Unformed Delta - language-learning

Against foolproof software

Fri, 17 Apr 2026 06:37:00 +0000

If you’re familiar with spaced repetition systems, this should work as a standalone post. This is the last post of a 3 post series. The first two posts explain WaniKani and jpdb.io in more detail.

WaniKani tries very hard to make learning Japanese foolproof. They don’t let you grade yourself or undo mistakes they want to protect you from “the illusion of knowing”. They don’t offer the ability to blacklist or permanently mark cards as known.

These design decisions probably make WaniKani more effective for the user profiles they think about the most. WaniKani’s testimonials highlight two kinds of users:

Japanese residents, who likely either need Japanese for work, e.g. “My job requires me to work with a number of Japanese companies.”, or strong intrinsic motivation, e.g. “I’ve lived in Japan for eight years”. For this kind of person, it’s easy to dedicate 1-2 hours every day to learning Japanese.
self-learners, with a lot of time to dedicate to learning Kanji. My friends who made it through much more of WaniKani than me, all fit into this category. Even for me, WaniKani worked well at first, before betraying my expectations in a couple of important ways.

First, WaniKani ended up growing into a much larger time commitment than I assumed from my first few weeks using it. WaniKani doesn’t claim how many hours per day it will take to learn Japanese in just over a year. But, because early on WaniKani’s mechanics for unlocking lessons prevented me from investing the ~30 minutes per day I wanted to, I assumed that the workload would continue to be manageable. Then, when reviewing flashcards from the past 4 months, the workload became overwhelming. jpdb.io’s better spaced repetition algorithm lets you build it into your schedule more consistently from the start, while also adapting better if you’re willing to commit less time in the future.

Second, WaniKani promised to teach me “2,000 kanji, hand-picked and cleverly ordered” plus “6,000 Japanese words, all carefully validated by a human to be common or useful”. Of course, I didn’t have a good sense of what vocabulary I wanted to learn, so this was part of what appealed to me when I chose WaniKani. If I’d been willing to commit the time to mine songs and texts for vocabulary I cared about, I probably would have started with Anki. But over time, the rigidity of the curriculum started to frustrate me. I longed to be able to exclude words, even considering a Firefox extension for WaniKani that would autofill the correct answer for words that I didn’t want to review.

Both of these problems were mostly a misalignment between my goals and knowledge, and the needs WaniKani’s developers assume their users have. But WaniKani is designed to be foolproof, without options or configuration that would make it less effective for certain kinds of users.

Aiming for foolproof makes a lot of sense when designing a product. Less options, configuration, and variables to consider makes the product simpler. This has huge dividends for scalability. A simpler product is easier to explain, optimize, market, and sell without needing to consider the needs of individual users.

But I wasn’t interested in WaniKani as a mindless consumer, I was interested in a tool to help me learn Japanese. Instead of aiming to be foolproof, jpdb.io trusts its users to accomplish their goals. This unlocks the possibility for it to be better tailored to any individual user’s specific goals.

Not everyone will carefully tweak all the configuration options to figure out what works the best for them. Good defaults will make sure that the average user still has a good experience. But almost everyone has specific complaints about the products they use or things they would want tweaked. It’s worth building a configuration option for at least the most common ones.

Even if users want something that you think would make the product worse, they’re probably trying to act in their own best interest. Sometimes, as the designer, you really do know better. But, at other times you don’t understand the user’s complaint well enough. If enough people request a setting you find confusing, perhaps it’s worth exercising some intellectual humility and adding it anyways, along with an explanation for why you think they shouldn’t use it¹. If you trust that your users are acting in their own best interest, they’re probably motivated enough to try to understand your reasoning.

Most users don’t need to be protected from themselves. Instead of making your software foolproof for your assumed user, design it to adapt to the needs of your specific users.

jpdb.io’s settings page actually does this several times. I haven’t changed any of the settings it recommends against. ↩

What makes a good spaced repetition system? (jpdb.io)

Thu, 16 Apr 2026 05:01:00 +0000

I recommend reading yesterday’s post discussing my follies with WaniKani the first time I tried learning Kanji.

Over a year and a trip to Japan passed before I made a serious effort at studying Japanese again. WaniKani had failed me in several ways, and I was reluctant to try a spaced repetition system that wouldn’t solve them.

I wanted something that would adapt to my existing Japanese knowledge, letting me avoid redundant flashcards. More importantly, I needed something that would respond more gracefully to getting overwhelmed or returning after a break.

jpdb.io checked both of these boxes. Instead of a fixed curriculum, jpdb.io allowed you to choose exactly which vocabulary you wanted to learn. Even better, it offered a wide variety of pre-built decks with vocabulary from anime, visual novels, and books. It also claimed to handle lapses and irregular time periods between flashcard reviews better by implementing an ML based spaced repetition algorithm.

jpdb.io’s flashcard system/lessons are simpler than WaniKani’s. There are only two kinds: kanji flashcards (which includes radicals) and vocabulary flashcards. Kanji flashcards come with mnemonics¹ ask you to visualize or draw the kanji or radical based on its name while vocabulary flashcards ask you to guess the pronunciation/meaning of the word in the context of an example sentence.

I was enjoying the Summer Pockets anime at the time and eager for more content, wanted to try reading the source visual novel as Japanese practice. jpdb.io configured to feed me the most common vocabulary in Summer Pockets, I raced through 30 minutes of flashcards per day, mostly marking that I would never forget vocabulary that I already knew.

Unlike WaniKani, flashcards are self graded — you choose whether remembering the kanji/vocabulary was hard/medium/easy or whether you forgot the flashcard nothing/something.

Anki allows using FSRS as its spaced repetition scheduler providing pretty much the same benefits as jpdb.io’s scheduler.

I was aware that without strongly committing to an answer, there was a risk that I would assume that I would assume my answer was correct, even when I hadn’t really known before seeing the solution. To guard against this I took the time to trace the kanji in the air and subvocalize the vocabulary’s meaning and pronunciation before clicking through to see the solution. Deciding how to grade myself was tricky at first. Sometimes when tracing out a kanji, I would draw the right strokes but in the wrong order. Because I was only tracing into the air, it was hard to judge how accurately I’d drawn the shape, so I opted to grade myself more conservatively.

jpdb.io’s spaced repetition algorithm also lived up to its promises. I’ve noticed two main benefits in practice. First, flashcards are pretty much always the right difficulty. My accuracy on jpdb.io ended up being fairly consistent with my accuracy on WaniKani at ~90%. Despite this, I rarely get a flashcard that I want to mark as easy; usually I have to think at least a second or two, before remembering the correct answer. Second, unlike WaniKani I don’t get flashcards wrong, only to weeks later on my next review still have no idea what the right answer is. Instead, jpdb.io usually shows me the flashcard again much sooner (i.e. the next day) when I get a flashcard wrong.

All of this remained true even when abandoning jpdb.io for longer periods of time, be it several weeks for a trip, or even once for several months. Just continuing to review my flashcards at whatever pace I manage when I got back, slowly reintroduces me to the flashcards I forgot, while giving me credit for remembering a longer period of time, for the flashcards I got wrong. It’s also been really nice to be able to tailor the vocabulary I’m learning to whatever Japanese media I’m consuming over the past year.

I haven’t practiced Japanese quite as consistently since switching to jpdb.io. But, I’ve been using it for at least a year, whereas WaniKani burnt me out after just ~5 months. It effortlessly slots into my life when I want it to, adapting to whatever is going on in my life at the moment, making it as painless as possible to pick it back up again.

Continued in Against foolproof software.

At least the most common ~1000 radicals/kanji have mnemonics. I found that learning the most common words in Summer Pockets, I fairly quickly started running into some kanji missing mnemonics. But jpdb.io lets you to define custom mnemonics, which has the added benefit of making it easier to remember your mnemonic (assuming you try reasonably hard to make one). ↩

What makes a good spaced repetition system? (WaniKani)

Wed, 15 Apr 2026 05:17:00 +0000

This is part 1 of a ~3 part post. This part explains how WaniKani works through my first experience and follies using a spaced repetition system.

Japanese is one of the world’s most difficult writing systems. Compared with English, Hiragana and Katagana are simple phonetic alphabets. Their pronunciation is always consistent, and across both alphabets there’s just ~92 distinct characters plus a handful of consistent diacritics and multi-character compounds. Much easier than English where “ghoti” could arguably be pronounced “fish”¹.

But to read Japanese, you also need to be able to interpret at least 2000 Kanji characters. Most Kanji characters can be pronounced at least two different ways depending on context: a native Japanese pronunciation (kun’yomi) in addition to the pronunciation derived from Chinese (on’yomi). For common characters there’s usually at least one or two more. In the worst cases, for example 生, an extremely common character that Japanese children learn in first grade, a single character can have as many as 6 different kun’yomi and 2 different on’yomi.

I knew all of this when I set out to learn the 2000 most common Kanji. I’d just returned from my first trip to Japan, and was more motivated to learn than ever.

WaniKani

WaniKani made getting started frictionless. Kanji can be broken down into radicals, bits of characters with a consistent meaning/shape. Each WaniKani level, starts off with lessons teaching you the names of the radicals. Each radical is taught with a mnemonic and often comes with a cute picture too:

Since there are no circles in Japanese characters, sometimes rectangles or squares have to make do. In this case, this big rectangle is the sun. The middle line is a cloud, moving across the sun. Picture it, but don’t actually look at the sun and burn your eyes!

After a couple hours, you prove that you remember the names of each of the radicals by typing them out, and then you’re done for a few hours. You review the radicals again, then again after a day, and finally if you’ve gotten the item right 4 times in a row, you’ve proven that you’ve committed the radical to your “medium-term” memory. The radical upgrades from the “Apprentice” to the “Guru” memory stage, unlocking Kanji that depend on that radical. For Kanji there’s two separate things you learn and are tested on: meaning and pronunciation. When a Kanji advances to “Guru” you unlock dependent vocabulary, again with separate meaning and pronunciation lessons. Once all but 3 Kanji in a level are “Guru”, you advance to the next level, unlocking the next set of radicals.

At first, it only took 5-10 minutes a day to do all of my reviews plus new lessons. I returned to WaniKani for my reviews after the exact review time intervals, to make sure that do my reviews as soon as possible. Because I was also maintaining a high accuracy, I was able to advance a level every ~8 days. Projecting forward, it wasn’t difficult to imagine getting to the last level, level 60 in a year and a half as advertised. I was starting to notice the vocabulary I was learning in the music I listened to, excited that so as long as I could match my current pace, I would have mastered Japanese in just over a year. How naive I had been.

The thing about spaced repetition systems, is that you’re trying to commit the Kanji and vocabulary to your long-term memory, remembering them indefinitely. So you keep reviewing again after 1 week, after 2 weeks, after 1 month, and finally, in WaniKani, after 4 months the item is considered “Burned” and you never see it again in reviews. Beyond that, ambient exposure through reading is probably enough to avoid forgetting it.

These longer-term reviews didn’t get in the way of my high accuracy rate, but the additional workload made my review sessions start to take longer and longer. By the time I was hitting my first batch of 4 month reviews, I was spending at least an hour a day on my 200ish of my lessons and reviews. To be confident that this was as hard as hard as it would get I calculated that about ~150 total vocab/radicals/kanji per WaniKani level. Some vocabulary were also starting to get fairly obscure. I’m still not exactly sure what outpatient means in English, but 外来 was one of the words I was learning in Japanese. An hour a day was a large enough time commitment that I was starting to question whether it was worth continuing.

When getting an item wrong, spaced repetition systems adjust their model of how well you know that item. WaniKani uses discrete memory stages: “Apprentice” 1-4, “Guru” 1 & 2, “Master”, “Enlightened”, and “Burned”. WaniKani places a lot of weight on how often you get an item wrong before you get it right. Getting an item wrong only once, then right on the next try because you carefully reviewed and committed the mnemonic to your short term memory, only moves an item back 2 memory stages ². This means you can get an item wrong after not having seen it in over 4 months, and then next see it again over two 2 weeks later. Just getting the item right on the first re-try isn’t enough to prove that you still know the item well enough to remember it in two weeks. This can make reviews feel extra frustrating because you’re being asked to review cards that you don’t actually have a good chance of getting right.

I started to dedicate less time to WaniKani, just 30-60 minutes per day instead of just over an hour. Some days I wouldn’t even finish all of my reviews, much less all of my new lessons. Consequently, I started to level up more slowly, and my accuracy started to drop. My lower accuracy further increased my workload, and I found that my accuracy on cards that I’d gotten wrong two weeks earlier stayed fairly low. Eventually I was rarely finishing all of my reviews. After traveling for two weeks, the mountain of reviews I returned to made continuing feel impossible.

After a couple of months futilely trying catch back up, my accuracy rate quickly falling, I let go completely, giving up not only on finishing WaniKani, but also achieving Japanese reading and writing proficiency.

Check out part 2 where I try again more intelligently.

“gh” as in cough, “o” as in “women”, and “ti” as in “nation”. ↩
The exact formula is documented in WaniKani’s knowledge base. ↩

How to speak more Japanese in Japan

Fri, 03 Apr 2026 04:43:00 +0000

For the intermediate Japanese learner visiting Japan there is a problem: Japanese people are too good at English. They’ll often begin an interaction in English or immediately switch to English if your Japanese isn’t perfect. If you know the natural way to continue the conversation in Japanese, they’ll probably appreciate your effort. In many cases, they’ll also switch back to Japanese. With practice, you can have longer, more meaningful interactions while traveling in Japan.

Let’s walk through a simple example:

You walk up to the empty register, carrying your tray with a curry roll and a melonpan, calling out “こんにちは”.

As the cashier walks up you say “この二ふたつください”.

Since you’re a foreigner and your pronunciation wasn’t the best, the cashier responds “For here or to go?”¹ while wrapping your individual rolls.

You respond “持もち帰かえりで”, saying that you want them to go.

The cashier, slightly surprised, continues “Do you need a bag?”² asking if you want an extra bag to carry the two already wrapped rolls.

You respond “袋ふくろは大丈夫だいじょうぶです”, confirming that you don’t need a bag.

The cashier continues, “七百二十円ななひゃくにじゅうえんです” saying that the total is ¥720.

You say “一千円いっせんえんで” handing them a ¥1000 bill.

They say “はい”, take the money, count out the change and then say “二百八十円にひゃくはちじゅうえんのお釣つりです。レシートは要いりますか”.

You respond “大丈夫だいじょうぶ、ありがとうございました”, take your individually wrapped rolls and leave as they echo “ありがとうございます”.

The cashier switched to Japanese by the end; often just being a bit stubborn is a enough to achieve this. Sometimes it ends up being easier: staff in small towns are less accustomed to tourists and more likely to just muddle through the whole interaction with you in Japanese. At other times it’s near hopeless: most often at hotels, but also in more touristy locations/situations, your conversation partner might not switch to Japanese no matter what you do. It’s still worth trying, but often giving in and finishing the conversation in English leads to a more pleasant interaction.

Here are some things that have helped me have longer, better interactions in Japanese:

Listen to the people ahead of you in line: usually they’re going through exactly the interaction you’re about to have. This is also good listening practice.
Practice upcoming interactions using LLMs: if there’s a harder conversation you anticipate having, you can practice with Claude before having it. After losing my toiletry bag I practiced with Claude before heading to the sento where I thought I lost it.
Sometimes less is more: for example, when ordering from a list of variants it’s more natural to just say “二番にばんお願ねがいします” to specify the second option. You won’t impress anyone by demonstrating your vocabulary unnecessarily.
Know your counters: at a museum you ask for 大人おとな二枚にまい tickets, at a food stall you ask for 一本いっぽん skewers, at a restaurant you’re seated as 二人ふたり people. Often this is your first interaction; nailing it can help set you up for success (and using them wrong can signal: I’m a beginner, switch to English). When in doubt, つ likely gets the point across well enough.
Know filler words: used appropriately えっと, あの, まあ, etc. can help you stall for time to formulate your response.
For longer interactions, aizuchi (相槌あいづち) is important: Using はい, なるほど, そうですね, etc. appropriately reassures your conversation partner that you’re following along. Make sure you’re actually following along though. If you aren’t, tell them you don’t by saying 分わかりません、もう一度いちどお願ねがいします.

Checking out at a bakery, ordering at a restaurant, or asking where to find 着火剤ちゃっかざい in a Daiso are all quite short/simple conversations to have. But I’ve found that having simple/transactional conversations builds the confidence to tackle the harder, more open ended conversations.

On my recent trip to Japan, when sitting down to sip some free tea at 一石栃立場茶屋いちこくとちたてばちゃや, a tea shop on the Nakasendo between Magome and Tsumago the owner asked me curiously: “Where are you from?”, “How long are you visiting Japan?”, and so on. I stubbornly responded in Japanese; he eventually started asking me questions in Japanese. After a while he asked “Why do you like visiting Japan?” and my Japanese vocabulary failed me. I responded in English, but by then the conversation had gone on long enough that it was effortless to shift back into Japanese afterwards. The conversation with the owner was memorable not because I managed to speak a lot of Japanese, but because we actually connected, comfortably mixing languages as needed.

It’s useful to know how the conversation would go if the whole interaction was in Japanese. In Japanese this might be 店内てんないでお召めし上あがりですか. ↩
In Japanese this might be 袋ふくろはご利用りようですか. ↩