Beyond Chat: How Language Models Enhance Academic Knowledge

Language models like ChatGPT are on everyone's lips, or rather: Claude Code. Everyone is talking about how these creatures are going to replace you, how people no longer search on Google, how programmers may no longer be needed (and for many things, it’s clear that they might not be necessary), etc. There’s also talk of cognitive destruction or replacement, and that could very well be the case; I don’t know, although the premises on which that argument is based seem dubious to me. Whether it happens or not, I want to talk about something else: how one (me, in this case), as a humanist, can make use of artificial intelligence tools (in a broad sense) in an interesting way for our purposes.

I do believe that these artificial creatures (if I may call them that) can fulfill the role of what Argentine academics have always seen as science fiction: support staff. Typing assistants (as can be read in some papers that were dictated by professors to their secretaries), editing assistants (like the ones we can’t afford), translators (like the ones that are rarely useful due to the specificity of our technical vocabularies), programmers we could never afford and would be reluctant to hire even if we had them for the kinds of tasks we would need them for.

These artificial creatures (if I may call them that) can fulfill the role of what Argentine academics have always seen as science fiction: support staff. Typing assistants, editing assistants, translators, programmers.

This isn’t primarily for those who already have a workflow established. I’m talking about those who haven’t yet touched these tools and have only interacted through a chat interface. The things I do are what I do when I can forget that what I’m doing is work and try to do it as well as I can; there’s no cost-benefit calculation as the central motivation. It’s a bit of curiosity and a touch of convenience, nothing more.

I don’t intend to write either a guide or a manifesto; I simply want to share some things that impressed me and that I enjoyed learning over the past year. They are new ways of doing what we’ve always done and, paradoxically, finally incorporating tools that simplify our work, allowing us to focus primarily on writing texts.

I don’t know how these tools I’m describing will look in a year. What I do know is that it’s worth taking an interest in some of them. Very soon, for many tasks, you’ll be able to simply install an open-source model and run it offline; you won’t need to be connected to any data center. Everyone is talking about the loss of autonomy that “AI” implies (and if you’re interested in the other side of that coin, Juan Brodersen demystifies the Dark Web and presents Tor as a legitimate privacy infrastructure against surveillance capitalism in Dark Web, Tor, and anonymity as a limit to Big Tech). I want to talk for a while about some of the possibilities that language models give us to reclaim or strengthen it.

And if you get stuck on something, give this article to your artificial creature and ask it to guide you.

Researching in the humanities for those who aren’t involved in it

What we do in the humanities is argue; humans produce language in the form of arguments just as spiders produce webs. Whenever you use a “therefore,” “consequently,” “because” (in one of its uses), and words like that, you’re marking textually or verbally that there are premises and a conclusion in what you’re saying. Those who study argumentation do so in many different ways, but in the humanities in general, we work with and produce arguments. That means we analyze, reconstruct, and elaborate arguments, counterarguments, and things like that. In particular, in philosophy, one can understand the tradition we are part of as a discussion that has been going on for two thousand five hundred years. From the outside, it can be seen as a great waste of time, but every person who tried to defend that it is a great waste of time ended up doing philosophy. My point isn’t to defend philosophy to anyone; it’s to express that philosophy is done through arguing, and arguing is done with words.

Setting aside the definitions we can give about what is or isn’t “artificial intelligence,” what we call ChatGPT and similar models are precisely language models (Large Language Models, or LLMs for short), and that’s why they are interesting and potentially useful. If a spider were invented with large models of webs, it would probably be interested; I don’t think it would stop making webs, but maybe it would delegate some annoying parts of the process. We all argue; all speakers of any language do; it’s part of the process. Those of us who work in the humanities are simply a professional variant of arguers; we are interested in arguments on certain topics (including the argument itself and the relationship it has with who we are). You don’t need to know more about what I do to understand what I’m going to say. Of course, if anyone is interested, I have no problem answering their questions.

What might interest you is quantifying what sizes of text we’re talking about. A presentation at a conference (which we participate in several times a year) has approximately 3000-5000 words (if we have 20-35 minutes to present), and we present that with either a handout (a bullet point outline) or slides (something like PowerPoint); either of those has to be done specially. Generally, after presenting something at a conference, the idea is to transform that idea into an article (paper or essay), which will have between 7000-12000 words (between 20 and 40 pages). A book or doctoral thesis in the humanities has between 60000 and 120000 words (approximately between 150 and 400 pages, but there are much longer ones). All of that includes a form of writing that looks like this:

“Jürgen Habermas argued (1996), against Richard Rorty (1995), that the assertion of something as true implies the duty to justify it to anyone. Rorty (2000) responded that...”

That is, we not only argue in the text but also use complete texts as premises in our arguments; the way to do that is to cite them through citation systems like APA (the one above, by American Psychological Association), and they function like traffic rules. That way of citing was invented by the APA, but there are many others. The point is to condense information like through hyperlinks. When making a statement in an article or essay, we are either taking what we say as true without needing to justify it (like saying “as everyone knows, the President of the United States is Donald Trump” or “the sky is blue” [but more specific statements for each area]), or we consider it justified based on reasons we provided earlier (things we take as true and expect our interlocutors to as well) by arguing, or we justify it through arguments, in the sense of conclusions from arguments, from others (or also bringing up what someone said and stating where we got it from, in case anyone doubts). Of course, it takes a lot of time and effort to understand the arguments we condense into a single paragraph, and nowadays we use tools for that.

Typical tools in the arsenal of a humanities researcher (and other things)

If we work with text, we use word processors like the one I’m using to write this. It could be Notepad on Windows, or Word, but any place that serves as a notebook works, as what we’re going to publish are text files; we produce the one that the journal or event in question requests. Each word processor simplifies some aspect of the work, and there are specialized ones. For example, those of us who do something in philosophy, something related to sciences that use specific formalism (logic, mathematics, natural sciences, social sciences like economics and sociology, etc.) might write things with many formulas. There, multipurpose tools like Word become cumbersome, and we write in formats like LaTeX, which doesn’t require you to write as the final format will look but as a rich text format, like HTML, or the one many of us used in discussion forums where italics were marked by writing “[i]this would go in italics[/i]”; LaTeX looks something like this:

Lo que sigue a ``quiero'' quiero \emph{escribirlo en itálicas}, y quiero escribir luego de \emph{esto}: $p \to q$ (\textbf{una fórmula lógica} [en negrita])

and it looks like this:
\text{What follows after ``I want'' I want } \textit{to write it in italics}\text{, and I want to write after }\textit{this}\text{: }p \to q \text{ (\textbf{a logical formula})}

Additionally, it’s very relevant to use citation managers. Manually entering and controlling all the information that goes in those parentheses is very labor-intensive. A reasonable workflow (but there are many ways to do it) is to first write a draft based on notes or from memory of what one wants to say, then rewrite it, then go check that what we say others say is indeed so, that we can defend such or such statement, and we have to be sure that our notes are good, etc. One way to save work is to use a citation manager like Zotero, or the one that comes with Word. That’s basically a database that allows you to simply enter the complete citation through a code, which is not only the one that goes in the text in parentheses but also the one that goes at the end of the text in the bibliography. Writing “Habermas (1996)” somewhere in the text tells someone who knows how to interpret academic texts (which, like any text, is a kind of machine with coordinated parts) that they can find at the end an entry that gives all the necessary data to locate the complete text in a library or database, which should include the author’s name, publication year, title of the publication, publisher, and sometimes the city of publication. Simplifying a lot. If this tired you out, well, yes, doing it a lot just makes it routine, not more interesting.

No one who works in this field gets too worked up about these things, but it takes time. A citation manager coordinates the reference in the text with the bibliography list and ensures that neither references we didn’t use nor others appear. One can make mistakes, and a lot, so it’s always necessary to review manually, but one makes fewer mistakes by automating part of the process.

The one I like the most is Zotero. It not only allows me to download the reference of a text I searched for on sites like Google Scholar (also known as the section of Google that isn’t filled with ads or unsolicited AI recommendations) with one click in my browser. I look for the text by Habermas that someone cited, I didn’t know it, let’s say, I go to that page, find it, click a button, and Zotero downloads all its bibliographic data, and if it’s freely available, also the .pdf. If it’s not freely available and I have access to that text, I can drag it directly into the Zotero entry. In Zotero, I can use the included .pdf reader (which doesn’t spam me like Adobe’s), I can highlight, I can take the highlighted parts and include them in the integrated notes of the application, etc. A marvel.

It’s very common for me to read a text in search of evidence in favor or evidence against what I want to say, and sometimes something simply strikes me as interesting. I highlight it, transfer it to my notes, and that can remain until I want to write something about it. Not to mention that the notes on the text can also contain everything I want in terms of my comments. Later (hours or years later), it can be the seed of something I write.

Everyone who does this has their own way of not going crazy with these annotations, and we all fail to some extent. If you're already starting to feel irritated, you're getting the point.

Install Zotero and download your first referenceGo to zotero.org and download it. It's free and open-source. Install it like any other program. Install the browser connector from Zotero's download page: it's an extension for Chrome or Firefox that adds an icon to your browser. Open Google Scholar and search for an article that interests you. When you're on the results page, click the Zotero icon. Zotero grabs the author, year, title, journal, DOI, everything. If the PDF is available, it downloads it. In three seconds, you have a complete bibliographic entry that would have taken you at least three minutes to do by hand. Open Zotero and check your library. There’s the entry with all the details. Double-click the PDF, highlight something, right-click, "Add to Note." You've just linked a note to its source.

Argumentation in the age of LLMs

Just as musicians tend to be obsessive fetishists about their instruments, and the highest quality ones can work magic in the hands of the most talented musicians; those of us who work, but not with our hands, but with words can become a bit fetishistic about our tools, even the intangible ones. Generally, the best thing a tool can do for a worker is to become invisible. That’s why I often recommend to my colleagues who aren’t as fond of computers as I am that they shouldn’t buy the worst computer they can find; they should try to get the cheapest one that allows them to use it without thinking about the fact that they’re using it. If your computer takes a long time to turn on, it’s a bad computer for writing; if it takes a long time to open Word, you might be using the wrong operating system, etc. I don’t care if this progresses, if a language model or some system becomes conscious; what can already be done is enough of a change to pay attention to.

So, whether you’re a professional or a committed amateur in writing, you have notes everywhere, highlighted books (both paper and .pdf), you spend your time discussing things in writing and verbally about something, including demonstrating things mathematically or showing through a formal model that inflation is (or is not) a monetary phenomenon (if you can agree on that, that would be great), it doesn’t matter; the more juice you squeeze out of what you do, the more opportunities and freedom of action you have. The idea that you can’t reach because there are others in the way (you have to earn the disciplinary right to say what you want to say) that you need to resolve first might be scattered throughout your notes, through your arguments; you want, at the very least, to organize them. Well, your artificial buddy can do that for you. Claude Code, but also Gemini CLI (which you can download for free), or if you have a powerful enough computer (maybe you’re using a relatively new Macbook Air and can run local models).

Those of us who work, but not with our hands, but with words can become a bit fetishistic about our tools, even the intangible ones. Generally, the best thing a tool can do for a worker is to become invisible.

Suppose that, like me, you have some occasional incentive to learn LaTeX but find writing on web platforms like Overleaf horrible. Well, LaTeX is simply modified .html, you could run it locally. But like any kind of code, it can blow up in your face (especially if you’re a beginner), in this case: not compiling, meaning: not giving you the .pdf. The equivalent of having a Word file break (which is perfectly possible). There’s a limit to how sophisticated you can do things well in a program like Word before it’s worth doing something in LaTeX. That is, if you learned to play guitar on an acoustic one, the day you want to switch to an electro-acoustic or electric, you’ll have to work a lot to be able to do the same things you did with something more basic. There’s a cost to pay for a tool with more versatility. Well, Claude Code or any command-line language model can fix your LaTeX file immediately. Not only that, you can use it to learn; you can instruct it to do the following: every time it fixes something for you, it goes and adds an entry in a notes file explaining what it did and how to learn to do it yourself. Also, if it’s not something you’re interested in learning how to do, but it took you a while to get the artificial buddy to do it right, you can instruct it to note down, for the future (on your machine), how to do it in the way it did when it found the hole in the mate.

This is ideal because it smooths out the learning curve (and learning faster is one of the best definitions of intelligence we have, also used by AI researchers), you can try out new tools that give you a lot of action power (which are not AI themselves) with the ease that, if you mess up something that doesn’t really matter to you (the technical functioning of the tool), you have an assistant who can fix it for you. It’s like starting to use a typewriter for the first time and having a typewriter technician at your disposal, starting to use a guitar and having a luthier able to tune it every time you need, etc.

How AI Shapes Us: McLuhan and the Grammar of LLMs

Technologies affect us on a much deeper level than our political opinions or musical tastes. AIgen changes our sensory relationship with the environment, with others, and with ourselves. Only through great effort can we understand what technique has made of us.

421 | Quality information for your cognitive dietMariano Vilar

Of course, it doesn’t always do what I ask it to do well. I don’t let it write anything more than reformulations of my drafts, I don’t allow these systems to handle my citations. Part of the work of using it is reading it with the same distrust you read any of your drafts at three in the morning. It’s not a colleague who knows more than you; it’s an assistant that does many things well and some things poorly, and you need to know which is which. If you don’t have criteria about the domain in which you’re asking it things, you shouldn’t ask it anything serious. That’s why I insist I’m not talking about replacing what you know how to do.

A parenthesis on teaching: none of this is for my students who still don’t know how to write in the relevant sense (obviously, they know how to write, but not to argue like professionals, and they don’t have the critical mass of their own texts to take advantage of an artificial buddy). They need to learn to think for themselves, and it doesn’t help me at all for Claude to think for me. That’s my job. Even if tomorrow my employers paid Anthropic to do what I do, I would continue doing it. Not because I think I do it better than some current or future model, I don’t care. I’ve never been tempted to hire a colleague I consider better to think for me, and I won’t be tempted to hire a model to do the same.

Remember: you (and I) don’t care about becoming an expert in the tool itself. You don’t care about reading the Word manual; you care about opening the box and starting to use the tool to do what you know how to do. So we come to the first application of LLMs for people like me: making many tools, which are not AI, accessible to you. As long as these things live in your console, they overlap with the idea behind installing Linux. But you can also have Linux on Windows through WSL (Windows Subsystem for Linux), which allows you to use various distros in a virtual machine and use applications like the ones we’re going to see as if you were in Linux, and have Claude Code, for example, run in your Linux virtual machine and not in the Windows system itself; but a Mac also runs a UNIX system, and the console experience is almost the same. You can’t imagine the number of small pieces of software that exist and were created by and for programmers in different areas that could be useful to you. A language model can help you use them.

LLM as an interface between you and new tools

It’s full of tools that would take you a tremendous amount of work to learn for the benefit they could provide you in a reasonable time. For example, I with LaTeX. Dedicating more work than any LaTeX fan wants to admit, I can perfectly write an article on my own, and I could already do it, but it was so distracting from my workflow, from my deep concentration states (what Americans call flow state) —which are, at the core, what brings us closer to the source of our intellectual addictions— that I preferred not to spend time on it. Well, since I started using Claude Code for other things, I now hardly use Word at all. I use it more to correct my students’ work than anything else. Here we have a breaking point in the usual discourses about LLMs: yes, I’m using a product from Anthropic, a U.S. startup valued in the billions, but it allows me to use open-source software like Visual Studio Code (maintained by Microsoft) to write in LaTeX (a "programming" language that is free) and stop using Microsoft Word, which charges me a lot of dollars to use and doesn’t let me do what I want with my files. I can write LaTeX wherever I want, in any plain text processor, and I can compile it in free software, with tools made by people who think about my problems as an academic first, and who don’t charge for that (and, if they do charge, they still don’t want to shove it down everyone’s throat). Software made out of love for the craft.

Interviews with the Computer Underground - Episode I: Santiago Roland, administrator of undernet.uy

We inaugurate a series of interviews focused on hackers, hacktivists, and digital activists who fight, propose, and build real alternatives to the.

421 | Quality information for your cognitive dietSoldan

Think of it this way: you can narrate some slides instead of putting together slides in PowerPoint. In fact, forget about PowerPoint. Write your ideas in a text file (I'll soon recommend one, Markdown), ask an artificial critter to create some LaTeX slides that are ready to compile, even if you don't use LaTeX for anything else. Want images? Find them or generate them with a model and drop them in the folder where you're working; your critter will use them. They are your slides, your ideas, but you didn't spend two hours making them, at least not starting them. You can tune your personal style and not stick to a pre-established format. You can do this by writing, which is what you know how to do because you're a professional human spider.

This is just one example; think of any software you want to use to do whatever you want; an LLM can help you use it, especially if it involves directly or indirectly editing text. Claude Code was made for programming, writing code, but all types of writing, as communication experts tell us, involve a code. Programming languages were created to give instructions to machines. A tool made for writing code indirectly allows you to talk to your computer like never before. So, Claude Code lives in text; if that's the case, why is it relatively bad at writing in Word? Because Word is software with hidden code; if you learn to use formats with explicit code, you'll find the middle ground between you and the type of critter that Claude Code is.

**Write your first .md file and compile LaTeX using a language template. Download Visual Studio Code and install it. It’s free. Open it, create a new file, and save it as draft.md. Write whatever you like: an abstract, a summary of something you read, anything. Install the LaTeX Workshop extension from the VSCode extensions menu (the grid icon on the left). Open a terminal in VSCode (Ctrl+ñ or Terminal → New Terminal) and ask your command-line model (Claude Code, for example): “Convert draft.md to a LaTeX file ready to compile as an academic article.” The bot will generate a .tex file for you. With LaTeX Workshop installed, click the compile button and you’ll have your .pdf.

That brings us to a possible middle ground, and the format in which I'm writing this. Enter the Markdown format, or, as we who want to be its friends call it: .md

Markdown, the digital text format of the future that has been around for a while

Let’s talk more about Markdown (or .md) before we continue. If you’re wondering where the artificial intelligence is in this article, well, trust me, we’re already getting to what matters to you and me—if you’re like me. I’m referring to .md as the digital text format of the future (I think there’s another text format of the future—in fact, I believe the text technology for teaching is pen and paper, but that’s not what we’re focusing on right now). Markdown is a beautiful text format. It’s not Word, where what you see while typing is what the reader sees, nor is it LaTeX, which is like writing a web page. It’s lightly formatted text. It’s the format used in WhatsApp, where putting something between underscores (or between a pair of asterisks) means writing it in italics, putting something between two pairs of asterisks means making it bold, and so on. Markdown sits in the middle of many text formats; marking something with a # means it’s a heading, ## a section, and so on. You can write it in any plain-text editor—whether it’s Windows Notepad, Notepad++, GNOME, nano, or VIM, etc.—anything. These editors will often have a way to display italics rather than the code you used to mark them, which results in something that’s like Word in practice, but much more versatile. The .md text file contains almost nothing but the characters you type into it. Beautiful. This has its own merits; it’s a noble, versatile format that doesn’t rely on anyone having access to proprietary software, and it’s lightweight—among many other virtues. I find it beautiful. But I find it even more beautiful to use it with the app I use to write in it and organize my notes in .md, and here I stand, because I’m going to talk about Obsidian. Obsidian isn’t open-source software, but it’s free, and the developers are committed to that. There are reasons to believe them. In any case, if they were to break their promise, you have almost nothing to worry about; everything you write in Obsidian is on your computer in .md format, you can edit it with whatever you want, and there’s plenty of open-source software to help you do just that.

Markdown is a beautiful text format. It's not Word, where what you see while writing is what the reader sees; nor is it LaTeX, which resembles writing a web page. It's lightly enriched text.

Some things you can do in .md: for example, I talked about Zotero, which is open-source, free, and the best academic .pdf reader on the market. At least, as far as I know. It's simply amazing. Well, the notes you can associate with each text you read in Zotero (or each bibliographic entry) in Zotero are in .md, the same format I suggest. You can also sync those same notes with plugins to edit them in Obsidian; that is, you can take notes in Zotero in .md, have that note linked to your Vault in Obsidian and edit it there, and Zotero will detect it, or copy and paste; it doesn't matter. You can also install the Zotero plugin in Obsidian and insert academic references in whatever format you want in Obsidian directly; you can start writing a draft in Obsidian using your bibliography in Zotero, etc. This is incredibly useful once you set up a workflow.

Think of it this way: if you're like me, you have hundreds or thousands of .pdfs scattered across your computer. Probably in a folder connected to some cloud so you can work wherever you need. Given how expensive academic books are and how hard it is to access updated academic bibliographic material in Argentina, it's surely one of your best-kept treasures. Now, to make it useful, how do you organize it? By author? By topic? By the paper you're writing? There's no easy answer, and we all adopt some method. With something like Zotero, it's unnecessary; the .pdfs are linked to your bibliographic database, and you can create as many sub-libraries as you want, organizing them however you like. If you write in LaTeX, you can export a .bib file; moreover, you can have a .bib file that automatically syncs on your computer managed by Zotero, updating it with every change you make and placing it in every folder where you're working in LaTeX through a dynamic link (a symlink in Linux, for example), something for which Overleaf charges you, and unlike an academic from the North, your university isn't paying for it.

Now, you have notes in .md, they're organized in Obsidian, but what do you do with that? You start linking them; you have your summaries, your notes, your notes written over months, years, etc., all in Obsidian (I'll mention how to do this now, since you don't have them in .md). Well, you can create topic indexes and make hyperlinks from one to another. You can also create a link to a note that is created by means of you creating the link by putting something between two pairs of brackets. This means that when you click on that, you'll go to a new note, already linked to the previous one. If you develop the habit, you'll no longer have just a set of notes, but a system of knowledge of your own, your own linguistic city, the boundaries of your professional or artisan language (if, as Nietzsche demanded, you don't live off philosophy, but for it), and even better, everything you add enriches it. There exists a group of theories about concepts and meaning that consider that concepts acquire meaning in a system, in a logical space. Obsidian can be a way to embody that idea. Your notes, your thoughts, are enriched through the connections you draw, often implicitly, but now explicitly. Obsidian not only allows you to do that; it rewards you for doing so by giving you a network view of your notes; mine looks like this (and I'm a late adopter and still working on my notes):

440
Each point is a note, each connection is a link. What could be isolated buildings in my conceptual city are, instead, a community. Whether you're interested in doing something like this or not, what I can guarantee is that it doesn't take much work. You might say: "Mauro, I have work to do; I don't have time to start my note system and transfer it to .md, not even if it just involves copying and pasting, much less establishing links between these things," and you'd be right. But here you can use Claude Code or Gemini CLI to tell it: read all my notes in my Obsidian vault and create intermediate indexes by topic, create thematic maps that lead to each one. You can do this regularly and innovate because these artificial critters can read the folders on your computer that you give them access to, so they can read each one. But you can also choose not to do it if you don't see the point. That doesn't mean you can't multiply the value you give to your Obsidian vault with a language model. Recently, Obsidian released a version that provides a command-line interface (or CLI). This means that if you open your computer's console, you can control Obsidian from there. But the beauty is not that you do it yourself (if you like CLIs, have fun); the beauty is that Claude Code, Gemini CLI, or Codex do this for you. You can tell them: "use commands obsidian --help and teach yourself to use the program, write a memory file with what you learned, use the application a bit and tell me what you learned," and voilà, Claude Code probably now knows how to use Obsidian to write, read, and whatever you want, leveraging all the information that you might not be interested in but that Obsidian can provide to Claude Code from your notes. With everything you do, you can tell CC: "check in my Obsidian notes if I wrote anything about this, if there's something I wrote that can be included in this draft, if so, add it and mark where it came from before and after." What goes in quotes is just that. You don't have to go to Claude's page and request anything; you don't have to copy and paste; you ask for something in Spanish, and you save the 10 minutes you would have spent finding and formatting what you did.

Download Obsidian and create your first vault. Go to obsidian.md and download it. It’s free and works on Windows, Mac, and Linux. When you open it, it’ll ask if you want to create a vault. Say yes, give it a name (something like “my-notes”), and choose a folder. An Obsidian vault is a folder. No magic, no mandatory cloud storage. Create your first note with Ctrl+N. Write whatever you want: an idea, a summary, a list. The file is an .md file that you can open with Notepad if you want. It’s yours. Create a second note. In either one, type the name of the other note between double square brackets: [[name of the other note]]. That’s it—you’ve created a link. Click it, and Obsidian will take you to the other note. Don’t touch any plugins or settings just yet. The default settings are fine. You’ll have time for that later.

And if you already have years of notes scattered across Word and .pdf files, your trusted artificial critter can transform them into .md to add them to your vault. Moreover, you can tell it to use other models to do so: if you teach it, Claude can use Gemini CLI (with commands like gemini --help or gemini -p) to read .pdfs that would make Claude choke, because Gemini has a much larger context window (what it can read before having a meltdown, so to speak) and has better capacity to read images (and it's very good at OCR, although these models usually just install specific software for that). Years of scattered notes can become entries in your vault in an afternoon.

Loops within loops within loops

At this point, you can start to see what you can do and what you can iterate on what you've done. You did or didn't create your basic web page, even for fun, but you don't want to rely on your artificial critter, and even if you want to trust it, you want to learn (remember, you're addicted to this if you're here). Well, add to your language model's instructions in your .md file that configures it for each project (or for your entire computer) that every time it does something interesting for you (tell it you're interested), it should create a tutorial with exactly what it did in your Obsidian vault in .md. Here's the beauty: Claude (and Gemini, and Codex) use .md files to self-configure and adapt to what you want from them; those files are available for you to edit manually or for you to tell your critter to edit, but you can also make it create one for you for your purposes. If you're anything like me, you enjoy, at least sometimes, learning through projects, not following a manual. You've left beginner manuals behind for at least a decade, and the little attention you have left makes opening one simply eliminate it. This is something else; it's a manual that starts from what you don't understand. Want to study it alone? Let Claude create it for that, and link the sources so you can check them. If something doesn't work, insist until it creates something autonomous, and you can generate the critical mass of skills to make mistakes on your own. Want to do it more assisted? Design it for that, and anything in between. Or whatever comes to mind, and if you think of something better, let me know.

Configure CLAUDE.md so your bot can learnIn your project folder (or your Obsidian vault), create a file named CLAUDE.md, or simply type /init in the console after opening Claude; it will scan your folder and generate an initial version that you can then edit. This file tells Claude Code how to behave when working there. Write instructions in Spanish (or English if you prefer to use these bots’ “native” language). For example: “This is my academic notes vault. When creating or editing files, use Argentine Spanish. Every time you do something interesting, create a tutorial note in the /tutorials folder explaining what you did and how I could learn it.“ Add whatever you want: ”Don’t modify my notes without asking me. If I ask for a draft, base it on my existing notes and cite where you got each piece of information." Save it. The next time you open Claude Code in that folder, it will read the file and behave according to your instructions. Edit it whenever you want; it’s plain text.

Now Claude, who lives in text, and you, who are a natural text producer, share a medium of information: .md. Even if a missile hits Anthropic's servers tomorrow, you still have exactly what you have; what your critter made for you didn't come from your PC (in the relevant sense, it's in plain text, not code). Moreover, since .md is not text with code, it's cleaner text so your critter doesn't get distracted by the codes made for the application in question (like Word) that are not your text. The text you see is, interestingly, basically the text that your critter consumes. They are language models, and you give them the language as coded as you want. This makes them function better, for me, or at least, it makes them spend fewer tokens, less electricity, and less water (if you're interested in that, and maybe you should be).

And the possibilities multiply. You can have your resume in Obsidian and tell Claude: "find my CV in Obsidian, create a LaTeX version in a style that meets these requirements, and compile the .pdf; use a Python script for the conversion that I can keep for next time." Done, even if you stop paying for Claude Code, you have a script in a programming language that you can run for free forever (ceteris paribus). But not only that, you can ask it to create a personal webpage that uses your CV, which you edit in .md, as a base for your CV uploaded for free to platforms like vercel, cloudflare, netlify, or directly GitHub pages. A sufficiently good academic webpage in an afternoon or a couple of days of light work, which updates simply by tweaking your .md text files.

Black Mountain College: A model for rethinking current universities?

Is it possible to have a university without hierarchies or exams? Black Mountain College was the epicenter of polymathy, the only educational experiment of the 20th century that used interdisciplinary disorder as a method. A radical model of education that today, in the full era of AI, is once again necessary.

421 | Quality information for your cognitive dietJulio Alonso

You've gone through twenty thousand procedures at your university, had to format project information to justify the funds you used to buy coffee for the conference where you managed to convince the big shot from your department to come without spending a dime, or the 10% of the trip that you almost entirely paid out of pocket to discuss your work on the other side of the country. Use templates to format these academic burdens in a way that allows you to speak in Spanish while the bureaucratese you need comes out the other side. Snap a photo of the SIGEVA entry that asks you to explain something, and use what you already have—your work, your project—to generate formatted text with an artificial buddy. Tell it not to write anything new, use it in a way that serves you.

And if you have teaching experience and come up with an idea that could be a game for your students, describe it in .md and give it to your artificial buddy. I've created several in the last twelve months using Claude Code. They work. You don't need funding from your university, to win a project and hire programmers, or to spend money that could change your life. If you have intuitions, ideas, things you want to try: go for it, there are no intermediaries for many of these things now.

A tool of tools

I want to pause on an example that encapsulates much of what I've been saying, because it involves several layers of what these artificial buddies can do. Last year, I was writing my doctoral thesis and working with an author, Robert Brandom, who uses directed graphs called "MUDs" (for Meaning-Use Diagrams).

It's easy to do with a thousand and one applications, but I wanted to create them always in the same way, exportable to code that I could insert into LaTeX and in a graphical interface that serves me. So I made pragma-graph. As a piece of software, it doesn't have distinctive value; it's not innovative from a coding perspective. As a tool for me: yes. No programmer would have made it, but it saves me time.

A lo largo de meses le fui agregando funciones, y fui viendo qué diferentes versiones de Claude podían hacer. Pero la vuelta de tuerca vino cuando pensé: si Obsidian puede tener una CLI para que un modelo de lenguaje la use, ¿por qué no mi app? Le pedí a Claude que extendiera pragma-graph para que se pueda instalar como herramienta de línea de comandos. Lo que sale al tipear pragma-cli --help es esto:

Usage: pragma-cli [options] [command]

CLI for Pragma Graph Tool — create and manipulate MUD/TOTE diagrams.

  Quick start:
    pragma-cli schema all                                    # discover all types
    pragma-cli --file d.json diagram create --name "My MUD" --type MUD
    pragma-cli --file d.json node add --type vocabulary --label "V₁" --x 100 --y 100
    pragma-cli --file d.json node add --type practice --label "P₁" --x 300 --y 100
    pragma-cli --file d.json edge add --source <V1_ID> --target <P1_ID> --type VP
    pragma-cli --file d.json export latex --raw > output.tex

Options:
  -V, --version   output the version number
  --json          Force JSON output
  --human         Force human-readable output
  --file <path>   Diagram file to load/save automatically
  --headless      Force headless mode (no GUI connection)
  -h, --help      display help for command

Commands:
  status          Show CLI status and connection info
  diagram         Diagram lifecycle commands
  node            Node manipulation commands
  edge            Edge manipulation commands
  entry           Entry point commands (TOTE diagrams)
  exit            Exit point commands (TOTE diagrams)
  export          Export diagram in various formats
  history         Undo/redo history commands
  schema          Type schema discovery (for LLM self-reference)
  help [command]  display help for command

La clave está en el comando schema. Ese comando le enseña a un modelo de lenguaje todo lo que necesita saber sobre la teoría de Brandom para producir diagramas correctos: qué tipos de nodos existen (vocabularios, prácticas), qué relaciones son posibles (PV, VP, PP, VV), cómo se conectan. El flujo de trabajo es: yo le digo a Claude qué grafo quiero, Claude lee el esquema que devuelve schema de pragma-cli, crea el diagrama con los comandos correctos, y exporta el resultado directamente a LaTeX para insertarlo en mi manuscrito. Antes podía pedirle gráficos a un modelo de lenguaje, incluso de estos, pero era jugar ida y vuelta varias veces hasta que los hacían como yo quería. Ahora los hacen según los parámetros de una app que existe aparte de ellos, y yo puedo editar manualmente lo que hacen. Todo esto se puede ver en acción en el video de abajo, y el código está para quien quiera hacer algo mejor.

The intersection between those of us who might want something like this (we specialize in these theories) and those who have the ability to program the various things needed to do it sufficiently well and sufficiently fast is probably empty. These tools wouldn't exist otherwise. A programmer didn't lose their job. This would be a waste of time for someone who can do this without wasting time.

With the same logic, I created epistrophe, an app for philosophy students at UBA to plan their careers. It's completely excessive and unnecessary, as many say philosophy is. But if someone wants something different, they can do the same and better, for this major or any other.

Which brings us to git; if you're interested in ensuring that nothing you've written gets lost, learn to use git and create a github account. It's the best cloud for text that exists, and it's free. You can have your Obsidian vault synced on any computer. If you log into your console on github, Claude can make commits and pushes to github for you.

Types, tokens, and the meaning of all this

Language, we who study it say, especially those who study it as linguists, includes types and cases (and instances). The sentence "Hello!" is the same type of sentence every time anyone uses it, but it's a different case each time it's used. Now, written text, and the ability to copy written text, allows us to distinguish between cases and instances. When you open this page, your browser will download a case of the sentence you're reading in the form of a case of the code that your device uses to show you the light spots made of contrast that your screen creates and that humans read as instances of letters, words, sentences, and paragraphs. Your written text in .md is a case of whatever you wrote, and it's a type for whatever you want to do with it in the future, in this sense.

What I'm writing now can be a resource for my artificial buddy and do with it whatever I want. I can create a webpage, I can turn it into a story, I can do many things. But these are things I do through this text-eating bacteria that is Claude Code.

What these technologies offer is a whole plexus of opportunities to recombine and extend what you've already done, to create more instances, to create more opportunities to do something with it. What the enthusiasts of serious language models say, even some of the scammers, the LLM-bros (who were once crypto-bros, and nft-bros, etc.), is that Spanish/English/German is the new programming language, and they mean what I've been saying. Language models have transformed inert text, except for a human, into something alive even in the digital realm. Now not only do we humans transform our text, the intangible matter of our rational and conscious thought, into other things, but our computers do too through language models that allow us not to translate, but to transform text into code, and code is images, sounds, interfaces, things.

SummaTechnologiae by Stanislaw Lem: Futurology and Cybernetics from Socialist Poland

The Polish writer developed a philosophical treatise on technology and cybernetics during the Cold War that anticipated current disputes over AI, virtual reality, and bioengineering. The unavoidable role of fundamental questions.

421 | Quality information for your cognitive dietJulio Alonso

No longer is it the coded language of formal programming like Python that comes to life in software and hardware; now it’s Spanish, it’s your idiolect (your version of Spanish) that can do it. In my opinion, language models aren’t conscious, nor are they intelligent in the traditional sense, but I’m not sure if these questions are even well-posed (not even for us). What I can say is that they are the clearest way we humans have designed to see that words resemble more living forms than gears.

Our words interact with their environment, and a language model is as revolutionary as the invention of food by living beings. There were things that would be food before there were living beings, but the first life form that used energy from the environment to do something invented what we call food. Grass existed before it was food; the first bug that ate it did something with that, creating more bugs like it. What I’m writing now can be an input for my artificial bug and I can do whatever I want with it. I can create a website, I can turn it into a story, I can do many things. But these are things I do through this text-eating bacteria that is Claude Code. I do it through the bug just like a football coach does something through their players. Just as humans coordinate to get things done, my instruction can culminate in an action taken by another. We’re talking about that, and it’s not just chatting with a model. Chatting with a model, if it were up to me, could stop being available. Chatting to do things with my text is what interests me. Talking to it to use more tools, to give me more options, to do more things.