top of page
AI processor with icons for security, growth, and marketing

You know the answer in the app. Why can’t you use it in conversation?

You know the answer in the app. Why can’t you use it in conversation?

Writer: Krzysztof Kosman
Krzysztof Kosman
3 days ago
4 min read

Four answers appear on the screen. One looks familiar, you select it, and the lesson moves on. Later, you are on a street in another country, trying to ask where to catch a bus. Nobody offers four sentences to choose from. You have to form the question yourself, then understand the reply.


Krzysztof Biedalak uses this example in our EdTech Dots conversation to describe a problem with language practice: how do we know whether an exercise prepares us to use what we learn? The SuperMemo co-founder has spent years working on learning systems. We discuss the Polish history of repetition algorithms, the limitations of learning apps, and what AI conversation can now make possible.



When the prompt disappears


Biedalak’s point is that we learn so we can use knowledge. With a language, that includes retrieving a word when we need it. Recognising that word in a list asks us to do something different.


He does not dismiss every multiple-choice exercise. SuperMemo also uses recognition when learners first encounter material, he explains. Later, it asks for an active response. The word being practised should come from the learner’s answer, rather than already being supplied in the question.


That distinction suggests a useful question to ask about an exercise: what must I produce without a prompt? It belongs alongside the completion screen. Finishing a task smoothly does not, by itself, tell us whether we can hold a conversation.


A Polish story that is easy to oversimplify


SuperMemo has operated since 1991. Biedalak describes it as probably the world’s oldest still-operating e-learning company. The ideas behind it developed earlier, in the computing community at Poznań University of Technology. Asked about his own contribution, he points especially to Piotr Woźniak and the other people without whom the project would not have existed.


The story extends beyond the company. An early published SuperMemo algorithm found its way into other products. Biedalak recalls that Anki used an older version for many years, and that Quizlet started with SuperMemo 2.


When discussing Duolingo, however, he distinguishes direct implementation from influence through the underlying theory. That qualification matters. Polish work on computer-scheduled repetition influenced tools used far beyond Poland, but this does not mean every app mentioned runs the same algorithm today. Nor does it mean SuperMemo invented the broader idea of spaced repetition.


Estimating what you remember


As Biedalak explains it, a repetition algorithm estimates the probability that a person will recall a particular piece of information. It uses that estimate to schedule the next review. The history of that learner’s responses matters, as does the material itself.


We do not remember everything equally well. New information might connect to something we already know, or remain unfamiliar. One word comes easily while another keeps slipping away. A single review calendar for everyone cannot capture the differences the system is trying to model for each learner.


Successful retrieval can allow the intervals between reviews to grow. The aim is not to keep looking at the same flashcards indefinitely. It is to maintain knowledge with a schedule that takes forgetting into account.


This also explains Biedalak’s approach to a leaderboard. SuperMemo tries to compare the knowledge learners are estimated to remember, rather than simply the time they have spent studying. This remains a model’s estimate, not a direct reading of someone’s mind. It does show how choosing a different measure changes the question we ask about progress.


A conversation nobody had to script in advance


Repetition does not address every part of learning a language. We also need conversations in which we do not know the next sentence in advance. Biedalak describes years of work trying to bring software closer to that experience.


In the interactive film Olive Green, users could follow different dialogue paths. But someone still had to prepare those possibilities. However elaborate the script, it was finite. Speech recognition helped people interact with the program, without itself providing an open-ended conversation.


Language models introduced another possibility: a reply can be generated in the context of what the learner has just said. For the SuperMemo team, this offers a way to pursue an idea that earlier technology could not adequately support. Biedalak discusses the development of the SuperMemo Live voice agent while acknowledging that such systems can still be awkward.


Here, the discussion of AI becomes specific. The opportunity is to practise saying something and responding to what comes back. Putting a chat interface inside an app does not settle how well that practice will work.


AI conversation has to fit inside a subscription


Biedalak brings the discussion down to the company’s economics. During the interview, he gives the example of a subscription costing PLN 36 a month including tax. Less than PLN 30 remains before the company covers the costs of providing the service.


External models cost money to use. Running infrastructure also costs money. The team has to reconcile the quality of the conversation with the expense of delivering it. That constraint gets less attention than impressive AI demonstrations, although it affects the product people actually pay for.


He offers no single recipe for resolving it. He describes the work of building a service whose technical capabilities can be delivered at an affordable price.


How much knowledge do we want to keep for ourselves?


Towards the end, I ask why we should learn at all if AI can provide more and more answers. Biedalak points to the personal judgement still needed to assess what a system gives us. He also considers learning as a form of personal development, even if its practical necessity changes.


He describes using technology while trying not to delegate his own self to an outside system. It is a personal position, not a confident prediction about how everyone will live in the future.


That brings me back to the bus-stop example. When choosing a learning tool, I would want to know what I can say once the answer options disappear. This conversation offers several ways to examine that question: active retrieval, the timing of reviews, and AI dialogue that can move beyond a pre-written path.


Watch the full EdTech Dots conversation with Krzysztof Biedalak: https://youtu.be/1f_g594VFEQ. The interview is in Polish, with English subtitles available.


bottom of page