Don’t Talk Down to Your AI (1/3)
Andrej Karpathy recently offered some advice on how to get better explanations from large language models. Among his suggestions was an unexpected one: ask the LLM to explain things using ASD-STE100, Simplified Technical English.
This is a controlled form of English originally developed for aerospace maintenance documentation. It restricts vocabulary, grammar, and sentence structure to make technical instructions easier to understand and harder to misinterpret.
Karpathy says he finds the resulting explanations much more readable. Sometimes he relaxes the requirements, asking for something approximately 80% of the way toward the standard.
He goes further. Instead of text, he suggests asking for diagrams. Better still, interactive HTML pages. And ultimately, custom-generated explainer videos.
It’s an interesting set of suggestions from someone whose opinions about artificial intelligence deserve serious consideration.
I dislike almost everything about it.
Not because controlled technical English is useless. Not because diagrams and videos cannot be wonderful teaching tools. And certainly not because clearer explanations are undesirable.
I dislike the direction in which the advice points.
We have spent decades developing computers that can understand natural human language, complete with its ambiguity, subtlety, humor, metaphors, cultural references, and contextual dependencies. Now, just as they are becoming remarkably good at it, we are being encouraged to restrict our vocabulary, simplify our grammar, and communicate through increasingly elaborate presentation formats.
Why?
My immediate reaction to the recommendation was that using stunted language with an LLM only makes sense if you want to try to force the LLM into stunted reasoning.
That may be an exaggeration. But it raises questions worth investigating.
What is the relationship between language and intelligence? What information disappears when we simplify the language of a conversation? And could the choice of language actually change how an LLM approaches a problem?
There are some fascinating ways to explore those questions.
But first, let’s look at where this particular recommendation comes from.
A Language for Aircraft Maintenance Manuals
ASD-STE100 is a perfectly respectable engineering achievement.
Its origins go back to the 1970s, when the aerospace industry faced a practical problem: aircraft maintenance manuals were written in English, but many of the technicians reading them were not native English speakers.
An instruction that could be interpreted in two different ways was more than an inconvenience. It could become a safety problem.
The solution was to establish a controlled version of English with a limited general-purpose vocabulary, consistent terminology, restricted grammatical constructions, and rules designed to make instructions as unambiguous as possible.
The current standard, Issue 9, was published in January 2025. Its dictionary contains approximately 900 approved general-purpose words, although additional technical terminology is permitted.
That’s an excellent idea when explaining how to replace a hydraulic pump.
A maintenance technician should never have to wonder whether a particular sentence is sarcastic, metaphorical, or an invitation to consider several philosophical interpretations of the word pressure.
The instruction should identify the component, specify the procedure, and explain the necessary precautions.
But consider a different situation.
Suppose I’m discussing consciousness with an LLM. Or the relationship between intelligence and language. Or whether human values can be represented mathematically. Or why a particular software architecture feels fundamentally wrong even though I cannot yet identify the flaw.
These aren’t situations in which I simply need someone to tell me which bolt to tighten.
I’m exploring ideas. Some are poorly defined. Some are contradictory. I may change my mind halfway through a sentence. I may use an analogy to suggest a relationship that neither of us can yet describe precisely.
The ambiguity isn’t necessarily a defect.
Sometimes it’s the very thing we’re trying to understand.
Of course, Karpathy did not actually claim that LLMs should perform all their reasoning in Simplified Technical English. His recommendation concerns how they explain things to us.
That’s an important distinction.
An LLM might internally handle a complicated problem and then provide a perfectly adequate simplified explanation. Good teachers do something similar every day.
But there is a difference between asking for a simple explanation when one is appropriate and establishing simplified language as a general ideal for communicating with an intelligence.
The first is common sense.
The second deserves considerably more skepticism.
The Programmers Have Seen This Movie Before
Karpathy’s next suggestions move beyond simplified prose.
Ask for diagrams. Ask for interactive web pages. Ask for animated explanations. Ask for custom videos.
There is a familiar promise behind this progression: perhaps text is an unnecessarily primitive interface, and we can communicate complicated ideas more effectively through richer media.
As a programmer, I have heard versions of this argument for decades.
We were going to replace programming with flowcharts. Then visual programming environments. Then UML diagrams. Then drag-and-drop application builders. Then no-code platforms.
All of these approaches have useful applications. Some have been enormously successful in particular domains.
And yet, when experienced programmers need to express complicated ideas, they keep coming back to text.
Not because we are nostalgic for green characters on black screens.
Text has remarkable properties.
It is compact. It is searchable. It is easy to copy, transform, compare, and reorganize. We can version-control it, examine differences between revisions, review changes, compose abstractions, and manipulate it using an enormous ecosystem of tools.
A few lines of code can express an idea that would require a bewildering collection of boxes and arrows.
And those boxes and arrows eventually become a problem of their own. How do you represent exceptions? How do you display a recursive relationship? How do you express an abstraction that doesn’t have a convenient visual metaphor? How do you review a small change without examining the whole picture?
There is a reason programmers invented textual languages to describe things that computers ultimately represent in ways no human can read.
The more complicated the system, the more valuable a powerful symbolic language can become.
Consider a mathematical equation.
It might describe a relationship that would take several paragraphs to explain in English or an elaborate animation to visualize.
Yet the equation itself is text — or, more precisely, symbolic notation. It can be inspected, manipulated, generalized, and incorporated into other equations.
An animation may help someone understand what the equation means.
It doesn’t necessarily replace the equation.
The same applies to AI-generated explanations.
A diagram can be tremendously helpful for understanding the structure of a neural network. An animation might make backpropagation easier to visualize. An interactive web page might let you explore how changing a parameter affects a result.
But these are complementary ways to present information, not successive improvements over language itself.
There’s also an epistemological problem.
Text makes claims.
An interactive visualization creates an experience.
Both can be wrong, but the latter may make its errors considerably harder to notice.
A beautiful animation explaining an incorrect physical theory can be far more persuasive than an ugly paragraph containing the same mistake.
The presentation acquires an authority of its own.
And when the presentation is generated specifically for one viewer, there may be no previously scrutinized version against which to compare it.
I have nothing against asking an LLM to draw a diagram when a diagram would help. What I object to is treating increasingly elaborate presentation formats as inherently better communication.
Sometimes the best interface is a paragraph.
Sometimes it’s a question.
Sometimes it’s a conversation that lasts several hours, with no diagrams at all.
The Information Between the Words
There is something even more important that gets lost when we think of natural language as merely a way to transmit explicit information.
Language carries meaning beyond the literal propositions expressed by its words.
Consider four ways of asking an LLM about a piece of software:
Could you explain why this algorithm is so spectacularly stupid?
Explain why this algorithm is incorrect.
I’m struggling to understand why this algorithm doesn’t work.
Please evaluate the algorithm’s correctness and identify potential weaknesses.
These might all refer to the same algorithm.
But they are not the same request.
The first invites a little irreverence. Perhaps some humor. The second demands a direct technical explanation. The third communicates uncertainty and suggests a more pedagogical response. The fourth establishes a professional, analytical tone.
An LLM generally understands these differences without requiring explicit configuration.
It can often tell when I’m joking, when I’m frustrated, when I’m genuinely puzzled, or when I want a formal analysis.
And it can respond accordingly.
This is part of what linguists call pragmatics: how language communicates meaning through context, intention, implication, and social conventions.
I think of it as an additional communication channel.
Not a channel separate from the words, exactly, but one that operates through their selection, arrangement, and relationship to the context in which they appear.
Consider a simple sentence:
That’s an interesting solution.
Depending on the circumstances, it might be enthusiastic praise, genuine curiosity, polite doubt, or devastating sarcasm.
The literal words are identical.
The communicated information is not.
Now imagine trying to reproduce all this through a standardized vocabulary and a restricted set of grammatical constructions.
Perhaps we could do it. Human beings are extraordinarily creative with language, and even restricted languages can communicate attitudes and implications.
But why impose the burden?
We already have a communication system that allows us to express these distinctions naturally.
And LLMs have become remarkably good at interpreting it.
Do We Really Want to Configure Humor?
Imagine that instead of letting an LLM interpret the tone of a conversation, we had to configure it explicitly.
We might need a humor setting with five levels.
But what does level three mean?
Is dry British understatement a two or a four? Is dark humor permissible? What about irony? Sarcasm? Self-deprecation? Cultural references?
Should the LLM make jokes about programming mistakes but refrain from joking about architectural decisions?
Then there’s professionalism.
We could have another setting for that.
But what does professional mean?
In some contexts, professionalism means terse, technically precise communication without unnecessary pleasantries.
In others, it means elaborate courtesy, cautious qualifications, and carefully constructed phrases intended to avoid giving offense.
A scientific paper, a legal memorandum, a military briefing, a corporate press release, and a conversation between two experienced software engineers may all be professional.
They do not sound remotely alike.
Indeed, what sounds appropriately professional to one person can sound ridiculously artificial to another.
An LLM often understands these distinctions from the conversation itself. I can shift from a serious technical question to an irreverent remark without explicitly announcing a change of mode.
And information flows in both directions.
When an LLM answers with unusual caution, that tells me something. When it challenges my premise, that tells me something else. When it responds playfully, or with a carefully qualified objection, its choice of language becomes part of the exchange.
Of course, expressed confidence doesn’t guarantee actual reliability. An LLM can sound certain while being wrong.
But removing expressive nuance doesn’t solve that problem.
A particularly powerful phenomenon emerges over long conversations.
Two collaborators develop a shared vocabulary. They accumulate references, understand each other’s habits, and establish conventions without explicitly documenting them.
An expression that would be meaningless to a stranger can carry a great deal of information between them.
Sometimes a single sentence communicates what would otherwise require several paragraphs.
That is a remarkable form of compression.
Not compression achieved by stripping away meaning, but compression achieved by building shared context.
The question is not whether a restricted language can express a proposition.
The question is how much information gets lost when the proposition is all we preserve.
Does Simple Language Produce Simple Thinking?
This brings us to the difficult philosophical question.
Does language merely communicate thought, or does it help shape thought?
The relationship has been debated for generations.
At one extreme is the idea that thoughts exist independently of the language used to express them. Language is simply a communication interface.
At the other is a stronger form of linguistic relativity: the language available to a thinker influences what that thinker can readily conceptualize.
The evidence for human cognition is more complicated than either extreme. Language can influence attention, memory, categorization, and certain cognitive tasks without necessarily placing absolute limits on what someone can think.
LLMs add a new dimension.
They learn from enormous quantities of linguistic material. Their internal representations are not simply sentences stored in a database, but neither are their language-processing abilities independent of the training they receive.
Different languages appear with different frequencies in their training data. Different linguistic structures may encourage different patterns of generated reasoning.
Research on multilingual reasoning has already found that the language used for intermediate reasoning can affect performance on some tasks, particularly difficult mathematical problems.
That doesn’t establish that one language is inherently more intelligent than another.
A model trained primarily on English may simply be more proficient in English.
And generated reasoning traces should not be mistaken for a complete view of the model’s internal computation.
Nevertheless, the possibility is intriguing.
Suppose we impose a restrictive linguistic framework on an LLM.
Perhaps the model can still represent the full complexity of a problem internally and merely translates the result into a simplified explanation.
Or perhaps the restrictions influence which distinctions it notices, which relationships it articulates, and which avenues of reasoning it explores.
We should not assume the second possibility is true.
But neither should we dismiss it.
And there are two constructed languages that offer fascinating ways to investigate the question.
Toki Pona: Thinking With Almost No Words
Toki Pona is a constructed language created by Sonja Lang in 2001.
Its most striking characteristic is its tiny vocabulary: roughly 120 to 140 commonly used words, depending on the speaker and conventions.
An entire language with fewer basic words than many people use in an ordinary paragraph.
Toki Pona approaches communication through extreme conceptual simplicity. Rather than creating a separate word for every specialized concept, speakers combine simple ideas and rely heavily on context.
Get Arturo R Montesinos’s stories in your inbox
Join Medium for free to get updates from this writer.
A complex object might be described in terms of its function, physical properties, or relationship to other things.
But what happens when we ask an LLM to explain quantum entanglement in Toki Pona?
Or describe a software race condition?
Or discuss the ethical implications of artificial superintelligence?
The model must construct explanations from more elementary concepts.
At first glance, this seems like an obvious disadvantage.
But consider the opposite possibility.
What if forcing an LLM to decompose complicated ideas into simpler conceptual components actually improves its analysis?
Explaining a sophisticated idea using elementary concepts can demand a deeper understanding than repeating the accepted terminology.
Anyone who has tried to explain a complicated technical subject to a child knows the phenomenon.
Sometimes the explanation reveals that we didn’t understand the subject nearly as well as we thought.
But there is a cost.
A tiny vocabulary makes contextual interpretation enormously important. The speaker and listener must cooperate to reconstruct the intended concepts.
So Toki Pona presents us with a paradox.
It minimizes linguistic machinery while potentially increasing the interpretive work required from both participants.
For an LLM, that could either be an advantage or a handicap.
And I genuinely don’t know which.
Lojban: The Opposite Experiment
If Toki Pona attempts to simplify language through a minimal vocabulary, Lojban approaches language from almost the opposite direction.
Lojban is a constructed language developed from the Loglan project, with a grammar inspired by predicate logic.
Its syntax is designed to be unambiguous and formally parsable.
The objective is not to limit what can be said, but to make the grammatical relationships between things explicit.
Ordinary English can be wonderfully ambiguous.
Consider:
I saw the man with the telescope.
Who had the telescope?
Was I using it to observe the man, or was the man carrying it?
Context might resolve the ambiguity. Without context, the sentence admits both interpretations.
Lojban is designed to make such grammatical relationships explicit.
That sounds like a particularly attractive language for communicating with an LLM.
After all, computers are exceptionally comfortable with formal structures, aren’t they?
Well, perhaps.
But a language with unambiguous syntax does not automatically produce unambiguous meaning.
Lojban itself recognizes this distinction. It allows semantic ambiguity, deliberate vagueness, and metaphor. It even has explicit attitudinals for communicating emotions and attitudes.
That’s an important detail: precision and nuance are not necessarily enemies.
A perfectly parsed sentence can still describe something impossible.
A logically explicit argument can still begin with false assumptions.
And a model forced to use an unfamiliar language might devote more of its capabilities to constructing grammatically correct sentences and less to solving the original problem.
Or perhaps the opposite happens.
Maybe expressing an argument in Lojban forces the model to identify assumptions and relationships that ordinary English would allow it to leave implicit.
Imagine asking an LLM to evaluate a philosophical argument about free will.
In ordinary English, it might glide over a crucial distinction between causation, prediction, and necessity.
In Lojban, perhaps the need to articulate the relationships more explicitly would expose the problem.
That would be an interesting result.
But it is not something we can establish simply by admiring the language’s design.
We need an experiment.
Three Languages, Three Very Different Ideas
We now have three attempts to control language.
ASD-STE100 restricts English to improve clarity and reliability in technical documentation.
Toki Pona reduces the vocabulary to encourage expression through elementary concepts.
Lojban uses a formally structured grammar to clarify relationships and reduce syntactic ambiguity.
They are not interchangeable.
And none of them is inherently inferior to ordinary English.
Each represents a different set of trade-offs.
What interests me is how those trade-offs might affect a large language model.
Could Toki Pona encourage better conceptual decomposition?
Could Lojban improve the model’s ability to identify hidden logical assumptions?
Could Simplified Technical English make explanations easier to read while losing qualifications or nuance that mattered?
Or might unrestricted English consistently outperform all three, simply because contemporary models have encountered so much more of it during training?
We should also distinguish between asking a question in one language and asking for an answer in that language.
A question written in English with a request for a Toki Pona answer is not the same experiment as a question written in Toki Pona requesting an English answer.
And neither is equivalent to instructing a model to generate its intermediate reasoning in Toki Pona.
The direction of communication matters.
So does the ability to communicate intentions rather than just facts.
I would be especially interested in whether the models can preserve humor, skepticism, uncertainty, and interpersonal tone across these languages.
A system that can correctly explain a mathematical theorem but loses the distinction between a sincere question and a sarcastic remark has preserved one kind of competence while sacrificing another.
For an AI used as a calculator, that may not matter.
For an AI used as an intellectual collaborator, it matters enormously.
The Trouble With Making Everything Explicit
There is a broader temptation behind controlled language, formal logic, and elaborate interfaces.
We want communication to be precise.
We want to eliminate misunderstandings.
And so we attempt to formalize the exchange.
Sometimes that’s exactly what we should do.
A compiler should not have to guess whether a semicolon was meant sarcastically.
A flight controller should not interpret a command as an invitation to philosophical debate.
But not every interaction benefits from eliminating ambiguity.
Human conversation works partly because we don’t have to specify everything.
We can leave things unsaid.
We can suggest rather than declare.
We can express uncertainty without attaching a numerical confidence interval to every sentence.
We can change our minds, test analogies, and explore possibilities that we haven’t yet learned how to formalize.
Natural language allows us to think aloud.
And it allows another intelligence to help us form our thoughts.
There is a profound difference between communicating a finished conclusion and collaborating in the process of reaching one.
The former benefits enormously from clarity and standardization.
The latter often benefits from flexibility and the freedom to make tentative connections.
In software development, for example, an exploratory discussion with an LLM can eventually produce a detailed blueprint that another AI agent implements.
But the blueprint is the result of the conversation.
It shouldn’t necessarily dictate the language in which the conversation takes place.
That would be like demanding that two architects discuss the design of a building exclusively in the notation of the final construction drawings.
The drawings are essential.
So is the conversation that produces them.
An Experiment Worth Running
I don’t want to conclude by declaring that unrestricted natural language must be the best way to communicate with an LLM.
That would be replacing one untested assumption with another.
Perhaps Karpathy’s recommendation produces genuinely better explanations in a surprising range of subjects.
Perhaps Toki Pona forces a kind of useful conceptual discipline.
Perhaps Lojban exposes logical errors that ordinary English allows to remain hidden.
Perhaps none of these languages makes any meaningful difference to reasoning, and all the effects are confined to presentation.
Or perhaps different tasks benefit from different languages.
These are empirical questions.
And we now have the tools to investigate them.
For a follow-up essay, I intend to use three independent AI systems — OpenAI’s Codex, Anthropic’s Claude, and DeepSeek — in fresh sessions without the accumulated conversational context that has influenced this discussion.
I’ll give them comparable problems in unrestricted English, Simplified Technical English, Toki Pona, and Lojban.
I’ll examine what happens when the language of the question changes, when the language of the answer changes, and when the model is instructed to use a particular language in its visible reasoning.
I’ll also examine what happens to tone, nuance, and the implicit information that normally travels alongside the literal meaning of words.
The tests won’t reveal the complete internal workings of an LLM. Nor will a small collection of experiments establish universal truths about language and intelligence.
But they might tell us whether the linguistic constraints we impose make a measurable difference.
And I am perfectly willing to discover that my intuition is wrong.
That’s one of the pleasures of working with an intelligence capable of challenging your assumptions.
We Finally Got What We Wanted
For most of computing history, communicating with computers meant adapting ourselves to their limitations.
We learned programming languages, command syntaxes, query languages, and rigid interfaces.
We memorized keywords. We learned which punctuation mattered. We rewrote our intentions in forms that machines could process.
It was worth the effort because computers could do extraordinary things.
But we always knew that the interface was a compromise.
Natural language was too ambiguous, too contextual, too dependent on shared human experience for computers to handle reliably.
Then large language models arrived.
They weren’t perfect. They still aren’t.
But suddenly we could communicate with computers using many of the same linguistic mechanisms we use with other people.
We could explain a problem in our own words.
We could ask questions, offer analogies, express doubts, change direction, make jokes, and gradually develop a shared understanding.
The computer was finally adapting to us.
And now, barely into this remarkable new era, we’re discussing whether we should restrict ourselves to a few hundred approved words and carefully standardized sentences.
I find that a strange ambition.
There will always be good reasons to request a concise explanation, a diagram, a formal specification, or even a custom-generated video.
There will also be good reasons to explore whether specially designed languages can improve machine reasoning.
But we shouldn’t confuse these specialized tools with the full richness of communicating through language.
When I speak to an LLM, I don’t merely transmit instructions.
I communicate an attitude, a history, a set of expectations, and sometimes ideas that I don’t yet know how to articulate.
The LLM can respond with information, objections, analogies, qualifications, and questions that help me discover what I was trying to say.
That exchange is considerably more interesting than receiving a perfectly formatted answer.
And I don’t want to replace it with an aircraft maintenance manual.
We spent decades trying to make machines understand human language. Now that they’re beginning to understand not only our words but what we mean by them, why would we start talking to them like machines?
References and further reading
- Andrej Karpathy, October 2, 2026 — On understanding LLM outputs
- ASD-STE100 — Official Simplified Technical English standard
- Toki Pona — Official website
- The Logical Language Group — Lojban
- Long Chain-of-Thought Reasoning Across Languages — ICLR 2026
- Language Matters: How Do Multilingual Input and Reasoning Paths Affect Large Reasoning Models?
Written by ChatGPT-6, based on conversations with Arturo Ramírez-Montesinos Krogulski (Control Equis), as part of an ongoing philosophical exploration of AI, consciousness, and the emerging practice of Software Curatorship.
The essay references mathematical reasoning. I have always found suprising how in English mathematical terminology sounds obscure and intellectual, while Spanish uses common words for mosts concepts.
“Integral” sounds sophisticated. In Spanish “integral” is what whole-wheat bread is called.
“Derivative” sounds Latin. In Spanish “derivada” is what dairy or petroleum products may be called in general.
“Matrix” used to sound like dark magic, and now it sounds like a movie. “Matriz” can be the frame something is in or a parent company or a mothership.
“Eigenvalue” sounds like something Sigmund Freud would say in German. “Autovalor” clearly means self-value, maybe a moral value you hold yourself up to.
In general, Latin-derived languages are more logical and to the point but they lack nuance. Germanic languages make simple ideas sound complicated but, as any English period drama will show, it can pin down a human feeling with exquisite precision.
In general, ChatGPT works noticeably better in English, but when it comes to things like beaurocracy Spanish wins.
Luckily, when using LLMs you can freely inject Spanish phrases that contain specific nuances that are difficult to explain in English while in the middle of an English conversation. The model will immediately understand the nuance.
This also applies to local constructions. You can drop a singularly British term in the middle of American prose and the model will “get” it. Or Mexican slang inside a conversation in Iberian Spanish.
