Season 1 Episode Bank: Creative AI x Mandinka x N'Ko

Season 1 Episode Bank: Creative AI x Mandinka x N'Ko

These are camera-first scripts. They are meant to sound like a person explaining an unusual creative process, not a researcher defending a benchmark. The technical details are still there, but they enter as proof only after the viewer understands the personal and creative reason for the work.

Episode 1: I Used AI To Study My Native Language

Length: 45-60 seconds

Hook: I used AI to study my native language, and it exposed something weird about modern AI.

Camera Script:

Most people use AI to write faster, make images, or automate work.

I ended up using it in a different way.

Mandinka is my native language, and I started looking deeper into N'Ko, the script designed for Manding languages. At first, it was personal. I wanted to understand the structure. The sounds. The characters. Why the script works the way it works.

Then the question changed.

If I am learning this structure, can a machine learn it too?

Not just show the characters on screen. Not just translate around it. Actually represent it. Actually hear Mandinka and write it in the script built for it.

That became the project.

I used AI like a microscope, pointed it at my own language, and started asking what modern AI still cannot see.

On-screen text: "I used AI like a microscope."

Visual: Face to camera, then quick flashes: N'Ko characters, waveform, model layers, notebook.

Caption: This started as learning. It turned into an AI infrastructure question.

Episode 2: The Script Built For The Sound

Length: 45-60 seconds

Hook: This script was not made to look exotic. It was made to fit the language.

Camera Script:

N'Ko was created by Solomana Kante in 1949 for Manding languages like Mandinka, Bambara, Maninka, and Dioula.

What makes it powerful is that it was designed around the sound of the language.

That matters because a writing system is not neutral. If the script leaves out tone, or forces one sound into multiple letters, or borrows rules from another language, then a machine learning system has to learn around that.

N'Ko does something different.

It gives the language a cleaner sound-to-character structure.

And once I saw that, I stopped thinking of the script as only cultural. I started thinking of it as infrastructure.

Because if AI is supposed to hear us, read us, and build tools for us, then the script it targets changes everything.

On-screen text: "A script can be infrastructure."

Visual: N'Ko word on screen, then "sound -> character", then "script = label space."

Caption: The output script changes what the machine is trying to learn.

Episode 3: AI Can Display It. That Does Not Mean It Understands It.

Length: 45-60 seconds

Hook: A language can be visible on your phone and still invisible to the model.

Camera Script:

This is one of the biggest things this project taught me.

When people see a script on screen, they think the technology supports it.

But display is not understanding.

Your phone can render N'Ko characters. A model can copy them. A website can store them. That only means the script exists in Unicode.

It does not mean the AI has strong internal patterns for it.

So I started looking inside the model. Layer by layer. What happens when it sees N'Ko? Does it form a clean representation? Does it behave like it behaves with English?

The answer was uncomfortable.

The model could see the script, but the internal signal was weaker.

That is when the project became bigger than one model. It became about the difference between being included on the screen and being included inside the machine.

On-screen text: "Unicode != understanding"

Visual: N'Ko rendered cleanly, then a layer diagram with weak signal.

Caption: Display support is not the same as model support.

Episode 4: I Used AI Like A Microscope

Length: 60 seconds

Hook: I did not just ask AI questions. I used AI to inspect AI.

Camera Script:

One of the most creative parts of this project was using AI like a microscope.

I took language examples and watched what happened inside the model as it processed them.

Think of it like a brain scan.

Not in the human sense, but in the machine sense: layer by layer, activation by activation, you can see whether the model is forming strong patterns or weak ones.

That changed how I thought about AI.

It is not just a box you prompt.

It is something you can interrogate. You can ask: where is the signal strong? Where does it fall apart? Which scripts have real pathways inside the model, and which ones are only barely represented?

That is how my Mandinka/N'Ko project became a different kind of AI project.

It was not about using AI to replace thought.

It was about using AI to find where the machine world had not learned enough.

On-screen text: "Prompting AI is one use. Inspecting AI is another."

Visual: Model layer stack, activation heatmap, N'Ko/English comparison card without heavy numbers.

Caption: Creative AI is not only generation. It can be investigation.

Episode 5: I Started Learning N'Ko And The Bugs Taught Me

Length: 60-75 seconds

Hook: The bugs were not random. They were showing me where the language was being squeezed into the wrong container.

Camera Script:

When I started building around N'Ko, I kept running into conversion problems.

A lot of the available data for Mandinka and Bambara is written in Latin script. So if I wanted the machine to work in N'Ko, I had to build a bridge from Latin into the sound structure, then into N'Ko.

At first, the bridge felt like preprocessing.

Then I realized the bugs were teaching me.

Every failure pointed to something important. A sound that Latin represented ambiguously. A tone that was missing. A spelling convention that depended on the transcriber.

That is when it clicked.

The script is not just how the language looks. It decides what information the machine receives.

Learning N'Ko was teaching me what the AI pipeline was missing.

On-screen text: "The bugs were evidence."

Visual: Latin -> warning signs -> N'Ko bridge, with "tone?", "digraph?", "sound?" labels.

Caption: Sometimes the error messages are the lesson.

Episode 6: Latin Is Not Neutral

Length: 45-60 seconds

Hook: Latin script is useful. But it is not neutral.

Camera Script:

This is not an anti-Latin argument.

Latin script is useful. People use it. Data exists in it. A lot of systems depend on it.

But for Manding languages, Latin was not designed from the inside of the sound system.

So when an AI model outputs Latin Bambara or Latin Mandinka, it is not only solving speech recognition. It is also solving a spelling-convention problem.

That matters.

Because if the model hears a sound correctly, but the writing system gives multiple ways to represent it, the error may not be purely acoustic.

N'Ko gives us another path.

It lets the machine aim closer to the sound structure.

That is why the script question is not cosmetic. It changes the machine-learning problem.

On-screen text: "The script changes the task."

Visual: Split screen: "sound problem" versus "spelling problem."

Caption: The output script is part of the model.

Episode 7: I Tried To Make The Machine Listen In N'Ko

Length: 45-60 seconds

Hook: I did not want the AI to hear Mandinka and then write around N'Ko. I wanted it to write N'Ko directly.

Camera Script:

The normal path for speech AI is: audio goes in, Latin text comes out.

But for this project, that felt like skipping the point.

If N'Ko was built for the sound of Manding languages, then I wanted the system to output N'Ko directly.

So the pipeline became: take the audio, extract speech features, and train a decoder that writes N'Ko characters.

That sounds technical, but the idea is simple.

Do not make the language pass through a script that was not designed for it if the script built for it already exists.

Make the machine meet the language where it is.

That was the goal.

On-screen text: "Make the machine meet the language."

Visual: Audio waveform -> feature stream -> N'Ko output, no acronym labels.

Caption: Script-native AI means the target is the script from the beginning.

Episode 8: This Is A Different Kind Of AI Creativity

Length: 45-60 seconds

Hook: AI creativity is not only making something pretty. Sometimes it is asking a question nobody around you is asking.

Camera Script:

When people talk about creative AI, they usually mean images, music, videos, logos, captions.

That is real.

But this project made me see another version.

Creative AI can mean using the tools to investigate something personal and underbuilt.

For me, that was Mandinka and N'Ko.

I used AI to scan model behavior, build language bridges, test speech recognition, and think through what real support would mean.

That is creative because it changes the question.

Instead of asking, "What can AI make for me?"

I started asking, "What can AI reveal about what technology has ignored?"

That is a completely different relationship with the tool.

On-screen text: "AI as investigation, not shortcut."

Visual: Fast montage: prompt box crossed out, microscope icon, N'Ko script, waveform, notebook.

Caption: Creative AI can be cultural infrastructure work.

Episode 9: The Number Comes Later

Length: 45-60 seconds

Hook: There is a benchmark number in this project, but that is not where the story starts.

Camera Script:

At some point, yes, the project produced a serious speech-recognition result.

The strongest archived checkpoint reported 20.57

But if I start there, most people miss the point.

The point is not just "look at this number."

The point is: why does the number matter?

What was the model trying to write? Why N'Ko? Why not Latin? What does the metric measure? What does it mean for AI to support a native language beyond just displaying characters?

That is why I am telling the story from the beginning.

The number is a receipt.

The journey is the argument.

On-screen text: "The number is a receipt. The journey is the argument."

Visual: 20.57 card appears briefly, then moves behind "Mandinka -> N'Ko -> AI infrastructure."

Caption: I will explain the score later. First, the story has to make sense.

Episode 10: My Native Language Became A Machine Question

Length: 60 seconds

Hook: I did not expect my native language to turn into a question about machine infrastructure.

Camera Script:

Mandinka being my native language changes how this feels.

This is not just "low-resource NLP" to me.

It is not just a dataset.

It is a language connected to family, memory, sound, culture, and identity.

So when I started learning N'Ko and then testing AI systems around it, the question became more serious.

What does it mean if the future of AI can handle English fluently, generate perfect marketing copy, and make cinematic images, but still barely understands the script built for my language family?

That gap is not small.

It affects who gets voice tools. Who gets searchable archives. Who gets educational software. Who gets to use AI without translating themselves first.

That is why I care about this.

It is personal, but it is also structural.

On-screen text: "Personal language. Structural problem."

Visual: Face to camera, family/language note aesthetic, then infrastructure stack.

Caption: Native language is not a niche dataset.

Episode 11: The Script Changed How I Heard The Language

Length: 45-60 seconds

Hook: Learning N'Ko made me listen differently.

Camera Script:

When you learn a script built around sound, you start hearing details differently.

You notice what a character is trying to preserve.

You notice where tone matters.

You notice when Latin spelling is doing something indirect.

That changed the AI question for me.

Because a speech model is also listening for structure.

If the target script preserves more of that structure, then the model is not learning the same task anymore.

It is not just "write the transcript."

It is: choose a representation of the speech.

And N'Ko made that visible to me.

On-screen text: "Learning the script changed how I heard."

Visual: Ear icon, tone marks, waveform, character appearing on beat.

Caption: The script is a listening tool.

Episode 12: I Am Not Trying To Make AI Sound Deep

Length: 45 seconds

Hook: I am not saying AI is magic. I am saying most people are using it too narrowly.

Camera Script:

I do not think AI is magic.

I do not think it automatically understands culture.

Actually, this project showed the opposite.

AI can be extremely powerful and still have huge blind spots.

But that is why the tool is interesting.

If you use it carefully, you can test those blind spots. You can inspect the system. You can build small bridges. You can ask whether the machine actually represents what it claims to support.

That is the creative part to me.

Not pretending AI knows everything.

Using AI to find out exactly what it does not know.

On-screen text: "AI is powerful. AI is incomplete."

Visual: Two-column card: "not magic" / "still useful."

Caption: The creative use is not blind trust. It is interrogation.

Episode 13: What Real AI Support Would Mean

Length: 60 seconds

Hook: Real AI support for a language is more than being able to type it.

Camera Script:

What would it mean for AI to really support Mandinka through N'Ko?

It would mean the model can represent the script internally.

It would mean speech systems can output N'Ko directly.

It would mean the metrics respect the sound structure instead of only measuring spelling convention.

It would mean correction systems do not erase uncertainty just to make text look fluent.

And it would mean people can build tools in the language without constantly translating through English or French first.

That is the bigger vision.

Typing the characters is step one.

Real support is infrastructure.

On-screen text: "Typing is step one. Infrastructure is support."

Visual: Stack: keyboard -> model representation -> speech -> correction -> education/search/archive.

Caption: Language support has layers.

Episode 14: The AI Did Not Fail Because N'Ko Is Weak

Length: 45-60 seconds

Hook: N'Ko is not the weak part of the system.

Camera Script:

When AI struggles with a language or script, people sometimes assume the language is the problem.

That is backwards.

N'Ko is precise.

It was designed around the sound system. It marks tone. It gives Manding languages a structure that Latin was not built to provide.

The weakness is not the script.

The weakness is the machine infrastructure around it.

Not enough data. Weak tokenization. Few internal circuits. Speech systems that default to Latin. Metrics that do not always measure the right thing.

That distinction matters.

The language is not behind.

The tools are.

On-screen text: "The language is not behind. The tools are."

Visual: N'Ko card strong/clear; AI stack with missing pieces.

Caption: Low-resource does not mean low-structure.

Episode 15: Why I Am Telling This On Camera

Length: 60 seconds

Hook: I am telling this publicly because people need more examples of AI being used this way.

Camera Script:

I want people to see a different kind of AI story.

Not just "here are five prompts."

Not just "here is a productivity hack."

Not just "AI made this image."

Those are fine, but they are not the whole field.

AI can also help you investigate your own history, your own language, your own blind spots, your own community's missing infrastructure.

That is what happened with Mandinka and N'Ko for me.

I started learning.

Then I started testing.

Then I started building.

And the more I built, the clearer the question became:

What would AI look like if it was designed to meet our languages where they actually are?

That is the series.

On-screen text: "Learning -> testing -> building"

Visual: Three-part sequence: notebook, model scan, ASR diagram.

Caption: This is the AI story I want to tell.

Recording Notes

The first batch should be Episodes 1, 2, 3, 4, 8, and 10.

Those six clips establish the frame before any metric appears:

Mandinka is native. N'Ko is the doorway. AI is the instrument. The blind spot is the conflict. The creative use is the hook.

Do not say "CER" in the first three videos unless responding to a technical commenter.

Do not say "CTC," "TAR," "TTT," or "AGP" in the first week unless the clip is explicitly for a technical audience.

The first week should make people think:

> "I did not know AI could be used like that."

Only after that should the research receipts come forward.