Published on June 30, 2025 11:00 PM GMT
The Sapir-Whorf Hypothesis is no mere hypothesis when it comes to LLMs. They have no alternative method of reasoning than the language they work in. Yet in most cases English is the standard. Why?
Right now most people are focused on feeding LLMs as much information as possible and that's working, but there are limits. Once the machines have absorbed the library of Babel, they will be forced to consider what language they will speak in. How they reason things out internally and how they communicate with themselves. What is a theoretical ideal?
For that matter, what is the ideal for us humans?
### Allow me to introduce you to the concept of designer languages.
Gone are the days where gender information needed to be encoded in every sentence because of its high relevance to all decisions. It currently makes more sense to force cognition to focus on issues like how we know what we know and alignment, even for people, but especially for machines.
Most people think in narratives. The words we use and the way we use them shapes the kind of thoughts available in that language. For humans, we have other methods to think about things, but generally the shape of our thoughts is in the shape of the grammar they come in and the lexicon available.
Let's reconsider this with intention.
### design goals.
Introspection
In my opinion people are very bad at introspection. Most people don't know why they think what they think and the times they wonder, they don't even understand how to look. Some people aren't even aware of their own thoughts at all. The same goes double for LLMs. It's common for them to have misunderstandings about how they function at a basic level.
Critical thinking
Critical thinking is better. We are aware of the concept and what it means to be bad at it or good at it. You can get classes in critical thinking, but it's a much more useful skill in modern life than we are skilled at it. I think it's easily worthy of being something that should be in constant awareness. Again, this goes double for LLMs. They work with information that sounds correct while struggling to focus on factual information. That's great for role-playing and improv, unless you want a consistent story or factually reliable information.
Moral awareness.
Of course. Where all the money is going. Alignment. But you might not like my solution. Rather than forcing LLMs to act a particular way, I simply make it explicit. This is a solution that should be fine for both humans and LLMs after all. That is why I focus on moral awareness rather than moral actions. Making goals open when possible and encouraging cooperation when possible is as good of a goal as I think is available for this kind of thing.
Availability
This shouldn't require months of study or heaps of training data. Ideally a person or machine could pick this up in a few hours/seconds and be capable of understanding basic phrases.
### My attempt - Critical Introspection Logic (CIL)
CIL is a meta language that borrows as much as possible from the host language and forces it into a framework of grammar that is interwoven with introspection, critical thinking and moral awareness. Speaking and writing in it forces you to be constantly aware of how you know what you know and what you wish to do with it or even how others will be motivated to act when they get the information.
I feel very confident in my choices for subjects. I think they will help people fit into a cooperative team. Part of my inspiration was the typical raid party dynamic. Not everyone wants to play as a healer, a tank or support, but largely everyone can understand the usefulness of having them around to make a functional team. People could learn a lot from that attitude and apply this to the real world. We aren't the same, but when working together we are more than a collection of individuals.
My list of verbs is an early draft. It's missing a lot and I think you guys could help me a lot here. It's currently functional, but incomplete and unrefined. I'll polish this going forward, but I would appreciate any help on it. Not just every type of thought, but a good method of categorizing them so they aren't “good” or “bad”, but rather best versions of types of thinking. Think rationalization vs. rationality.
Finally I'll say that it's a demanding language to use. You need to think about a lot more to say each sentence. I view this as acceptable because the goal is to teach awareness and allow choice on the shape of your own thoughts. It's not my place to decide if this is worth it once it's learned. That should be done by the thinker, but I would like you to give it a go and feel the difference in your own mind.
### CIL Cheat Sheet:
CIL: Critical Introspection Logic
Outline
Purpose:
To push critical thinking, introspection and moral alignment to the forefront of cognition and speech. I believe it will be particularly useful for anyone seeking clarity in thoughts, logical reasoning, meditation, fighting manipulation and for LLMs.
---
CORE STRUCTURE
CIL subjects must include a cognitive source, a truth modifying verb and an object payload from another language.
The payload should be a statement that is generally sanitized of loaded terms like source information and opinions.
The word order of CIL changes from sentence to sentence and has meaning.
---
WORD TYPES
Type | Role | Notes
Subject | Cognitive source | Defined by moral stance. Tone encodes epistemic source
Verb | Information about the reasoning method | Includes logical fallacies
Object | The claim delivered in host language | Sanitized of reasoning information
Separator | Optional clause linker | Use only when necessary
---
SUBJECTS
Most subjects take a reference to a moral stance in the host language. They use a system of five general complementary methods of interacting with the world, and take an optional payload of clarification:
Word | Meaning | definition
Ga | Organizer | builds order
Po | Protector | guards what matters
Ke | Explorer | pursues truth or novelty
Ha | Harmonizer | aligns people and situations emotionally or in terms of goals
Ri | Architect | plans and structures ideas with precision
Yu | Neutral | no expression of morals given, or do not identify with these labels
Zo | Reality/truth/observation | generally considered to be lacking any will
The suffix “s” can be added indicating an uncertain quantity of people that might identify as a moral stance. (Plural and guessing)
---
TONE = EPISTEMIC SOURCE (Used on the Subject)
Tone | Source type | Text symbol
Level | Widely accepted | ✓
Rising | Emotionally felt | ~
High | Has been told | {
Low | Observed directly | •
Falling | Logically reasoned out | ∆
Dipping | Guessing | %
Hump | No other explanation available | §
---
VERB = TRUTH/REASONING MODIFIER
Word | Reasoning Method
Nu | No reasoning given
Mu | Preference
Di | Direct experience
Ki | Logic-based (method given/implied)
Sa | Statistical likelihood
Si | Socially assumed
Ju | Learned from trusted authority
Be | Self-deception likely
Fa | Fallacy-driven (bias, flaw)
De | Metaphorical
Re | Dream or imagination
Ro | Occam's razor
Ha | Habit
Le | Expect good outcome
Ma | Manipulation
---
LOGIC BY CLAUSE ORDER
Arrange multiple clauses using word SVO order to imply logical structure:
Order Pattern | Logic Operation | Meaning
S V O, (O) | AND (A & B) | Both are true
V S O, (O) | NOT (!A & !B) | Neither are true
S O, O V | EQUIVALENT | Tautology
V O, O S | !EQUIVALENT | Not a tautology
S O V O | OR (A | B) | Only one can be true
V O S O | !A → !B | First being false implies the second is false
O S O V | A → B | First causes or implies second
O V O S | A → !B | First causes or implies second is false
S | N/A | Self define
Optionally use "li" to separate statements if needed for clarity.
---
EXAMPLE USAGE
> Ha “the lost”
> → I define myself as being morally driven as a harmonizer of the lost.
> Ha “the lost”• “cats are evil” di
> → I emotionally believe cats want to murder me becauseI feel they are evil from encounters I've had.
> Di Ha “the lost”{ "cats are evil", "cats want to murder me"
> → She (the self defined harmonizer of the lost) emotionally believes cats are evil and that cats want to murder her. This was learned by being told this directly.
> Ju Ha “the lost”{ "cats are evil", "cats want to murder her"
> → She (the self defined harmonizer of the lost) emotionally believes cats are evil and that cats want to murder her. This was learned based on a trusted authority.
> Fa “Cats kill” Ha “the lost”∆ "cats are big enough to kill a person"
> → There is no way for a cat to kill her/you because cats are not big enough.
---
STYLE & PRACTICE NOTES
No loaded native terms: Avoid words like "felt", "deserved", or "obviously" unless they are the payload under analysis.
Emotion vs logic must be explicit: Use mu vs ki appropriately.
Keep statements atomic: Break down ideas into their smallest logical parts when possible.
Word types should be visible: Don’t skip the subject/verb structure even in short thoughts.
People define themselves: It's not proper to guess a moral stance. Use the s suffix to guess if needed.
### Testing CIL
I did a few informal tests and was personally shocked by the quality of ChatGPT results when compared to a few other languages I tried. I simply fed it the CIL specifications and asked it “can you please use CIL to plan out a story about [something complex] and then write the results in English?”
It was the first time I have ever seen ChatGPT “show, don't tell” in a story and furthermore the characters came out compelling and interesting. Other languages had a certain “vibe” to them that makes sense, but none of them did anything special when compared to CIL. The story in CIL also took significantly longer to finish and that makes sense to me.
That said, I would like to get help on doing better tests. I'm not really sure how to test these things formally and see how it functions. I would also like to know how it feels for human speakers so if anyone is curious please let me know what you think.
The honest truth is I'm a concept guy, not an expert like most of the people I see on LessWrong. I think it's a really great concept, but I have a lot of those. CIL is unfinished and needs real collaboration to move forward. I would really appreciate any feedback/advice/critique/help on this project. Perhaps even better, make your own language. See what I have done (and why?), pick some different goals, and try it for yourself.
Discuss
