Jeroen Sangers logo Jeroen Sangers

Posts in "Digital garden"

An indexed PKM of sources and personal syntheses makes a truly personal AI possible

The usual way to give an LLM your documents is to upload them, or to plug in retrieval (RAG). For every question the model hunts down the relevant fragments in the raw sources and pieces together an answer. It works, but nothing builds up. A question that combines five documents gets answered from scratch again next time.

The alternative is a PKM in two layers. The sources sit untouched in a raw layer. Above them lives an indexed layer of summaries, concept notes and cross-references that gets updated with every new source: contradictions flagged, the synthesis revised. Knowledge is compiled once and then kept current. When a question comes in, the LLM reads the index first and then the relevant notes. Up to a few hundred notes that is enough, with no search infrastructure at all.

What turns this into a truly personal AI is the feedback loop. A good answer, whether a comparison, an analysis or a connection you spotted, goes back into the system as a new note. Your own explorations pile up the same way your sources do. The system then holds what you read and what you thought about it, and every next question draws on both.

A PKM built from sources plus atomic notes already has those two layers. The LLM only needs to do the maintenance: indexing, cross-referencing, checking for consistency. Once the system grows large enough, fine-tuning is the next step, so the knowledge lives in the model’s weights instead of its context window.

AIL 2: Human Created, Major AI Augmentation. About AIL

AI lowers the threshold for building tools, but that is not where business transformation stalls

Anyone can now have a tool built in five minutes by asking a model, without writing a line of code. That feels like a breakthrough if you build tools for a living. Most people do not. As Benedict Evans puts it, a great lawyer spends the day thinking about the case and the client, not about which software would make that work faster. So the bottleneck sits somewhere else: in seeing that there is something to automate at all, and in working out what the right solution actually is. That second part is rarely obvious. It can take years and several failed attempts before someone finds the right shape for a problem.

Even with the right tool in hand, the work is not done. A process touching fifty to five hundred people across departments and systems stops being an individual decision and becomes a purchasing project that runs for months. Software moves on a sliding scale between improvised (a spreadsheet, an email) and institutionalised (a system with audit, security and ownership), and a task slides toward the institutional end once there is enough revenue or risk attached to justify it. AI leaves that scale intact. It moves where the thresholds sit and which choices you have.

This is why giving everyone ChatGPT never turns into company-wide change on its own. It is the same pattern as putting a PC with a spreadsheet program on every desk in the early eighties: a small group took off with it, the rest used it now and then. The real transformation lives in the slow, concrete work of going through every existing process, one at a time.

Productivity doesn't come from working longer, but from letting your brain reset

Your brain does not multitask, it switches, and every switch leaves a residue of attention behind that eats into your cognitive capacity. That is where uninterrupted blocks earn their keep: inside such a block you work on one thing, with no notifications and nothing on the side.

A block like that has a shelf life. After a stretch of sustained attention your sharpness drains away on its own, and a workday without breaks, an open-plan office and a chat that never falls silent hold you in that state of low attention and constant stress. Pushing on feels productive at that moment, while the quality of your thinking has already dropped.

So blocks and breaks do not contradict each other, they work on two timescales: no switching within the block, a reset between blocks. In practice that means stepping away briefly every 60 to 90 minutes without new input, plus a clear stopping time at the end of the day, because it is in those quiet moments that your brain consolidates and creativity comes back. What wrecks your attention is the task switch, not the break.

The Instagram algorithm tests new content on people who don't follow you yet

When you post something, the platform first shows it to a small test audience of people who do not follow you, and it watches closely: do they swipe past, how long do they stay, do they react at all. If they drop off, the algorithm concludes that strangers have no interest in this and stops showing it. That is the mechanism behind the feeling that your reach dies. It is not a punishment. It is a test you lost.

That changes the question you should be asking. Not “is this content good”, which is a judgement made from your own taste or that of your regular audience, but “does this work for someone who has never seen me before”. If you optimise only for the people who already follow you, you are optimising for an ever smaller slice of your own reach.

The consequence is that generic, broadly phrased content loses structurally. A stranger has no reason to keep watching something he has seen a hundred times before, phrased in exactly the same way. Being specific and recognisable for your own position works better with strangers than with existing followers, because that is precisely what survives a test swipe.

You optimise a system by looking at the whole flow, not by making the fastest part faster

A process is a chain of steps, and it is tempting to improve the step that is easiest to see, or the one that seems most obvious. But making a step faster only helps if that step is the bottleneck. If it is not, the work piles up in front of the next step, which is still the real constraint.

The question that matters is not how to make one part faster, but what a change does to the whole flow. Skip that question and you optimise inside an illusion of progress: a fine result at one station, while the total lead time stays the same or gets worse.

This holds for personal effectiveness as much as for a production line. Finishing a task faster only helps if that task is where the work gets stuck. Otherwise you are just moving the queue.

Self-control is a learnable skill, not an innate trait

We usually read self-control as a character trait: you have it or you don’t. Walter Mischel’s marshmallow test fed that picture, because the children who waited longer turned out, decades later, to earn more, be healthier and have stronger friendships. Kentaro Fujita points at the finding that got skipped over: those children were being taught strategies, and the strategies worked. Three-year-olds stare at the marshmallow on the assumption that staring helps. Five-year-olds have already worked out that covering it up or looking away works better. By thirteen, the children who understood the rules of self-control had fewer behavioural problems. The route is known too: children pick it up largely by watching what adults do.

The distinction that makes this possible is that willpower and self-control are two different things. Willpower is the blunt version, suppressing the impulse, and training it produces small and erratic effects in the research. Self-control is the toolkit around it: looking away, reframing the temptation, viewing yourself from a distance. Those strategies do transfer, so you can get much better at self-control without ever gaining more willpower.

That also changes what a slip means. Not that your character falls short, but that this particular tool failed here, and the question becomes which tool you try next.

Structure is cheaper than intelligence

Ask a language model something about your own life and it has to guess. It never saw your data, so it reconstructs what is probably true, and every so often it hands you something that never happened. No amount of clever phrasing fixes that. Guessing is what generation is.

A well structured knowledge base works the other way around. It looks things up and gives back what is there, with the source attached and with how certain that source is.

Underneath sits a ratio worth remembering: roughly 98 percent of a “Digital assistant” system is data, 2 percent is AI. And that 2 percent does translation, not thinking. It understands that “called Joe about the fieldlab” is a phone call, that Joe is a person and that fieldlab is a project. Nothing more is asked of it.

Every layer of structure you add makes the AI layer thinner.

What looks like intelligence usually turns out to be retrieval. Who should I call is a sort on contact strength and time. This pattern stands out is a count. What did I do last time is a lookup in history. Query, weigh, show. A simple search over a clean schema beats a brilliant model over messy data.

That reframes where the work goes. The interesting question stops being which model you use and becomes how well your own material is organised. Naming things consistently, linking them, recording the type of a thing rather than leaving it as loose text: each of those shifts work out of the model and into the schema, where it stays put and costs nothing to run.

You still need a large model while the schema is being discovered. That is the architect designing the house: which tables relate to which, where the structure has gaps, which patterns hold. Once the house stands, you can live in it without the architect. What remains is small enough to run on your own laptop, no cloud and no dependency on a vendor who might change the terms next year.

So the useful measure of a personal system is not how smart its assistant is. It is how much of your life the assistant can simply look up. Every time you tidy a corner of that, you need slightly less intelligence to make sense of it.

Time management training leaves volition alone

Time management training is popular. The evidence that it works is thin.

The critique that landed hardest for me: these trainings teach planning and structuring, and they leave volition alone. Volition is the capacity to protect an intention against everything competing with it. Distraction, a colleague at your desk, the pull of an easier task.

Planning and structuring your activities may simply not be enough to keep an intention alive once the day starts pushing back.

Kuhl and Fuhrmann pointed at what is missing: strategies that work on motivation rather than structure. Finding meaning in a task. Making it more attractive to start. Deciding in advance what you do when the pull arrives.

That reframes what a good system is for. Not a better container for your work, but something that keeps you willing to do the work you already decided mattered. Structure gets you to the desk. Something else keeps you there.

Implementation intentions correct the planning fallacy

A to-do says what. An implementation intention says when, where and how.

“Write the proposal” is a task. “I write the proposal tomorrow at nine, at the kitchen table, before I open my mail” is an implementation intention. Gollwitzer named them in 1999, and the difference turns out to matter more than it looks.

Koole and van ’t Spijker found that people who wrote their plans this way completed more of them and were interrupted less often.

They also help with the planning fallacy: our habit of estimating how long something takes by imagining it going well. No traffic, no interruptions, no discovering halfway through that the file is somewhere else. Specifying the circumstances forces the obstacles into view before you commit to a time.

It costs about ten extra words per task. That is a cheap correction for a bias that has been wrecking estimates for as long as anyone has been making them.

Unplanned tasks get finished more often than planned ones

More from the same diary study. When the researchers looked at what people actually completed, unplanned tasks won.

Tasks that showed up during the day were rated as more urgent and more important than the ones people had planned in advance. And they were completed at a higher rate.

Recency bias and time discounting explain most of it. Something standing in front of you right now weighs more than something you decided was important last Thursday.

This undermines an assumption sitting under most productivity advice: that writing your priorities down protects them. It does not. Your plan gives you direction and an overview, and both are worth having. What it cannot do is win an argument with whatever just walked into your inbox.

Which means the plan needs something next to it. A protected block, a closed door, a rule about what you do before you open your mail. The list alone was never going to be enough.