Do not apologize for replying late to my email

You don’t need to apologize for taking hours, days, or years to reply to one of my emails.

If we are not close collaborators, and if I didn’t explicitly tell you I was waiting for your answer within a specific timeframe, then please stop apologizing for replying late!

This is a trend I’m witnessing, probably caused by the addiction to instant messaging. Most of the emails I receive these days contain some sort of apology. I received an apology from someone who took five hours to reply to what was a cold and unimportant email. I received apologies in what was a reply to a reply I had sent only a couple of days earlier.

Apologizing for taking time to reply to my email is awkward and makes me uncomfortable.

It also puts a lot of pressure on me: what if I take more time than you to reply? Isn’t the whole point of asynchronous communication to be… asynchronous? Each on its own rhythm?

I was not waiting for your email in the first place.

As soon as my email was sent, I probably forgot about it. I may have thought a lot before writing it. I may have drafted it multiple times. Or not. But as soon as it was in my outbox, it was also out of my mind.

That’s the very point of asynchronous communication. That’s why I use email. I’m not making any assumptions about your availability.

Most of the emails I send are replies to emails I received. So, no, I was not waiting for a reply to my reply.

My email might also be an idea I wanted to share with you, a suggestion, a random thought, a way to connect. In all cases, I’m not sitting there, waiting impatiently for your answer.

Even if my email was about requesting some help or collaborating with you, I’ve been trying to move forward anyway. Your reply, whenever it comes, will only be a bonus. But, except if we are in close collaboration and I explicitly said so in the email, I’m not waiting for you!

I don’t want to know all the details of your life.

Yes, you took several days to reply to my email. That’s OK. I don’t need to know that it’s because your mother was dying of cancer or that you were expelled from your house. I’m not making those up! I really receive that kind of apology from people who took several days to reply to emails that look trivial in comparison.

Life happens. If you have things more important to do than replying to my email, then, for god’s sake, don’t reply to it. I get it! I’m human too. If I sometimes reply to all the emails I receive for several days, I may also archive them quickly for weeks because I don’t have the mental space.

If you want to reply but don’t have time, put the burden on me

If I’m asking you something and you really would like to take the time to reply to my email, it is OK to simply send one line like

Hey Ploum, I don’t have the time and mental space right now. Could you contact me again in 6 months to discuss this idea?

Then archive or delete my email. That’s fine. If I really want your input, I will manage to remind you in 6 months. You don’t need to justify. You don’t need to explain. Being short saves time for both of us.

You don’t need to reply at all!

Except if explicitly stated, don’t feel any pressure to reply to one of my emails. Feel free to read and discard the email. Feel free to think about it. Feel free to reply to it, even years later, if it makes sense for you. But, most importantly, feel free not to care!

We all receive too many messages in a day. We all have to make choices. We cannot follow all the paths that look interesting because we are all constrained by having, at most, a couple billion seconds left to live.

Consider whether replying adds any value to the discussion. Is a trivial answer really needed? Is there really something to add? Can’t we both save time by you not replying?

If my email is already a reply to yours, is there something you really want to add? At some point, it is better to stop the conversation. And, as I said, it is not rude: I’m not waiting for your reply!

Don’t tell me you will reply later!

Some people specialize in answering email by explaining why they have no time and that they will reply later.

If I’m not explicitly waiting for you, then that’s the very definition of a useless email. That also adds a lot of cognitive load on you: you promised to answer! The fact that you wrote it makes your brain believe that replying to my email is a daunting task. How will you settle for a quick reply after that? What am I supposed to do with such a non-reply email?

In case an acknowledgement is needed, a simple reply with "thanks" or "received" is enough to inform me that you’ve got the message. Or "ack" if you are a geek.

If you do reply, remind me of the context

If you choose to reply, consider that I have switched to completely different tasks and may have forgotten the context of my own message. When online, my attention span is measured in seconds, so it doesn’t matter if you take 30 minutes or 30 days to answer my email: I guarantee you that I forgot about it.

Consequently, please keep the original text of the whole discussion!

Use bottom-posting style to reply to each question or remark in the body of the original mail itself. Don’t hesitate to cut out parts of the original email that are not needed anymore. Feel free to ignore large parts of the email. It is fine to give a one-line answer to a very long question.

I’m trying to make my emails structured. If there are questions I want you to answer, each question will be on its own line and will end with a question mark. If you do not see such lines, then there’s probably no question to answer.

If you do top posting, please remind me briefly of the context we are in.

Dear Ploum,

I contacted you 6 months ago about my "fooing the bar" project after we met at FOSDEM. You replied to my email with a suggestion of "baring the foo." You also asked a lot of questions. I will answer those below in your own email:

In short, that’s basic mailing-list etiquette.

• Mailing list etiquette (www.mediawiki.org)

No, seriously, I don’t expect you to reply!

If there’s one thing to remember, it’s that I don’t expect you to reply. I’m not waiting for it. I have a life, a family, and plenty of projects. The chance I’m thinking about the email I sent you is close to zero. No, it is literally zero.

So don’t feel pressured to reply. Should you really reply in the first place? In case of doubt, drop the email. Life will continue.

If you do reply, I will be honored, whatever time it took for you to send it.

In any case, whatever you choose, do not apologize for replying late!

• And use plaintext email!

Offpunk 3.0 "A Community is Born" Release

For the last four years, I’ve been developing Offpunk, a command-line Web, Gemini, and Gopher browser that allows you to work offline. And I’ve just released version 3.0. It is probably not for everyone, but I use it every single day. I like it, and it seems I’m not alone!

• Offpunk, the offline-first command-line browser

Something wonderful happened on the road leading to 3.0: Offpunk became a true cooperative effort. Offpunk 3.0 is probably the first release that contains code I didn’t review line-by-line. Umerdify (by Vincent Jousse), all the translation infrastructure (by the always-present JMCS), and the community packaging effort are areas for which I barely touched the code.

So, before anything else, I want to thank all the people involved for sharing their energy and motivation. I’m very grateful for every contribution the project received. I’m also really happy to see "old names" replying from time to time on the mailing list. It makes me feel like there’s an emerging Offpunk community where everybody can contribute at their own pace.

There were a lot of changes between 2.8 and 3.0, which probably means some new bugs and some regressions. We count on you, yes, you!, to report them and make 3.1 a lot more stable. It’s as easy at typing "bugreport" in offpunk!

From the deepest of my terminal, thank you!

But enough with the cheering, let’s jump to…

The 11 most important changes in Offpunk 3.0

• Instructions to install Offpunk
• Dowload Offpunk 3.0

0. Use Offpunk in your language.

Offpunk is now translatable and has been translated into Spanish, Galician, and Dutch. Step-in to translate Offpunk into your language! (awesome work by JMCS with the help of Bert Livens)

• How to help translating Offpunk (offpunk.net)

1. Openk as a standalone tool

"opnk" standalone tool has been renamed to "openk" to make it more obvious. Openk is a command-line tool that tries to open any file in the terminal and, if not possible, opens it in your preferred software, falling back to xdg-open as a last resort.

People using opnk directly should change it everywhere. Users not using "opnk" in their terminal are not affected.

2. See XKCD comics in your terminal

"xkcdpunk" is a new standalone tool that allows displaying XKCD comics directly in your terminal.

3. Get only the good part and remove cruft for thousands of websites

Offpunk now integrates "unmerdify," a library written by Vincent Jousse that extracts the content of HTML articles using the "ftr-site-config" set of rules maintained by the FiveFilters community.

• Unmerdify on Codeberg

You can contribute by creating or improving rules for your frequently visited websites.

• Site Patterns | FiveFilters.org Docs (help.fivefilters.org)
• fivefilters/ftr-site-config: Site-specific article extraction rules to aid content extractors, feed readers, and 'read later' applications. (github.com)

If no ftr rule is found, Offpunk falls back to "readability," as has been the case since 0.1. "info" will tell you if unmerdify or readability was used to display the content of a page.

To use umerdify, users should manually clone the ftr-site-config repository:
git clone https://github.com/fivefilters/ftr-site-config.git
Then, in their offpunkrc:

set ftr_site_config /path/to/ftr-site-config

Automating this step is an objective for 3.1

4. Offpunk goes social with "share" and "reply"

New social functions: "share" to send the URL of a page by email and "reply" to reply to the author if an email is found. "Reply" will remember the email used for each site/capsule/hole.

5. Browse websites while logged in

Offpunk doesn’t support login into websites. But the new "cookies" command allows you to import a cookie txt file to be used with a given http domain.

From your traditional browser (Firefox, Librewolf, Chromium, … ), log into the website. Then export the cookie with the "cookie-txt" extension. Once you have this "mycookie.txt" text file, launch Offpunk and run:

cookies import mycookie.txt https://domain-of-the-cookie.net/

This allows you, for example, to read LWN.NET if you have a subscription. (contributed by Urja)

6. Bigger, better images, even in Gemini

Images are now displayed by default in gemini and their display size has been increased.

This can be reverted with the following lines in offpunkrc:

set images_size 40
set gemini_images false

Remember that images are displayed as "blocks" when reading a page but if you access the image URL directly (by following the yellow link beneath), the image will be displayed perfectly if you are using a sixels-compatible terminal.

7. Display hidden RSS/Atom links

If available, links to hidden RSS/Atom feeds are now displayed at the bottom of HTML pages.

This makes the "feed" command a lot less useful and allows you to quickly discover interesting new feeds.

8. Display blocked links

Links to blocked domains are now displayed in red by default.

This can be reverted with the following lines in offpunkrc:

theme blocked_link none

9. Preset themes

Support for multiple themes with "theme preset." Existing themes are "offpunk1" (default), "cyan," "yellow" and "bw." Don’t hesitate to contribute yours!

10. Better redirects and true blocks

"redirects" now operate on the netcache level. This means that no requests to blocked URLs should ever be made (which was still happening before)

And many changes, improvements and bugfixes

- "root" is now smarter and goes to the root of a website, not the domain.
Old behaviour can still be achieved with "root /"
- "ls" command is deprecated and has been replaced by "links"
- new "websearch" command configured to use wiby.me by default
- "set default_cmd" allows you to configure what Offpunk will do when pressing enter on an empty command line. By default, it is "links 10."
- "view switch" allows you to switch between normal and full view (contributed by Andrew Fowlie)
- "help help" will allow you to send an email to the offpunk-users mailing list
- "bugreport" will send a bug report to the offpunk-devel mailing list
- And, of course, multiple bugfixes…

• Come and join the Offpunk community!

The Disconnected Git Workflow

Using git-send-email while being offline and with multiple email accounts

WARNING: the following is a technical reminder for my future self. If you don’t use the "git" software, you can safely ignore this post.

The more I work with git-send-email, the less I find the GitHub interface sufferable.

Want to send a small patch to a GitHub project? You need to clone the repository, push your changes to your own branch, then ask for a pull request using the cumbersome web interface, replying to comments online while trying to avoid smileys.

With git send-email, I simply work offline, do my local commit, then:

git send-email HEAD^

And I’m done. I reply to comments by email, with Vim/Mutt. When the patch is accepted, getting a clean tree usually boils down to:

git pull
git rebase

Yeah for git-send-email!

• Guide to setup git send-email (git-send-email.io)

And, yes, I do that while offline and with multiple email accounts. That’s one more reason to hate GitHub.

• How GitHub monopoly is destroying the open source ecosystem (ploum.net)
• We need to talk about your GitHub addiction (ploum.net)

One mail account for each git repository

The secret is not to configure email accounts in git but to use "msmtp" to send email. Msmtp is a really cool sendmail replacement.

In .msmtprc, you can configure multiple accounts with multiple options, including calling a command to get your password.

# account 1 - pro
account work
host smtp.company.com
port 465
user login@company.com
from ploum@company.com
password SuPeRstr0ngP4ssw0rd
tls_starttls off

# personal account for FLOSS
account floss
host mail.provider.net
port 465
user ploum@mydomain.net
from ploum@mydomain.net
from ploum*@mydomain.net
passwordeval "cat ~/incredibly_encrypted_password.txt | rot13"
tls_starttls off

The important bit here is that you can set multiple "from" addresses for a given account, including a regexp to catch multiple aliases!

Now, we will ask git to automatically use the right msmtp account. In your global .gitconfig, set the following:

[sendemail]
   sendmailCmd = /usr/bin/msmtp --set-from-header=on
   envelopeSender = auto

The "envelopesender" option will ensure that the sendemail.from will be used and given to msmtp as a "from address." This might be redundant with "--set-from-header=on" in msmtp but, in my tests, having both was required. And, cherry on the cake, it automatically works for all accounts configured in msmtprc.

Older git versions (< 2.33) don’t have sendmailCmd and should do:

[sendemail]
   smtpserver = /usr/bin/msmtp
   smtpserveroption = --set-from-header=on
   envelopesender = auto

I usually stick to a "ploum-PROJECT@mydomain.net" for each project I contribute to. This allows me to easily cut spam when needed. So far, the worst has been with a bug reported on the FreeBSD Bugzilla. The address used there (and nowhere else) has since been spammed to death.

In each git project, you need to do the following:

1. Set the email address used in your commit that will appear in "git log" (if different from the global one)

git config user.email "Ploum <ploum-PROJECT@mydomain.net>"

2. Set the email address that will be used to actually send the patch (could be different from the first one)

git config sendemail.from "Ploum <ploum-PROJECT@mydomain.net>"

3. Set the email address of the developer or the mailing list to which you want to contribute

git config sendemail.to project-devel@mailing-list.com

Damn, I did a commit with the wrong user.email!

Yep, I always forget to change it when working on a new project or from a fresh git clone. Not a problem. Just use "git config" like above, then:

git commit --amend --reset-author

And that’s it.

Working offline

I told you I mostly work offline. And, as you might expect, msmtp requires a working Internet connection to send an email.

But msmtp comes with three wonderful little scripts: msmtp-enqueue.sh, msmtp-listqueue.sh and msmtp-runqueue.sh.

The first one saves your email to be sent in ~/.msmtpqueue, with the sending options in a separate file. The second one lists the unsent emails, and the third one actually sends all the emails in the queue.

All you need to do is change the msmtp line in your global .gitconfig to call the msmtpqueue.sh script:

[sendemail]
    sendmailcmd = /usr/libexec/msmtp/msmtpqueue/msmtp-enqueue.sh --set-from-header=on
    envelopeSender = auto 

In Debian, the scripts are available with the msmtp package. But the three are simple bash scripts that can be run from any path if your msmtp package doesn’t provide them.

You can test sending a mail, then check the ~/.msmtpqueue folder for the email itself (.email file) and the related msmtp command line (.msmtp file). It happens nearly every day that I visit this folder to quickly add missing information to an email or simply remove it completely from the queue.

Of course, once connected, you need to remember to run:

/usr/libexec/msmtp/msmtpqueue/msmtp-runqueue.sh

If not connected, mails will not be sent and will be kept in the queue. This line is obviously part of my do_the_internet.sh script, along with "offpunk --sync".

• How I use Offpunk (by Ploum) (offpunk.net)

It is not only git!

If it works for git, it works for any mail client. I use neomutt with the following configuration to use msmtp-enqueue and reply to email using the address it was sent to.

set sendmail="/usr/libexec/msmtp/msmtpqueue/msmtp-enqueue.sh --set-from-header=on"
unset envelope_from_address
set use_envelope_from
set reverse_name
set from="ploum@mydomain.net"
alternates ploum[A-Za-z0-9]*@mydomain.net

Of course, the whole config is a little more complex to handle multiple accounts that are all stored locally in Maildir format through offlineimap and indexed with notmuch. But this is a bit out of the scope of this post.

At least, you get the idea, and you could probably adapt it to your own mail client.

Conclusion

Sure, it’s a whole blog post just to get the config right. But there’s nothing really out of this world. And once the setup is done, it is done for good. No need to adapt to every change in a clumsy web interface, no need to use your mouse. Simple command lines and simple git flow!

Sometimes, I work late at night. When finished, I close the lid of my laptop and call it a day without reconnecting my laptop. This allows me not to see anything new before going to bed. When this happens, queued mails are sent the next morning, when I run the first do_the_internet.sh of the day.

And it always brings a smile to my face to see those bits being sent while I’ve completely forgotten about them…

Why there’s no European Google?

And why it is a good thing!

With some adjustments, this post is mostly a translation of a post I published in French three years ago. In light of the European Commission’s "call for evidence on Open Source," and as a professor of "Open Source Strategies" at École Polytechnique de Louvain, I thought it was a good idea to translate it into English as a public answer to that call.

• Pourquoi n’y a-t-il pas de Google européen ? (ploum.net)
• European Commission issues call for evidence on open source (lwn.net)

Google (sorry, Alphabet), Facebook (sorry, Meta), Twitter (sorry, X), Netflix, Amazon, Microsoft. All those giants are part of our daily personal and professional lives. We may even not interact with anything else but them. All are 100% American companies.

China is not totally forgotten, with Alibaba, TikTok, and some services less popular in Europe yet used by billions worldwide.

What about European tech champions? Nearly nothing, to the great sadness of politicians who believe that the success of a society is measured by the number of billionaires it creates.

Despite having few tech-billionaires, Europe is far from ridiculous. In fact, it’s the opposite: Europe is the central place that allowed most of our tech to flourish.

The Internet, the interconnection of most of the computers in the world, has existed since the late sixties. But no protocol existed to actually exploit that network, to explore and search for information. At the time, you needed to know exactly what you wanted and where to find it. That’s why the USA tried to develop a protocol called "Gopher."

At the same time, the "World Wide Web," composed of the HTTP protocol and the HTML format, was invented by a British citizen and a Belgian citizen who were working in a European research facility located in Switzerland. But the building was on the border with France, and there’s much historical evidence pointing to the Web and its first server having been invented in France.

It’s hard to be more European than the Web! It looks like the Official European Joke! (And, yes, I consider Brits Europeans. They will join us back, we miss them, I promise.)

• The Official European Joke (ploum.net)

Gopher is still used by a few hobbyists (like yours trully), but it never truly became popular, except for a very short time in some parts of America. One of the reasons might have been that Gopher’s creators wanted to keep their rights to it and license any related software, unlike the European Web, which conquered the world because it was offered as a common good instead of seeking short-term profits.

While Robert Cailliau and Tim Berners-Lee were busy inventing the World Wide Web in their CERN office, a Swedish-speaking Finnish student started to code an operating system and make it available to everyone under the name "Linux." Today, Linux is probably the most popular operating system in the world. It runs on any Android smartphone, is used in most data centers, in most of your appliances, in satellites, in watches and is the operating system of choice for many of the programmers who write the code you use to run your business. Its creator, the European Linus Torvalds, is not a billionaire. And he’s very happy about it: he never wanted to become one. He continued coding and wrote the "git" software, which is probably used by 100% of the software developers around the world. Like Linux, Git is part of the common good: you can use it freely, you can modify it, you can redistribute it, you can sell it. The only thing you cannot do? Privatize it. This is called "copyleft."

In 2017, a decentralized and ethical alternative to Twitter appeared: Mastodon. Its creator? A German student, born in Russia, who had the goal of allowing social network users to leave monopolies to have humane conversations without being spied on and bombarded with advertising or pushed-by-algorithm fake news. Like Linux, like git, Mastodon is copyleft and now part of the common goods.

Allowing human-scale discussion with privacy and without advertising was also the main motivation behind the Gemini protocol (whose name has since been hijacked by Google AI). Gemini is a stripped-down version of the Web which, by design, is considered definitive. Everybody can write Gemini-related software without having to update it in the future. The goal is not to attract billions of users but to be there for those who need it, even in the distant future. The creator of the Gemini protocol wishes to remain anonymous, but we know that the project started while he was living in Finland.

I could continue with the famous VLC media player, probably the most popular media player in the world. Its creator, the Frenchman Jean-Baptiste Kempf, refused many offers that would have made him a very rich man. But he wanted to keep VLC a copyleft tool part of the common goods.

Don’t forget LibreOffice, the copyleft office suite maintained by hundreds of contributors around the world under the umbrella of the Document Foundation, a German institution.

We often hear that Europeans don’t have, like Americans, the "success culture." Those examples, and there are many more, prove the opposite. Europeans like success. But they often don’t consider "winning against the whole society" as one. Instead, they tend to consider success a collective endeavour. Success is when your work is recognized long after you are gone, when it benefits every citizen. Europeans dream big: they hope that their work will benefit humankind as a whole!

We don’t want a European Google Maps! We want our institutions at all levels to contribute to OpenStreetMap (which was created by a British citizen, by the way).

Google, Microsoft, Facebook may disappear tomorrow. It is even very probable that they will not exist in fourty or fifty years. It would even be a good thing. But could you imagine the world without the Web? Without HTML? Without Linux?

Those European endeavours are now a fundamental infrastructure of all humanity. Those technologies are definitely part of our long-term history.

In the media, success is often reduced to the size of a company or the bank account of its founder. Can we just stop equating success with short-term economic growth? What if we used usefulness and longevity? What if we gave more value to the fundamental technological infrastructure instead of the shiny new marketing gimmick used to empty naive wallets? Well, I guess that if we changed how we measure success, Europe would be incredibly successful.

And, as Europeans, we could even be proud of it. Proud of our inventions. Proud of how we contribute to the common good instead of considering ourselves American vassals.

Some are proud because they made a lot of money while cutting down a forest. Others are proud because they are planting trees that will produce the oxygen breathed by their grandchildren. What if success was not privatizing resources but instead contributing to the commons, to make it each day better, richer, stronger?

The choice is ours. We simply need to choose whom we admire. Whom we want to recognize as successful. Whom we aspire to be when we grow up. We need to sing the praises of our true heroes: those who contribute to our commons.

• Picture by Frans Snik of a LEGO version of the European Extremely Large Telescope. Yes, LEGO is a European invention too.

Giving University Exams in the Age of Chatbots

What I like most about teaching "Open Source Strategies" at École Polytechnique de Louvain is how much I learn from my students, especially during the exam.

I dislike exams. I still have nightmares about exams. That’s why I try to subvert this stressful moment and make it a learning opportunity. I know that adrenaline increases memorization dramatically. I make sure to explain to each student what I was expecting and to be helpful.

Here are the rules:

1. You can have all the resources you want (including a laptop connected to the Internet)
2. There’s no formal time limit (but if you stay too long, it’s a symptom of a deeper problem)
3. I allow students to discuss among themselves if it is on topic. (in reality, they never do it spontanously until I force two students with a similar problem to discuss together)
4. You can prepare and bring your own exam question if you want (something done by fewer than 10% of the students)
5. Come dressed for the exam you dream of taking!

This last rule is awesome. Over the years, I have had a lot of fun with traditional folkloric clothing from different countries, students in pajamas, a banana and this year’s champion, my Studentausorus Rex!

My all-time favourite is still a fully clothed Minnie Mouse, who did an awesome exam with full face make-up, big ears, big shoes, and huge gloves. I still regret not taking a picture, but she was the very first student to take my words for what was a joke and started a tradition over the years.

Giving Chatbots Choice to the Students

Rule N°1 implies having all the resources you want. But what about chatbots? I didn’t want to test how ChatGPT was answering my questions, I wanted to help my students better understand what Open Source means.

Before the exam, I copy/pasted my questions into some LLMs and, yes, the results were interesting enough. So I came up with the following solution: I would let the students choose whether they wanted to use an LLM or not. This was an experiment.

The questionnaire contained the following:

# Use of Chatbots

Tell the professor if you usually use chatbots (ChatGPT/LLM/whatever) when doing research and investigating a subject. You have the choice to use them or not during the exam, but you must decide in advance and inform the professor.

Option A: I will not use any chatbot, only traditional web searches. Any use of them will be considered cheating.

Option B: I may use a chatbot as it’s part of my toolbox. I will then respect the following rules:
1) I will inform the professor each time information come from a chatbot
2) When explaining my answers, I will share the prompts I’ve used so the professor understands how I use the tool
3) I will identify mistakes in answers from the chatbot and explain why those are mistakes

Not following those rules will be considered cheating. Mistakes made by chatbots will be considered more important than honest human mistakes, resulting in the loss of more points. If you use chatbots, you should be held accountable for the output.

I thought this was fair. You can use chatbots, but you will be held accountable for it.

Most Students Don’t Want to Use Chatbots

This January, I saw 60 students. I interacted with each of them for a mean time of 26 minutes. This is a tiring but really rewarding process.

Of 60 students, 57 decided not to use any chatbots. For 30 of them, I managed to ask them to explain their choices. For the others, I unfortunately did not have the time. After the exam, I grouped those justifications into four different clusters. I did it without looking at their grades.

The first group is the "personal preference" group. They prefer not to use chatbots. They use them only as a last resort, in very special cases or for very specific subjects. Some even made it a matter of personal pride. Two students told me explicitly "For this course, I want to be proud of myself." Another also explained: "If I need to verify what an LLM said, it will take more time!"

The second group was the "never use" one. They don’t use LLMs at all. Some are even very angry at them, not for philosophical reasons, but mainly because they hate the interactions. One student told me: "Can I summarize this for you? No, shut up! I can read it by myself you stupid bot."

The third group was the "pragmatic" group. They reasoned that this was the kind of exam where it would not be needed.

The last and fourth group was the "heavy user" group. They told me they heavily use chatbots but, in this case, were afraid of the constraints. They were afraid of having to justify a chatbot’s output or of missing a mistake.

After doing that clustering, I wrote the grade of each student in its own cluster and I was shocked by how coherent it was. Note: grades are between 0 and 20, with 10 being the minimum grade to pass the class.

The "personal preference" students were all between 15 and 19, which makes them very good students, without exception! The "proud" students were all above 17!

The "never use" was composed of middle-ground students around 13 with one outlier below 10.

The pragmatics were in the same vein but a bit better: they were all between 12 and 16 without exceptions.

The heavy users were, by far, the worst. All students were between 8 and 11, with only one exception at 16.

This is, of course, not an unbiased scientific experiment. I didn’t expect anything. I will not make any conclusion. I only share the observation.

But Some Do

Of 60 students, only 3 decided to use chatbots. This is not very representative, but I still learned a lot because part of the constraints was to show me how they used chatbots. I hoped to learn more about their process.

The first chatbot student forgot to use it. He did the whole exam and then, at the end, told me he hadn’t thought about using chatbots. I guess this put him in the "pragmatic" group.

The second chatbot student asked only a couple of short questions to make sure he clearly understood some concepts. This was a smart and minimal use of LLMs. The resulting exam was good. I’m sure he could have done it without a chatbot. The questions he asked were mostly a matter of improving his confidence in his own reasoning.

This reminded me of a previous-year student who told me he used chatbots to study. When I asked how, he told me he would tell the chatbot to act as the professor and ask exam questions. As a student, this allowed him to know whether he understood enough. I found the idea smart but not groundbreaking (my generation simply used previous years’ questions).

The third chatbot-using student had a very complex setup where he would use one LLM, then ask another unrelated LLM for confirmation. He had walls of text that were barely readable. When glancing at his screen, I immediately spotted a mistake (a chatbot explaining that "Sepia Search is a compass for the whole Fediverse"). I asked if he understood the problem with that specific sentence. He did not. Then I asked him questions for which I had seen the solution printed in his LLM output. He could not answer even though he had the answer on his screen.

But once we began a chatbot-less discussion, I discovered that his understanding of the whole matter was okay-ish. So, in this case, chatbots disserved him heavily. He was totally lost in his own setup. He had LLMs generate walls of text he could not read. Instead of trying to think for himself, he tried to have chatbots pass the exam for him, which was doomed to fail because I was asking him, not the chatbots. He passed but would probably have fared better without chatbots.

Can chatbots help? Yes, if you know how to use them. But if you do, chances are you don’t need chatbots.

A Generational Fear of Cheating

One clear conclusion is that the vast majority of students do not trust chatbots. If they are explicitly made accountable for what a chatbot says, they immediately choose not to use it at all.

One obvious bias is that students want to please the teacher, and I guess they know where I am on this spectrum. One even told me: "I think you do not like chatbots very much so I will pass the exam without them" (very pragmatic of him).

But I also minimized one important generational bias: the fear of cheating. When I was a student, being caught cheating was a clear zero for the exam. You could, in theory, be expelled from university for aggravated cheating, whatever "aggravated" could mean.

During the exam, a good number of students called me panicked because Google was forcing autogenerated answers and they could not disable it. They were very worried I would consider this cheating.

First, I realized that, like GitHub, Google has a 100% market share, to the point students don’t even consider using something else a possibility. I should work on that next year.

Second, I learned that cheating, however lightly, is now considered a major crime. It might result in the student being banned from any university in the country for three years. Discussing exam with someone who has yet to pass it might be considered cheating. Students have very strict rules on their Discord.

I was completely flabbergasted because, to me, discussing "What questions did you have?" was always part of the collaboration between students. I remember one specific exam where we gathered in an empty room and we helped each other before passing it. When one would finish her exam, she would come back to the room and tell all the remaining students what questions she had and how she solved them. We never considered that "cheating" and, as a professor, I always design my exams hoping that the good one (who usually choose to pass the exam early) will help the remaining crowd. Every learning opportunity is good to take!

I realized that my students are so afraid of cheating that they mostly don’t collaborate before their exams! At least not as much as what we were doing.

In retrospect, my instructions were probably too harsh and discouraged some students from using chatbots.

Stream of Consciousness

Another innovation I introduced in the 2026 exam was the stream of consciousness. I asked them to open an empty text file and keep a stream of consciousness during the exam. The rules were the following:

In this file, please write all your questions and all your answers as a "stream of consciousness." This means the following rules:

1. Don’t delete anything.
2. Don’t correct anything.
3. Never go backward to retouch anything.
4. Write as thoughts come.
5. No copy/pasting allowed (only exception: URLs)
6. Rule 5. implies no chatbot for this exercice. This is your own stream of consciousness.

Don’t worry, you won’t be judged on that file. This is a tool to help you during the exam. You can swear, you can write wrong things. Just keep writing without deleting. If you are lost, write why you are lost. Be honest with yourself.

This file will only be used to try to get you more points, but only if it is clear that the rules have been followed.

I asked them to send me the file within 24h after the exam. Out of 60 students, I received 55 files (the remaining 5 were not penalized). There was also a bonus point if you sent it to the exam git repository using git-send-email, something 24 managed to do correctly.

The results were incredible. I did not read them all but this tool allowed me to have a glimpse inside the minds of the students. One said: "I should have used AI, this is the kind of question perfect for AI" (he did very well without it). For others, I realized how much stress they had but were hiding. I was touched by one stream of consciousness starting with "I’m stressed, this doesn’t make any sense. Why can’t we correct what we write in this file" then, 15 lines later "this is funny how writing the questions with my own words made the problem much clearer and how the stress start to fade away".

And yes, I read all the failed students and managed to save a bunch of them when it was clear that they, in fact, understood the matter but could not articulate it well in front of me because of the stress. Unfortunately, not everybody could be saved.

Conclusion

My main takeaway is that I will keep this method next year. I believe that students are confronted with their own use of chatbots. I also learn how they use them. I’m delighted to read their thought processes through the stream of consciousness.

Like every generation of students, there are good students, bad students and very brilliant students. It will always be the case, people evolve (I was, myself, not a very good student). Chatbots don’t change anything regarding that. Like every new technology, smart young people are very critical and, by defintion, smart about how they use it.

The problem is not the young generation. The problem is the older generation destroying critical infrastructure out of fear of missing out on the new shiny thing from big corp’s marketing department.

Most of my students don’t like email. An awful lot of them learned only with me that Git is not the GitHub command-line tool. It turns out that by imposing Outlook with mandatory subscription to useless academic emails, we make sure that students hate email (Microsoft is on a mission to destroy email with the worst possible user experience).

I will never forgive the people who decided to migrate university mail servers to Outlook. This was both incompetence and malice on a terrifying level because there were enough warnings and opposition from very competent people at the time. Yet they decided to destroy one of the university’s core infrastructures and historical foundations (UCLouvain is listed by Peter Salus as the very first European university to have a mail server, there were famous pioneers in the department).

By using Outlook, they continue to destroy the email experience. Out of 55 streams of consciousness, 15 ended in my spam folder. All had their links destroyed by Outlook. And university keep sending so many useless emails to everyone. One of my students told me that they refer to their university email as "La boîte à spams du recteur" (Chancellor’s spam inbox). And I dare to ask why they use Discord?

Another student asked me why it took four years of computer engineering studies to get a teacher explaining to them that Git was not GitHub and that GitHub was part of Microsoft. He had a distressed look: "How could I have known? We were imposed GitHub for so many exercises!"

• How GitHub monopoly is destroying the open source ecosystem (ploum.net)

Each year, I tell my students the following:

It took me 20 years after university to learn what I know today about computers. And I’ve only one reason to be there in front of you: be sure you are faster than me. Be sure that you do it better and deeper than I did. If you don’t manage to outsmart me, I will have failed.

Because that’s what progress is about. Progress is each generation going further than the previous one while learning from the mistakes of your elders. I’m there to tell you about my own mistakes and the mistakes of my generation.

I know that most of you are only there to get a diploma while doing the minimal required effort. Fair enough, that’s part of the game. Challenge accepted. I will try to make you think even if you don’t intend to do it.

In earnest, I have a lot of fun teaching, even during the exam. For my students, the mileage may vary. But for the second time in my life, a student gave me the best possible compliment:

— You know, you are the only course for which I wake up at 8AM.

To which I responded:

– The feeling is mutual. I hate waking up early, except to teach in front of you.

La ligne de commande communiste, le code Baudot et le comte ChatGPT

Il y a quelques mois, un lecteur m’a fait découvrir un incroyable jeu historico-politique entièrement réalisé en AsciiArt: « Le comte et la communiste », de Tristan Pun.

• Le comte et la communiste (debeauharnais.itch.io)

La simplicité des graphismes m’a fait plonger dans l’histoire comme dans un excellent livre. Jouant une espionne/servante, vous êtes amené à découvrir l’envers du décor d’un château aristocrate pendant la Première Guerre mondiale. De manière très intéressante pour le sous-texte sociopolitique, le temps passé dans le château est proportionnel au nombre de lessives. Car, bien qu’espionne, vous êtes avant tout une servante et il va falloir se remonter les manches et faire la lessive !

Si je vous en parle avec tant d’enthousiasme, c’est qu’à un moment du jeu vous trouvez une bande de papier contenant un message au format Baudot. Le code Baudot est l’ancêtre de l’ASCII et le premier encodage de caractère utilisé sur les télégraphes automatiques : le message est directement reçu sur des bandes de papier, à la différence du Morse qui nécessite un opérateur humain à la réception.

Dans le jeu, le message est trouvé sur une bande de papier servant de signet qui ressemble à ceci.

Sur une plaque, vous trouverez toutes les informations pour déchiffrer le code Baudot.

Il ne reste plus qu’à déchiffrer le message… Mais, très vite, j’ai trouvé ça laborieux. Et si je demandais à ma fidèle ligne de commande de le faire à ma place ? Après tout, la lecture de l’excellentissime « Efficient Linux at the Command Line », de Daniel J. Barret, m’a mis en confiance.

Décoder le message Baudot en ligne de commande

Attention, je vous préviens, ça va être très technique. Vous n’êtes pas obligé de vous infliger ça.

Vous êtes toujours là ? C’est parti !

Une fois le code en Asciiart copié/collé dans un fichier, on va l’afficher avec "cat". Même si ce n’est pas strictement nécessaire, je commence toujours toutes mes chaînes de pipe avec cat. Cela me semble plus clair.

Ici, la difficulté est que je veux accéder aux colonnes : je veux reconstruire un mot en prenant les premières lettres de chaque ligne, puis les secondes de chaque ligne, etc.

La commande Unix qui se rapproche le plus de cela est "cut". Cut permet de prendre le Xème caractère d’une ligne avec l’option -cX. Pour la quinzième, je fais donc "cut -c15". Tout de suite, je réalise que je vais avoir besoin d’une boucle. Considérant que les lignes font moins de 100 caractères, je peux faire un

for i in {1..100}; do" avec un "cut -c$i

Et voilà, je sais désormais isoler chaque colonne.

Dans l’asciiart, le binaire du code Baudot est représenté par des "o" et des espaces. Par soucis de clarté, je vais remplacer les espaces par des "l" (ça ressemble à un "1" binaire). La commande Unix pour faire des substitutions de caractères est "tr" (translate): tr " " l (l’espace est entre guillemets).

Toujours par souci de clarté, je vais effacer tout ce qui n’est pas un o ou un l. Tr permet d’effacer des caractères avec l’option -d et de prendre "tous les caractères sauf ceux de la liste" avec l’option -c (complémentaire). Je rajoute donc un tr -cd "ol".

La chaîne dans ma boucle ressemble donc à:

cat $1|cut -c"$i"|tr " " l|tr -cd "ol"

Problème : mes groupes de 5 lettres sont toujours verticaux !

J’ai tenté l’approche de supprimer les retours à la ligne avec tr -d "\n", mais cela produit des résultats bizarres, surtout que le dernier retour à la ligne est lui aussi supprimé.

Du coup, la commande qui me semble tout indiquée pour faire cela est le contraire de "cut" : "paste". Mais, essayez de lancer la chaîne dans "paste" : rien ne se passe !

Et pour cause : paste est initialement conçu pour joindre chaque ligne de plusieurs fichiers. Ici, il n’y a qu’un seul fichier : le flux d’entrée. Paste ne peux donc rien lui joindre.

Heureusement, l’option "-s" permet de dire à paste de tout mettre sur une seule ligne. Je suis complètement passé à côté pour une raison très simple : la page man de paste est incompréhensible.

-s, --serial
copier un fichier à la fois au lieu de le faire en parallèle

Je défie quiconque de voir le rapport entre la page de man et la fonction réelle. J’ai heureusement eu l’intuition de tester avec "tldr" au lieu de man.

Join all the lines into a single line, using TAB as delimiter:
paste -s path/to/file

Avouez que c’est déjà beaucoup plus clair !

Un p’tit test me révèle que c’est presque bon. Tout est sur une ligne. Sauf qu’il y’a désormais des TAB entre chaque lettre. On pourrait les enlever avec tr. Ou simplement dire à paste de ne pas les mettre en utilisant à la place un séparateur nul:

cat $1|cut -c"$i"|tr " " l|tr -cd "ol"|paste -sd "\0"

Ma boucle for m’affiche désormais chaque lettre sur une ligne. Il ne me reste qu’à traduire le code Baudot.

Je peux, par exemple, rajouter un pipe vers "sed s/llloo/a/" pour remplacer les lettres a. Et un pipe vers un nouveau sed pour la lettre b et ainsi de suite. Ça fonctionne, mais c’est moche et très lent (chaque sed lançant son propre process).

Lorsqu’on a beaucoup de règles sed, autant les mettre dans un fichier baudot.sed qui contient les commandes sed, une par ligne :

s/llloo/a/
s/oollo/b/
s/loool/c/
...

Je peux appeler ces règles avec "sed -f baudot.sed".

Le code Baudot a une subtilité : y’a un code qui permet de passer en mode "caractère spécial". Je ne prends pas la tête, je me contente de remplacer ce code par "<" et le code pour revenir en mode normal par ">" (ces deux caractères n’existant pas dans le code Baudot). De cette manière, je sais que toute lettre entre < > n’est pas vraiment la lettre, mais le symbole correspondant. Un <m> est en réalité un ".". Également, il y a de nombreux espaces avant et après les messages, qui ont été convertis en autant de "l". Là, j’ai fait un bon gros hack. Au lieu de mettre un "l" normal dans le code Baudot, j’ai mis un L majuscule. Ensuite, à la toute fin, je supprime les "l" restant avec une règle globale : "s/l//g" (le "g" indique de changer tous les l sur une ligne, même s’il y’en a plusieurs). Puis, je remets en minuscule le "L" avec "s/L/l". Oui, c’est un bon gros hack. Ça fait l’affaire.

Mon fichier baudot.sed ressemble alors à ça :

s/ooloo/</
s/lloll/ /
s/ooooo/>/
s/llloo/a/
s/oollo/b/
s/loool/c/
s/lollo/d/
s/llllo/e/
s/loolo/f/
s/oolol/g/
s/ololl/h/
s/llool/i/
s/loloo/j/
s/loooo/k/
s/ollol/L/
s/oooll/m/
s/looll/n/
s/oolll/o/
s/oolll/p/
s/olooo/q/
s/lolol/r/
s/llolo/s/
s/ollll/t/
s/llooo/u/
s/ooool/v/
s/olloo/w/
s/ooolo/x/
s/ololo/y/
s/olllo/z/
s/l//g
s/L/l/g

Mon message est décodé. Mais, bien entendu, il s’affiche verticalement.

Ça ne vous rappelle rien ?

Un bon vieux paste -sd "\0" remet tout à l’endroit (et cette fois-ci, je n’ai pas eu à chercher).

Dans le jeu, le seul symbole utilisé sera le point, qui est ici devenu un "<m>" ou "<m" s’il est à la fin du message. Soyons propres jusqu’au bout et rajoutons deux petits sed. On pourrait également faire un autre fichier sed avec tous les caractères, mais le jeu ne comporte finalement que deux messages.

Dommage, là j’étais chaud pour plus.

Mon script final est donc :

#!/bin/bash
for i in {1..100}; do
	cat $1|cut -c"$i"|tr " " l|tr -cd "ol"|paste -sd "\0"
done|sed -f baudot.sed|paste -sd "\0"|sed "s/<m>/./"|sed "s/<m/./"

Avec le recul, ce script est beaucoup plus simple et efficace qu’un script Python. Le script Python prendrait des dizaines de lignes et agirait sur une matrice de caractères. Les outils Unix, eux, agissent sur des flux de texte. J’y trouve une certaine élégance, un plaisir particulier. Le tout m’a pris environ 30 minutes, dont une bonne partie sur l’erreur de "tr -d \n".

Demander à ChatGPT de décoder le message Baudot

Comme mon compte Kagi me donne désormais accès à ChatGPT, je me suis dit que j’allais faire l’expérience de lui demander de résoudre le même problème que moi. Histoire de comprendre ce qu’est le « Vibe coding ».

Au départ, ChatGPT est perdu. Il ne sait clairement pas traduire les messages, m’assène de longues tables de soi-disant code Baudot (qui sont parfois correctes, mais pas toujours) et me raconte l’histoire de ce code (qui est généralement correcte, mais que je n’ai pas demandée). C’est verbeux, je dois lui dire plusieurs fois de faire cours et d’être efficace.

Je lui demande de me faire un script bash. Ses premières tentatives sont extrêmement longues et incompréhensibles. Il semble beaucoup aimer les scripts awk à rallonge.

Persévérant, je demande à ChatGPT de ne plus utiliser awk et je lui explicite chaque étape l’une après l’autre: je lui dit qu’il faut parcourir chaque colonne, la redresser, la convertir, etc.

Logiquement, il arrive à un résultat très similaire au mien. Il choisit d’utiliser "tr -d \n" pour supprimer les fins de ligne, mais, comme je l’ai dit, ça ne fonctionne pas correctement. Je lui passe, car je n’ai moi‑même pas compris l’erreur.

Je constate cependant une amélioration intéressante : là où j’ai demandé les 100 premières colonnes, chatGPT mesure la longueur avec :

for i in $(seq $(head -1 $1|wc -c)); do

Concrètement, il prend la première ligne avec "head -1", compte les caractères avec "wc -c" et construit une séquence de 1 jusqu’a ce nombre avec "seq".

C’est une excellente idée si on considère que la première ligne est représentative des autres. Une fois le "redressement" en place (il m’a fallu des dizaines de prompts et d’exemples pour arriver à lui expliquer, à chaque fois il prétend qu’il a compris et c’est faux), je lui demande de traduire en utilisant le code Baudot.

Au lieu de ma solution avec "sed", il fait une fonction bash "baudot_to_letter" qui est un gros case:

baudot_to_letter() {
  case $1 in
    00000) echo " " ;;
    00001) echo "E" ;;
    00010) echo "A" ;;

Fait amusant, les caractères ne sont pas en ordre alphabétique, mais dans l’ordre binaire de la représentation Baudot. Pourquoi pas ?

Je valide cette solution même si je préfère le sed parce que je n’avais pas envie de faire du bash mais utiliser les outils Unix. Je lui ai dit plusieurs fois que je voulais une commande Unix, pas un script bash avec plusieurs fonctions. Mais je lui laisse, car, au fond, sa solution est pertinente.

Maintenant que toutes les étapes ont été décrites, je teste son code final. Qui ne fonctionne pas, produisant une erreur bash. Je lui demande une nouvelle version en lui copiant-collant l’erreur. Après plusieurs itérations de ce type, j’ai enfin un script qui fonctionne et me renvoie la traduction du message suivante :

LLECJGCISTGCIUMCEISJEISKN CT UIXV SJECUZUCEXV

Le nombre de lettres n’est même pas correct. Je n’ai évidemment pas envie de débugger le truc. Je demande à lors à ChatGPT de lancer le script lui-même pour me donner la traduction :

À ce point de l’histoire, j’ai, montre en main, passé plus de temps à tenter d’utiliser ChatGPT que je n’en ai mis pour écrire mon propre script, certes imparfait, mais fonctionnel. Je me retrouve à faire des copier-coller incessants entre ChatGPT et mon terminal, à tenter de négocier des changements et déboguer du code que je n’ai pas écrit.

Et encore, pour ChatGPT, j’ai nettoyé le message en enlevant tout ce qui n’est pas le code Baudot. Mon propre script est beaucoup plus robuste. J’en ai marre et j’abandonne.

Moralité

ChatGPT n’est impressionnant que par sa capacité à converser, à prétendre. Oui, il peut parfois donner des idées. Il a par exemple amélioré la première ligne de mon script que j’avais bâclée.

Mais il faut avoir du temps à perdre. À ce moment-là, autant replonger dans un bon livre dont on sait que la majeure partie des idées sont bonnes.

ChatGPT peut être utile pour brainstormer à condition d’être soi-même fin connaisseur du domaine et très critique avec tout ce qui sort. C’est dire la débauche d’énergie pour finalement très peu de choses.

Vous croyez être plus efficace en utilisant l’AI, car vous passez moins de temps à « penser ». Mais les études semblent montrer qu’en réalité, vous êtes plus lent. Avec ChatGPT, on est tout le temps "occupé". On ne s’arrête jamais pour réfléchir à son problème, pour lire différentes ressources.

• Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (metr.org)

Et lorsque ChatGPT vous fait vraiment gagner du temps, c’est peut-être parce qu’il s’agit d’un domaine où vous n’êtes pas vraiment compétent. En ce qui concerne la ligne de commande, je ne peux que répéter ma suggestion de lire « Efficient Linux at the Command Line ». Vraiment !

Comme le rappelle très bien Cal Newport: qui a le plus intérêt à ce que vous soyez un ouvrier incompétent qui se contente de pousser aveuglément sur les boutons d’une machine au lieu de faire fonctionner son cerveau ?

• Be Wary of Digital Deskilling (calnewport.com)

Peut-être que passer quelques heures à espionner les aristocrates en étant obligé de faire des lessives vous fera réfléchir à ce sujet. En tout cas, ça en vaut la peine !

Bon amusement, bonnes réflexions et… bonnes lessives !