Da7em

Tech Expert & Content Creator | AI & Open Source | Ambassador @cognition | DMs open for collabs. 📧 [email protected] | Support me:

[Da7em @Da7_Tech]

All this in just the last week. Insane.

[Da7em @Da7_Tech]

I can't believe I hit 20M monthly views.

As this community grows, I promise to give back big.

I'll back your projects and plans with funding and mentorship.

Big things are coming.

[Da7em @Da7_Tech]

@zyh6036 I don't think so.

[Da7em @Da7_Tech]

Kimi's the only Chinese lab that actually got to Fable level.

So yeah, I've got high hopes their next model hits Opus 5.5, maybe even beats it.

[Da7em @Da7_Tech]

I don't care about benchmarks anymore.

Don't trust them at all, they mean nothing to me.

[Da7em @Da7_Tech]

@Jack_BigBigha Hi, not yet.

[Da7em @Da7_Tech]

It's coming.

The best open model in the world.

[Da7em @Da7_Tech]

@PosaniRishi Wait for it

[Da7em @Da7_Tech]

@qq1082 Absolutely

[Da7em @Da7_Tech]

@henkidama I think so.

[Da7em @Da7_Tech]

@knowixbuilds They did it before.

[Da7em @Da7_Tech]

@zhodonx Absolutely

[Da7em @Da7_Tech]

@florin_dev No

[Da7em @Da7_Tech]

Your own example makes my point.

Look at what the forms actually do to s-l-m, whose core is safety and wholeness:

salima: to be safe.
sallama (Form II, causative): to make safe, to deliver something safely, to hand it over. Greeting someone is literally saying salām to them, wishing them safety.
tasallama (Form V, the reflexive of II): to be handed something, to receive it. II is handing over, V is the other side of that same handing.
aslama (Form IV, causative): to hand yourself over. That's submission.
istaslama (Form X): to give yourself up. That's surrender.

None of that is random. The root gives the field, the form gives the relation. The meanings aren't fully derivable, but they're narrowed to a small, predictable space, and the relation between II and V holds across thousands of roots. Learn it once and every root you meet reuses it.

Now do English: safe, greet, hand over, receive, submit, surrender. Six unrelated words, six arbitrary mappings, and nothing on the surface tells you they belong together. That's not one go/went. That's the whole vocabulary working like go/went.

You ended with: Arabic doesn't remove memorization, it organizes it. Exactly. Organized memorization is compression, and compression is how a model learns. The question was never zero memorization. It's how much each example teaches you about the next one. In English, learning "surrender" teaches you nothing about "receive." In Arabic, it teaches you about every Form X verb you haven't seen yet.

Weak and hamzated roots follow sound rules that apply across all of them. Which plural pattern a noun takes is chosen from a closed set, and the root survives inside it. That's bounded uncertainty, not open-ended memorization.

So yes, you still have to learn things. But you learn them inside a system that pays you back with every new word.

[Da7em @Da7_Tech]

Since a lot of you asked, let me explain what I meant.

AI models today need insane amounts of data and compute to learn language, and I think a big part of that cost comes from training them on languages that hide their own structure, like English. Chinese has the same problem, which is why US and Chinese labs are stuck in the same race for more data and bigger clusters. English is full of irregular forms like go/went and mouse/mice that follow no rule and have to be memorized. Its spelling doesn't match its sound, as in through, though, and tough. Related meanings share no visible link, so write, book, and office look like strangers. Relationships between words are hidden in word order, where moving one word can flip the meaning. Phrasal verbs like get up, get over, and get by can't be understood from their parts, words like set and run carry dozens of meanings the model has to guess between, and the history is broken, since Old English is unreadable today.

The fix is to train the core model on a language where the structure is explicit, so it learns more from less, and then add other languages on top so it can speak with anyone. That language is fully vowelized Arabic. Its root system means kataba, kaatib, maktoob, and maktaba all grow from one root, so the model learns a whole family of meanings at once. Its regular patterns let the model infer words it has never seen. Case endings mark every word's role on the word itself, so relationships are written instead of guessed, and because meaning holds when word order changes, the model learns meaning rather than position. New words are derived by clear rules instead of random borrowing. Texts from the 7th century are still readable today, giving 1,400 years of continuity, and they come with an annotated heritage of commentaries, lexicons, and formal reasoning methods, meaning the model gets the text plus how to understand it.

Nobody has seriously tested this yet. Personally, I'm aiming to train a model this way once I have the resources.

[Da7em @Da7_Tech]

What if AI's biggest bottleneck isn't compute or data, but English itself?

[Da7em @Da7_Tech]

@ZryMiller In short, it's a mathematical language with strong roots, so a model can make better predictions with it.

[Da7em @Da7_Tech]

It's coming.

The best open model in the world.

[Da7em @Da7_Tech]

I want to get into local models.

[Da7em @Da7_Tech]

Since a lot of you asked, let me explain what I meant.

AI models today need insane amounts of data and compute to learn language, and I think a big part of that cost comes from training them on languages that hide their own structure, like English. Chinese has the same problem, which is why US and Chinese labs are stuck in the same race for more data and bigger clusters. English is full of irregular forms like go/went and mouse/mice that follow no rule and have to be memorized. Its spelling doesn't match its sound, as in through, though, and tough. Related meanings share no visible link, so write, book, and office look like strangers. Relationships between words are hidden in word order, where moving one word can flip the meaning. Phrasal verbs like get up, get over, and get by can't be understood from their parts, words like set and run carry dozens of meanings the model has to guess between, and the history is broken, since Old English is unreadable today.

The fix is to train the core model on a language where the structure is explicit, so it learns more from less, and then add other languages on top so it can speak with anyone. That language is fully vowelized Arabic. Its root system means kataba, kaatib, maktoob, and maktaba all grow from one root, so the model learns a whole family of meanings at once. Its regular patterns let the model infer words it has never seen. Case endings mark every word's role on the word itself, so relationships are written instead of guessed, and because meaning holds when word order changes, the model learns meaning rather than position. New words are derived by clear rules instead of random borrowing. Texts from the 7th century are still readable today, giving 1,400 years of continuity, and they come with an annotated heritage of commentaries, lexicons, and formal reasoning methods, meaning the model gets the text plus how to understand it.

Nobody has seriously tested this yet. Personally, I'm aiming to train a model this way once I have the resources.

[Da7em @Da7_Tech]

Kimi is the only Chinese lab that can deliver Opus 5.5-level performance.

[Da7em @Da7_Tech]

the irregularity comparison doesn't hold. Arabic broken plurals and weak verbs follow a limited set of patterns and sound rules. go/went follows nothing. A model generalizes the first and memorizes the second, and memorization is where the data and compute go.

On tokenizers: a tokenizer can only expose structure the language already has. Arabic decomposes cleanly into root and pattern. No tokenizer can recover go from went. The language supplies the structure, the tokenizer reveals it. You need both.

More rules isn't the problem. Unpredictable rules are.

On data: the goal is a core model trained on less but cleaner text, not one that competes with the whole internet on volume. Vowelized classical and educational text exists, automatic diacritization keeps improving, and we already know data quality can offset a lot of quantity.

Why not a constructed language? Because none of them carry centuries of real human thought. Arabic is near-regular and comes with 1,400 years of continuous, annotated text.

[Da7em @Da7_Tech]

@DemetreMinchev This is really interesting.

[Da7em @Da7_Tech]

@neojohnson113 No custom language will unleash its potential more than Arabic will.

[Da7em @Da7_Tech]

@GeorgieW1776 That's what I'm talking about. Arabic can fix the problem with English because it works on precise algorithms.