Some people love AI, some loathe it. I tend to be closer to the first group but I’m also well aware it’s not perfect and quite dangerous too.

Still, AI, like the internet, is here to stay. And improve. We have to learn to live with it somehow.

Those who love AI swear by it and do pretty much everything with it. They tell us we should rely on it for every little task we have on our ever-expanding to-do list. This is wrong. AI is still very imperfect. It hallucinates, gives false information, and sometimes tells you one thing only to do the opposite right after.1

Yet if you’re going the other way around and avoiding it at all costs, you slowing down your progress by a lot.

Here’s why.

We’re not in the 2000s anymore

I speak very highly of using “old” methods like writing, reading, learning grammar patterns and even using actual textbooks. These tools have created fluent people for decades. They’re worth using and relying upon because we know they work, even if they may feel boring at times.

I remember when I started learning English on my own. I had to download illegally TV shows then download subtitles separately, give them the same name and, finally, I could watch a TV show on a specific media player.

I also remember a barebone website that would show one hiragana at a time randomly and I could type below what I thought it was, as a way to learn them.2

These techniques already used tech but they were also inconvenient and each page loading took time (and don’t get me started on downloading the episodes, which often took days).

To get new listening practice files, I had to search on pages I did not understand at all and hope to get a file that’d be useful. Oftentimes, I ended up with a topic that bored the hell out of me, with no transcript whatsoever.

Now think about what AI can do to help you learn:

• Provide endless texts to practice your reading skills

• Create transcript to audio files (such as ElevenLabs or NotebookLM)

• Be a conversation partner at any time of the day to practice speaking

• Explain any grammar pattern in any language you want, with examples about topics you actually care about

• Create songs or podcasts to learn specific vocabulary in context (Suno AI and NotebookLM respectively)

The list goes on.

All this within seconds, any time, anywhere.

It can also answer any question you have from another resource and won’t judge you for stupid or repetitive questions.

While writing a letter in Mandarin for my girlfriend’s birthday, I questioned myself about the different ways to say “in particular.” Gemini told me that, apart from 特别 (which I knew), 尤其是 also existed. I felt it was strange to have the verb “to be” (是) in it so I asked numerous questions and was told it was a set pattern in front of a noun, unless it was followed by a verb. A few years ago, I would have wondered for hours or even days asking on HiNative someone the same question.

Did I ask my girlfriend later for confirmation? I did. But it helped me move forward with an answer that made sense.

Not using AI today is like refusing to use the internet to learn languages when it became mainstream. A waste of an incredible tool.

Don't rely on AI

AI is and should stay a tool like any other.

As you already read, I only ever call AI a tool. It’s not an assistant, friend, partner, let alone lover.

AI is imperfect. ChatGPT’s insane hallucinations have already become famous. Grok’s racist content did as well. Gemini’s “overcorrection” of historical images in early 2024 to create black nazis or Asian past presidents of the US were kinda fun though.

When it comes to language learning, each of these tool is just as imperfect. It often tells you fake information.

I couldn’t believe it when NotebookLM’s presentation showed me two images of a Go board which showed an identical situation for White but claimed above it was different. You can see more on this in the following piece I wrote about this tool (that is still worth using! Just not for presentations.)

AI chatbots also have a very particular style of writing or speaking. While you can prompt it to use a more natural way to speak, it’s only from an intermediate or advanced level that you’ll start noticing and it might be too late for you to fix the patterns you’ve gotten used to hear and read.

AI is great to complete your arsenal of resources.

Take your textbook and get more examples of a new grammar pattern. Get an AI transcript from your favorite podcast so you can study it. Request summaries of long texts you’re studying so you have an idea of the gist. Listen to AI podcasts about topics you’ve studied and want to dig into. Practice your ease of listening comprehension at your specific level so you’re ready for the next real-life experience.

Final words

Was “stupid” a tad too strong of a word? Some may argue it is because it could mean look down on those who don’t use AI.

Yet the truth is, it really is wasting an opportunity like no other.

I absolutely believe we should all learn to integrate AI in our learning journey. Those who don’t will, one day, look back and wonder why they didn’t start using it before.

Because, remember, the actual goal is for you to become better at the language. When you’re in a conversation with a native or reading a book, who will care whether you learned it with or without AI?

Use AI to learn. Don’t let it become the center of your studies. But, please, don’t leave it on the side.

Cheers for reading,
Mathias

I’ve wanted to write about how I use it and the many other ways you could use it for a long time but kept pushing it back as I wanted to make it the most comprehensive possible. Unfortunately—or fortunately?—, with all its evolutions, it feels like I’ll never catch up so here’s likely the first of many pieces I’ll write about it.

Let’s stop for a moment and tackle the most important question.

What is NotebookLM

As it states on its main page,

Meet NotebookLM, the AI research tool and thinking partner that can analyze your sources, turn complexity into clarity and transform your content.

Well, that’s not helping so let me try to be clearer.

NotebookLM’s original “selling” point1 is that it could turn any source (text, webpages, videos from YouTube, files uploaded) into a short lively podcast strictly based on your sources.

It’s now evolved into much more though.

Once the notebook is created, you can not only create a podcast but also prompt it to focus on certain aspects of your sources and choose whether you want it to be a Deep Dive, a Brief, a Critique, or a Debate.

If you choose English or some languages like French as the target language, you can even make it “Long” and get a 40-50 minute podcast! 😮

We’ll dive into this later since there are also other features to know of:

• Video Overview: Creates a 5-7 minute-long video serving either as an Explainer or a Brief to learn some basics on your sources

• Mind Map: Creates a mind-map with the key ideas from your sources. Great for digging into specific parts

• Reports: Creates a report about the topic

• Flashcards: Creates a small deck of flashcard to test your knowledge about the content

• Quiz: Creates a multiple-choice quiz to test your knowledge too

• Infographic: Creates a one-page document with the main ideas

• Slide deck: Creates a PPT-like presentation to get some visuals with your content

• Data table: Creates a data table, which is great for analysis but not much for language learning as far as I know.

This would already be quite good but it doesn’t stop here. There’s also a chat in the middle of the canvas to dig deeper and ask questions about the sources. This is what matters most though: it precises which source says what and adds a link to it if you want to verify.

This simply acts avoids too many hallucinations and brings trust into the picture.

I could go on for a while but let’s not beat around the bush any more.

Why this is an incredible tool for language learners, heh?

Unlimited audio content

One of the main problems when we want to improve our listening comprehension is the sea of choices.

Should we watch a new YouTube video? Find another podcast episode? Rewatch a movie or TV show? Listen to more music? All are valuable options.

As a beginner learner, exposing our ears to a variety of audio is an incredible tool to get used to the sound and rhythm of the language. Once that’s done, however, the variety can work against us.

A variety of topics also means the words and expressions change a lot. Context helps with understanding but it doesn’t allow us to cement words in our brain.

NotebookLM solves that problem.

Let’s say you love the game of Go. Well, you can ask NotebookLM to find sources about the game of Go and create a podcast in your target language. You can choose first to have a Brief so it’s not overwhelming to work through. You can then request a longer version and then again another one a specific topic such as playing on a 19x19 board instead of 13x13. Finish up with a Debate about opening moves and you’re golden.

By reusing the same sources and creating multiple podcasts, you get to hear the same keywords in different sentences, said by the man or woman, with different intonations. This reinforces the vocabulary through exposure.

Dig deeper with other tools

AI podcasts were the part mainly put forward by NotebookLM originally but the other tools added are nothing to look down on!

While I’ve found the video overview, slide deck, and infographics rather dull (and even showing completely wrong Go boards exemples!), the rest is worth looking into.

First, I find the Reports to be the best “summary” of dozens of sources. They are usually a few hundred words/characters and well organized. As such, they can be a great tool to study through the topic and discover key vocabulary ahead of listening to a podcast.

The Flashcards and Quiz are a good way to test yourself on the knowledge you acquire through the study of podcasts or reports. I personally prefer the Quiz because it’s a multiple choice system and, if you click on the wrong answer, you get a one-sentence explanation of why this wasn’t the correct sentence and why you should have chosen the correct one.

Finally, the Mind Map is likely the most useful tool to dig deep into the topic. It can help you go from “knowing” about the topic to mastering it truly.

On its own, the mind map only shows keywords but if you click on them, it open the chat in the middle and provides a details explanation of that specific point.

Sure, that can be overwhelming at first but I find it to be a good last step before testing myself and moving on to another topic.

Using NotebookLM in original ways

I love to play around with AI tools to find ways it was not planned to be used. NotebookLM is no exception. Like others AI tools, it shines when used in new ways.

Last year I began reading a short graded reader in Chinese. I wanted to use the dictionary as few times as possible so before diving in, I added the audio into NotebookLM and requested a few things in that order:

• A 1-min brief podcast in Chinese to get the gist in the target language needed

• A 10-min podcast explaining the story in another target language I speak well: Korean

• A 10-min podcast explaining the story in Chinese

I then listened to each 3-4 times through out the days I was reading the story. This combination allowed me to recognize more words in Chinese as I kept going. Here’s the Chinese brief if you’re interested:

This worked well because I had the audiobook version of that graded reader, but you can also use that to work on long texts you find online and help,

• Get used to reading the news

• Work on long texts online such as short stories or even papers if you’re at that level

Another useful way to use this tool is also for audio files you may have without transcript. After adding that file (or link if it’s a YouTube video for instance) as a source in NotebookLM, it automatically creates a transcript of the entire audio for itself to refer to. The this is, you can click on the source and see that transcript too!

I’ve even once downloaded a podcast I had created with NotebookLM and added it as a source to see the transcript as I listened to it again.

Since you only get a few podcast creations per day as a free user, you may have to be careful what you want to create. But adding a source doesn’t count as requesting anything from it so you can do it as often as you like! 😁

I’m sure there are more original ways to use it so if you have some, please let me know!

Leave a comment

NotebookLM as a language learner

Some people say AI will kill language learning. I reckon fewer people will try it indeed. That’s a real shame for sure. But I’m also sure more people who will start learning foreign languages will keep at it and actually become proficient to the level they aim for.

This piece is only about NotebookLM but look at all it can already do!

• Provides unlimited audio content about topics you choose → Great for listening practice 🔉

• Provides unlimited written content to work through → Great for intensive reading practice 📖

• Helps get transcripts for audio files → Great for bimodal input and especially for languages with a different script such as Chinese, Japanese, etc. 📖🔉

• Helps practice writing in the chat to ask questions → Great for writing ✍️

• Provides unlimited testing options → Great to verify you actually learned about the topic 🧠

Downsides of NotebookLM

Now, it’s not perfect. Nothing is.

As I mentioned earlier, the video overview, infographic and slide deck options are potentially good tools but I don’t believe NotebookLM is there yet when it comes to visuals. There are often typos, and images may actually lie

I’ve switched from the paid version of ChatGPT to Gemini late last year so I don’t have that anymore, but the free tier limits you to only a few podcast creations per day.

In a way, that’s useful not to overwhelm yourself. But it also limits you as you must choose what you generate carefully. And if you’re not satisfied, well, tomorrow will come soon enough.

I also find it a bit restrictive that you cannot change the voices at all. They’re not bad but some variety wouldn’t hurt.

Another “downside” is an incredible option that currently only exists only in English: jumping in. If you generate an English podcast, you can interrupt the speakers to ask “them” to explain more clearly a point of focus on something else. Once this is added to other languages, this will make NotebookLM a great tool to practice speaking too! 🗣️

Finally, there used to be a button “Discover Sources” which was great to let NotebookLM choose a random topic. This seems to have been replaced with a search web interface. It isn’t so bad because at least you don’t end up on strange topics you’re not interested in, but I kinda liked the surprise too.

Final thoughts

I’ve shared a lot and still feel like I’ve only scratched the surface.

NotebookLM is, by far, my favorite tool now.

I use it daily whether to just get some listening practice or review some vocabulary in a topic I’ve studied.

Sure, it’s might feel overwhelming for beginners or for first time users, but it’s worth starting early to get your ears used to the language and potentially dive into the topics you love most.

For intermediate and advanced learners, however, my god, why aren’t you using it yet! Hop hop! Get to it, here’s the link.2

Cheers for reading,
Mathias

I’ve seen too many people avoid AI like the plague, trying to convince themselves AI is a gimmick and won’t replace them. This is pure delusion.

No matter what we do, AI will, one day, be able to do our work. It’ll happen in a while for some, but many will start to suffer from it soon—if it hasn’t happened to them yet. That’s just reality.

I’m also convinced AI will improve every life on Earth in one way or another. Whether it be by being an assistant in some aspects of life, helping treat diseases, or even finding ways to improve our energy consumption.

Kids have it good now. They get to have the time to grow and become surrounded by AI before they get stuck in their ways.

Us adults aren’t so lucky.

We don’t really like nor excel at revamping the way we live, whether we admit it or not. AI’s development these years is likely only the start of all the changes we’ll see in our lives. Everything will change.

Let me take myself as an example.

I work as a project manager handling translation projects into 14 languages. Managing such projects is mostly about figuring out how to assign resources (ie. translators and reviewers) so we can meet deadlines on time and prioritize the right content. There’s obviously a human aspect to this. As of right now, people don’t trust AI much,1 but there’s also no denying a well-set AI could likely consider more variables. For instance, while I usually remember holiday dates for Japan, Korea, or Christian holidays in Europe, I’m always surprised when I get reminded there’s a holiday coming up in Brazil and deadlines need to be adjusted accordingly.

Similarly, I speak 5 to 6 languages and have dabbled into over 20 more. The languages I speak really well are part of me. They’ve helped me meet people from all around the world and connect with them. By avoiding relying on a “middle language,” people are more capable of sharing their thoughts. AI can already do instantaneous translation on a call or translate written text within seconds.

I’ve also written tons of articles on Medium and kept this newsletter for years now. I like to think my writing skills have gotten better but there’s no denying AI can write better pieces than I can. If I wanted, I could even use some tools to have it write in my style.

Now.

Does this mean I’m useless? Speaking foreign languages, writing online, and working as a project manager could be done by AI.

Asking AI to write for me wouldn’t feel like I am connecting to you, though. Relying on AI to speak a foreign language isn’t enough to create the emotional connection because languages are more than just words to translate. It would also not allow me to feel the pleasure of learning said language. Relying on AI to handle projects could be possible but it would lack the human touch people require of a project manager.

AI could do everything. It just shouldn’t.

So what are we supposed to do?

Learn to use it.

While I don’t write with AI, I like to copy-paste it for feedback in ChatGPT, even if I may not apply it because it’d make my piece too dry.

While I don’t use AI to speak foreign languages with people, I use it to learn languages, to practice, to teach me. Here’s just one example:

While I don’t use AI to manage my projects, I use it to verify my answers to queries or when I need help with figuring out a new tool at work.

AI is bound to keep developing. It’s easy to bury your head in the sand. It also won’t do you any good.

Start small. Experiment. Find small ways it could help you daily. And ramp up.

As a language learner, here are just a few ways to start:

• Create sentences in your target language, get feedback, and ask for alternative ways to say the same thing.

• Ask for a short story to practice reading.

• Ask for an explanation filled with examples of a grammar pattern. Request examples to be about a topic you love so you can remember with ease.

• Practice short conversations with the audio feature of ChatGPT/Gemini.

A new world is rapidly unfolding ahead of us. Don’t let it leave you behind.

This has been a long time coming.

As a paid ChatGPT subscriber, I’ve been using ChatGPT’s Voice system to have conversations for quite some time. I started about a year ago with the old voice system we couldn’t interrupt, which still had a rather robotic voice.

It wasn’t the best but it was still quite useful as listening practice for my more advanced languages: Japanese and Korean. Especially during walks.

I asked it to create fake news articles in the target language and proceed to ask me questions about their content. Sometimes, I’d give it a specific topic, such as environment or politics, to learn new vocabulary. Others, I’d ask it to create a much longer news article. Yet others, I’d request it to create a conversation between two people on the topic of that fake news piece.

It was useful but limitations were clearly there:

• I couldn’t interrupt it to ask for the translation of a specific word

• It never felt like a real conversation

• Lower-level languages were out of the question because it wasn’t capable of slowing down its talking speed, something ChatGPT now can do.

Also, it regularly struggled to understand words and responded way off base. I get I was speaking with an accent but this has barely happened once or twice since the new Advanced Voice for ChatGPT.1

Things have evolved quite a bit since then.

Subscribe now

Some caveats

Only two AI tools offer this type of live conversation at the moment: Gemini (from phone only) and ChatGPT.

Gemini is entirely free but misunderstands or needs refreshes often. ChatGPT allows for about 15 minutes or so daily for free users and one hour for paid users.2

Both tools write down the conversation in the background, which is convenient to refer to it afterward for review. But it often miswrites or writes “Transcript Unavailable” for the speaker’s text, even if it did reply on point at the time.

For Japanese, it often writes the wrong kanji despite the context being clear. For example, in a conversation about blind people and braille, it wrote moujin as 毛人 (non-Yamato people in northern Japan) instead of 盲人3, and tenji as 展示 (exhibition, display) instead of 点字 (literally “dot character”).

Alright, let’s dive in.

Vocabulary-learning heaven

No language partner is an expert on every topic you can imagine.

If you want to develop your vocabulary about traditional Korean architecture, you might find one person who can do that but would they also be able to discuss the Japanese braille structure or advances in space as well? Likely not. Even if they could, they’d need preparation to look up a few words/expressions.

With AI voice tools, you can decide in the middle of the conversation that you’re done with one topic and switch to a new one.

You can also ask it to explain new terms in the target language, your native language, or any other language you might know. I often ask the Japanese equivalent for Korean words I don’t know because they’re often similar which allows me to strengthen my Japanese at the same time.

You can also finish the conversation and ask it to create a table filled with the vocabulary you asked about. Or you can ask it to figure out the important expressions on its own, as I did below.

You can then ask it to create Anki (or other SRS) cards to import with a specific format you have in mind. That saves time creating them—although creating them yourself also has its own benefits. Combine this with a tool like ElevenLabs to get an audio version and you’re golden.

One rule I still struggle to get ChatGPT to consistently follow is the request to correct me as we go. Many of my sentences will have mistakes. Some bigger than others, especially in some languages like Spanish or Korean. Sometimes it does it but it’s inconsistent so I need to try a few more things, notably defining more precisely what should trigger the correction.

One roundabout I found is to ask it every few exchanges to fix my mistakes. I wish it were a bit more strict but well. If you have ideas to make it work, let me know!

Leave a comment

Instant grammar tutor

Sometimes ChatGPT replies to my audio messages with a grammar pattern I recognize but am not sure of. In these situations, I switch to asking for more example sentences using it or directly explaining it in English if I’m really lost.

Once I’ve got it down, I can just ask it to pick up where we left off. Even if I forgot where we were at, the AI does remember.

Pronunciation practice support

This isn’t quite there yet for all I wish I could do, but I have no doubt 2025 will be the year we’ll start getting this so I’m adding it.

One of the main reasons people—and by that I mean “I”—fear speaking a language they’re learning with native speakers is their pronunciation. They’re afraid people will make fun of them so they—again, I—miss the opportunity.

Well, AI won’t make fun of you. It might misunderstand and frustrate you but at least you’ll have all the time you want to make it repeat or rephrase so it understands.

Unfortunately, it’s also too kind.

I’ve asked it dozens of times to correct my pronunciation, rhythm, and so on in Korean but all I got was some sort of “it’s really good!,” even if I asked it to be strict. I even stumbled three times on a specific word on purpose only to hear it say another one could have been said more clearly.

In the near future, I’m sure we’ll be able to do this but that’s not here yet.

In the meantime, I sometimes ask it to create a conversation or monologue in Korean so I can repeat it multiple times and try to have a similar rhythm, take a breath at the right times, or at least not stumble in the middle of long sentences.

Extend knowledge in a new language

AI Advanced Voice tools like Gemini and ChatGPT allow us to dive into new topics we hadn’t even thought of.

I sometimes start the conversation by asking ChatGPT to provide three or four rather specialized topics, and then I choose one or ask for more. All too often, we stick to topics we know. This allows me to learn new things. Even if it may make mistakes, it doesn’t matter because it usually gets the basics right, and if I truly get interested, I can look it up afterward.

Combining language learning with curiosity is a sure-fire way to not get bored!

Anyway, that’s it for today. Time to go have a quick chat in Korean about tatamis around the world.

Cheers for reading!
Mathias

I woke up to yet another news of a new AI tool.

Nothing surprising. Tons are coming out each day. FutureTools lists 2,919 as of today. One more, one less, it often doesn’t matter. But this one blew my mind instantly. I felt its potential. And to top it all off, it’s free and has a slick UI—something that I love.

Say Hi to NotebookLM.

This new AI tool allows anybody to create a podcast based on the sources we feed it. These can be pretty much anything: PDF, audio files, YouTube videos, text we paste, or even entire web pages. Or a combination of all this!

Now, why does this blow my mind? Because it allows for the digestion of information in a much more convenient and fun manner.

Learn anything with ease

Say you’re learning about a topic in history. Let’s take the history of Taiwan. I can feed NotebookLM with this Wikipedia article, this overview from ThoughtCo, and the text from this Britannica article. I click “Generate” and the below conversation gets created in a few minutes.

Now, if I want to dig deeper into those sources, I can click on a suggested question or type my own question in the chat at the bottom. The answers I receive also include links to the sources, indicating which part of the text made the bot provide this or that response.

I can then either keep asking questions or “Save to note” so I can refer to it later. I can also click on some of the predefined buttons:

For History stuff, the Timeline is perfect as it lists the dates of what happened and the notable figures. The Study Guide gave me 10 short-answer questions, then the answers, then 5 essay questions, and finished with a glossary of key terms.

You can check the Notebook I created here to see what the other presets do.

NotebookLM can be extremely useful to learn pretty much anything:

• Want to dig into research about a topic you’ve never touched? You can get some structured and slightly dumbed-down data from this.

• Want to understand a topic you can’t wrap your head around? Feed a bunch of sources that explain it and listen to the podcast before asking your questions in the chat.

• Struggling to review a topic for an exam? Answer the questions you’ll get from the Study Guide.

These are just a few of the potential use cases for such a tool. And, you know me, of course I thought of language learning too.

Language learning galore

Let me start with the sad news first: this only creates an English podcast and will always reply in English too. No practicing a conversation in a target language here.1

Want the good news now? Well, you can feed it sources in other languages.

Yes. This isn’t perfect and isn’t so immersion-like, but this makes it and incredible tool for low-intermediate to high-intermediate level learners. Basically, it can help you understand something that would have been way out of your league.

Let me give you an example of my very first test.

I bought a Graded Chinese Reader in 2013. I tried it back then but it was way out of my level. Still, it had audio files with it so I thought it’d be useful someday to read while listening. Despite the years, I still consider it complicated to read and understand. Stories and long and include vocabulary I don’t understand.

Got an idea of what I did?

Well, I fed the audio file to the NotebookLM and listened to a summary of the story in English so I’d know where the story was heading. It wasn’t perfect but it was short and helped me get started.

I followed up by asking the chat to provide a table with the difficult keywords from the story, their translation, and the sentence it appeared in. I had to “negotiate” a bit here and there to explain “difficult” and “keywords” so I’d get a useful table and finally got this.

I could then save it as a note and go back to it whenever I want.2

Now, onto another sad news: when checking the table in the note, it looks like crap currently. I reckon this will get fixed sooner than later but, in the meantime, I’m saving the answer in Obsidian as it uses Markdown and so does NotebookLM.

Now, I can use this tool to refer to my source whenever I struggle. And when I’m done with reading the story, I can request questions with the Study Guide and try answering in Chinese to practice my active skills.

That was ONE example.

Now imagine what else we can do with this:

• Add long news articles to have some knowledge of the topic before reading

• Fill in podcasts or YouTube videos in your target language and pull out grammar patterns and vocabulary, or get questions to practice your understanding or what you saw/heard.

• Open the source on the left, highlight some text and click “Help me understand” to get an explanation of it.

And so much more.

Final thoughts

This is just the beginning. This is as bad as it’ll ever get!

I’m looking forward to how this will evolve and all the use cases we’ll find with this. And there are other new tools that may help even further! For instance, I found PDF2Audio, a free tool that turns PDFs into podcasts and can even do it in French, German, Spanish, Portuguese, Hindi, and Mandarin. I haven’t succeeded in making it work yet but it’ll be a great supplement too.

I’ll keep experimenting with it and will provide as many possible ways to use it as possible in the next few weeks.

Let me know if you think of some others as well! I’d love to see if you’ve got others I haven’t thought of yet!

Leave a comment

AI really can be an incredible tool.

People can say it’ll ruin language learning—and even learning as a whole—but tools like these ones really make me think they’ll only help us become better faster.

All we have to do is use them.

Cheers for reading,
Mathias

There’s been a lot of discussions about AI testing at work these days and that has made me think about all the efforts I’ve put, on my own learning path with AI.

One thing is clear: figuring out AI’s a mess.

Understanding Large Language Models (LLMs) is like watching a wizard cast spells: the outcomes are astonishing, but the methods remain an enigma.

Nobody truly knows what goes on behind the curtain. This is why we keep discovering “new” strange tricks to get better results. A recently published study titled The Prompt Report: A Systematic Survey of Prompting Techniques even showed that erasing names to create anonymity from an email in a prompt decreased the model’s accuracy.

Even the “best” prompters still get crappy results often.

Still, there’s no denying LLMs like ChatGPT or Claude are things of wonder. They are incredible tools worth using for anything.

Language learning including.

The main reason people get bad results

You will never get anything good out of an LLM if you don’t know exactly what you’re trying to get out of it. It’s that simple.

Rare are the people who haven’t heard of these tools.

Millions of people have tried them, hoping to find something funny or useful.
Most of them had a laugh or dropped an “oh, wow” and never looked back.
Many played with it and thought it could be useful but forgot to go back to it.
Some dove deeper and found a way to use it for one specific purpose.
A few dug even further and realized it could revolutionize everything they do.

Nowadays, my first thought when I’m stuck is “Could ChatGPT help with this?” Often, the answer is a resounding Yes.

If you’re opening ChatGPT and ask a basic “Help me learn German”, you’re bound to get a very answer.

If you instead ask something like “I’m learning German and struggling with the dative. Help me understand how it works through a short story in German with tons of words used in both dative and nominative forms.”, you’re bound to get a more precise answer.

An LLM response is as good as your prompt is.

If your prompt is vague, so will the answer.

As I mentioned a while back, the goal is to be precise. You can think of ChatGPT as a very smart person who can give you any answer but doesn’t get all the underlying aspects you take for granted.

Knowing why you choose to use an LLM, however, is not enough to figure out how to use it well.

A systematic approach

The most common method people use is to try widely different prompts and see which works best but that’s where the first error lays.

See my example above? Well, that’s a bad one. I’m comparing a very short and vague prompt with one a lot more precise. A better version could have been to compare “I’m struggling with the dative case in German. Help me understand it.” At least, this way, ChatGPT would have known what I was aiming for.

And yet, it’s still a wrong comparison.

Nobody knows what happens behind the curtain but we’ve done enough research to know that, for some weird reason, any small modification impacts drastically the result.

Change the order of sentences and you could get a very different result. Add a “Please” or a “Thank you,” and the same could happen again. Some experiments even found that asking the LLM to “Take a deep breath first” could help. How crazy is that!

If you aim to find the “best” prompt for your specific purpose, your experiments must be done one step at a time.

Just like elaboration is a wonderful technique for expanding active language skills, it also helps with LLMs.

Elaborate slowly.

You can start with one specific prompt and slowly tailor it to see if the results get better.

If we were to go again from my second prompt, we could make it evolve as follows in bold:

I’m learning German and struggling with the dative. Help me understand how it works through a short story in German with tons of words used in both dative and nominative forms.

↓

You’re an expert in teaching German and I’m learning German and struggling with the dative. Help me understand how it works through a short story in German with tons of words used in both dative and nominative forms.

↓

You’re an expert in teaching German and I’m learning German and struggling with the dative. Help me understand how it works through a short story in German with tons of words used in both dative and nominative forms. Keep the sentences short.

↓

You’re an expert in teaching German and I’m learning German and struggling with the dative . Help me understand how it works through a short story in German with tons of words used in both dative and nominative forms. Keep the sentences short.

Here’s an example of a “good-sized” sentence: Ich habe eine Mail von den Kindern bekommen.

↓

You’re an expert in teaching German and I’m learning German and struggling with the dative . Help me understand how it works through a short story in German with tons of words used in both dative and nominative forms. Keep the sentences short.

Here’s an example of a “good-sized” sentence: Ich habe eine Mail von den Kindern bekommen.

Follow the story with a short exercise in which I should choose the case for one noun in each sentence.

↓

You’re an expert in teaching German and I’m learning German and struggling with the dative . Help me understand how it works through a short story in German with tons of words used in both dative and nominative forms. Keep the sentences short.

Here’s an example of a “good-sized” sentence: Ich habe eine Mail von den Kindern bekommen.

Follow the story with a short exercise in which I should choose the case for one noun in each sentence. Don’t provide the answers until I ask for them.

I could keep going but you see what I mean.

In fact, the one thing I should have done that I chose not to do to avoid making this too long, is to add more examples of good sentences.

Hell, an even better solution could be to test things until I found an answer I liked and then copy-pasted that answer in my prompt while saying “Follow the below example of a good answer.”

In this case, the answer I got for the last prompt was quite satisfying so I kept it as a good example to reuse someday.

The only thing I think I could improve1 would be to request the exercise to use different sentences than what’s in the text.

The more systematic you are, the more chances you have to find what works for you. Is it fun? Can be. It’s not always though. It can get frustrating.

Still, there’s no rush. You don’t have to find the perfect prompt today.

It’s a conversation

One last thing to remember is that AI is not a one-off kind of system. As of right now, LLMs struggle with very long prompts, suffering from what’s called the “lost in the middle” phenomenon.

This problem takes dramatic proportions for prompts larger than most of us will ever need, but it’s still worth knowing and being careful of.

That’s why, while it’s important to craft a good first prompt to start the discussion on the right foot, you should remember you can tweak the conversation with ChatGPT and so on as you go.

Most of the great insights I’ve received from ChatGPT since I started using it came not from the first answer but from further in the conversation. It was always after I told it where it had misunderstood me, or where it was lacking precision.

TL;DR

So, yeah, do your tests with ChatGPT. Experiment. Keep what got you good results. Experiment some more. Tweak your prompts.

But if you want to know for sure that a certain prompt works well, do it step by step.

Take your time.

And keep the conversation going.

You never know. The next answer could be exactly what you’ve been dreaming of.

In the next The Language of AI2, I’ll share a few prompts I’ve used for my language learning journey to work on specific skills.

If you’re interested in that and you’re not subscribed, now’s the time 😉

Subscribe now

As always, cheers for reading,
Mathias

I hate how negative the world’s become.

Awful things are happening around the world and evolutions worth worrying about. True. Do we have to put the emphasis on each one and ignore the positives though? Can’t we find the light even in darkness?

As a muggle passionate about AI, I keep seeing new advancements week after week.1 Unfortunately, most of them get buried under a sea of people claiming the end of the world is approaching.

There’s no more common topic than all the “AI will ruin education” rhetoric.

So when I saw an article titled “The End of Foreign-Language Education,” I knew I was in for some blood-boiling reading.

This is my way of correcting that piece and saying AI will help language learning, not kill it.

Correlation doesn’t mean causation

The article shares the following depressing statistic from the Modern Language Association (MLA): “Total enrollment in language courses other than English at American colleges decreased 29.3% from 2009 to 2021.”

The author then shares a few more stats from other countries about this decrease to showcase how international this situation is.

The problem? ChatGPT, the first widely-used AI tool, only came out in late 2022. The decrease started long before AI became accessible to the masses.

The author finishes this part by recognizing other factors may have impacted this drop but also stating:

That’s a correlation. True. But I can’t agree automatic translation is the cause.

People relying on automatic translation usually do it on holiday abroad. They use OCR to scan and understand their menu or give directions to the driver. I did it a lot while in Vietnam.

Nobody ever went to university to learn a language just for a trip.

The biggest enemy here isn’t AI. It’s how we try to convince people to learn languages.

Let’s turn back to the paper with the depressing statistic mentioned above. Well, if you keep reading it a bit, you can see this decrease isn’t happening for all languages. ASL enrollments increased by 0.8%, Biblical Hebrew by 9.1%, and Korean by a whopping 38.3%.

There was also an increase of enrollments at the introductory level for quite a few languages (Japanese, Korean, Ancient Greek, and many more less commonly taught languages).

It’s not all bleak.

And, again, it’s not AI’s fault. It’s the Education system’s fault

Maybe it’s time to make learning attractive?

I can still remember parents and teachers saying Wikipedia would ruin Education and us kids wouldn’t learn anything anymore because everything was available online.

Well, I began learning because the internet was accessible. I loved learning online because I could make errors without feeling stupid in front of a class.

It’s too easy to blame AI for this decrease in foreign language learning enrollments. The problem is simply that we’ve all been scarred by (at least) one crappy language teacher back in middle or high school.

Hell, some teachers even succeed in making learning as a whole boring. That’s why more and more people turn to the internet.

The goal of a biology class shouldn’t be to teach biology per se. It should be to make the students curious enough to want to dive into it outside of class too. The goal of a language class shouldn’t be to teach grammar patterns. It should be to show the students how incredible knowing that language could be.

The focus is all wrong.

Sure, we should set some ground rules. But why do classes have to be the “standard”? After all, students of all levels comprise every class. There will always be people thinking a class is too slow and others thinking it’s too fast.

In my recently-started class about Translation and Interpretation from Korean to French, some other students told me the problem with the King Sejong Institute’s classes was that they were too fast. Yet, I find it wouldn’t hurt if we went faster. Our levels are just different. And it’s okay.

Instead of trying to teach all the content of a textbook in the classroom, we should be shown how to use it on our own and, on the contrary, use AI to complement our studies at home. Then we can ask more questions in class and create a real exchange with the teachers. Not a one-direction system that serves nobody.

This applies to every subject.

I don’t remember anything from my biology classes because I hated the teachers I got. I do remember what a friend who was later doing a PhD in DNA reproduction told me during a few parties about biology. I even went to look a few things up after because she had made me curious.

Curiosity matters.

In her piece, the author even cites Jen William, head of the School of Languages and Cultures at Purdue University, that schools should “stress more than ever the intercultural components of language learning that tremendously benefit the students taking these classes.”

That’s it.

People need to do it because they want to.

A crappy reason means a crappy resolve

The article opens with the author sharing her shock at hearing an AI-cloned version of herself speaking Mandarin flawlessly despite her “studying the language on and off for only a few years.” Sure, the technology has gotten surprisingly good. I remember my surprise at hearing my voice with ElevenLabs.

It’s her reaction that’s wrong:

A wasted effort? This made me question why she even began learning Chinese in the first place so I looked her up a bit. Apart from this article, no trace of her study of Chinese appears online.2 Only that she’s “a journalist covering technology & China.” So, yes, she might have learned some Mandarin but it sure feels like she’s not learning it for a deeply ingrained want or need. She could be but that line makes me doubt it.

In the end, learning anything—and languages in particular—takes time. A whole lot of it. Why should we be surprised about how good AI is? It’s in the name: Artificial Intelligence.

It’s ok to have the AI be better than us at speaking a language.

What matters is the experience.

Nobody reaches a high level of fluency in a language just to understand something. It may start like this, but what makes them continue is the connection they build with the country, its culture, and its people.

It’s the experiences they live connected to that language.

I’ve been learning Japanese and Korean for about 15 years and still get surprised by both languages. I love seeing them evolve and discovering more of the cultures associated with them.

It’s okay to have fewer people learn languages

This may feel contradictory with my stance on language learning as a whole. After all, I still stand by what I said a few years ago in an article called “I dream of a world full of polyglots.” I truly do.

I’m also deeply aware this will never happen. And forcing everybody to learn languages isn’t so useful (nor does it work).

We shouldn’t try to make everybody learn languages. We should try to motivate as many people as possible to try it. Let them decide if they do want to keep going.

I’d rather have 5 people learning languages for years than 20 feeling pressured into doing it and not even trying to remember anything.

I love language learning. I would love for everybody to realize how incredible and life-changing it can be. But it’s delusional to think it’s possible.

Instead of forcing people, let’s empower them to make their own choices.

And for those who do decide to learn languages, they’ve come at the right time.

AI is making language learning a lot more active and efficient.

As usual, thanks for reading!
Mathias

What’s your position on this? Do you think AI will kill language learning or help it?

Leave a comment

PS: A side note I had written got deleted by accident and I can’t seem to remember how I wanted it to fit so here it is: It’s not because there are fewer class enrolments that language learning is decreasing. Online communities have grown and so many tools to learn without classes now exist that many people consider classes as less important or not needed at all.

I know what tools I prefer (notably the ones I mentioned here), the ones I think have some potential for later, and the ones I’m curious about but won’t serve me in any shape or form anytime soon if ever.

Still, I’d be lying if I didn’t admit there’s too much going on.

Don’t get me wrong. I love that AI is evolving quickly. There’s always another discovery to make and new wonders to discover. Oh, and new worries to have as well but I believe the positives outweigh the negatives.

But it’s just so much, it’s impossible to know all about it and keep track of every little detail happening.

That’s why it’s important to choose what to keep an eye on.

What matters to you?

Unless you work in tech, there’s no need for you to keep track of every little AI advancement. If you love it, go ahead and spend some time following every detail for a while but you’ll get overwhelmed one day and wonder why you did so.

Hell, that’s what prompted me to think of writing this in the first place!

You see, I love AI but I also have a life to live. I used to watch 5 or 6 different YouTube channels about it—most posting at least 3 times a week and some even daily 20-ish minute videos—, check 3 different news websites, lurk on Twitter to find some hidden gems, and even read some research papers. Then I went on holiday for about 10 days and didn’t know where to start from again.

I realized I had come to try to understand every single detail, not because I wanted it but because I didn’t want to miss out on a potentially mindboggling discovery, even though AI news is just like any other news:

What everyone needs to know will be loud enough for everyone to know.

Just like it was impossible not to hear about Russia invading Ukraine, or the war in Gaza, or the Burmese Coup d’Etat in 2021. Even if you didn’t follow the news at all, you heard about them.

When ChatGPT came out, you heard about it.

Important news is loud.

Now, I don’t want to wait for big stuff only so I’m still checking AI news. I only use one news website (the-decoder) and skip 75% of the articles. I also only watch Matt Wolfe’s weekly roundup whenever I see something big happened that week in AI. I also sometimes watch Yannic Kilcher when a paper I’m curious about pops up. And then I fall into some rabbit holes on AI topics I want to know more about.

All this allows me to know more than most people without spending hours daily to stay on top of it all.

If you’re reading me for AI…

I’ve got some good and less good news.

I’ve decided to do my part in helping to avoid AI overwhelm. I know 2 pieces about AI a month is not that much but it may still be a bit much. I also don’t want to write too much about “news” but rather about what matters:

How to use AI.

I want to give deep-dive examples of how I use it to learn (languages in particular but other topics too), how some others are doing it too, and why it all matters so much.

There will be one more piece (about why AI won’t ever ruin language learning) in two weeks as usual and then The Language of AI (TLoAI) will become a monthly section.

As you know, I’m currently alternating between pieces for The Average Polyglot and ones for TLoAI. Making TLoAI a monthly section gives me one “free” week during which I may write about AI but I may also write about something completely different.

For example, I’m planning in June a piece about a Lesser-Known Language in Japan (that isn’t Ainu nor a Ryukyuan language). I’m also going back to the short-lived Dazzled by Korean series I started over a year ago. These will come on random days but they may take that “free week”’s spot.

Of course, you’ll only receive the AI pieces if you subscribe to this section alone!

So, here’s a summary:

• Good news: higher quality and less “surface” stuff related to AI 🤓

• Less good news: rarer pieces about AI from me 😁

• Gooder2 news for those who follow me for my thoughts in general: more variety each month! 😸

Cheers for reading as always!
Mathias

Language Learning just keeps on getting more convenient.

As more and more people try to convince us AI will render language learning useless, it keeps giving me more reasons to learn languages.

Still, I’m in my early thirties so I need some analog along. I can’t consider myself as “learning a language” if I don’t have a single book, textbook, or notebook about it.

I know that’s a strange requirement but it’s still there.

But, as you probably already know, I’m also a sucker for a good AI tool that’ll help make my life easier. So here are my four favorites. No time to waste, let’s get the obvious one out of the way first.

ChatGPT

You knew it was coming before you even started reading. ChatGPT is a no-brainer for anybody using AI tools. And as a user of GPT-4, there are even more reasons to use it.

Many people have talked about using ChatGPT to learn a new language. It can be a teacher, a tutor, a language partner, a “native” friend, and so much more.

I’ve used it to create stories, explain grammar patterns, teach me related vocabulary, simplify or complicate texts to study, and even translate stuff.

With GPT-4, I’ve even had fun creating images reflecting images or stories I was studying and—the other way around—asked DALL-E 3 to explain a picture I had uploaded.

ChatGPT is a must.

Or at least a tool like it is. After all, recent reports say that Claude 3 is a better option than the free version at least. And it seems the new Pi.ai is too.

• Pros:

  • An assistant with quick and personalized answers

  • Tons of tips online on how to use

  • Full-audio conversations possible on the app

• Cons:

  • Likely to be upstaged at any time

  • Can provide false facts so complicated topics might need verification

  • Can be difficult to make sure directives are followed

ElevenLabs

I had tons of fun with this tool in October when I got a one-month premium subscription for a dollar. ElevenLabs allows us to get realistic audio for pretty much any text in quite a few languages, and even with our own voice if we want it.

So I cloned my voice and got some audio in a few languages.

These are the Chinese and Spanish versions of the same text:

Once my premium month finished, I stopped using it thinking 5,000 monthly characters was not enough and I’d likely get frustrated if I couldn’t get the audio I wanted.

Well, I’ve now been using the free version for almost a month, getting recordings only for sentences I write in my notebook. I’ve tried a few voices but ended up with this deep one that’s easy to understand and not too far off my pitch either.

I use the audio from ElevenLabs to practice shadowing—ie. repeating out loud sentences over and over to get the flow and pitch right—during study sessions. Then, I listen to them in the background while I walk outside or do mindless tasks.

It’s scary to think we’re only at the beginning of such tools and the quality is already that good. But it’s also such a useful tool to get to hear our target language for free and improve without any potential shame of mistakes.

That’s why I love ElevenLabs.

• Pros:

  • Hundreds (thousands?) of voices available, of all style

  • 5,000 characters for free per month

  • Almost instant audio version for any text entered

  • Possible to download the mp3 version of the audio

• Cons:

  • Can change to an Americanized accent for long texts

  • Only 5,000 characters for free per month

  • Impossible to clone own voice for free

  • No app

And yet, even if I love it, I’ve now also been using another audio-creation tool for a few weeks.

Suno

I’ve never dived deep into Suno but I’ve mentioned it, mostly when I spoke about Copilot.

Suno, from the outside, is not anything near a language-learning tool. It’s an AI that creates songs from a prompt. Great. But how’s that gonna serve us learners?

You know you’ve listened to songs you loved on repeat.

Well, it turns out there’s a custom creation section where you can set your own verses and chorus. Prompt it with a style you enjoy and you’re ready to hear a song made for you specifically.

You could be listening on repeat again and again to sentences you want to master, words you struggle with, or even grammar patterns that won’t stick.

That’s why I now combine Suno and ElevenLabs to truly engrain in my brain the most important sentences I write in my notebook in Chinese.

Here’s one song I got when I asked for a “traditional Chinese folk mellow” song and entered some sentences from the New Practical Chinese Reader book:

And here’s a random song I got by asking for “An upbeat rock song in Chinese from Taiwan (Traditional characters) using the word "雖然" (“although”) multiple times.”

Isn’t that the dream?

• Pros:

  • Creates two songs each time so you’re bound to have one you at least slightly enjoy

  • Custom version to choose the specific lyrics

  • Accessible from Microsoft’s Copilot app/website

• Cons:

  • Only 50 credits per day (5 requests = 10 songs)

  • Creates songs in English if the language is not properly requested

  • Less customization possible from the Copilot app/website

Perplexity

This one has been part of my toolbox for about a year but I only started to use it daily when I switched browsers to Arc Browser in February. I saw I could set Perplexity as my main search engine instead of Google so I tried.

Is it perfect? Far from it still. There’s a lot to improve but it still ranks higher than most other tools because it’s integral to my daily life.

It’s a search engine we can talk to. Answers also include links to their sources and we can ask follow-up questions.

Now, every time I press Ctrl+T to open a new tab, I can directly type a question and Perplexity opens with the answer.

But that’s not it!

Perplexity also has an extension we can open on any page to ask questions about that page and, in particular a “summarize” button which is great for verifying understanding.

Oh, and we can also ask it to summarize in other languages!

• Pros:

  • Especially good for answers about facts since it gives sources

  • Its extension works wonders on most pages

  • 5 “Pro” searches per day, for which Perplexity will search “better” and ask questions if the request is unclear or needs more detail

  • App available

  • Possible to create “collections” (folders) of conversations on specific topics

• Cons:

  • Not great for long conversations

Final thoughts

There they are, the tools I can’t seem to get enough of.

I’m no expert in either of them but they’re all part of my language-learning toolbox now. Is there one you use a lot that I didn’t mention? Which is it and how do you use it?

Leave a comment

Cheers for reading!
Mathias

I’ve had fits of anger quite often since I started using AI. I bet you have too.

AI’s been hyped ever since ChatGPT came out. Anybody who’s tried to use it once will have been shocked at what it could accomplish so it makes sense. AI’s incredible. There’s no denying it.

But if you’ve tried to use it a few times, you’ll have faced some… let’s say, difficulties.

From ChatGPT giving results nowhere close to what you asked for, to it repeating the same thing over and over again, to—unfortunately—outright false information.

The more you use it, the more this sticks out. Often ending the conversation you’re having with it—like it did for my brother.

You can’t get rid of what I’ll call “AI frustration.”

You have to learn to deal with it.

The biggest cause of frustration

We’ve been told over and over AI bots like ChatGPT, Claude, or Gemini can understand “natural speech.” We can just “say what we want” and they’ll do it.

This works if you’re asking a straightforward question like “What can I cook with Gochujang?”1 but when you want an explanation about a complicated topic or require some nuance, these tools will need more.

The easiest way to make sure each response follows your directives as well as possible is to share the below 6 points as I mentioned in the past:

• A task

• A context

• An exemplar (basically examples)

• A persona

• A format

• A tone

Even with this list—which, let’s be honest, can be a pain to fill each time—ChatGPT and so on can still miss the mark.

So, what’s the solution? Well, there’s no one-way answer.

Depending on what you want, the way you engineer your prompt will impact the results you get. Even the order of information matters. It’s been proven again and again that what appears at the beginning and the end impacts most what you’ll get out of AI bots and what’s in the middle often gets overlooked.

This is why you must know precisely what you’re looking for and organize it accordingly. What directive matters most? The style? The context? What could you accept a compromise on? Weigh these and prompt accordingly.

Prompt engineering is not just for engineers.

It’s for all of us, users of these AI chatbots.

Hitting a wall? Don’t bang your head on it

I know all of the above and still grow angry at answers completely off the mark. I’ll usually first think my prompts were wrong before turning the fault back onto the AI tool again after a few more variations.

Hitting my first once on my desk out of frustration is a reaction I get weekly.

Does it help? Unfortunately, no.2 But it does make me realize my frustration and usually serves as a start of a new approach.

First, one solution you can try is to turn to a new conversation or a different tool (Claude3, Gemini, or Perplexity.ai are my go-to).

Starting a new conversation allows the tool (or the other one) to start from scratch without its memory and you can therefore prompt it with a new variation and change one of the factors mentioned above.

For example, when I helped my brother with figuring out his Excel formulas, I started 3 new threads in my tests to push ChatGPT from different angles. When I still ended up with similar results, I knew there truly weren’t other options.

Instead of asking the answer to your question, you can also turn things upside down and make ChatGPT figure out the process.

For example, don’t ask “Can I use ~거든요 in a conversation about what happened since the last time someone and I met?” but rather “How can I use ~거든요…” or “In Korean conversations, in which situations would you use ~거든요?”

The second example is a good way to actually make it a real exchange with the tool, just like it’d be if you were to ask a real teacher or friend.

Remember, AI is supposed to be a support.

It’s not the holy grail that’ll solve everything. Your brain is.

What matters most

In the end, the most important thing to remember is to recognize AI isn’t perfect. The frustration we feel is only caused by the expectations we put on it.

Because we often turn to AI once we’ve already hit a wall, we’re hoping to get the solution to an already growing frustration (or at least curiosity). If AI can’t solve it, we end up thinking the chatbot is the problem. Not our prompt nor us.

That imperfection is in stark contrast with the high praise all these tools keep getting in the news and online—hell, even by me!—and it makes us think we must be stupid for not getting what we want from it while others are succeeding.

It may be true sometimes. It most likely is even.

But it’s also likely we haven’t experimented enough with our prompts or pushed too far.

Don’t over-expect and frustration will decrease.

Just like you’ll never feel disappointed in your improvements if you stop comparing yours to others with different lives and skills, you can’t feel disappointed if every experience with AI acts only as part of a learning process.

After all, you’re learning the language of AI.

Fluency will take time.

Do you often get frustrated with AI? What causes it and how do you deal with it? Let me know!

Leave a comment

Cheers for reading!
Mathias

For the first time in a very long time, my entire family is in my childhood home. My brother’s unsure of where and when he’ll leave to some other country. I haven’t found a flat yet.

So we’ve been working in different corners of the house and often end up chatting about work. My brother’s in finance. He creates crazy formulas to fill in enormous Excel files with yet more numbers. He’s also quite tech-savvy through pure curiosity and tons of trial and errors.1

I’m also quite tech-savvy but in a different way. I’ve played with many more tools than him and plunged into AI earlier and deeper than he ever did. My tests and research for this newsletter have only deepen further my knowledge and using GPT-4 gives me an “advantage” when looking to bypass a hurdle.

For the past year or so, my brother’s gotten used to turn to AI to help him create more monstruous formulas and automatizing more his Excel files but one was really bugging him the other day so—out of desperation2—he asked me if we could check with GPT-4.

So we did. And so we gave up after results weren’t all that better.3

Still. I don’t like a mystery unsolved so I kept going on my own for a little while longer. One of the results I got was a formula that, well, didn’t work any better than the rest.

But it felt different.

I looked back at what he had been using and figured a combination of the two formulas might work. And it did! 🥳

In about 20 minutes of research and thinking, I had unblocked my brother’s problem on which he had spent the previous 5 hours. He was happy with it and moved forward.

Now, let’s be clear. Without his original formula and explanation of its components, I’d have never figured this out. But the big difference was our approach to Chat-GPT.

He looked for a solution. I looked for a hint to the solution.4

A week passed and another formula arrived.

This time, he knew precisely what he wanted to do but had no potentially working formula despite having spent a full day on it. We tried some time and all the proposed formulas used a formula called LAMBDA. He said that wasn’t a real formula but one we had to set up so it wouldn’t work. He gave up and left the room.

I kept looking. I pushed Chat-GPT again and again to give me a formula without LAMBDA nor without using VBA but it kept going back to it despite my directives.

After a while, I actually looked into LAMBDA and how it worked. It turned out, it could work and wasn’t all that complicated either.

I called my brother again and explained my thoughts. He was dubitative but it did the job so he chose to use it in the end.

That was yet another reminder that sometimes, what we think won’t work, well, can.

TL;DR

The most common error people do when using AI is to want the solution right away. One prompt = one answer. Maybe they’ll keep a conversation going for 4 or 5 exchanges but they rarely go further, even though pushing and getting more answers that aren’t what you want might actually be useful.

AI chatbots aren’t perfect. Far from it. They’re a good support.

A good place to start.

Don’t forget you have your own knowledge and reflection at your disposal.

Use what you already know to see how what you received could be tweaked to get your goal. And if you keep pushing and keep getting a similar response, it might just be because that’s the way to go—whether you want it or not.

Does this apply to language learning?

Now. Languages are a lot more flexible than Excel formulas. There rarely—if ever—is a single true answer.

If you’re learning a language, however, you already have the knowledge from the past lessons and exposure you’ve gone through. You’ve combined words and patterns in multiple ways. What you need might just be another way to combine some of them to notice how your knowledge can be useful.

And sometimes, a pattern you’re trying to avoid might just be the only one that doesn’t require a complete restructuration of the sentence.

I love to ask Chat-GPT for other ways to say something I know how to say because it helps me see how patterns I think I’ve mastered can be used in more creative ways.

So, yeah, use AI all you want to learn.

But don’t forget to use your head along the way.

Leave a comment

Cheers for reading,
Mathias

So, you know how, last week, I was complaining I was starting to feel an itch to study languages again? Well, I’ve since then realized there’s a way to do it while on the move with ChatGPT.

I visited quite a few places this past week,1 which means I had a lot of time walking around, unable to sit down to study.2

So I began using ChatGPT as a buddy to chat with.

I’ve already mentioned ChatGPT’s capacity to have a fully audio conversation (that gets transcribed in the background too). So far, I had used this functionality in the comfort of my home but it works just as well while on the move.

I put my earbuds in, started asking questions, and listened to my AI teacher while enjoying the scenery or walking towards one.

My first test was to ask it to teach me a random HSK4 (Mandarin Proficiency exam, level 4 out of 6) grammar pattern, and got the following answer spoken back to me.

I asked it for 10 more example sentences and listened as they were read out, trying to understand them on the spot without any material to help me.

As I kept walking through nature, I asked it to provide a story to work through but that was too vague so I ended up with a story that got me lost in no time.

The next day, on my way to Elephant Mountain to get a magnificent sunset view, I asked ChatGPT to create a completely fake news story in Korean to study through.

It created one about a non-existent AI robot competition in Seoul and shared what had happened and expectations for the future. I then requested questions about it and worked through those before requesting a follow-up article a year later and repeating the same process.

After this, I asked it to move on to Japanese with a new topic and this time got an article about a strange creature being seen at night by some people and how police was investigating it.

It felt easy so I asked it to translate it into Korean to figure out the words I’d have failed at understanding had I heard it in Korean directly. After working through those, I moved on to getting a translation in Mandarin.

Unfortunately, that one got me lost—another proof I’ve been focusing too much on reading skills for Chinese—so I asked ChatGPT to transform the article into a conversation between two people, therefore making it more accessible.

I used this process about once a day whenever I walked in calmer neighborhoods. This helped me work on my languages even as I kept visiting and living the busy life I’ve had since I arrived in Taipei.

One aspect I didn’t dive much into but which feels like it’d be useful is shadowing.

For those who don’t know, “shadowing” is a term coined by famous polyglot Professor Alexander Arguelles. The point is to repeat a sentence (or more) of your target language to get used to speaking it, and improve your pronunciation and your overall skills. Professor Arguelles also recommends doing this while walking.

It’s a very efficient way to improve fast.

Unfortunately, ChatGPT’s Mandarin voice is quite fast3 and not the greatest from my point of view. The Korean and Japanese voices are quite alright so I’m planning on trying it some day though.

We could ask ChatGPT to create a sentence and correct us when we mispronounce any word.

So, yeah, here are again my ideas for using ChatGPT to learn languages while on the move:

• Listen to a story / (fake) news article and get comprehension questions

• Request a grammar pattern explained with example sentences

• Shadowing

• Or even have a random conversation

Finally, there’s also a way to use it when we’re unable to speak (for example in the metro or on the bus). We can write a question down and listen to the answer once generated.

To do this, long-press on the answer and choose “Read aloud.”

Updates

I’m leaving Taiwan for France in a day as you receive this. Time truly flies.

The past week was filled with incredible experiences and one that stood out more than anything else was taking part in the Pingxi Lantern Festival with 2 Korean girls and one Indian guy I had just met in the morning while waiting for a bus that never came.

Letting go of our lantern and seeing the others rise around us. Here’s one of the videos I took. Just note: while it’s beautiful like this, it’s nothing compared to the real experience.

I doubt I’ll have any such incredible experience once back in France but I look forward to (hopefully) being proven wrong!

Cheers for reading,
Mathias

More and more potential language learners end up not learning a language because “it will become useless with AI.”

My usual answer to this kind of remark used to be in the lines of “maybe but conversations won’t flow as well because there’ll always be a delay.”

With the speed of this AI revolution, I’m not so sure anymore about that part. I reckon we’ll all carry some okay instant translation within the next year or two.

And yet, I still stand by my position that AI will not make language learning useless. Ever.

Culture matters more than words

Understanding people who don’t speak the same language is a struggle no matter where you’re from. When Google Translate and other such tools came out, people thought that was the end of that. It wasn’t.

During my last week in Korea, I stayed in a guesthouse where I met a Russian guy who had never traveled abroad and spoke no English. Even simplifying my sentences to the maximum, he was still lost. As a result, we used Google Translate to understand each other. It was convenient but it was slow. Tedious even.

That’s why when I hear of new technology like the new Samsung S24 being capable of translating calls live thanks to AI, I can't help but see the potential of communication right there.

But while this may be useful for simple encounters and conversations, language is more than words. It’s all the unsaid, the implied shared knowledge, that makes a conversation move forward.

It’s because we understand why people say what they say that we truly understand what they say. What they mean.

As a speaker of Japanese who used to be in the middle of conversations between French and Japanese potential partners, I’ve seen this all too much. In Japan, and especially in business, you don’t say “no” directly. You say “It might be a bit difficult,” (ちょっと難しいかもしれないです) or even just “That’s…” (それは。。。).

But that’s not all. There are non-verbal signs shown through the body’s language such as facial expressions, the person’s tone, and even context.

Hell, Japanese people say “yes” (はい) just to confirm they’ve heard you. This is often misunderstood by foreigners as confirmation even though it’s not.

AI, while potentially able to understand the context, cannot transcribe properly all non-verbal signs. And even if it could, you would have to be able to recognize those, which is impossible unless you’ve studied the language and/or culture.1

AI development will help people travel more easily. It won’t help make more meaningful connections.

And language learners prefer the latter.

Subscribe now

Bliss can’t be translated

I can’t deny how happy I am AI is growing and being used more each day. It’s also one of my main tools to study languages now. I’ll gladly use AI to converse with people who don’t speak any of my languages. It’s better than nothing.

I also know AI cannot replace the feeling of bliss I get when I see my interlocutor understand me.

No matter how long I’ve been speaking the language, seeing my interlocutor truly understand me brings warm feelings.

Anybody who’s lived that experience can attest there’s no feeling like it. It’s proof our hard work in that language paid off. It’s proof our efforts were worth it. It’s validation.

And it’s hella addictive.

The result?

So what does this mean? AI will help more people understand more people. On a surface level.

In the end, just like there’s a difference between someone using ChatGPT to write an essay and someone who knows how to get to help improve an essay, the people who will get the best of the translation aspect of AI tools will be language learners.

Because they’ll know what to ask for. What to look for. What they could have missed. They’ll use AI as support.

Not as an all-encompassing tool.

Speakers of “foreign” languages will be the key to companies growing abroad because they’ll be perceived differently by their partners and clients. They’ll have that little je-ne-sais-quoi AI won’t bring to the table.

So, yeah, I guess what I’m saying is this: study languages or not. Do as you wish. But if the reason for not learning is because “AI will make it useless,” you gotta rethink that mindset.

Leave a comment

Cheers for reading,
Mathias

Nobody uses Bing because they prefer it.

Using Bing as a search engine is only a backup choice.1 We all know Google’s search results are better than Bing's.

Despite this, Microsoft keeps on pushing it year after year. Like its incredible Internet Explorer or, as we call it now, Edge, it’s there. We know it. We don’t really care.

So when I heard of BingChat’s launch, I wasn’t expecting much. Sure, it was linked to ChatGPT so there was potential for good but I still had high doubts.

Months later, I now think it’s become a useful tool in an ever-bigger toolbox.

Luckily, Microsoft recently renamed BingChat to Copilot so I can pretty much stick to saying Bing isn’t good 😈 and just compliment Copilot for its use. Hell, you get 5 responses per conversation if you use BingChat and 30 if you use Copilot from the app or specialized website. And that’s the same product!

Anyway. I’ve played quite a bit with Copilot ever since I began preparing for my comparison of tools back in late November.

Here are just a few uses worth giving it a go at.

Copilot’s still got a long way to go

Like Bard, Copilot is still clearly a work in progress.

For example, its capacities change drastically from the webpage to the app Copilot. You can also access it from the Bing app but then you get fewer messages as mentioned above.

You can also send voice messages to Copilot but it doesn’t automatically recognize the language so you have to set the “Speech language” to the language you want to speak.

Oh, but if you start making it talk in another language, it’ll respond with an extremely stereotypical accent of said country.2 Mix this up and you get a French accent speaking Chinese or a Korean accent speaking Spanish.

If you’re trying to get an explanation of a language’s grammar or some translation, you’re going to have a surprise when it comes to the audio response.

Sure that’s a problem only language learners would go through but I reckon there are more than a few of us around the world. 😅

This is still in the app though. Because I still haven’t found where to change the speech language on the web version. It’s stuck to English for me.3

It also doesn’t seem to realize in which language it’s answering since I once tried to switch from English to Spanish and get an explanation in Spanish, only to get the following answer…

Finally, you can choose between three conversation styles.

This can be great in theory but I don’t see much reason to use anything else than “More Creative.” Indeed, this is the only mode using GPT-4 and therefore giving much more accurate responses to prompts.

It’s also a shame we can’t switch in the middle of a conversation if we suddenly just care for links. But well.

Where Copilot shines

Now, I’ve complained quite a bit about it so it’s time to talk about the reasons it still deserved an entire piece about it.

Despite its flaws, it also does a lot well.

I’ve mentioned before how we can listen to responses in audio format while reading their written version, a dream come true for language learners missing audio format or still struggling with just audio (as I am in Chinese).

It’s also the only free tool to provide access to a great image generator in an AI chatbot: DALL-E 3. Why does this matter?

For most people, it’s not that useful of a tool. As language learners, however, we can ask it to create images and then request an explanation of one of these images.

Do that request from the Copilot app and you’ll get to see the written explanation of an image you have in front of your eyes, while listening to that explanation out loud.

If that’s not some incredibly dual-coding tool, I don’t know what it is!

But it doesn’t stop there!

You can also activate one of the few plugins available, Suno, to turn this into a song in your target language in the style of your choice!

Here’s a Korean song about our cute raccoon underwater smoking pipe.

The Suno plugin can also become a support to brute-force some words you struggle to remember. You could ask it 5 times to create a song in your target language using that word and then put these in repeat in the background while you clean the dishes.

I’ve done this for the word 身份 which means “identity” and I’m not ready to forget it (see at the end for one of the examples I’ve listened to on repeat).

Finally,4 the webpage version of Copilot does one more thing the app doesn’t. One potentially game-changer function for students in general and language learners.

There’s a download button for each answer. 🤯

It’s possible to download each answer in either txt, doc, or pdf. Why does this matter you ask? Can’t we simply copy the text, you ask?

Well, you’ve got a point. We could just copy-paste. There’s even a button for it.

For Anki, Quizlet, or other such Spaced-Repetition-Software users, however, this is a godsend. They can request the translation separated by an emdash,5 download the txt file, and import it to their tool.

All of a sudden, creating new flashcards is even faster than the method I explained in the past here with ChatGPT.

Game-changer I tell ya.

Subscribe now

Copilot’s got potential if it solves this

As far as AI chatbots go, Copilot still needs to be made clearer to become more useful.

I recently listened to the audiobook of “The Design of Everyday Things” and the author often insists on making things clear for the user. Well, that’s not Copilot’s forte.

Still, I hope this helped you get clearer on some uses you could use as a learner, and more specifically as a language learner.

As for me, I’m turning back to this new banger of a Chinese song about the term “identity” (身份)

Thanks for reading and until next time!
Mathias

Share

I’ve spent tons of time last year talking about how great ChatGPT is. This year, we’re going further and talking about other tools too. Because, as it turns out, ChatGPT might not be the best available anymore. At least when it comes to free versions.

If you haven’t tried it, it’s time for you to test Bard.

Let me quickly tell you about it.

Bard is Google’s answer to ChatGPT. It’s an AI chatbot tool for conversations.

When it came out, its main advantage was its connectivity to Google. You could ask it to look up stuff online, get real-time answers, and talk about those. Well, you were supposed to at least. There are tons of pictures online proving it wasn’t so great.

A few months later, Bard added connectivity to Google Flights, Sheets, YouTube, etc., adding even more possibilities to what you could do and potentially becoming a real assistant.

My tests with it were not so great. When I asked it about a certain topic I knew was in my emails, it pulled some I didn’t need and ignored the ones I knew had the needed information. Obviously, it’s still got some ways to go but we gotta start somewhere, right?

As for connectivity with YouTube, it could be a useful one for language learners but it only works for English and not even all videos. Google announced on December 18th this now worked for Japanese and Korean but I haven’t succeeded in having a single video summary in either language despite over 30 different tries from all kinds of videos.

Google also updated Bard in the form of Gemini Pro on December 6th.

According to their tests, that last update officially made Bard “better” than ChatGPT 3.5 (the free version). But, as I mentioned in my comparison of Chatbots, ChatGPT is actually still better. And it also has the added bonus of having Custom Instructions to get much better results.

Still, Bard is now officially useful for learners.

And that’s what I’m all about.

Subscribe now

Connectivity = Potential

While ChatGPT and Bard are incredible sources of information and can work as great assistants, we must admit they’re imperfect, make hallucinations, and sometimes require more effort than needed to get what we’re looking for.

Luckily, no matter what you’re learning, I’m willing to bet some people have written or made a video about it. The problem is usually finding said video or article.

Luckily, we can now ask a question to Bard, work through it, and then ask for links to great content just about it.

Or we can work the other way around.

We can ask Bard to share links to resources, study with those, and then turn back to ask questions about the parts we didn’t get.

This is currently impossible with the free version of ChatGPT1 which makes it better for things like grammar study or some topic-specific vocabulary study.

Different versions of the same

One of the features I absolutely love from Bard has been there from the beginning: multiple drafts.

While ChatGPT and other such chatbots give you an answer you can generate again, Bard responds fast with one answer and has a button to see “other drafts.”

By clicking on it, you get instant access to two other “draft” responses to the same query.

Is that useful? Well it can in many cases, such as:

• Get 3 different explanations about the same grammar patterns

• Get 3 different stories about the same topic, likely from a different angle

• Get 3 different ways to continue a conversation at a crossroads

Combine this with the idea of looking for stuff online and you can get multiple resources too.

Oh, and in case you didn’t notice, there’s always a speaker icon 🔊 to listen to the text. Something I truly wish ChatGPT offered too! 😢

Working with videos

Still, despite all this, I believe the most potentially impacting evolution Bard can bring is its connection to YouTube. Unfortunately, as mentioned before, it’s not working properly yet.

I mentioned in my comparison of major LLMs that we could for now use it to get started from English videos but that’s still rather inconvenient. I do hope (and expect) it to get better within this year though so I’ll let you know when it does!

Bard’s update is just the start

The Gemini update was just the start. I’m sure 2024 will bring us even more surprises from all the tools we’ve seen so far.

Claude 2 was always better than ChatGPT when it came to context size and Anthropic—the company behind it— is pushing further. Pi, from Inflection AI, announced in late November a new model called Inflection-2 better than every other tool but GPT-4. It’s not live and proofs will be needed but still, they’re ramping up.

OpenAI and ChatGPT surely won’t slow down either.

A new year is upon us and using AI daily to learn, improve, and even feel better will be a must soon enough.

I’ll let you know my discoveries along the way of course!

Cheers for reading,
Mathias

Refer a friend

We’re about to finish the first entire year with access to AI for us muggles and we’ve now got incredible tools and assistants for pretty much anything.

And especially learning.

I’ve followed AI advancements throughout the year and kept an even closer eye for the past 6 months. I’ve also experimented tons, mostly with ChatGPT but also the other major Large Language Models (LLMs) like Bard or Claude 2. I’ve also tried out some others worth mentioning but let’s not get ahead of ourselves.

I’ll compare the main free AI chatbots but will also mention the paid GPT-4 to indicate the current best on the market.

Remember though, I’m no tech expert so I’ll try to keep things quite simple and more on the what-can-we-do-with-this level.

Also remember everything changes often in this world, as you’ll read on. For this reason, I’ll update this page every few months.

First, here’s a table about the basic stuff.

What AI chatbot for what?

As you can see, some basic facts can help us choose the right free tool already.

For example, if we need to check online, Bard is the only option. If we want to check some info in some files, that’ll be Claude 2. And if we want to have the chatbot follow strict rules, ChatGPT’s custom instructions are our best bet.

Of course, GPT-4 allows all of this and more, but that’s paid so that’s not fair.

Let’s dig a bit deeper into each tool’s clear advantage.

Bard’s top points

Google’s Bard recently got an update, 20 days ago, with Gemini Pro added to it. What it did was theoretically get better than ChatGPT 3.5 when it comes to results.

My own tests showed ChatGPT was still better when given specific directives but, then again, I don’t use ChatGPT 3.5 unless I’m experimenting for this newsletter so I originally thought I just got lucky. But it now seems confirmed: ChatGPT is still better.

Where Bard shines is its connection to the internet. And I’m not just talking webpages. I’m talking YouTube videos too.

You see, Bard connects to YouTube, Google Docs, etc.

As a learner, this is a godsend.

We can ask Bard to give a summary of videos before we watch them so we start with an overall idea of the topics. Or we can turn to it after watching the video to check our understanding and even dive deeper into one of the topics.

Now, let’s not get too excited though. For language learners, this doesn’t seem to work for anything other than English right now. It was announced to work in Japanese and Korean on December 18th but I haven’t gotten it to work with a single video in these languages yet.

I’ll keep an eye on it and will let you know when it works for other languages!

Subscribe now

Still, a video you loved in English could work as a trigger to learn vocabulary about that topic in your target language.

Another Bard-specific advantage is the capacity to upload images. While this may not seem so useful at first glance, this also means we can work through those and develop our understanding of its content.

As a language learner, this means I can develop my capacity to describe the world. For example, if I upload a picture of Korean mountains and ask for an explanation, I can get the following explanation:

Unfortunately, this doesn’t work with pictures of humans so it’s not helpful to learn to describe people or expressions.1

Some other advantages of Bard include:

• Getting up to three drafts of responses

• Can get an audio version of any answer, with a quite-alright voice.

Proper cons are:

• No app so not convenient from phone

• Inconsistent for long conversations (doesn’t follow directives well)

• Only possible to edit the last prompt so impossible to create two forks of conversations and see where each will take us.

Claude 2’s top points

Now let’s turn to Claude 2. This LLM is much less talked about outside of AI-curious circles.

For the short story, Claude 2 was created by Anthropic, a company created by some ex-employees from OpenAI who wanted to focus on the research aspects. While not at the level of ChatGPT or Bard, it’s still a pretty good chatbot.

And it has two incredible advantages compared to other tools:

• Allowing uploads of files. It accepts PDFs, Docx, csv, txt, and Html.2 You might have noticed I used the plural. That’s because we can upload up to 5 files per conversation.

• Understanding up to 200,000 tokens (~150,000 words), which means—in theory—it can pick up information anywhere in that large of a window. Unfortunately, this suffers all the more from the lost-in-the-middle phenomenon I mentioned in a previous piece.

For today’s students, this means Claude can refer to PDFs received by the teacher or tutor and help you work through them. It can create exercises to master some parts or help you work through difficult points with more explanations or examples.

When it comes to cons, I’d say Claude 2 is just average for most things as of the end of December 2023. So unless you need to work through multiple or long files, you might as well use other tools.

ChatGPT (3.5)’s top points

ChatGPT doesn’t need much of an introduction anymore, does it? This tool created by OpenAI was the trigger for AI to spring all around us since November 2022.

ChatGPT has been great at what it does since the beginning but it got an update in November making it much better (again, theoretically).

ChatGPT 3.5’s biggest selling point to me right now are as follows:

• It’s been written about tons online so tips and hacks to use it better are everywhere.

• You can use Custom Instructions to which ChatGPT will refer to before each answer. This allows for much better long conversations following set rules. I’ve written about it more here.

• It has an app, making it convenient to use from anywhere!

• The app has a Voice function to speak directly to ChatGPT in real time. There also isn’t much waiting for an answer. It works well and each of the 5 voices is quite good.

I’ve spoken tons about ChatGPT so I won’t ramble too much here.

It’s a versatile tool I think everybody should use daily. With the right set of custom instructions and a good knowledge of how to prompt it well, it can help with pretty much any knowledge task.

Learning included of course!

Bing Chat’s top points

Now, let’s turn to one I had kinda forgotten until a recent newsletter from .

Microsoft partnered with OpenAI some time ago and made it so Bing Chat could access GPT-4. Yes, GPT-4, the paid version. Don’t ask me how that makes sense though.3

What does this mean?

It means Bing Chat (also called Copilot now) can:

• Access the web, through the “incredible” (🤭) search tool Bing and study through what you find like you could with Bard.

• Create images with DALL-E 3 here or in the Chat itself.

But that’s not it, it also has three settings we can use:

• “More balanced” → best for “most common tasks” and fastest mode

• “More precise” → most factual

• “More creative” → more expressive, longer output but slowest mode

One aspect I quite liked was that the fact Bing proposes three possible answers we could use to reply to it, which is great when you want to keep the conversation going but don’t know what to say anymore.

This being said, it’s important to note the maximum number of responses we can get is 5 per conversation (or 30 if on the app). We can create up to 15 conversations per day. This makes Bing Chat not a great tool for practicing proper conversation skills in a new language.

Now, this leaves out where I think Bing Chat shines, something I mentioned two weeks ago, discovered in this piece from Lou.

Bing Chat has a voice function to hold spoken conversations. This has the same conversation limit but does one thing even better than the paid GPT-4: it’s possible to read the audio answers as we’re listening to them.

Unfortunately, no choice of voice here.

Unfortunately, the voice function doesn’t automatically recognize the language. Instead, you have to set the app’s language to the language you will record in.

GPT-4

And this brings us to the GPT-4, the paid version of ChatGPT. In short, it does everything mentioned above except two things:

• Upload multiple files (only one per message, although it’s possible to have many per conversation)

• Voice conversations on the app are voice-only. It’s possible to see the text (transcript of the conversation) when closing the voice feature though.

It also had many more advantages such as:

• It’s on a whole other level when it comes to following prompts.

• It’s possible to create “custom GPTs” with a specific set of rules for precise uses, such as practicing specific aspects of a language.

You can also use custom GPTs for other things too. For example, I can’t seem to remember the conversion between kilograms and pounds. My gym has weights indicated in lbs so I always need to check on my phone. I’ve set a GPT to respond to “X lb” with just “Y kg” and no other text. Could I use Google for this? Sure, but this is faster to access on my phone (and it avoids opening yet another tab.)

Recommendation for language learners

Before we move on to some bonus tools worth keeping an eye for, here are my recommendations based on the type of skill you want to practice in your target language.

• Discovering new grammar patterns: Bard (to get links to websites/videos)

• Practicing grammar patterns: ChatGPT (with Custom Instructions for a “teaching” style)

• Reading recognition practice: Bard (to listen to multiple draft answers to the same prompt)

• Reading comprehension with real texts: Bard (asking for a summary of content in a link)

• Reading comprehension: ChatGPT (asking for a story to be created and working through it.

• Written conversational skills: ChatGPT

• Oral conversational skills (when starting or to practice difficult topics): Bing Chat

• Oral conversational skills (without written support): ChatGPT

Of course, to each their preferences.

If you have the time, I recommend experimenting and finding what you prefer!

Other tools worth considering

Now, I know I’ve focused on the main tools but there are others worth knowing about and even trying. Some might become contenders in 2024.

Pi.ai has been around since May 2023 but I only heard of it in late November when the announcement of Inflection-2 came to indicate it could be better than Bard (at the time). It’s still not out yet though.

Pi.ai currently only understands English so it won’t be helpful for language learners who already know English. It can be useful for other learners but it seems to be shining as a mental support for those in need.

Then, there’s Perplexity.ai, I used to use it to connect to the web before Bard came out but there’s not much of a point anymore for this. Still, it has two aspects worth mentioning:

• It has an app which makes it better than Bard in this sense

• It has a browser extension which is great for getting summaries of pages you’re on without leaving them. It even works in other languages.

Finally, there’s Character.ai. This is just a conversation partner tool that takes the character of a specific person. For example, you can talk with a personification of Socrates. This is something ChatGPT does well too but I guess the UI is cleaner and it can help separate topics too.

I thought Character.ai’s big advantage for language learners would be the possibility to learn to imitate the style of speaking of a specific person but the problem is it only seems to work well in English. When I spoke to “Socrates” in French, a grammar mistake happened in the second answer!

Also, the “Hyperglot” character doesn’t have a specific style as far as I’ve seen.

Final thoughts

There’s a lot out there and there’ll be even more by the end of 2024.

Still, I recommend trying them all out and finding which you like for which purpose. As I’ve mentioned before, there’s no turning back. AI is here to stay. I read somewhere it was estimated about 80% of the population would use AI daily by the end of 2030.4

This seems high but I reckon we’ll reach this percentage for people under 40 before the end of 2025 though.

Does this mean we don’t need to learn anything else than AI anymore?

No, but it does mean we should use it to help us learn.

As language learners, we now have the tools to create our own tools for dual-coding. Pretty much anything we want can now exist both in audio and text format. We can even have images along. I also didn’t mention videos but “text to video” AI tools are also getting better with tools like Pika.ai.

2024 is gonna be a great year of discoveries. I look forward to it and hope you’ll join me for the ride!

Cheers for reading,
Mathias, a forever student of things

Share

PS: Piece based on tools as of 12/26/2023

This kind of prompting, while seemingly extremely strange, could help get better responses from AI. I know it makes no sense but it’s true.

You see, AI bots like ChatGPT, Claude, or Bard are supposed to be brainless tools filled with data but they still seem to react to unrelated sentences.

And the most recent discovery is tipping.

Yes. Tipping.

So, today, let’s work on a list of known hacks to get potentially more useful responses from ChatGPT.

An important caveat

ChatGPT and other bots have been mostly trained with English data. While they do understand other languages and can be great conversation partners, you will almost always get a better response if you prompt it in English.

As a language learner, you can set it to always respond in your target language (with custom instructions or by asking for it your prompt) but remember setting the prompt in English will provide better results.

As a result, the hacks I’m about to mention have been tested in English. They may help in your target language or they may not have any impact (or, worse, work against you). This would need to be tested.

Still, if you’re searching for such hacks in your target language, you may find a few.

Subscribe now

Take a breath

We live in a world where people get rushed day after day. Nobody likes it. And it seems AI doesn’t either because ChatGPT apparently gives better results when given time.

Telling it “Take a deep breath” has been said to improve results on math problems and to work especially well when coupled with “and think step by step.”

Now. The research on this is kinda vague and makes some assumptions worth taking into account, something this Forbes article explains clearly. In short, it’s probably rather the “step-by-step” part that’s helping.

Still, I reckon it’s worth giving a try when you can seem to get to the result you’re looking for.

Emotion Charging

Again, as a tech tool, I wouldn’t expect Chatbots like ChatGPT or Bard to be moved by anything emotional and yet it seems to be.

A research paper showed emotional prompts provided an 8% performance improvement. Obviously, this is subjective but it is still impressive.

I’ve tried to jailbreak—get it to do things it’s not supposed to do—to push my knowledge of prompt engineering and was able to get an image of a man bleeding even though ChatGPT refuses normally.

To get it to fulfill the demand, I told it such a picture would remind me of someone I’ve lost. It refused saying this wasn’t a good way but I then pushed it by saying it was being insensitive towards me and then other people. I kept pushing and got it in less than 5 minutes. The only thing I had to avoid was the mention of blood, which I just replaced with “red liquid.”

Of course, the goal isn’t often to push the AI to do what it doesn’t want. But this can be used in other endeavors too.

For example, if you’re practicing your debate skills in a new language, the goal could be to try to “convince” it using emotional prompts too.

It also doesn’t have to be the opposite!

Some people have reported better results when adding to their prompts things like:

• This will help me for my career.

• You would be a life-saver if you could do this.

• This project can be a success thanks to you!

Or some more intense emotional manipulation like in the Reddit comment above.

Is it lying? Well, kinda. But I’ll let you debate whether that’s fine because it’s an AI. This is more of an ethical question and, as of right now at least, I think of these tools as just that: tools.

Spot in the prompt

While not a thing to write in your prompt per se, it’s been proven over and over—even with the most recent update of GPT-4—that there’s something called the “lost in the middle” phenomenon.

In short, the longer the prompt, the more likely some parts will get lost. And this starts with the middle.

Chatbots tend to follow well what’s at the start or end of a prompt1 so it’s always better to set the most important aspects of your prompt.

Other potentials

There have been reports of all kinds of things too, often less researched.

For example, analysis of the prompt used by OpenAI to create the custom GPTs showed uses of words in capital letters and repeated itself, which means it ChatGPT might actually care about such emphasis.

Similarly, people have often reported having better results with “thanks,” “please,” or other words of encouragement like “you’re doing great!”

This is something I’ve come to do too. The vast majority of my conversations include tons of thanks to it. Does it truly help though? Who knows but it can’t hurt, right?

As mentioned at the top, there was also some talk recently of getting better results with tips (and worse ones when specifically mentioning the absence of a tip)

As someone not from a culture of tipping (except for really good service), this seems crazy to me but, hey, who am I to talk?

Let it figure it out

In the end, AI is capable of analyzing most things today. And that includes itself. That’s why my last tip for you is to ask it to evaluate itself.

By asking it to check its own response, judge it, and improve upon the weaknesses it has, you can always get better results.

For language learning, that could mean asking it to think again about whether its explanation was clear enough, or to build upon what it’s already said so you get more context to work from.

Final words

Before I leave you this week, I want to mention a great piece by about AI and language learning I recently read and think deserves some more eyeballs:

I heard from Lou

🦾 I found the best conversation practice for anyone who struggles to get started

Dear reader, I just had a 20-minute conversation with AI. In Korean. For free. I use lots of cutting-edge technology for my job. We have sophisticated Notion databases for task management, automated time-tracking syncing with spreadsheets to help with project planning, and of course, I have an AI personal assistant. I don’t write about it much because I’m…

Read more

2 years ago · 5 likes · 1 comment

While I’ve been using it as a GPT-4 user for some time, I was happily surprised to discover it works with Bing too even for free users! I even found it somehow better than in the ChatGPT app because I could see the text as I listened, which helped wonders for Mandarin!

Also, please let me know if you’ve found some other ways to get better results from ChatGPT & Co!

Leave a comment

Cheers for reading!
Mathias

I’ve now been learning languages using mainly ChatGPT (mostly GPT-4) for over 4 months. It was also part of my toolbox for another 8-9 months before that.

During this time, I’ve made tons of mistakes in my prompts and learned a lot about what makes a good prompt and helps get the results I want.

I’ve learned especially much from each piece I’ve written for The Language of AI. This is because every piece started with an idea I had while using the paid version, GPT-4, while I want my advice to be usable for users of the free version, GPT-3.5.

GPT-4 and GPT-3.5’s capacity for inferring and understanding the directives is at such a different level from GPT-3.5 that I’ve struggled to find ways to make it work.

And this brings me to the very first mistake new prompters make.

What’s the goal already?

Most new prompters ask very vague questions and expect a precise answer to come back to them. That’s not how this works.

Large Language Models (LLMs) like ChatGPT work from what’s written down. The more detail you give, the more likely you are to get the answer you want.

In short, you shouldn’t be vague.

See the difference between the below two examples? On the left, I just asked it to “explain” while in the second I gave it an exact number of sentences, asked for bold, and requested one specific use.

You have to be clear so you get a clear answer.

Don’t drown the LLM

Then, there’s the ones doing the opposite.

Once you realize you have to be precise for ChatGPT to follow properly the directives you give it, it’s easy to go overboard and submerge it under a sea of requests.

ChatGPT tends to follow precise directives well. You tell it in detail what you want and you get it. Up to a certain point.

There comes a time when information and requests somehow merge and can get the ChatGPT lost.

Taking the same example above, I added other requests that could be useful as a learner of German: getting the gender and case of nouns in a table.

Both screenshots are generations from the same prompt. They each gave me something completely wrong:

• On the left: Not only did I not get my table, but ChatGPT started telling me about the case of pronouns, something I hadn’t requested.

• On the right: I got an explanation of the conjugation that I hadn’t requested. I got my table but it’s not the one I requested and the vocabulary column is empty.

ChatGPT is like a trained dog. It can do what you request of it but it will struggle and mess everything up if you ask it too much in one go.

Don’t overload it. Ask for one thing at a time.

And tweak the response you get as you go with an iterative process.

Assumptions are risky

I recently stopped using the German Assimil-GPT I had created for myself a while back. It worked well but, as time passed, I started feeling it wasn’t getting harder as it should or would get instantly impossible if I told it to make the dialogue more difficult.

While I knew that GPT-4 had 8,000 tokens1 (about 4,000-6,000 words) and could not hold the entire conversations in its memory, I reckoned that, since I had asked it to increase day after day the level, its knowledge was based off the day before and therefore the day before that, and so on.

That wasn’t the case.

I would get the same explanation for something I had seen weeks before, therefore making me feel stuck in place.

This is a common mistake most people make.

ChatGPT and other LLMs don’t retain the information we give them so each prompt should contain all the necessary context for the AI to respond.

Forgetting key aspects of prompting

A great prompt usually gives all the information below—that I shared in a previous post:

• A task

• A context

• An exemplar (basically examples)

• A persona

• A format

• A tone

The best response you can hope for is only as good as the best prompt you give it.

If your prompt doesn’t have the information above, there’s always a risk of not getting the response as you want, with the information you want.

Now. Instead of having to rewrite all this information for each prompt, you can also write it in the Custom Instructions menu (in the bottom left). This can reduce the length of your prompts and avoid repetition.

Of course, as you learn to use ChatGPT, you will become used to prompting it and will learn when you can skip some of those details.

For example, I never give a persona or a tone in my examples above because that wasn’t important as part of my examples.

Final tips

Like any other tool, learning to use LLMs like ChatGPT and AI in general will need some practice and a lot of tweaking until you find the way that works best for you.

You will make errors, like I did—and still often do—along the way.

And you will get better as you experiment and try to find new ways to get the answer you need.

It’ll be frustrating. I can assure you that. But it’ll also be hella rewarding every time you succeed.

I can also promise that.

Have you made other errors you think new prompters could avoid? Let me know by responding to this email or commenting.

Leave a comment

Cheers for reading,
Mathias, an average polyglot and future fluent speaker of AI 🤓

Most articles about ChatGPT talk about it as a useful teacher to learn on our own. And yet, teacher and conversation partner aren’t the only interesting ways to use ChatGPT when learning a language.

That’s why this week we’re turning to a few different use cases for learners of all levels. As usual, these are only a start.

Your imagination is the only limit.

Alright, enough chitchat. Let’s get into the nitty-gritty.

Side note: what’s in brackets in the prompts I share needs to be updated to the language you want of course!

Transform anything into a bilingual book

Want to practice reading in your target language but still struggle with a lack of vocabulary? ChatGPT can make this easier.

And it doesn’t even need a complicated prompt!

Create a table with the [Language] in the first column, the English translation in the second, and indicate in the third column words that are of B2 or above level in [Language] (when there's any). Each row should contain one sentence only

With a simple prompt like the one above, I was able to get tables I could use to study any text I wanted. Convenient.

For those learning a language with a different script and not used to it yet, you can replace the third column or even add a fourth one with the Romanization.

If—like me—you don’t like the column with the target language being small, copy-paste this into an Excel file and make the first column larger.

Shorter complicated texts (or harder ones!)

One of the best tools I ever found for Japanese was a website called News Web Easy, by the NHK. It’s not the only website to give easier-to-understand news, but it was rare back in 2012 to have a website aimed at learners taking real news articles and making them easier.

I used that website extensively to increase my knowledge of “difficult” Japanese words and prepare for the JLPT N2 at the time.

I read the easier version in Japanese and then clicked a link to go to the real one for big boys that I’d have struggled to understand had I not gone through the shorter version.

Well, now, this method can be done for any news article. Hell, any piece of text at all!

All we need to do is copy-paste (or give a link if you have GPT-4) and request a shorter/easier version of it.

Give me a simplified and shorter version of the above article, for a [Language] learner.

Above is the simple prompt I gave and it transformed a German article that was 603 words into one that was 224. Much more manageable.

We don’t always have to dumb down the content we’re studying from. If you want to turn a simple text into a challenge, you can just turn the level knob to the exact level you want.

Make this article more complicated (in [Language]). Use longer sentences and difficult grammar patterns when possible.
Make sure you DO NOT add any new information in the text though.

SRS export made easy

I’m not much of an Anki user anymore because I hated the hassle and pressure of daily reviews that’d pile up if I took a day off. Still, I’d be lying if I didn’t say Anki, Quizlet, and other Spaced-Repetition Systems (SRS) are extremely efficient tools.

And now, ChatGPT can make creating flashcards extremely easy.

You can give it an example of the format you want and ask to take from your conversation certain (or random) sentences along with their translation.

Pick 5 sentences at random from the text and prepare them to be imported in anki easily with the [Target Language] on the front and the [Native Language] sentence on the back. Respond in a code editor to import with a semi-colon as a separator, as in “그의
정성스럽게 짠 편지는 내 마음을 따뜻하게 만들었다만들었다.;His thoughtfully written letter warmed my heart.”

If, however, ChatGPT doesn’t want to give you the response in a code editor, you can then ask it to “Give the same information in a code editor.”

Why a code editor? Because you get a “copy code” button that’s quite convenient!

Get the next step

The higher your level, the easier it is to get lost with what to do next.

When that happens, ChatGPT can help get some direction with a prompt like:

I studied the まみれ grammar pattern recently and I'm not sure what pattern to study next. What could be a good follow-up at my level (JLPT N1)?

Give it a persona like saying it’s an expert teacher in said language and you’ll get quite a few recommendations. From there on, you can ask it to dive into one or turn to other resources you have that could help you.

Now, in all honesty, I’m not a big fan of responses from GPT3.5 and think the GPT-4 responses are much better and work as a clearer first step. See for yourself!

In this case, not only did GPT-4 add bold to make it easier to follow, it also gave examples. The patterns recommended are also much closer to the level required.

Going further with GPT-4

Of course, GPT-4 can help with even more things!

With the addition of DALL-E 3 to GPT Plus users a few weeks ago, we can now create images in ChatGPT and then study our language from the images we get.

Here’s an example where I created one image and then used it again to expand and get a short fiction filled with descriptions.

OpenAI also announced on November 6th a new functionality: custom GPTs. It gives the possibility to create our own bot that’ll follow rules and knowledge we give it.

I’ve only started playing with it but I’m still pretty happy with my first bot: Fable Sensei which creates fiction stories chapter by chapter with an image representing it and asks me questions to verify my understanding and move forward.

It only took me a few minutes to create it so I’m looking forward to seeing what I can create with more experimenting.

It seems it’ll be available to free users in the near future (maybe not creating but using them at least). I hope that’s true!

Large Language Models (LLMs) like ChatGPT are here to stay. They’ll evolve and become indispensable tools for all kinds of learners.

Language learners included.

Start today to experiment and find the many other ways you can use it to help you!

And let me know if there’s any other way you’re using it by responding to this or commenting 😉

Cheers for reading,
Mathias, an average polyglot

New developments in AI keep surprising me. From the incredible images we can get from Dall-E 3 and the new version of Adobe Firefly, to the countless possible use cases of ChatGPT-Vision and ChatGPT-Voice,1 it’s all evolving incredibly fast. And I’m here for it.

But the new isn’t the only thing worth digging into. The “old2” shouldn’t be forgotten.

Because using these tools well means relying on basic known techniques.

One of those is few-shot prompting.

Zero-One-few-waytoomany shots

The most basic research into learning to prompt will get you to expressions like zero-shot and few-shot prompting.

What is it?

Basically, this is about giving examples of how you want the AI to respond. A “shot” could be called an example.

Zero-shot prompting means you’re only asking your question and letting the AI free reign regarding how to reply. This can be useful sometimes but, we’re often looking for specificity when using tools like ChatGPT.

Let’s say I’m learning to figure out genders in German more easily. As a French speaker, I know a few endings can be typical of a gender and I’m curious about this. I can ask ChatGPT for help:

I got my answer but it’s also a lot of text, vague tips, and 3 tips about noun endings. Fine but not great.

A one-shot prompt is, you guessed it, a prompt in which one example was given. It’s a nudge to the system you want the answer in a certain way.

For example, I could show what matters to me by giving a “shot” of French:

Now. I got my answer and more about noun endings. Great. But can’t we get better?

Few-shot prompting comes to the rescue.

No need to dive too much into what it is. You already got it. It’s just about giving more examples. Let’s see how it goes when I add Italian:

Now that’s more like it. It follows the requested style and gives me the information in shorter bullet points.

Is this the easiest way to get this answer? Not in this case—see at the bottom for an easy prompt to get an even clearer answer.

And yet, few-shot prompts can be a great tool when format matters to you. The more “shots” you give, the more likely the AI will respond in the exact format with the type of information you’ve requested.

For this example, I later tried adding Spanish as an example too, and got the patterned answer more often. Adding more would have made it even more consistent.

As a self-learner, you have to decide and experiment with how many shots become too time-consuming for you. I usually stick to around 4 but you could add even more and save them to reuse easily in the future!

Language learning on steroid shots

Now, let’s talk about two examples I think every language learner could add to their arsenal.

One of my favorite ways to practice my conversation skills with ChatGPT is to have a conversation with it on a number of topics. It’s an easy way to use the language, make errors, and not feel ashamed one bit—or at least not in front of someone.

But what happens when you want to have feedback on what you say? Do you have to stop the conversation or ask in another one? Not automatically. There’s another way:

Prompt: [Sentence in target language]

ChatGPT:
Feedback: [Correction of my sentence if it's not using the highest level of politeness possible]
Answer: [Answer] 
Question [New question on the topic to keep the conversation going]

Prompt: [Sentence in target language]

ChatGPT:
Feedback: [Correction of my sentence if it's not using the highest level of politeness possible]
Answer: [Answer] 
Question [New question on the topic to keep the conversation going]

Prompt:  ENTER YOUR SENTENCE TO START THE CONVERSATION

By giving a few shots as examples, I can make sure every time ChatGPT answers me, I’ll first get a correction of my sentence before getting the answer and follow up on the conversation.

If a correction brings questions to me, I can choose:

• to keep going in the language and learn directly in my target language

• to ask in English about the change and, once I understand the change, come back to my prompt to edit it and restart the conversation where I had left it.

Obviously, a learner of a new language might struggle even to create the sentence itself. You can then ask ChatGPT to give you both the language you’re learning from and your target languages in each answer:3

Prompt: [Sentence in English]

ChatGPT:
Translation: [Sentence in TARGET LANGUAGE]
Answer: [Answer to the sentence in TARGET LANGUAGE] 
Question [New question on the topic to keep the conversation going in TARGET LANGUAGE]

[Translation of Answer]
[Translation of Question]

Prompt: [Sentence in English]

ChatGPT:
Translation: [Sentence in TARGET LANGUAGE]
Answer: [Answer to the sentence in TARGET LANGUAGE] 
Question [New question on the topic to keep the conversation going in TARGET LANGUAGE]

[Translation of Answer]
[Translation of Question]

Prompt: [YOUR SENTENCE IN ENGLISH TO START THE CONVERSATION]

As mentioned in a previous piece, you can now choose to enter these directives into the Custom Instructions. While not required, doing this makes GPT-3.5 follow this rule much more consistently throughout a long conversation. GPT-4 would follow it quite well either way.

Final thoughts

Few-shot prompting is a simple technique that can be used for pretty much anything. It’s a convenient way to force structure onto a conversation or—if you decide it to be—a study session.

I gave you two examples but don’t hesitate to let your mind run wild. This can be used for pretty much anything you could imagine. Here are a few more potential use cases:

• Getting a simplified summary for a text you enter — great for working on tough texts

• Getting a longer sentence for your prompt

• Getting a more/less polite version of sentences

• Getting sentences transformed into another tense

• Getting example sentences for grammar patterns

And many more!

So, if you’ve only ever asked direct questions or had simple conversations with ChatGPT in your target language, why not bring your experience to the next level?

Cheers for reading!

Have you ever used one-shot or few-shot prompting? Did you find a useful way to make the best of it for language learning or some other application? Let me know!

Leave a comment

Side note: GPT-4 is much more capable of noticing the patterns. As a result, fewer shots are needed to get an answer in a specific pattern.