Alex Power's blog

These tasks are one end of the spectrum of "language model" tasks.

A "language model" takes language and does operations on it. Not thought. The cross-language translation and "sentence decomposition" rely on a simpler understanding of knowledge. ✨ it is much easier to say "I can think of a castle" than it is to build one

We do need some element of the second type of model.

Which is going to be an LLM/software centaur version.

the LLM here does need to be able to do thought tasks. not just language-model.

🔥 there is a reason we don't put a dictionary in charge of the country.

The sentence decomposition logic is almost done.

The design:

• A sentence is entered into the database, often (but not always) in English.
• It is translated by LLM operation 1 into other languages: English, Chinese, French, Spanish, Liithuanian, Ukrainian, Kannada, Bengali. ⚙️ all PIE, except Chinese. This is unintentional but welcome. It remains within the areas of linguistics where I am comfortable.
• We check the database for word matches. This is used to create a list of candidate lemmas. We avoid the need to use LLMs to do the word-meaning resolution problem this way. The "correct" meaning will be present in most/all languages 💡 not all sentence translations will have the same lemmas. I like bread and Man patinka duona ⚙️ Lithuanian - literally translated as "Bread is pleasing to me" are the same sentence without a 1-1 lemma match
• Once we get the list of candidate lemmas, we ask the LLM to do the sentence decomposition.

For a sentence, the decomposition identifies:

• What specific lemma / sense-of-meaning of a word is in use. Is it to lose a race or to lose one's keys? A river-bank or a financial bank?
• What tense is the word in? Present/past? Is it a conjugated form? This should align with the lemma's information for the grammatical form, but there is no requirement for it to do so.

We do not yet have a full "sentence diagram" out of the words. The traditional Reed-Kellogg sentence diagram is uninteresting -- it only applies to English, is outdated, and remains confusing. However, the data missing to be able to generate this will be relevant later on.

I have been trying to run some very simple API calls using Claude to operate the new Barsukas HTTP API.

The flow is a 20-word prompt, reading 3-4 files from disk, and making a 4-line Python script to call the API.

This costs 22 cents. ⚙️ which feels like too much

Why so much? There is something like 8 KB of system prompt and 14 KB of system tools. All of which should be cached/built-in to the model, but is instead just charged to the user. And several rounds of "cached reads" as the tool pauses for inputs.

Fortunately, the $20/month subscription gives something like $300/month of credits. When the cost is inflated by a factor of 5, it helps that the currency is also inflated.

On the other hand, the sheer volume of struggles Claude Opus is having with this task (simple directions like "only read the api/ dir" get ignored about 1/4 of the time) does make me want to look for a different harness, which can pivot between model providers. ⚙️ it's not as bad as it was six weeks ago, but it is still very disappointing for the "recommended/expensive model" that it can't do this well.

a recent bug in Greenland: I was asking ChatGPT 5.4 mini to do a per-word breakdown of a sentence. But, sometimes it left words out. Other times, it would include the trailing period as a word, despite being instructed not to.

The cause of the bug: the prompt also asked ChatGPT to get the "word count" of the sentence. It is bad at this.

So, if it said The dog is on the log has 4 words, it would stop after 4 words and ignore the log when processing the sentence.

Of course, Python can easily do this. 💡 and, in the future, AI systems will know to just use Python to get this information rather than trying to guess it themselves. But here, it wasn't even used; I don't remember/know why a "word count" got added to the response.

I compared the Cambridge YLE list to the Trakaido wordlists. What I found was that most of the missing words were either deliberate ⚙️ modal verbs like "can", as well as pronouns, are handled outside the wordlists, correct 💡 ice skating, jellyfish, or toothache do not need to be on the list, or irrelevant 💡 having "grandfather" instead of "grandpa" is fine ... though it would be good to improve the synonyms.

Click title to read full message...

One of the under-appreciated benefits of using LLMs for coding is that they are experts on most topics.

Writing Trakaido would have been impossible if the LLM did not already have familiarity with Lithuanian. 💡 It was also familiar with dozens of other languages.

Languages are a strong point for large language models, but there are many other software projects that can benefit from having domain experts available to all developers at all times at a minimal cost.

This applies to non-coding projects as well. Trying to use LLMs to help file my taxes was ... still a bit optimistic in early 2026 🌎 there were several mistakes it made. But it generally understood what was going on. Without context.

💡 I have discussed before the question of context v. built-in training data. Should the machine be able to play chess without instructions (or external tools)? What about backgammon? Or a game I just made up? Is there any reason to believe/expect the universal game-playing machine will be a Large Language Model?

Other costs include:

• $99/year to be an Apple Developer
• $10/year for a domain name
• $30/month for web hosting on Digital Ocean.

Of course, the largest expense remains my own time 💡 None of the coding tools are good enough to do a project longer than about 2000 lines of code without expert supervision. Yet.. My time is not metered, and thus is somewhat invisible. But, at the low end, I would probably quote Fifty Thousand Dollars to do a similar project for someone else as paid work.

Compared to what it would take to build a similar app 5 years ago, this is fairly inexpensive. However, it is still a barrier to entry substantial enough to explain why there has not been a 🔥 more substantial flood of apps into the App Store.

Trakaido started out of a personal need. I wanted to learn Lithuanian, and Duolingo did not have it. The other apps on the app store were not satisfactory for various reasons.

As part of the process, I have developed my own pedagogy regarding language learning.

While I have found that the comprehensible input approach has some value, especially for grammar conjugations and the like, for building initial vocabulary and bootstrapping one's understanding of the language, I have found there is no approach better than rote memorization. This requires roughly 250 words, depending on:

• How one counts word forms
• How motivated one is to acquire vocabulary
• Whether one memorizes words such as numerals

Once a user has a basic understanding of some vocabulary, sentences can be presented that are mostly understandable: grammatical forms, verb conjugations (which are often more sensible audibly than in spelling), and (conjunctions, prepositions, and the like) are then acquired more naturally in an approach similar to comprehensible input. 💡 In particular, I have found that memorization of verb conjugations is time- and effort-intensive for little practical benefit either in recalling them or in using them to generate sentences.

Trakaido is live in the App Store! ⚙️ https://apps.apple.com/us/app/trakaido/id6755203308

Click title to read full message...

Learning a new language is a long game. You need a tool that meets you where you are — whether you're on your phone during a commute, at your desk with a full keyboard, or offline on a plane. That's what Trakaido is built to do.

What Is Trakaido?

Trakaido is a vocabulary and language learning app available on the web, iOS/macOS, and Android. It's designed around one core idea: the best way to build lasting vocabulary is through varied, well-timed practice — not grinding the same flashcard deck until you're bored.

You can start learning immediately in demo mode without creating an account. When you're ready, sign in to sync your progress across all your devices.

Languages You Can Learn

Trakaido currently supports four languages across all platforms:

- *Lithuanian* — including full noun declensions and verb conjugations

- *Chinese* — with pinyin support

- *French*

- *Spanish*

Additional languages (German, Italian, Dutch, Portuguese, Swedish) are available on the web app for early exploration.

You can use Trakaido in your own native language too. The app's interface and word translations are available in English, Kannada, Bengali, Ukrainian, Polish, and Romanian, with more on the way.

How the Learning Works

Trakaido is built around a *tier-based spaced repetition system*. The app tracks which words you know well, which you're still learning, and which you haven't seen yet — then schedules review at the right time to lock knowledge into long-term memory.

Journey Mode: Your Daily Practice

*Journey Mode* is the heart of Trakaido. Every session, it automatically selects words for you based on where you are in your learning progression. You don't have to decide what to study — the app figures that out. It mixes in new words gradually while making sure older words don't fade.

Within a Journey session, the app varies the type of activity to keep things from getting repetitive:

- *Flashcards* — see the word, recall the meaning

- *Multiple choice* — pick the right answer from options

- *Listening comprehension* — hear the word, identify it

- *Typing practice* — spell it out yourself

- *Sentence practice* — see words in context

Focused Practice Modes

When you want to drill a specific skill, Trakaido has dedicated modes for that too:

- *Drill Mode* — choose any set of words and activity type for targeted practice

- *Blitz Mode* — timed speed rounds for a quick challenge

- *Vocabulary List* — browse and search your word lists

- *Grammar Mode* — structured lessons on conjugations and declensions

Study Materials: You Choose the Scope

Words are organized by *level* (beginner through advanced) and *group* (thematic categories like greetings, food, travel). You decide which levels and groups to include, so your practice stays focused on what's relevant to you right now.

Native Audio from the Start

Every word in Trakaido comes with high-quality audio. You'll hear native pronunciation whenever you encounter a word — in flashcards, multiple choice, and listening activities alike. The iOS app pre-caches audio so it works even when you're offline.

Available on Every Platform

*Web app* — works in any browser, on any device. No install required. Start at trakaido.com and pick your language.

*iOS/macOS* — a full native app with offline support, haptics, and iCloud sync. Built for iPhone, iPad, and Mac.

*Android* — a native app with background sync and home screen widget support.

All three apps share the same learning model and word data, so switching between devices feels seamless.

Privacy and Data

Trakaido takes a simple approach: demo mode requires no account and stores nothing remotely. When you sign in, your progress syncs to the cloud — and your data stays yours. There are no ads.

For Educators

If you teach a language and want to track student progress or assign spaced repetition practice as homework, Trakaido is building educator-specific features. Reach out at educators@trakaido.com to share what you need.

---

Trakaido is available now at [trakaido.com](https://trakaido.com). Pick a language and start your first session — no credit card, no setup, no excuses.