I Built an AI Language App Because Everything Else Felt Wrong

After spending 20 years studying computational linguistics and watching language learners struggle with methods that contradict how brains actually work, I finally did something about it

A confession: I've spent most of my career thinking about how machines learn language.

Twenty years studying computational linguistics, natural language processing, large language models. Watching neural networks internalize Portuguese, Japanese, Arabic—not through grammar drills or vocabulary flashcards, but through exposure to patterns in high-dimensional semantic spaces.

The models don't "study" language. They absorb it through repeated exposure to comprehensible input, building internal representations that let them predict what comes next. No conjugation tables. No memorization. Just pattern recognition operating at scale.

Then it hit me: this is exactly how humans acquired their first language as children.

And it's not how we're being taught our second.

The Disconnect

I've watched countless people spend years on Duolingo, Rosetta Stone, Babbel. They dutifully complete lessons, memorize verb conjugations, translate sentences. Some achieve moderate success. Many quit.

The traditional approach treats language as a puzzle to be solved consciously—learn these rules, memorize these words, then assemble them into speech. It works for some learners. But it felt fundamentally misaligned with everything I understood about how neural systems—biological or artificial—actually acquire linguistic capability.

When I train a language model, I don't give it grammar rules. I expose it to massive amounts of text where it can understand most of the content but encounters new patterns at the edges. The model builds internal representations through repeated exposure, developing an intuitive grasp that enables generation rather than just recognition.

This isn't a metaphor. The mechanisms are genuinely parallel—both biological and artificial neural networks use pattern matching across high-dimensional spaces to internalize linguistic structure.

So why were we teaching humans with methods that contradicted how brains work?

Finding Krashen

The answer, it turns out, existed decades before neural networks proved it computationally.

Stephen Krashen, a linguist working in the 1970s and 80s, developed a theory called Comprehensible Input. His central claim: language acquisition happens naturally through exposure to language you can mostly understand (comprehend) but that contains elements slightly beyond your current level—what he termed "i+1."

Not through studying grammar. Not through memorization. Through absorption.

Reading Krashen felt like encountering the biological version of training principles I'd been applying to language models for years. The parallels were uncanny. Both systems required:

• Repeated exposure to mostly comprehensible content
• Gradual increases in complexity as patterns internalized
• Context clues (visual, situational) to support meaning
• Patience with ambiguity rather than demanding immediate understanding

The computational linguistics literature had been rediscovering what Krashen proposed based on observing human learners. We'd just been doing it with matrices and gradient descent instead of classrooms and textbooks.

The Tool That Didn't Exist

Once you see the connection, it becomes obvious what's missing.

YouTube has Spanish content. Netflix has Portuguese shows. But finding material at your exact level—comprehensible enough to follow but challenging enough to acquire new patterns—is like finding a needle in an increasingly vast haystack.

The problem is accessibility: native content is too hard for beginners, while beginner content is often artificial and boring. The sweet spot—material you can understand 70-80% of while being challenged by the remaining 20-30%—rarely exists for most learners at most levels.

Meanwhile, I had spent years working with generative AI. These systems could create content at any complexity level, with visual context, in any language, depicting any scenario. The technology existed to generate perfectly calibrated comprehensible input for any learner at any level.

Nobody had built it.

Building LinguaMama

LinguaMama started as a thought experiment: what if you could generate short video content—coffee shop conversations, daily routines, travel scenarios—narrated at exactly your proficiency level?

Not translations. Not subtitles. Actual narration describing what's happening on screen, calibrated so you understand most words but encounter enough new patterns to drive acquisition.

The key insight: you don't watch each video once and move on. You watch it ten times. Twenty times. Like listening to a complex piece of music until you can predict every note before it plays. That's when the pattern becomes internalized. That's when your brain develops generative capability—the ability to produce language, not just recognize it.

Traditional learning teaches you about language. Comprehensible input through repetition lets you acquire language the way your brain evolved to do it.

I spent months prototyping with generative AI models, testing different approaches to video generation, narration complexity, visual clarity. The technical challenges were significant but manageable. The conceptual challenge—trusting that repetition without explicit instruction would actually work—required unlearning decades of educational assumptions.

The Uncomfortable Part: Marketing

I write about computational linguistics, graph theory, consciousness in machine learning. I analyze how high-dimensional semantic spaces enable emergent reasoning. That's my comfort zone.

Marketing a language learning app? That feels like crossing into territory where I'm significantly less fluent.

But here's the thing: LinguaMama works because it's built on principles that are scientifically sound. Krashen's comprehensible input hypothesis isn't pop psychology—it's backed by decades of research. The parallel to how neural networks acquire language isn't hand-waving—it's grounded in genuine mechanistic similarity.

This isn't snake oil. It's not "fluent in 30 days" or "this one weird trick." It's a tool that delivers comprehensible input in a format your brain can actually use to acquire linguistic patterns.

The challenge is explaining that without either sounding like every other language learning gimmick or drowning readers in technical detail about latent semantic spaces and gradient-based optimization.

So I'm trying something different: just telling the truth about what it is, how it works, and why I built it.

What LinguaMama Actually Is

It's a platform that generates short videos in your target language—currently supporting 16 language variants across 9 languages that friends, family, and early users have requested: English (US/UK), Spanish (Spain/Mexico/Puerto Rico), Portuguese (Brazil/Portugal), German (Germany/Austria), French (France/Canada), Italian, Hungarian, Croatian, Thai, and Mandarin Chinese. Our AI-generated approach makes it remarkably easy to add authentic new languages, and we're planning to expand to over 100 language variants over time.

The major European languages (English, Spanish, French, German, Italian) are excellent. Portuguese works very well in both Brazilian and European variants. Some languages present unique challenges—Hungarian grammar is notably complex, but the output quality is still good. Thai and Mandarin Chinese are newer additions with promising quality. Croatian works well for those learning South Slavic languages. If you're a native speaker of any language we support and would like to help review output quality, we'd love to hear from you.

Each video depicts everyday scenarios: ordering coffee, navigating a market, morning routines, conversations. The visual content provides context while narration describes what's happening, calibrated to one of six proficiency levels (Tourist through Guide).

You pick a scenario. You choose your level. You watch it daily for 10-20 minutes. You don't pause to look up words. You don't take notes. You just... watch. Again and again. Until your brain internalizes the patterns.

When that content becomes too easy (you understand 80%+ without effort), you level up or switch scenarios.

That's the entire method. No tricks. No optimization hacks. Just comprehensible input through repetition, letting your brain do what it evolved to do.

The Beta Reality

I should be transparent about where things stand.

LinguaMama is in beta. We have about 20 episodes currently available, with more being added regularly. That's nowhere near the "thousands of hours" of content needed for complete fluency. But since you're watching each episode 10+ times, 20 episodes actually provides weeks or months of material for most learners.

The site is currently wide open. All features are available to everyone. You can use it entirely for free if you want—I genuinely want people to test whether this approach works for them.

If you want to support development, there's a Navigator subscription option. But right now, during beta, there are no "premium-only" features. Subscribing just means you're supporting the project financially while we build out the full platform.

There will always be a free tier. Eventually, premium subscribers will get additional features (more content, advanced tracking, personalized recommendations). But the core comprehensible input delivery—watching videos at your level—will remain accessible to everyone.

The GLOBO2026 Story

If you do decide to subscribe to the Navigator package, you can use coupon code GLOBO2026 for one year free.

Why GLOBO2026? It's an inside joke that traces back to how I discovered comprehensible input actually works.

There's a YouTube video—one of Dennis Borisov's comprehensible input demonstrations—that became oddly central to my language learning journey. In it, a teacher holds up a globe. The narration, delivered in careful Portuguese for language learners, goes something like: "This is a teacher. The teacher is speaking. Globo."

Something about that moment—the simplicity, the globe, that word "globo"—made us burst out laughing every time. The video became known simply as "Globo" in our household. We'd watch it daily. Someone would say "GLOBO!" as a signal to sit down and watch it together.

That video, absurd as it sounds, proved that comprehensible input through repetition actually works. Watching the same content repeatedly wasn't boring—it was how patterns internalized. "Globo" became shorthand for the entire approach.

So GLOBO2026 is both homage and goal: by 2026, I want comprehensible input to have taken me as far as it can. Dennis Borisov's work showed me the method was viable. Pablo Roman's Dreaming Spanish site demonstrated it could scale. LinguaMama is my attempt to bring that same approach to more languages with AI-generated content calibrated to your exact level.

If you believe in the approach and want to support the project out of pure kindness, the most helpful thing is actually signing up without using the coupon code. That provides direct funding for continued development even though the premium features don't exist yet.

But I'm not going to guilt anyone into paying for a beta product. Use it free. Use it with the coupon. Use it however makes sense for you. The goal right now is getting feedback from real learners about whether this approach actually works outside my own head.

What I Actually Want from You

Honest feedback.

Try LinguaMama for a few weeks. Pick an episode. Watch it daily. Don't stop to analyze—just let the repetition do its work. Then tell me:

• Does this feel like it's working?
• Is the content engaging enough to watch 10+ times?
• Are the proficiency levels calibrated correctly?
• What's missing?
• What's unnecessary?
• Does the comprehensible input approach make sense once you experience it?

I've spent two decades studying how neural networks acquire language. I've read the research on comprehensible input. I understand the computational linguistics theory. But I need to know if this actually helps real humans acquire real language competency.

That's what the beta period is for. Validation. Iteration. Finding the gaps between elegant theory and messy reality.

The Vulnerable Part

Building something and asking people to try it feels uncomfortably exposed.

I'm used to analyzing systems, not marketing them. Used to writing about abstract concepts, not asking for feedback on something I created. The academic instinct is to wait until everything is peer-reviewed and bulletproof before presenting it publicly.

But LinguaMama isn't an academic paper. It's a tool meant to be used. And the only way to know if it works is to let people use it and report back honestly.

So: linguamama.ai

Sign up for a free account if you want to track your progress. Subscribe with GLOBO2026 if you want a year free. Or just use it without an account if you prefer.

Most importantly: tell me what you think. What works. What doesn't. Whether the comprehensible input approach makes sense once you're actually using it rather than reading about it.

I built LinguaMama because traditional language learning felt misaligned with how brains actually acquire linguistic capability. Whether I built the right solution is something only actual learners can determine.

Ready to find out?