Anki Meets ChatGPT for Snacks in Paris for Lunch

5–7 minutes

I have been developing my French, so I rely on several primary resources.

Myself

This includes grammar books, novels and the rest of the traditional fare, some of which I have listed below. Books remain useful, despite their regrettable inability to generate notifications, harvest engagement or insist that one more chapter will transform your life.

Audio: NotebookLM summary podcast of this topic.
NotebookLM Infographic on this topic.

YouTube and Such

Social media is a time sink, and YouTube is no exception. This is not to argue that I derive no value from either. I do. It is merely to observe that the cost-benefit calculation may not always be in my favour.

That said, I encounter many useful tips, ideas and sources of general exposure to French. These help with learning, relearning and noticing gaps in what I thought I already knew.

There is, however, a risk. People teach and learn differently. Some teachers also believe they understand how they themselves learned and construct a method around that retrospective account. Memory being the immaculate scientific instrument that it is, this can create difficulties.

The first is obvious. A teacher may misattribute the cause of their own success and teach what is merely a covariant feature of the process. The method may have accompanied their learning without producing it. Remove the principal variable and the supposed method may accomplish very little.

A related problem arises when learners do not yet understand how they themselves learn. They may spend hours following a method that is suboptimal or even counterproductive. On the other hand, the method might happen to suit them perfectly. It can be difficult to know in advance.

A teacher may also be ideal for a learner at A1 or A2, but less appropriate at B1, B2 or beyond. By then, however, familiarity has become comfort, and comfort can masquerade as progress. This seems worth bearing in mind.

Even so, YouTube contains an enormous amount of material at nearly every level. The problem is not scarcity but selection.

Anki

Anki is excellent for repetitive learning tasks. It is essentially a flashcard system on steroids, although with fewer alarming side effects.

Many decks have already been constructed and made available online. I happen to use Anki for language learning, but it would be useful for almost any flashcard-oriented task, including academic study or memorising recipes, should that be your peculiar ambition.

This post is little more than a reference to Anki for context, but I recommend exploring it. There are also plenty of YouTube resources explaining how to use it for different purposes.

Anki supports text, images, audio and video, so it is genuinely multimedia rather than merely a digital imitation of index cards.

I have been using it to retain and expand my French vocabulary and grammar because I live somewhere French is not especially available. I would be better situated if I wanted to learn Spanish or Portuguese, and perhaps several other languages. For now, I shall pass.

Many of the cards I have imported come from decks with names such as “Top 1,000 Common Words”, “100 Conjugated Verbs” and “50 Useful Travel Phrases”. After enough examples involving “She has an apple”, “She gave him an apple” and “He would like her to give an apple to his father”, my mind wandered towards Snow White and poisoned apples.

In English, one may speak of a poison apple, treating poison as an attributive noun, or of a poisoned apple, in which poisoned describes what has happened to it. French does not ordinarily preserve the first construction unless one is deliberately inventing a literary compound. The natural French expression is:

Elle a mangé une pomme empoisonnée.
She ate a poisoned apple.

I asked ChatGPT about this. I shall spare you the full forensic inquiry into the apple. What interested me more was its suggestion that I diversify the examples in my deck.

iTalki

I use iTalki for live conversation with native speakers. Although I am not formally endorsing the platform, I recommend investigating it and have found it useful.

I have interacted with only two teachers, but I enjoyed working with each of them for different reasons. Lessons are relatively inexpensive, although rates vary considerably between teachers. They are not as cheap as YouTube or Anki because they are not free, but, as an economist, I also take opportunity costs into account.

Spending hours on a free platform without acquiring any durable knowledge still entails a cost. Free is not synonymous with costless, despite what several business models would prefer us to believe.

I am not trying to sell anyone on iTalki or a similar service, but one can sample it fairly cheaply. A lesson might cost about the same as a cup or two of aggressively marketed coffee. Since I do not drink coffee, I already have the necessary budgetary fiction in place.

ChatGPT and Claude

I also use ChatGPT and Claude frequently. In this case, ChatGPT suggested several variations that retained the grammatical structure while making the sentences slightly less lifeless:

Elle a mangé une pomme empoisonnée.
She ate a poisoned apple.

Elle a mangé une pomme pourrie.
She ate a rotten apple.

Elle a mangé une pomme volée.
She ate a stolen apple.

Elle a mangé une pomme interdite.
She ate a forbidden apple.

Elle a mangé une pomme étrange.
She ate a strange apple.

This is a minor use of a large language model, but a useful one. It can produce controlled variations, introduce vocabulary I have not encountered often and help prevent repetitive exercises from collapsing into grammatical wallpaper.

It can also explain why one construction sounds natural and another does not, generate examples at a particular level and create drills around a specific weakness. Naturally, its answers still require scrutiny. Confidently delivered nonsense remains nonsense, even when it arrives with immaculate formatting.

Taken together, these resources serve different functions. Books provide structure and depth. YouTube supplies exposure and ideas. Anki makes repetition tolerable. iTalki provides live interaction. Language models generate variations and explanations on demand.

None is sufficient by itself. Used together, however, they form a reasonably effective learning environment, provided one continues to distinguish activity from progress. The first is easy to measure. The second is what all the activity was supposedly for.


NB: I wrote this by hand and passed it through ChatGPT to clean it up. Let’s call it 95% human authored and call it a day.

Image: Gemini Nano Banana Pro’s first mis-render. This and the cover image are intentionally stereotypically satirical.

Discomfort of You

2–3 minutes

NB: The following post will be entirely human-generated, with no LLM involvement – except for the cover image rendered by Gemini. It will also be brief.

I am (still, perpetually) learning French, and I use Anki to help me. I recommend it for standard and non-standard flashcard uses. Recently, I asked ChatGPT to generate a CSV with phrases for the front and back of the cards, for example:

Front: I will be able to visit France next year.

Back: Je pourrai visiter la France l’année prochaine. (futur simple)

I found the pack useful, but a few days later (today), I asked for more than first-person perspectives, as well as additional tenses. I entered this prompt, which is what triggered this blog entry:

‘You’ rendered this for me. Might you generate a similar file that I can use to import to Anki that includes versions other than first-person singular? Future tenses would be helpful as well.

Notice the ‘you‘ in scare quotes – inverted commas.

It’s no secret that I am partial to the episodic selves of Galen Strawson in contrast with diachronic selves, which is to say that I don;t believe that the you or the I or yesterday – or even a second ago – is intrinsically the same person. I’ve explained this elsewhere, so I’ll leave it here.

As much as I feel this way, I don’t feel put off referring to myself as myself or Ime, myself, and I – or to you as you, but I am more consciously uncomfortable with addressing an LLM as you, as demonstrated.

I know full well that each prompt and response is to a different instance of the LLM – a different episode. Unless the LLM accesses an earlier conversation in memory – and I use this term loosely here, too – it would not retain context, and the absence of you-ness would be painfully obvious.

I don;t believe that the self of people is functionally any different. Obviously, humans are carbon-based rather than silicon-based, and storage and retrieval operate differently beyond the substrates, but the mechanism is metaphorically similar.

Does anyone else feel dis-ease in addressing an LLM in the second person? Or is it just me – or me?

Ear Training

1–2 minutes

Me: I got an admission: I never enjoyed musical ear training – trying to name a pitch, interval, or chord.

You: That’s nice. So what?

Me: Well, let me tell you…

I’ve been doing a similar exercise… also involving ears. I’ve decided to engage in IPA phonetic ear training as part of my language curriculum, as it were.

I’ve created an Anki flashcard pack, of – as well as other things –phonetic symbols to match to the sound and vice versa. It’s harder than it sounds. Like pitch, if I play an A (Do) I can tell what an E (Sol) sounds like, a perfect fifth; but I can’t produce an E from vapour: If I hear it absent of musical information, I can’t name it; neither can I produce it without a reference. This is a limitation of relative pitch.

On a guitar, I can play an E relative to other strings, but I can’t tell you whether the A is pitched to 440 (top) or 432 (bottom).

440 Hz
432 Hz

Of course, if you tell me the top sound is pitched to A-440 and ask if the second one is higher or lower, I can tell you that. Hooray for me. But if the A-432 was actually A-431, you’d have had me tricked.

You: Where’s this going?

I experience the same challenge in my IPA studies. In context, if I hear an open and closed O sound – ɔ and o – I can tell you which is which, but I haven’t yet mastered the ability to utter these in the wild. I might be able to manage a nasal O – ɔ̃ – but we still haven’t arrived at the neighbours – ɵ, ɞ, ɤ, and so on. Source. Here’s a random or at least arbitrary IPA site.

I wonder if you people have perfect pitch in this regard.