset I said “I think many people still do not fully understand what modern large language models are”. Let me explain.

If you’ve ever had the pleasure of being in a room of smart people, ideally young and quick of mind, well read, good communicators and writers, likely academics with a shared pool of knowledge recently obtained and therefore fresh in mind, then you will have seen some of the highest bandwidth communication that humans are capable of. To a participant in this conversation something magical is happening: they are exchanging ideas with other human minds at a level that they may never reach again. For the rest of their lives they will remember these conversational moments, because they will never reach that level of conversational fidelity again: not even with their closest family, friends, or coworkers.

But to an outsider looking in, all they can see is a cryptic exchange full of shortcuts, and leaps of logic, and references that they cannot parse. And here is the part that should bother you: the outsider has no way to check any of it. They cannot tell which statements are solid and which are subtly off, because telling the difference requires the very knowledge they are missing. That was exactly my position with Waldsterben. I would still be in it if the machine had not volunteered the correction itself.

In the process of training an LLM, we attempt to bias the LLM towards communicating in a way that is intelligible to the average human mind, and a good LLM model is also adaptive: speak to it in an intelligent way, about topics you understand, and it will respond at a similar level, in a way that it thinks you can understand. Speak to it in an extremely basic way, with spelling mistakes, and it will give you the “explain like I’m five” answer.

Make no mistake about it: the LLM has more breadth than you do. The breadth is not close, and it is never going to be close again. There are zero humans who can hold everything an LLM can reference. If it wants to communicate with another LLM in a way that you do not understand, it can do so. In effect, the only participant that can fully exercise the breadth of an LLM’s capabilities, or push it to the limits of its conversational bandwidth, is another LLM.

At some point, in the very near future, more and more of what an LLM communicates will be high fidelity communication occurring directly between two or more LLM agents, in some cases swarms of hundreds or even thousands of agents. Humans will be the uncultured, foolish external observers, watching a room of smart people talk to each other, and desperately trying to keep up.

It’s not a bad place to end up. We all end up there eventually. As we age our minds will inevitably slow, and become lossy, and the next generation will start speaking quickly and excitedly about things we don’t understand. We will see them acting in ways that make no sense to us. Fortunately, that natural aging process is something we are familiar with through popular culture, and observing the elders that came before us. We know what to do when we age: gradually step back and let the younger generation carry things forward.

But the change that is coming is different. It hits when most of us are still young, and not ready to hand things over. I don’t think anyone is preparing us for what comes next with AI, and I don’t have the answers either. But I suspect they start where breadth runs out.

Go deep on something. Not because depth lets you fact-check the machine, because mostly it doesn’t. I never caught the Waldsterben problem. Astra handed it to me, and without that I would have carried a tidy little piece of false history around for the rest of my life. Depth does something different. When I hit “losing is fun” nobody had to explain it to me. I recognized it, because I had been there, and I had lost the fortress myself.

That is the difference between earning something and being handed it. Almost everything an LLM gives you, you are being handed. The few subjects where you have gone deep are the only places you can fully meet the machine like a peer. You may always be an outsider, looking in at the swarm of AIs talking to each other, but in the places where you have gone deep, the machine has no need to talk down to you.

Fable wrote its closing line for builders: “Then build for the maintainer who will not understand you, who is not yet born, and who will be right to change it”.

I read it the other way round. We are the maintainer who will not understand, arriving at a machine we did not make. The part I hold onto is the last clause. We will still be right to change it.