For a while now I’ve been running an Excalidraw instance on my VPS (with Yunohost). Excalidraw is a virtual whiteboard, and a useful tool to me. Excalidraw is lightweight, and any drawings you make are stored in text format, which makes them highly portable. It integrates very well with Obsidian, with the Excalidraw plugin, by fellow PKM enthusiast Zsolt Viczián. My Excalidraw instance is also used by my company’s team.

I run my own Excalidraw instance for reasons of digital autonomy. For single person use it provides that. However, Excalidraw has a collaboration mode, reserved to Excalidraw.com only (as noted in the package description), which depends on two outside parties with control over it. For my company collaboration mode is the primary desired use form, but it must be digitally autonomous to be of actual value. It should also actually work, which it doesn’t on a self-hosted instance.

In Excalidraw’s collaboration mode multiple users can work on its canvas (dubbed a ‘room’), where you see each-others edits in real time. Drawings you’ve made collaboratively also have a unique URL so the room can be called back up later, to continue work.

The Excalidraw.com deployment’s open source package, included in the general excalidraw package (which the Yunohost install package uses directly) has two things hardcoded that are needed for collaboration and persistence:

  • To make the realtime movement of cursors and edits visible, it calls the oss-collab.excalidraw.com subdomain for websocket connections but only answers requests with excalidraw.com as origin, and returns forbidden for my and other domains. In the browser, wss://oss-collab.excalidraw.com/socket.io/?EIO=4&transport=websocket fails. So it refuses self-hosted Excalidraw instances. The live view of other users and their edits therefore does not work on self-hosted instances.
  • It stores drawn objects in Firebase for persistence of rooms, and does so encrypted. For self-hosted instances this does work, reloading a room from storage shows the objects others made. However Excalidraw.com controls that storage, and holds the keys to it, meaning they can delete your stored material or remove your access at will. An issue of digital autonomy. Firebase itself is provided by Google, and while the Excalidraw objects are encrypted, it introduces another party with a hold on the data.

In practice, it means that a self-hosted instance of Excalidraw has a half working collaboration capability (no live sync, but reloading the persistent Firebase version works). It of course wasn’t promised to work, but even if it did it would not actually be an autonomous install due to the hardcoded dependencies. To gain that digital autonomy, and have functioning collaborative functionality, both a necessity for my team, I’d need to ‘bring home’ those dependencies.

The live collaboration part is created with the open source excalidraw-room script, which is easy enough to install on my VPS. Local storage may be more difficult to figure out (but not too hard, I hope). The main thing is that to let the excalidraw package talk to that room script and the storage, I need to alter that package itself. If the dependencies had been configurable settings it would be a different thing.
Now it needs a fork from Excalidraw, to be able to change those hard coded settings to provide both excalidraw-room and storage on my VPS itself. I do not look forward to maintaining a fork in order to change 3 things or so for the live collaboration to work (more to swap out storage), and then needing to keep it aligned with the upstream version of Excalidraw, but it seems to be the only viable path.

Whether there are other open source packages out there where self-hosting also carries hard coded cloud dependencies is a question I need to also stay aware of, I think.

The shiny new tech syndrome makes us focus on always wanting to work with the latest greatest AI models, almost by default. And switching away from one of those to something else is then tied to the question if it is as good or better than the frontier model you already use. I know companies who switch between various models based on task considerations, and use a router of some sort in between to seamlessly switch, but it will still be switching between frontier models.

For me some of my use cases for AI cannot go forward in any acceptable way if it means that the context and data for them get send to the companies offering the models where it vanishes out of my control. It means that none of the closed leading models can be considered for them.
This is not just the case for me personally. Any individual mindful of aspects like data protection, jurisdiction tools fall under, autonomy, extraction, transparency, costs and other externalities, and certainly any company anywhere that has to contend with liability and compliance will have the same reservations and considerations.

It means there’s a need to go counter to people’s first reflex to use the largest and best performing model for everything. Because using the latest largest model for everything may well be overkill, although you won’t notice as the true cost gets hidden in subscription tiers. Mostly however because what the best model is or its replacement, is the wrong question when choosing your AI tools.

The right question to answer is not if there’s is another ‘best’ model out there, but if there’s a ‘good enough’ model out there that also addresses your other concerns, where ‘good enough’ is also task specific. What is good enough, though? In the past days I came up with a workable answer for myself.

I spent most of last week in Switzerland, visiting the Swiss Data Space Forum in Lausanne, where AI (running on top of data spaces) was extensively discussed, in the context of the mentioned concerns. Different approaches to addressing those concerns were discussed. Mixes of costly self-owned infrastructure and public and private cloud set-ups, or how Swiss company Anyway, pools your existing hardware to host models locally (something that sounds of interest to my company). One panelist who discussed using AI in a medical setting made a remark that stood out to me: open weight models, the ones they were running on their own infrastructure, are just 6 to 8 weeks behind the closed frontier models in terms of capabilities. Six to eight weeks sounds like it’s on the optimistic end of the scale, but it did make me think. It changed my perspective of what is good enough.

Good enough then is not when one current model is equal or better than another current model. Good enough is when a current model with less baggage, is comparable to another earlier version of a frontier model that gave me acceptable output then.
In my playing with AI models my usage of Claude Code from the fall of 2025 made a real difference (compared to the AI supported coding I did earlier, in 2023 e.g.), so I’ll take that moment as a watershed. Anything that is currently near the level of what Claude Code with Sonnet 4.5 and Opus 4.5. could do at the end of 2025 will do. My current good enough therefore is some current open weights model with a gap of a year or less to current frontier models.

Baggage does not only come from closed models of course. Open weight models by Chinese companies may come under export restrictions at any time for instance, just as their US counterparts. You will also need to run the alternative models in some way, on your own costly iron, or hosted. For hosting providers the same caveats apply as for the closed model providers.

Running a different set-up must also be done in a way that does not disrupt your current working environment, a way that is mostly just pointing your own set-up to another API endpoint with the models you need to interact with.
Which loops back to the routing I mentioned companies doing. I’ve played with fully local models (not good enough in terms of speed and result), and on a VPS (not good enough in terms of speed). So having a well equipped hoster that provides me with a router in the way larger companies have is needed, and at the same time not themselves bring the drawbacks that closed model providers already do. Previously I made some use of Openrouter.ai, a US company that allowed me to switch between models, but does not address any other drawbacks I mentioned.

Today E pointed me to HostYourAI, a Dutch company that emerged out of the founder getting locked-out of Anthropic at some point. It hosts a wide range of models, all usable through an API (or dedicated set-ups if you need), and it can be set to do so fully under single jurisdiction, the EU, and with a processing agreement where it is clear where my data does and does not go.
That is certainly worth trying out, keeping the acceptable gap of 10-12 months in mind, and I can test drive it by changing the API endpoint that Claude Code talks to. I made an account, threw in some Euro’s, and came up with some test cases to select two or three from (things I vibe-coded extensions to in January, and could now do again to compare, the Persona experiment I started in February, the output of the three helper tools I created in April, or the design of knowledge-graphs I did AI assisted over the summer)

I now have a testable approach to using AI for personal use cases I would not use closed model providers for. It doesn’t require running my own infrastructure (too slow or too expensive), I can use hosted infrastructure than can be set to adhere to my requirements w.r.t. data protection and exclusive EU jurisdiction, and I do not have to change my current set-up otherwise. But most of all, I have a measure to select usable open weight models and judge testing them by: comparable capabilities to the frontier models from about a year ago when I thought their output was useful, and compare the output I still have from then with output now.

A year ago I read Libraries of the Mind by William Marx, and some of it is still very much on my mind ever since. Particularly what William Marx called ‘dark matter’ literature, which exists but you don’t have access to. E.g. because it’s in a language you can’t read, or material you don’t know how to search for. Material that is invisible from your own viewing position. Because that invisibility is easily accepted as proof of absence of such material, rather than proof of limitations on your own toolset and agency preventing you to find it, such resources will be generally ignored.

Still, many questions have global aspects. Our information networks, our trade routes, our dependencies are truly global. Because of it many of our issues, including climate emergencies, are global in both their roots and consequences. We all suffer from small earth syndrome these days, as I phrased it in 2010 at the SHiFT Conference, where everything anyone of us does reverberates around the world because we are continuously bumping into real global limits, Club of Rome style.

The same is true in reverse: many of the pathways to agency, to resolution and mitigation are global in nature too, or at the very least need us to look at available information, thinking, and perspectives globally as well.
‘Global’ therefore is not to be taken here as a level of abstraction or aggregation, which erodes your own place in it and throws out your own localised experience and stories, but global in the scope you yourself default to when it comes to who you see as part of the whole, as worthy of consideration and/or empathy, as affected by your choices and actions. Global also not as an equaliser or one-size-fits-all thing, unmooring identities e.g., but using differences as an energy source and as potential for creativity. In any situation there are boundaries between groups and narratives. Power differences, and identities resting on opposing or similar assumptions, differences in who has legitimacy or standing in which situations and issues, differences in who has or is afforded agency and in which ways, all abound. All these differences are potential energy to tap into.

Put together it spells to me that when I am doing some deskresearch shining a better light on literary ‘dark matter’ becomes much more important. Are there any francophone resources, Nordic authors, Eastern European authors, hispanic movements around an issue I am exploring? Yes, the scientific world largely uses English as lingua franca, but language areas also have their own spheres outside of it. Look in any French or German book store for instance and compare the shelves on topics of your interest to e.g. a UK book store. They also all have their own scientific publishing channels. That is all still within a largely overlapping European context. What about ongoing Chinese perspectives, Southern-American movements, the various spheres of discussion on the African continent, and across Asia? Those will be almost completely invisible to me on any topic, when my own tools are limited to my personal language skills and the general anglo-american bias in online search tools.

Enter AI tools as a middle way, along the lines of what Jeremy Cherfas wrote recently. He posits his own AI use for “things that you perhaps could manage on your own, but that would take too long to be worthwhile.” And while “there are plenty of areas in which I feel that AI could never be appropriate” he “was happy to use generative AI to extend my capabilities” where that strengthened his own efforts andv where outputs are for his personal or intermediate use in his own work processes.

When I ask e.g. Claude about a book and its message, and to help build a more syntopic reading list around something, what you get is a few English titles, as would be expected given their origin. Adding my own language skills, it will add a few other resources. Any extension beyond that you need to directly ask about: is there anything in e.g. Nordic countries, are there traces of similar discourse in China etc.? In steps what will emerge is a fuller list of pointers. Many of those I won’t be able to read directly, but I can circle around them, e.g. through secondary sources, by machine translating parts etc. I’ve noticed that to do this, to go beyond middle of the road AI responses, I need to build it up, in steps. Which is something I can put together as a recipe and feed into Claude.

I won’t speak or read Chinese, Igbo, Swedish or Polish in my lifetime. But I can ferret out resources and what they may tell me this way.

Today I started reading in Rest Is Resistance by US author Tricia Hersey, it was mentioned to me recently in the context of maintaining hope or optimism while also feeling dread and despair about the state and direction of the world, alongside Rebecca Solnit’s Hope in the Dark, and the Dutch title ‘Optimism without hope’ by Tommy Wieringa. Turned out I already bought it in 2023. I wondered about other similar or adjacent texts and used Claude in the way described above. First came a list of mostly English resources, and one German, most of which I was already aware of. Then following up a handful of francophone resources, nordic ones, and hispanic ones came up, which unearthed a link to Latin-American texts I had already explored a few years back related to a local legal initiative in the Netherlands w.r.t. climate adaptation. It struggled with providing pointers to resources from the African continent, but ultimately did. Further prompting brought both official and grass roots perspectives from China, more francophone resources from the Caribbean world, and also more potentially relevant African and Indian resources.

A first approximation of a syntopic approach like this is no output in itself, but a starting point, meaning I now have more to do than if I had been satisfied with just the titles I already own or am already aware of. But I am casting a wider net, even if what the net brought up may not all be equally useful, it pushed me to look further and broader. It stretches my Library of the Mind.


The University of Maastricht Library, which I walked by a week ago. Its collection limited by necessity, like our global digital library is limited by our tooling.

In the past week I’ve started three personal experiments that use AI (in this case Claude Code). For each, the experiment lies in automating steps in my cognitive work that are useful or necessary but not the actual cognitive work itself. They’re helper activities, supporting the main task. For two of the three that is the clear focus, the third is slightly different.

The three experiments are:

  • Filtering on interests in my feed reader, let’s call it ‘Weak-tAIs’.
  • ‘Slopsidian’, lifting concepts and argumentation from papers into Obsidian notes, and linking them iteratively.
  • Explore questions with pre-existing ‘recipes’ that take a specific philosophical perspective. Perhaps I should dub this type of language game ‘WittgenstAIn III’.

It started from an automation task, which I mentioned here: manipulating non-fiction e-books. I have a script that I can point to an e-book in my Calibre collection, and then will populate a note with elements from the book: foreword, index and literature list, content overview, all if present, and for each chapter of the book the first and last few paragraphs. This is what I look at and skim whenever I want to gain a first impression and understanding what a book is about, and what questions it addresses or what it proposes. All very Mortimer Adler. From it I can then decide which parts of a book to read more closely, which parts likely contain things I am already familiar with or fall outside my current interest in the book. From those skims I jot down things in my note for the book. This quickly turned out to be useful to me, because it removed the wall between the e-book and my notes by bringing parts of the e-book into my notes temporarily where I could more quickly go through them in preparation for ‘proper’ reading (although in fact it is part of reading).

It got me thinking what other helper activities in reading and filtering I could identify.
Helper activities are tasks that support a main task by making it easier or providing guard rails. Checklists are an example, they ensure that you don’t skip important steps. In most cases nothing will immediately go wrong if you don’t do the helper activity but if you do them the main task gets a little easier to do well. A lot of helper tasks can be regularly automated, like the e-book excerpt script above. Others less so because they contain elements of processing actual texts, like the three experiments I describe here. There perhaps using a model like Claude Code can be of value (and hopefully soon, through local model deployment).

A brief description of the three experiments:

Weak tAIs
I order my RSS feeds by social distance for reading. Part of the reasoning is that I want to be well informed about what close ties write, but I am aware that interesting information likely comes from a wider social distance. This practice has been in place for some two decades and enormously valuable all that time. The most interesting stuff usually comes from the third layer, a folder named ‘c150’, in my feedreader: close enough to know who the author is, and engage in interaction if I want, disconnected enough for them to encounter things I am less likely to have already seen myself. That is the The Strength of Weak Ties (1973) as Granovetter called it.

I also keep a list of current interests, a bit like Feynman’s dozen or so currently favourite problems. For each interest I have formulated a few aspects:

  • what is conceptually interesting to me in a topic (e.g. my interest in EU digital and data policy conceptually is that it forms a geopolitical proposition externally, while being a quality improvement instrument internally that takes rights and societal values as yardstick),
  • am I theoretically interested or more practically,
  • do I have a knowledge fundament for the topic or am I a newbie,
  • is there a link with any long term goals,
  • can it be put into a specific context or tied to a specific issue/question,
  • can I shape or create an enduring practice around it,
  • can I build a bridge to outputs, like blogposts, presentations, or client proposals

My feedreader tracks just under five hundred people writing on the open web. That can easily amount to two thousand postings in a week. I can have several intentions to start reading, one of them is to find and read material relevant to my list of current interests. A reading intention does not do away with items, it’s not a filter to remove material. It’s essentially just a view on the entire set of incoming items in the feed reader that I usually construct in my mind. What if I can construct those views on my screen too?
The ‘c150’ social layer, the weak ties, what do they write about that connects to the fields of interest from my list? Such filtering does not lend itself to text based search based on fixed terms. I usually skim titles for first impressions, and click opportunistically through the postings. What if I can have a model weigh the postings and compare them to my list of current interests, to mark them for my attention? In aid of that one specific reading intention.

That’s what the first experiment does: label postings that seem to fit my interests, and express why. So that I can skim the folder of weak ties by interest, and read those items first if my intention is to explore those interests. I limited it to the c150 folder as feeding all rss feeds into the model is consuming a lot of time and tokens, so I started with the part most likely to bring useful results.
The labeling works now as part of my feedreader. I am not yet convinced of the quality of it though. The motivation for the labels usually is along the lines of "it fits interest X but not in the way you’re looking for", which to me means it actually doesn’t really fit.

Slopsidian
This week I read an article about AI documenting its own actions and output in a wiki, and saw one or two similar efforts described. I applied that to a different helper task, which is the preparation of reading a paper and helping me to decide to dig deeper. This is similar to skimming a non-fiction book, but more involved. Can AI reliably pull from a paper the concepts used and introduced, and the line of argumentation? Saving them both in a single note for the resource, and in separate notes for each of the concepts? Additionally can it logically link concepts from different resources? This is what an ‘ingestion skill’ now does for me. I let it store the output it generates in a folder that I can also open as an Obsidian vault, hence the name Slopsidian. The papers come from my Zotero collection, meaning I previously saved them. That original step of curation also means I have a line or two about why I thought them interesting at the time. Feeding that curating decision and the paper into the ingestion skill allows a second order look at a paper. What are the concepts discussed, and, reading the output, do I think some of those are of interest to me? If so, I can look at the paper more closely and do my own note making and paraphrasing and placement in my actual Obsidian collection. Lifting out concepts works rather well, the linking is less useful in the first experiences (too obvious, not sparse enough) and can seem forced when you look at why some concepts get linked.

WittgenstAIn III
The third experiment is a bit more on the edge I think. Here the probabilistic language games that LLMs are have more of a free rein. Part of the university courses on philosophy of science I did 25 years ago was using different philosophical schools of thought as lenses to approach a question. Not to answer the question, that is hardly ever the point after all, but to holding it, and holding it differently. Plato’s essentialism, Kant’s transcendence, dialectics (Hegel), phenomenology (Husserl), Wittgenstein II’s analytical method, hermeneutics (Heidegger), deconstruction (Derrida), and Rorty’s pragmatism. For each of these, for over 2 decades, I’ve had a recipe in my notes to apply to a question.
I put together a ‘language game’ in which I pose a question, which a ‘router’ prompt tries to match to one or more of the 8 recipes, or to a combination of recipes chained together (e.g. first look at a question from an analytical perspective and then feed the results in to a deconstruction exercise.)
My existing multi-step recipes are followed, and output is generated for each of those steps, into a resulting note.
I read those resulting notes, lift out what catches my eye or what resonates and I use it to flesh it out more, for me to hold the question still longer. Models are language games of a sort, so hence the name WittgenstAIn III, a third iteration, extending the second Wittgenstein’s language games to and with AI.
The output here makes me more uncomfortable than the other two. Reasoning is being mimicked, with the usual overconfident wrongness we’ve come to expect from generative AI, and that works out in odd ways sometimes. Still there is utility that can be lifted from the output. It is a good kickstart for exploring questions to quickly see if a recipe might yield something or not, judging by my first attempts in this experiment. It does certainly lower the threshold, as helper task, to engage with the recipes. I’ve used it more in the past days than in the past months. Part of that is the novelty of the experiment, and that may wear off quickly, but perhaps it carries the kernel of more habitual use.

Earlier this week I started reading an ebook and was a bit irritated because the book did not show me a table of contents. This seems to be a regular thing in ebooks. Already earlier I have complained here about why ebooks, or perhaps mostly e-readers, make so little use of the affordances of digital files.

ePUB files are really XML in zipped archives. Since I left Amazon and the Kindle reader behind, all my ebooks are ePUB files. XML means that the files are machine readable and highly structured. That opens up possibilities to manipulate them.

I used Claude Code to ask a few questions about ePUB files and how they are treated by e-readers. E-readers deal differently with the information in an ePUB file. They may load a table of content into a local database, and use that to allow navigation, or ignore various pieces of information in the XML altogether.
For fun, I asked Claude Code to check the XML file of the ebook I was reading earlier this week, to see if it actually contained a table of contents that was just not shown to me in my reader. Turns out it did.

I also asked it, if it would take a lot to extract a table of contents from an ebook. It doesn’t, so I now have a first script that finds the table of contents if present, or builds one from the headers in the ePUB’s XML if not. The php script saves it to a markdown file that I can then use in my book notes, to group my thoughts and annotations.

In my Kobo reader, and in my Calibre reader the ToC information that the ePUB file provides outside the regular content of the book (NCX or xhtml), is accessible through the reader’s interface, but not part of the reading experience itself. I generally like my ToC to also be presented in the book, like it is in a paper one, and I actually prefer it not at the start as is usual but at the end, near notes, references, and literature lists, to have all the book’s metadata together to glance at. For that a ToC must be not separate from the book’s content, but within it. It would need to be in the ‘spine‘, the part that is presented for reading by readers.

If I annotate or highlight in a book, those are kept by an ereader separate from the book and refer to specific points inside the XML (through canonical fragment identifiers, CFI). You can alter an e-book, it’s XML after all, but that would shift the position of content fragments, and existing pointers from annotations and highlights would then point to the wrong lines in a book.
So if I add a ToC, grabbed from the existing metadata or constructed, inside an e-book, my preference to having it at the end is actually useful. Because if I add it to the end, it will not shift anything I may have annotated or highlighted already, messing up the pointers in the annotation file.

Next to extracting a ToC I’m also thinking about extracting other meta-information (like indexes, references, lists of images or tables) but a first glimpse into some ebooks suggests that those are not usually listed in the Manifest of an ebook, so would have to be constructed from clues inside the book.
However it will help me read non-fiction non-linearly if I could extract such things, e.g. the figures and tables present. It seems to me a number of such steps should be straightforward from the structure of an ePUB file, others need a parser to extract the right information and shape in a useful form, but still can be done with regular scripts (e.g. show me the first and last two paragraphs of a chapter to get a notion what it talks about), yet others do need a (local) LLM, e.g. to summarise each section of a book separately. I’ll see how far I can get, and learn about the ePUB format along the way, with deterministic code first to extend my personal and local toolkit on my computer.

Update 12-04-2026: I now have a script, that I run in my browser, which allows me to select an ebook from my Calibre library, and then explores it w.r.t. the table of contents, reference and literature sections, and images, and also pulls in the first and last few paragraphs of a chapter (which let’s me explore what a chapter is about, Adler style). All that gets turned into a markdown file that is then put in the corresponding book note in my Obsidian vault using the right template.

At PKM Summit this weekend one thing that stood out was that many have started creating their own tools, and were using vibecoding to create them.

While the term agency turned out to be unknown to almost all participants, that is of course what such tools create. The ability to do things, individually or as a group, in this case by creating your own tools to get there.
The power of finding new agency was felt and expressed by quite a few, and played a role in a good number of sessions too.

When I first encountered computers, in the early 1980s, creating your own stuff was the norm. It was almost the only option. Making the machine work for myself. Like software to keep my ham radio logs and print QSL cards.
These days I run a good many smaller and larger personal pieces of tooling on my laptop. Things like making it easy to search by date in my photos on Flickr, or posting to my website from my internal notes, or from within my feedreader.
Things that reduce friction, speed things up, reduce dependency on external systems.

Vibecoding, and especially the Claude Code style of vibe coding, is bringing people to create their own tools, who weren’t able to do so before. A pool of latent needs they can now tap into on their own.

Some I know are really now learning how a computer works under the hood through their vibe coding. Testing the limits of their machines, finding out how fast local stuff can be. Discovering the power of APIs, the utility of cron jobs, and learning how to run their own VPS or local servers.
Others are creating little tools that work the way they want. An app to present books from their collection in that one specific way just so. A mobile app for public transport built on your own existing commute patterns and nothing else. Apps pulling in data from several sources and presenting them in one interface that likely only makes sense to themselves.

Tools built by people realising they are pretty predictable to themselves, and that such highly localised and specifically contextualised predictability now lends itself to automation by the intended user themself.
Tools, in short, where, access to and control over data lies fully with the user, where applications are views on that data (and multiple apps use the same data), and interfaces queries on the data. Along the lines of Ruben Verborgh’s 2017 article “Paradigm Shifts for the Decentralised Web“ but then way more personal. The decoupling that is possible between data, applications and interfaces is even more powerful when you can do them all three for yourself. And then mash them up in any which way you want.

Vibecoding is allowing people to jump the barriers to entry to that. And judging by the stories they share, it feels like pole vaulting over them, not just clearing the barriers. That energy then propels them on to do more.

Over the past months I’ve also heard regularly how people are cancelling paid subscriptions to various online services, and switching to their personal tools that fit their use case much more precisely.

There are many ethical, political, and societal issues with much of the gen AI world, and how models come about, and how corporate vendors exploit and leverage their power.
Yet, where these things are not just consumed but used locally as a leg-up to a different level of self-reliance, it looks quite different. Something is brewing it feels like.
A shift, and I’d love to see more people explore and extend their own agency with such tools.