Why Krendo doesn’t have AI features.

Every other extension in the language category has shipped at least one AI feature in the last eighteen months. Krendo hasn’t, and isn’t planning to in v1.x. This is the honest reason — not a slogan, not a marketing position we’ll quietly reverse next quarter.

What we mean by “no AI”

To be specific about what we’re ruling out:

• No AI tutor or conversation partner inside the extension.
• No generative grammar explanations on the dictionary panel.
• No automatic example sentences invented by an LLM.
• No “personalized lesson plan” that depends on a language model.
• No “ask the AI about this word” modal.

The translation itself goes to Google Translate’s public endpoint, which is — technically — a machine learning model. Future multi-engine support will likely add DeepL and may add bring-your-own-key LLM providers like Claude or OpenAI. That’s a different thing: better translation quality is on the table, AI as a learning loop is not.

What we don’t want to be

In 2024 and 2025 the App Store and the Chrome Web Store filled up with “AI language tutor” apps. Most of them are thin wrappers around GPT calls with a couple of streak mechanics on top. The complaints follow a pattern: the AI gives plausible-sounding but wrong grammar explanations, the “personalized” lesson is the same generic content as every other user’s, and the feature lives behind a subscription that compresses real language-learning into a metaphor for engagement metrics.

We have nothing against AI as a technology. We have a lot against being one more product in that pile. The reason most of those products feel bad to use is that they treat language learning as a SaaS funnel and the language model as the magic that justifies the subscription. The actual job that needs doing — turning the reading you already do into vocabulary you remember — doesn’t need an LLM. It needs a clean capture loop and a sensible review algorithm.

What “no AI” costs us

We want to be honest about the trade-offs, not just praise the choice.

• No grammar explanations.If you save a word and the dictionary entry isn’t enough — you don’t understand why the verb conjugates this way, or what register the word lives in — you have to look elsewhere. We don’t invent an explanation; we don’t give you a wrong one either.
• No auto-generated example sentences.The only sentences attached to your saved words are the ones you actually encountered. That’s a feature for long-term retention (real context beats invented context) and a small inconvenience when the word has only been saved from a thin context like a headline.
• No personalized lesson plan.Krendo doesn’t tell you what to study next. The review queue surfaces what’s due, and you decide whether to do five minutes or twenty. The structure is up to you.

What it saves us — and you

• Hallucinations don’t enter your vocabulary deck.Wrong grammar explanations, fabricated idioms, and made-up example sentences all become real problems once they’re stored as cards. We don’t generate any of those, so we can’t poison your retention.
• No subscription to justify model costs. Running a useful LLM per user costs money — non-trivially so. Tools that ship AI features under a free tier either run a cheap model, restrict heavily, or burn money. Both paths usually end in a Pro upgrade pop-up. Krendo doesn’t have that pressure because the architecture doesn’t require an LLM provider.
• The product stays simple to use and explain. Select, save, review, export. No model picker, no token quota, no “upgrade for better explanations,” no disclaimer about hallucinations. You don’t need to read documentation to know what the tool does.
• Privacy by default.Saved words and notes don’t need to be sent anywhere for the product to work. They live in your browser. That’s only possible because we don’t need to feed them to a model.

What would change our mind

We’re not ideological about this. If we found a specific AI feature that would genuinely help vocabulary retention without the costs above — one whose value is clearly bigger than the failure modes — we’d consider it. The bar is high:

• It would have to be opt-in, not on by default.
• It would have to be a clearly bounded helper feature, not a centerpiece. Adding an LLM-powered conversation tutor would push us into the category we’re deliberately outside.
• It would have to respect the local-first commitment. Either run client-side, or send only the specific thing being augmented, with explicit consent each time.
• We’d explain it in writing before shipping it, including the trade-offs. Same way we’re explaining not shipping it now.

A note on translation engines vs “AI features”

The line we’re drawing is between “using AI to translate” and “AI as the learning loop.” Better translation engines — DeepL, an LLM with your own API key, anything that produces better source-to-target text — are infrastructure improvements that make the core loop work better. An AI tutor or generative grammar explainer is a different category of feature with different failure modes.

Multi-engine translation with bring-your-own-key support is on Krendo’s roadmap, partly because the undocumented Google endpoint we use today is fragile and partly because users with specific language pairs (Japanese, Chinese, translation-heavy reading) benefit from engine choice.

The short version

We’re building a tool that takes reading and turns it into vocabulary. The reading does the work, the review algorithm closes the loop, and the writing already in your browser is where the lesson lives. Adding an AI layer on top would dilute that — we’d become a tutor, which isn’t what we want to be, isn’t what the language-learning audience needs another of, and would cost the simplicity that’s the whole point.