Wissensraum
Auf Deutsch lesenTopics Wissensmanagement
What it is about
From the inbox folder to shared memory: storing knowledge so that AI can work with it
Alexander Heusingfeld is back as a guest, this time without Jens. The question at hand is one the two of them have been passing back and forth in voice messages for weeks: how do you build a knowledge space that an AI system can really work with? The starting point is simple. What an agent achieves depends not only on the model and the harness, but on the context you give it. With good context, a casually spoken prompt is enough. Without it, even a long one accomplishes little.
Before getting to tools, it's about data. Alex has the new agent offerings from OpenAI and Meta's assistant Muse in view and asks whether people realize what happens to their entire life, health data included. There are plenty of examples of the power of linked data, such as movement patterns from location data that show where someone works and sleeps. Mark sees the biggest risk in companies as a new kind of shadow IT: sentences and documents end up in AI systems whose data processing agreements nobody has settled. Then a story from private life. Someone who uploads their pet's insurance contract for analysis may, in the unlucky case, be shown a passage from a stranger's dental insurance contract, complete with name and contract number.
Mark's alternative to the built-in memory features of the big apps starts with chaos: a folder called Inbox, holding emails, chat messages, PowerPoints and transcripts. A skill converts this into Markdown, builds a wiki with a knowledge graph from it, and adds a file that defines how the system should answer: with a source or without, with the internet or without. That is how SkillSafe came about. The knowledge ends up in a directory that runs offline and anywhere an agent can read skills. Alex highlights what many RAG systems cannot do: qualify the relationships between nodes. Whether someone is a partner, a child or a colleague, we know from everyday life. In a knowledge graph, it first has to be spelled out.
Then it gets practical. Mark packed two matters with a combined 28 GB of payload data into one SkillSafe each and had the system look for the differences. He turned the talks from Apple's developer conference into a skill and sent an agent through his own projects in Azure DevOps with it. The result was a list of what is outdated and where new features are an opportunity. Alex brings in the layers: organization, area, team, person. And versioning, familiar from everyday life without anyone calling it that. Grandma isn't coming on Thursday after all but on Wednesday, and everyone should have the current state.
Two more building blocks come in. ContextMe doesn't determine knowledge but relevance: once a day, it looks at mails, Teams messages and projects to see what currently occupies Mark, professionally and privately. An app Mark is building, which is currently in review, uses this. It collects screenshots, breaks them down with a local model into text, dates, links and QR codes, and only notifies him about what matters. His daughter's riding tournament needs no push notification. An interesting Git repo perhaps does.
The last big block concerns collaboration. The SkillSafe workshop keeps knowledge alive rather than frozen, and so it has to answer questions that a single skill never faces: Does only the new information apply, or the old too? Who checks that everything is still consistent? What happens when two agents write at the same time? Alex works with colleagues through a shared OneDrive folder. From imported meeting transcripts, the workshop found contradictory decisions, asked for votes, created a template for decision records and pulled tasks over from Jira. Much of this was not built in at all.
Finally, about getting started. Mark describes how, after a weekend, he had a usable solution while others had spent months on concepts. Alex, by his own account a chronic overthinker, argues for a protected framework in which the worst case is that little happens, and for small experiments. His tip: the command-line tools of Claude and Codex can now see each other's sessions. If you're stuck, open a new window and get help there.
—
Think Different. Think AI. with Mark Zimmermann and Jens Scharnetzki.
Listen: Apple, Spotify, YouTube, RSS
https://think-ai.podigee.io
New episodes first in the WhatsApp channel
https://whatsapp.com/channel/0029VbDJ4ZlGufIrpquad004
If this episode was useful to you: rate us once on Apple. That decides whether new people find us.
https://podcasts.apple.com/de/podcast/think-different-think-ai/id1828021699
Feedback and guests: via the show page.
Transcript
00:00:00Welcome to Think Different. Think AI., the podcast by Mark and Jens.
00:00:07Two tech-loving minds who don't just talk about artificial intelligence, but live it.
00:00:14Here you get clear classifications, real practical insights and a fresh look at what's possible.
00:00:20Down to earth, critical, and always with a wink.
00:00:24Something to think about, to smile about, and above all to join in on.
00:00:29Welcome to Think Different, Think AI.
00:00:36Actually, I wanted to start completely differently.
00:00:39I wanted to start with "Think different, Think Alex."
00:00:42And welcome our guest right away.
00:00:45Sometimes introduced as a camera kid, sometimes as co-host,
00:00:49sometimes together with other people,
00:00:52but still happy to be our guest on his own, again and again.
00:00:56Today we talk about a lot when it comes to knowledge.
00:00:59And when we talk about knowledge and brains and such,
00:01:03Alex, as I said, can't be far away.
00:01:05Alex, great that you're here.
00:01:08Thanks for the invitation, and for having me back.
00:01:11Apparently the comments on the last podcast weren't that bad,
00:01:15even though we talked for an hour and twenty or so.
00:01:18Yes, maybe it wasn't that bad.
00:01:20Maybe the people who listened to it
00:01:23fell into some kind of waking coma
00:01:25and couldn't even write that it was too long, or people listen to it
00:01:31in stages.
00:01:32Or as a sleep aid.
00:01:33Yes, as a sleep aid.
00:01:34We talked earlier about all kinds of aids, and I said it sounds like a compression stocking.
00:01:40At this point, maybe some self-promotion: I'm building a podcast app
00:01:44that doesn't play whole podcasts anymore. Instead, you say which topics
00:01:49interest you, and it checks which episodes or seasons
00:01:54or special episodes covered the topic, then pulls the passages out of the transcription and chapter marks
00:02:00and builds a new podcast out of them, perhaps one that's just about Jev, for instance.
00:02:05That's really funny: one day you think you'd like something like that,
00:02:14and a day later Claude has hacked it together for you.
00:02:19In software development you upload to Apple so often, because you want to test it on
00:02:25TestFlight, that at some point TestFlight sends you the message
00:02:30that you've made too many uploads and are now blocked for 24 hours. I've learned a lot
00:02:34with this project. But today we don't want to talk about podcast projects
00:02:38on the podcast. That would be something like a Möbius strip or so. No idea,
00:02:43I don't know much about that, sorry to all the mathematicians and so on.
00:02:46Wrong podcast, we're talking about AI here, and about focus, Alex.
00:02:51That's interesting, because my hobby app is basically, I'm having Codex right now
00:02:59build something that a very respected colleague we both know recommended to me,
00:03:04namely a focus tool.
00:03:06It asks: what's your context right now, for the next 25 minutes or for the appointment,
00:03:13and what do you need for it?
00:03:15Hide everything first. I don't know if you know this, with 570 tasks and 73 tabs open,
00:03:24what do I need right now for the next topic, to get through the next topic. That's
00:03:31a bit of a theme, I think, that you mentioned to me in our pre-chat, under the keyword
00:03:36complexity of knowledge work. I'd like to digress briefly, as is customary.
00:03:43So when you said focus and app, and that you should keep your focus,
00:03:51for whatever reason I had to think of Die Hard with Bruce Willis,
00:03:57where he has a sign around his neck, and I briefly imagined him having "Focus" written on it,
00:04:02but that wasn't the case, and I don't want to repeat
00:04:05what was on the sign. I'm pretty sure otherwise I'd get some kind of
00:04:08letters. We're still too early in the episode for that. You know, we're still
00:04:12much too early, because we haven't lost that many listeners yet. We've come together today
00:04:17here because we have a topic that keeps coming up for us, either in
00:04:24our little one-message exchanges or in the discreet voice notes
00:04:28that I sometimes send, longer than my newsletters,
00:04:33which I write, but that's another topic too. We have to make longer podcast episodes,
00:04:39so your voice notes can get longer accordingly.
00:04:42Oh, I see.
00:04:43My problem is always that I want to refer back to something you said at minute 8:16,
00:04:50that's the thing.
00:04:52Yes, maybe you should build an app for that, some kind of commenting feature for voice notes.
00:04:59We could call it "Einmessage" in the... wait.
00:05:01Yes, I think that name is already protected.
00:05:04Greetings go out to Annette.
00:05:06The whole topic of Second Brain, how do you build one? This came up for me as well
00:05:12not in preparation for this episode, but through a conversation today,
00:05:17which I had with my boss, who also raised with me, this
00:05:22"T-Motor": whether we could take another look at how to
00:05:25set that up nicely. And I have one or two GitHub projects,
00:05:30so that the self-promotion is completely finished and hopefully wrapped up.
00:05:33I don't think so, we still talk about it quite a lot.
00:05:37Well, we'll see. But then you may talk a bit too. You know, when you talk about it yourself it's a bit tricky.
00:05:43Because I think we probably agree this far: when you talk to AI systems, what differs is not only the model or the harness you use it in, but, mind you, the context you give the system.
00:06:00So that even with a casually spoken prompt it delivers a much better result,
00:06:08than someone else, who is missing a lot of context around it, but maybe
00:06:13so the prompt was long, and that's why we wanted to talk about Second Brains today.
00:06:17Right, let's talk about it.
00:06:18Exactly. And I'll claim that neither of us is an absolute expert in information management.
00:06:27We've simply tried out a lot, and I've already said that I
00:06:33don't know whether it's a kind of paranoia, but when I look at presentations like
00:06:39yesterday's from OpenAI's dots, for example, I ask myself, do people actually see
00:06:47what happens to their data there? I mean, Cambridge Analytica was an eye-opener,
00:06:53we know Meta probably does a few things with the data, and whether you'd then want Meta's
00:06:59personal assistant Muse to entrust your entire life to, including health data,
00:07:07and who knows what else. Everyone decides for themselves. But it would also be quite
00:07:13interesting to understand what that actually means. What parts does it consist of, and
00:07:19what alternatives do I actually have. And as you said, you've built something like
00:07:25SkillSafe, you have the SkillSafe workshop, you've built a Knowledge Distiller that
00:07:30can ultimately build a knowledge graph, and you've built ContextMe, and last week
00:07:37you also talked about Jev. So there are loads of building blocks, and I understand
00:07:46completely why people say, okay, on the one hand I've got all these building blocks,
00:07:51and on the other side I've got my personal knowledge, oh wait, that's not even
00:08:01personal, because I have topics that are more private,
00:08:06friends, family, hobbies, whatever, and work. With work I'm bound by contracts
00:08:13and compliance, whatever. So I'd have to make sure the data and the tasks,
00:08:18responsibilities and goals I have are organized in such a way that it
00:08:23complies with the company. On the other hand, it would also be good if I didn't forget
00:08:28the kids' dentist appointments. So, you've laid that out nicely.
00:08:34Thanks. Well, maybe not wrong, and the question is simply: how do you not get lost in
00:08:43it all. It feels like everything around us gets more complicated, and there are more and more variables,
00:08:49and I completely understand that a lot of people say, no, really not. Oh,
00:08:55OpenAI's dots, yes, yes, let's do it. So before we really get into the topic,
00:09:02and maybe you could, I mean, we're an audio podcast, but it might help us
00:09:07with structure if you bring up that nice little diagram again,
00:09:11so we can move along it, how you do this with knowledge
00:09:15work. But while you're looking for the document for the old white man, on your side
00:09:20of the monitor, I wanted to say something more about the dots. Basically, I found
00:09:25the dots cool, the whole idea that the agent doesn't just work on your
00:09:31computer, the agent has its own environment, the agent has a way to keep working
00:09:37even when I've closed my laptop, or to build its own tools. But I found
00:09:44a few things there really efficient. One is that they said
00:09:48there's a collaboration space, and you can use it together in teams and so on,
00:09:53but you can also let them talk to each other. I briefly wondered
00:09:58whether the fine print says you can stay at home, yes, the dots
00:10:02will sort it out among themselves. And the other thing is, all the data you give, now, I think it's
00:10:09of course understandable that the more data systems have, the better they can
00:10:15answer. And you also have the case, back then it was still called Facebook, I think,
00:10:20Meta, the renaming, which came later. The power that lies in your data also surprises
00:10:31me quite regularly, what conclusions you can draw when you connect a lot of
00:10:38things together. And that's a bit of what this whole thing is about,
00:10:44what we want to discuss in this episode. Another good case for why data
00:10:50might be worth protecting: greetings to Flübcke, a VW data scandal involving the
00:10:56geodata. You can also derive movement patterns from location data. Where
00:11:03does that person probably work? Where do they sleep? During the day, though, they're often in the
00:11:11forest, what's in that forest? Well, if it were only the forest. Yes, yes, exactly. And so
00:11:19from my point of view it should prompt some thought, to say, okay, everything I do somewhere,
00:11:26do, say, can and will be used against me in court. No, just kidding.
00:11:34But the question is, how do you
00:11:39how do you, I like this phrasing,
00:11:43about the data, but how do you manage to keep track of what actually
00:11:49goes where? Well, I can at least speak for myself, in a company context this is
00:11:56already expected, that you know what you're doing and how you handle the data
00:12:00responsibly. Some people are even paid for it. I'm not, but are you
00:12:07simply a natural talent? Now you should talk to them about a salary negotiation.
00:12:11I think at that point, regardless of which tools you've been cleared to use in a
00:12:19company, the biggest risk alone is what's developing in companies,
00:12:25what we used to call shadow IT, which I think becomes much, much more critical with AI.
00:12:34Because I don't want to play it down, yes, to all the colleagues listening who'll say, well, he didn't really say that.
00:12:41But if you've typed something into a Google window, nothing against Google, but you may have entered a sentence there,
00:12:50triggered a search, and then, well, that's that.
00:12:52So when you enter sentences into AI systems today whose data processing agreement you haven't settled,
00:13:01where you think, come on, it's not that bad, just throw the sentence in, you mustn't forget,
00:13:08the more people do it, the more it adds up, right?
00:13:12So you're handing out units of a certain IP range just by uploading documents into systems,
00:13:20how quickly it's able to put together not just a nice answer from a document,
00:13:24a nice picture, a little video, a short summary, and we don't really
00:13:31know what becomes of it. From my private life I had a case recently,
00:13:36where I had to look something up, or wanted to, whether the vet's
00:13:43insurance for our dog had an update on a certain matter. And then I couldn't
00:13:51find it, so I took all the papers, put them into a scanner app,
00:13:55a fast duplex scanner, uploaded the PDF to ChatGPT, and said,
00:14:00come on, do it. For my dog's contract data I didn't think that was so bad.
00:14:05Unless somebody calls to say the dog has data protection rights too,
00:14:10dogs still count as property, I think, but I don't want to debate that.
00:14:14Then I uploaded it, and ChatGPT came back and said, yes, there it is.
00:14:18So I'm blind now, I can see that, I know it, but it's not on page 2.
00:14:25And ChatGPT said, yes, it's on page 2.
00:14:27Okay, please show me the text excerpt.
00:14:29But not from the document.
00:14:31Wait, here it comes, yes, I don't see that.
00:14:34And then it offers me: shall I show you the document with the highlighted passage?
00:14:38And then there was a health insurance contract,
00:14:43not for the dog,
00:14:45but with name and contract number.
00:14:49And then you stand there and think,
00:14:51where on earth did it get that from, you ask.
00:14:55And then it says, let's sing it too:
00:14:57that should never have happened.
00:14:58And then you think, well, firstly, right,
00:15:02so a lot failed there, right.
00:15:05Firstly, right, that's what showed it to me.
00:15:06Secondly, where did it come from? Not from old books they shred to scan,
00:15:11it can't have come from that. And it was even from the same insurer. So
00:15:20they were also the right ones, because there are companies with certain color values,
00:15:24that recognize documents when they reach you in their postal state,
00:15:30you know them. And it all matched, but it was
00:15:34not a dog. The person wasn't called Hund either. But it was about teeth. Well, at least that. And
00:15:45probably he wanted insurance too, if you know someone. And probably
00:15:50he didn't want me looking at it on my computer. I'll assume that for now.
00:15:54And from that side it's a big deal, if you potentially
00:16:01upload something like that. Regardless of whether you're allowed to or not, the whole point is, when
00:16:07data is made available to an AI, it's a curse and a blessing. It can
00:16:13be very powerful and also very risky.
00:16:16In a company context we now have the option to say, we use, for
00:16:23example, one of the big hyperscalers, who say, well, we guarantee that
00:16:30the models are hosted with us and don't leave any EU country. They have a
00:16:37redundancy, no idea, Stockholm, Paris, Frankfurt, and it stays there. Unless you
00:16:46at certain times used Fable, or they at least guarantee it. There are also
00:16:50some who guarantee it contractually, but those things still sit somewhere
00:16:54overseas, which is funny too. Those exist as well, and I think that's great.
00:16:58Good, but let's say, in theory it's all fine, but you lose
00:17:05convenience, because with the standard thing, oh yes, I have a
00:17:14ChatGPT app and a memory gets created there. In that moment I
00:17:19simply don't have that anymore. On the other hand, personally I see it as an advantage, because
00:17:26with a memory, yes, what if we, for example, work on a topic together and we both on
00:17:36this, let's not call it memory, but maybe a knowledge trail or knowledge
00:17:42trace, and we both contribute to it? That's a use case that, from my point of view,
00:17:49the standard systems don't support? Well. Let me put it bluntly.
00:17:55The standard systems are very strongly oriented toward individuals,
00:18:01commodity-oriented, and the whole question of what requirements a company like that has,
00:18:06what requirements teams of two or more people have,
00:18:12what requirements you have? Even requirements in a family context.
00:18:17Along the lines of, I share, I'd offer something like family accounts and maybe also say what counts as family knowledge or not. They haven't got there yet.
00:18:27At the moment the race is faster, higher, further, give me your data. Today a model comes out so that tomorrow the next model comes out.
00:18:35But this topic of collaboration, I think, is still in its infancy.
00:18:42Exactly, I see it gradually coming, that once you have a team account, at some point you can say, okay,
00:18:48for this entire team of 200 people,
00:18:54for this whole team you can then
00:18:57define a shared memory, but we wanted to talk about alternatives.
00:19:02So I'd suggest we dive into this use case and say,
00:19:07okay, personal memory is fine.
00:19:10Can we briefly say how we're doing it right now, because we both use, I think,
00:19:17the ContextMe plug-in, right? I like that we both use
00:19:22the ContextMe plug-in, so maybe just as a... So we're not using the same thing,
00:19:26we're using... I've installed it and you've installed it, but...
00:19:29Hold on a second, I have to switch off yours, it always gets transferred to me, trust
00:19:33the skill you've loaded. Always check which one you're using,
00:19:38That's also funny, because depending on that, there was a nice security scan
00:19:43recently, where people said that some skills hide things,
00:19:47where they might call up a website.
00:19:50In MCP servers.
00:19:52Pretty funny.
00:19:53How many people do you know who look at the installers for such a skill?
00:19:58I mean, when you copy and paste, they usually have on their websites
00:20:06Yes, you just copy this command and paste it into your terminal command line.
00:20:12And every time it's some bash thing, run this shell script, and I think,
00:20:20okay, should I maybe download it first and check what gets executed?
00:20:24Oh, come on.
00:20:27Ah, it'll be fine.
00:20:32Exactly, nothing has ever happened.
00:20:36No, let's stay focused on the use case. Say we both work together in the team now. How would you set that up?
00:20:45That's a real trick question, if you come with the argument that we both use, let's say, the same technology and at least the same basic origin.
00:20:54I'd like to get into the topic of when people say, I'm building
00:20:58a Second Brain, a knowledge space, for themselves or for the team, where then
00:21:03first comes that entry hurdle: what do I need to watch out for, what do I need to
00:21:09think about what goes in there.
00:21:11I'd like to start there first, before we maybe get into collaboration
00:21:16or the organizational levels.
00:21:18And at this point, hard to believe, chaos is your best friend.
00:21:24Because when you grab an inbox and say, I'm throwing in whatever's on my mind,
00:21:32then at least that's my beginning.
00:21:34I'll come to context in a moment, in case that helps.
00:21:37But in the first step, from the history, when I started with SkillSafe,
00:21:44the best result was that I said, look out.
00:21:47I've got a whole pile of topics here, process documentation, descriptions, emails, PowerPoints from a project.
00:21:56Then I created the folder called Inbox, threw things in and built a system that pulls something out of the inbox,
00:22:05converted to Markdown, because I, old white man that I am, I stand
00:22:11by it and say Markdown is still, I think, the best format ever for an AI system, and
00:22:18not HTML. I never jumped on that HTML trip, because there's so much
00:22:24clutter in it that an AI isn't really interested in, that only humans care about, or
00:22:29whatever. And then SkillSafe came along and said,
00:22:34okay, I'll structure it in the Open Knowledge Format from Google, and store it in a kind of LLM wiki, the way we heard about from Karpathy,
00:22:44so that the skill has a kind of structure, so that the AI system that loads the skill
00:22:53can consume this knowledge and give me answers based on it.
00:22:57So at some point I realized, damn it, with what you put in, it's maybe quite funny
00:23:03to explain to the system what I actually expect when I ask it something. So do I want
00:23:08a long-winded answer from you, or bullet points, do I want you to
00:23:13always name the source, do I want you to, if you're unsure,
00:23:17look things up on the internet or not, and then, alongside the wiki that emerges in the skill,
00:23:23something like a SOUL.md, so it knows how it should react. So,
00:23:29if I don't know something, should it fetch something from the internet? Yes, no. Or if I
00:23:33know something, should it always name the source, in the form of which document I
00:23:37originally pieced it together from? And how should it cite that? And I found that
00:23:41pretty cool for a long time, because loyal podcast listeners have noticed it too.
00:23:46Skills are kind of the big magic of teaching the system to do things repeatably,
00:23:51and to retrieve knowledge too, repeatability of things. And
00:23:58and then you stand there, and people like you come around the corner and say, yes,
00:24:02that's us, when you work with different departments.
00:24:06That's us, when you work with different people.
00:24:09And when it came to SkillSafe, I notice I'm already rambling again like an old man with verbal diarrhea,
00:24:15I thought, okay, damn it, he's not wrong,
00:24:19because there is, whether it's at home or not,
00:24:23the motto: family knowledge above all, or someone who, for example, has a company's compliance knowledge
00:24:30above all, corporate guidelines and whatever else there is, you could
00:24:36throw a SkillSafe at that, but then it would be nice if such a SkillSafe were aware of
00:24:40where it's organizationally deployed. Along the lines of, what is
00:24:46my scope, can my knowledge be overwritten by someone else, or am I
00:24:52not so overarching that other rules have to subordinate themselves to me. That may
00:25:00already be in the name with a corporate guideline, but in an AI system that's pretty
00:25:04irrelevant, what the file is called. Context definitely helps, so the system knows how to sort itself
00:25:10in, and so SkillSafes got a property. First, you can
00:25:15address other SkillSafes, that's given to you. Second, another SkillSafe can
00:25:20have its organizational assignment passed along.
00:25:23Along the lines of, I'm a personal SkillSafe, I'm a team SkillSafe,
00:25:27I'm a project or department SkillSafe, or I'm an organization SkillSafe.
00:25:31You can already tell, all of these things have a somewhat business character.
00:25:35That was the first attempt.
00:25:37Did that surprise you?
00:25:39Nobody at home deals with this stuff otherwise, right?
00:25:42Well, what do you mean, deals with?
00:25:44I think, well, let's say,
00:25:46in the family you don't have that, right?
00:25:49You deal with it...
00:25:51In the family, that counts for real.
00:25:53No, but...
00:25:55That's a different family, more like the mafia, greetings to Italy.
00:25:58You don't have that many different... Yes, exactly.
00:26:00I was just going to say, you don't have that many different groups.
00:26:02If you were to say, well...
00:26:05There's information that should stay within the family.
00:26:08There's information I want to share with the grandparents.
00:26:12But certainly not everything we discuss in the family.
00:26:15There's also information I maybe only want to
00:26:18share with my partner and not with the kids. So, when something like that comes in,
00:26:23then you can say, well, there are actually already different groups for different
00:26:27knowledge spaces. You can of course make it as complex as you like, but at first you can also
00:26:33start simply, the way you just described. The nice thing, but we're getting into the
00:26:38north of these SkillSafes, is that you can, it's kind of a skill, distribute it
00:26:43And you can also maintain it like a skill, with the help of the skill.
00:26:49And you can also, let's say, freeze it.
00:26:51Yes, version states, along the lines of the skill that's available,
00:26:58aha, he's smiling, he wants to jump in, I'll finish the sentence before he says anything.
00:27:02You can bring it to version states.
00:27:05And when you then distribute the skill and make it available to the AI system,
00:27:09you can assume, okay, this is in there and that isn't.
00:27:12And that's the version, not some interim state or anything, a certain reliability.
00:27:19Or would you like to disagree with me there?
00:27:21No, no, that's fine.
00:27:23At this point I'd like to ask a question to the listeners of the podcast.
00:27:31Be honest, please, who among you uses version control or versioning for your files?
00:27:39And who says, knowledge management, yes, I've heard of that, that was the neighbor, he was a librarian.
00:27:49I notice, Mark, that I often make assumptions about what technical background people have.
00:27:58And I'm always surprised how many colleagues and friends
00:28:05use AI agents, on their computers, privately or professionally, some even use their private
00:28:14AI agents on their work computer, without really knowing what's
00:28:20actually happening underneath, and how do you secure fine data, and how do you version it?
00:28:25And this versioning, that's not, so you'll perhaps have to briefly explain,
00:28:30why you want that, also for a knowledge store like this.
00:28:34I also think it's lovely how you gave the example of versioning,
00:28:40another anecdote, people who use Photoshop know that Photoshop
00:28:45has different layers.
00:28:47And layers give you the chance to change things, let's say, non-destructively,
00:28:53because you move things back and forth and the background stays intact,
00:28:58and you move elements back and forth in front, and you get that through the
00:29:02layers.
00:29:03And I never learned Photoshop, and sometimes I sit in front of Photoshop and then you stand
00:29:09there and say, merge all layers.
00:29:12Every graphic designer's toenails curl up immediately when you say to merge everything onto one layer,
00:29:18but let's leave that aside.
00:29:20Versioning is, if I'm honest, something I actually
00:29:28use very rarely, only in the very few cases. Because with me, I mean,
00:29:38something like a contract, that I file away, and so on, I don't get
00:29:43new versions of it that often, new editions. Yes, I have a contract for my car insurance
00:29:48and that's that. The highest feeling is that I might get dropped and sign a new
00:29:52car insurance at some point. Yes, okay, fair point, but that's the
00:29:57kind of case where nothing happens, I don't know. Every ten years I add a new thing to it.
00:30:01I actually have little of this versioning at home, but in the
00:30:06company, knowing what applied when, what the basis was, which decision was made at which
00:30:13period and point in time, what was a quotation mark for the validity of the law and
00:30:19money, a guideline and statement, and also binding nature and reliability, when you look today at
00:30:27a matter, so you know in what context it took place, what the
00:30:33rules were, what was valid and binding at the time, and what is it today?
00:30:42I'd argue that we implicitly do this in private life too, when you say, hey, we had
00:30:51that appointment, but I can't make it, I'm sorry I can't come, because I
00:30:56have to go to the parents' evening.
00:30:57That's surely implicit, so an implicit example would be a different scenario
00:31:03would be.
00:31:04I was just wondering whether you should file the parents' council elections in GitHub.
00:31:09Sorry.
00:31:10Yes, no.
00:31:11There are definitely people who set up a meal plan and say, no idea,
00:31:21what we have to buy, and then we rebuild again, and because Grandma comes Thursday,
00:31:27oh no, she isn't coming Thursday after all, she's coming Wednesday now, those are the kinds of changes.
00:31:33Only that's not exclusive versioning.
00:31:36But anyway, what I was getting at is, you open these different levels,
00:31:43where you said, in your SkillSafe there's a level 1, in a company you would
00:31:49say, that's organization-wide, level 2 is then a smaller group, could
00:31:56be the entity, department, whatever, level 3 would then be a team level
00:32:02and level 4 is the personal. And at first I thought, wow, that's such
00:32:09complexity, you'll never need that. And then there are situations where you notice, yes,
00:32:15colleagues made a decision that contradicts a decision at
00:32:23organization level, say, the Chief Information Security Officer
00:32:29said you're not allowed to do something. But the team did it anyway. Or the head of
00:32:35Corporate Communications said we'll only use the following wording from now on, and you notice
00:32:43your team unfortunately still used something old in the app. Those are classic cases
00:32:50where you want to, and in my view that has not so much
00:32:56to do with carelessness, but rather with information gaps. For me it's about making sure everyone
00:33:01always has the currently valid information. And so I'd argue that in situations like this,
00:33:09where you want to make sure everyone knows, hey, Grandma isn't coming
00:33:14on Thursday, but on Wednesday, you version the data and make sure that
00:33:19only the most current state is there. And if someone then says, well, I thought
00:33:24Grandma was coming Thursday, it's clear, yes, that was the original plan,
00:33:29but it changed, on Monday evening, because Grandma said she's already coming
00:33:34on Wednesday. That saves a lot of trouble. It definitely saves a lot of trouble if it's
00:33:41easy to trace it like that, especially with partners,
00:33:49and with colleagues or, well, partners.
00:33:54Can you tell me? No.
00:33:57I basically only talk about people who are present.
00:34:00Hoping I haven't lost everyone by now, I'd come back briefly to the
00:34:04starting point, which is, what is this little bit of, let's say, knowledge trail for.
00:34:07How does knowledge relate to others, so that we have some foundation for it.
00:34:11But originally I came in with the sentence,
00:34:13unstructured chaos is my friend, that's how it really started,
00:34:17throwing everything into the inbox. And when I say everything, I mean chat messages,
00:34:21emails, documents like PowerPoints, transcriptions, whatever. And the astonishing thing is,
00:34:29that the topic account can do something with that quite well. Then I made SkillSafes for the
00:34:33most varied topics. That means, quite concretely, you had, for example, a Claude
00:34:40Code, and you installed this, the SkillSafe or the SkillSafe workshop plug-in, on it.
00:34:46Well, at that point I still had SkillSafe.
00:34:50And then you simply said, what you give it, basically.
00:34:54Here's a directory, build something out of it, or if you already have something, supplement the existing one.
00:35:03Exactly, yes.
00:35:04I can remember that.
00:35:05And then came discussions, hmm, and how do we structure this differently?
00:35:10Yes, exactly.
00:35:11But what's maybe funny at this point, greetings to all the friends
00:35:14who always put RAG systems somewhere in the wallpaper, along the lines of, yes, you need
00:35:19a RAG system now, which answers questions as an expert system, for, I have
00:35:25a strange data set and have questions about it. Basically, that's possible
00:35:31to do with a SkillSafe. So I don't want a RAG system...
00:35:35Because your, your, your, no, that, yes, implicit praise, sorry, but what I find really great is that you don't just have a RAG setup, you have above all a knowledge graph, and you can qualify the connections between the nodes, not just information at the nodes.
00:36:01Maybe you know this: you don't just say, I have a relationship with another person,
00:36:09but you maybe also want to know what kind of relationship it is, what it expresses.
00:36:13In task management you say, this one ticket depends on the other in the project.
00:36:19It has to be done before the other one can be done.
00:36:23Among people, you want to know: is that a partner, is that a friend,
00:36:30is that a child, is that a colleague? So that's normal.
00:36:36No, it's several of those.
00:36:38Yes, the child is the colleague.
00:36:43Lovely.
00:36:44Take the trash out, no, the work one says, doesn't matter. What I mean is, in private
00:36:51life we're completely used to qualifying such relationships between these nodes in a knowledge
00:37:00graph. We just don't put it that way. So a term like
00:37:06knowledge graph is simply foreign to us. But actually, implicitly, we know it,
00:37:14that we, well, sorry for the word, entities, so nodes, which can
00:37:22be documents, which can be people carrying some information, that we
00:37:28have certain attributes, have properties. We're used to qualifying the
00:37:33relationship. And in a RAG system, in most RAG systems, it's not
00:37:39common. Above all, even if you implement it there, setting it up, we
00:37:47do then chances and whatever, and then it's queried, and all of that takes time,
00:37:51effort and money. And you have to plan what you're doing, and that's also
00:37:55a complexity you can see. I once created, with SkillSafe, 28 GB of payload data, to compare two
00:38:07matters. So for each matter I built a SkillSafe and told the system, in the cloud: you've got two SkillSafes here, go and see where they differ. It then chewed through both knowledge graphs and told me at the end of the day where the differences are, in a very reasonable time, and nobody sets anything up just for something like that.
00:38:37You don't do that. And that was also the motivation for the
00:38:45SkillSafe, when I understood. Damn it, it can't just contain work instructions,
00:38:50but also spread knowledge, hold knowledge, and
00:38:53be responsible for questions and show connections between each other. That's
00:38:59also a capability you simply give it, because at the end of the day it's all Markdown files
00:39:05And yes, it's ultimately in a directory, and you hand over all the knowledge and have it available offline in the end too.
00:39:15You have it on the laptop, you can install it anywhere you have Claude Code.
00:39:20You don't even need code, even Copilot can do that, so it works everywhere.
00:39:26Because at the end of the day it's really nothing more than a directory structure with Markdown documents and sufficient links between them,
00:39:34where you have to be a bit careful with the SOUL.md and such,
00:39:38so that it's read properly.
00:39:40But as they say, in most cases,
00:39:42you tell the thing, here's a directory,
00:39:44work through it, answer my questions, and that's that.
00:39:47And yes, you can, but you can also,
00:39:51and I've actually done that in one case,
00:39:53say what's especially important to you during the import.
00:39:57So what should it please include in the knowledge graph?
00:40:01Not just saying, ah, okay,
00:40:03a document, and there's certain information in it, but you can
00:40:06also tell it what you want to pull out, whether you
00:40:10want decisions as their own elements or nodes,
00:40:15certain kinds of insights, when you say, I have a
00:40:20research task, and I have these questions that I need
00:40:24answered, works great. Before I get to context in the
00:40:30conversation. Maybe one more use case. I think we have so much,
00:40:36we talk, you know, about nodes and we talk about the inbox and we talk about
00:40:39blah blah blah. I think sometimes you should also give a few more
00:40:43examples. Where does it actually help you? I have one example of my own, because you have
00:40:48different matters and you want to compare the differences. The other
00:40:51example was, you want to keep knowledge for an expert system,
00:40:54so that you don't constantly have to query some database somewhere and
00:40:58ask, how does this look, and set up a big database, but that in your
00:41:02agentic system you can play out your chat-window question and answer, but also use it in the
00:41:07agent context to act on it. And one example of context was, your drive
00:41:13has picked up on it a lot. I'm not related to Apple and I'm not in-laws with them
00:41:17and I don't have any shares and I don't get any money from them, but I find
00:41:21the company quite good anyway, and the products too. And accordingly I
00:41:25always follow their developer conference, and so I went and got the
00:41:29whole transcripts, that's the spoken word of the presenters, which over a week
00:41:34or across the last few days, I think, was all published the same day.
00:41:40But you've got this whole bag of transcripts on the most varied topics, from the most varied
00:41:45frameworks, from the most varied programming paradigms, from the most varied rules that
00:41:50you build software with, for example, or for the Apple operating system.
00:41:53And for Apple's operating system topics, and so on.
00:41:56And there I simply went and said, look out here, there's YouTube,
00:42:01Apple has published all of that there, so please download all of it and throw
00:42:04it all into a directory, and then I built the SkillSafe out of it.
00:42:09And then had a WWDC skill.
00:42:13And now you could ask this skill what Apple has brought out new, the management,
00:42:19and what you have to watch out for, what's deprecated now, that was the one thing.
00:42:22But the other thing was, I could tell the system, that is, tell the co-work.
00:42:28Look, here you have an Azure DevOps, that's a lot of our projects.
00:42:33You see that in the word map, you see what people are working on right now.
00:42:37You see bugs there, you see the knowledge documentation.
00:42:41But you also see the source code there.
00:42:43And then I went and said, look here.
00:42:46Do you see all of that?
00:42:47Just take a look.
00:42:49For your view, use this SkillSafe. And then the thing went and looked at the
00:42:55projects one by one, one by one, and then told me afterwards, look out.
00:43:00In that project, the following things are deprecated. In that project you got the feedback,
00:43:06this offers this functionality from iOS 27, and so on. So it basically made a
00:43:13mapping with the knowledge that Apple had conveyed, with active challenges
00:43:19you have in projects right now, or where changes are coming up, and it
00:43:23basically superimposed them, and then you sit there, you can barely
00:43:28wipe the smile off your face, because at that speed, dependencies,
00:43:33changes, opportunities and risks come out, and if the
00:43:39answer of one or the other listener is, it thought of 1000 percent of all topics,
00:43:44yes, I know, 1,000 percent gets difficult when 100 percent is basically the limit. But
00:43:49no, probably not. But it probably thought of a lot more
00:43:55than I did. And above all, you had a system and could go to the product owners and
00:44:01say, look out, things are changing here. Here's the SkillSafe, ask it.
00:44:07Talk it through with your project again, discuss it again with your developers, and so on.
00:44:12And I find this question, how do I write a work instruction as a skill, is one thing.
00:44:19But how do I store knowledge so that the skill can work even better,
00:44:23so that you can work even better with your system?
00:44:25There's so much gold hidden there, because AI systems are already smart.
00:44:33But the more context they have about what's occupying you, what you want to do right now,
00:44:40where you need help right now, the more it helps them focus and bring you
00:44:47support.
00:44:48And we're not even at collaboration with people yet.
00:44:50We're also not at what they promised here with their dots.
00:44:54We're simply at the question of how the system gets equipped so that
00:44:59it helps you here and now, that it understands you, that it supports you, that it
00:45:06starts actions that are simply more than if I'd just asked
00:45:12for another classic case, research what Apple has released on the internet and tell me what
00:45:16that means for my project?
00:45:17You'd get an answer too, but I'd bet on it, it would be
00:45:22by a long way worse.
00:45:24And once you've got used to building a SkillSafe like that, and you can
00:45:29also build your capacity wickeys or whatever, however you like, but once you've started
00:45:34throwing things in there to make the AI system more digestible, usable,
00:45:41you see what data is, the gold of the future, blah, blah, blah,
00:45:46yes, we've heard those stories very often, for a very long time,
00:45:49but you get a sense of, yes, that's right.
00:45:53And I'll even top that.
00:45:55I have, as I said, Apple and so on.
00:45:58I'd like to chime in briefly.
00:46:00I just want to remind people once more.
00:46:02This is the alternative to the memory solutions that the standard apps are bringing along right now.
00:46:12And what Mark just said about the new scales, that's not only in use with him,
00:46:16but with me too, where we say there's information that is not
00:46:22volatile, but should be available for a few more sessions, or for a somewhat longer
00:46:29runtime. And for use cases like that you can very well pack it into such a skill,
00:46:34into such a knowledge draft, because I assume you delegate something to an agent
00:46:40platform, right? New buzzword, but everyone offers all kinds of things, you can
00:46:46piece it together now even at the hyperscalers, a bit of hot air with it, but you can
00:46:54simply hand such a SkillSafe along as a skill and say, by the way, this is your main context,
00:46:59and this SkillSafe, you can, which you haven't said yet, you can also repackage again and again,
00:47:03so you say, okay, first I thought, we'll do, no idea,
00:47:08all the groups in the company, all company-wide policies we pack into one skill, and
00:47:14then you notice, ah yes, good, the purchasing policy and the controlling policy are now
00:47:20perhaps irrelevant for the software developer, so let's make
00:47:26a knowledge store that's relevant for software developers, and one for controllers, or
00:47:32one for purchasing policy.
00:47:33And you can repack them again and again, and at first I thought, oh well, then
00:47:38I'd be throwing tokens out of the window every time.
00:47:41No. You can take the original SkillSafe or the original knowledge, as a
00:47:48base, right?
00:47:49Because you have the knowledge graph, which is an abstraction, but this knowledge is
00:47:54the relevant part.
00:47:55Unless you say, no, if I'm now extracting for the controller, then
00:47:59different rules apply to the extraction, then I want
00:48:03different nodes and relationships in the knowledge draft.
00:48:06But again, that's of value to you too. If you say, no, I want to reprocess this information for the controller, then it's of value when exactly that happens.
00:48:19And now let me ask, how well do the standard systems support that, the ones you can buy outside for big money and big convenience?
00:48:29Not that I want to advertise for your skills, but they're available, and I have to
00:48:37occupy myself with them a bit at the start, although the documentation, even though it's
00:48:41in German, I find quite incomprehensible.
00:48:44You can already get dissed for German documentation.
00:48:47That's what it is, I find that nice, that pleases me, but ...
00:48:50I can't really diss German documentation, I don't even know, is "dissen" a German
00:48:54word?
00:48:55Am I allowed to do that?
00:48:56Whatever.
00:48:57Whatever.
00:48:58What I was getting at, I always try, you named the use cases, each of you
00:49:06can download it, so essentially they can point their Claude Code,
00:49:11or Codex or Gemini or whatever, at the address.
00:49:14Yes, sorry, exactly, because we're not sponsored, your podcast,
00:49:20but you can take it and simply explore, nothing bad happens there.
00:49:26This data, depending on what subscription you have, or whether you say it's on
00:49:33Bedrock, AI Foundry, GCP Model Garden, wherever.
00:49:38Whatever you choose, wherever your GPUs run.
00:49:44Or locally. You decide. You have the option.
00:49:49Just try it out.
00:49:50And I was just trying to catch my breath, but I'd originally also said, once you've
00:49:57played around with it, you get a feeling at some point, I'm saying feeling here,
00:50:04for what else you can use it for? Where else does it help you, and one application, as I
00:50:10just... I honestly simply installed it and said, I'll just let it run
00:50:16in the background and ask it things in between, and now something where I
00:50:20could use SkillSafe. Look at our last few sessions, if I could optimize this somewhere
00:50:26with SkillSafe. And its answer would be, yes, I've already done that. But even now
00:50:32I'd still like to come back to the topic briefly, because what you say
00:50:37is of course true, that you can say, dear system, look, here you have a lot of
00:50:42options, now, never mind, that applies to all skills and plug-ins, that you can
00:50:48ask the system where it might have helped me if it had been used, or
00:50:53can you give me a recommendation. Where I'm currently using it too, and the app is unfortunately
00:50:58not in the store yet, where I also noticed, first with initial curiosity and then, because
00:51:05I can't get my hands on it anymore, that I already have the app on my phone,
00:51:09it's an application that does nothing other than pack the screenshots I take
00:51:13into a SkillSafe. Because I'm quite, we had it in the
00:51:20pre-discussion, yes, greetings to people who can't stand it
00:51:24when a mail icon shows more than the number 1 in red for unread mails, yes.
00:51:31I'm at a large four-digit number, but only because
00:51:35I now and then just say, select all and mark as read, done, so I don't
00:51:39have that issue, I see something, whether it's on social media or somewhere else, yes,
00:51:45and then I take a screenshot on the device I'm on. And a screenshot is,
00:51:51well, it's a photo, right? A photo of the screen content, there can be text in it,
00:51:56there can be anything else in it. Or QR codes, or whatever. And then I built
00:52:02a small application that does nothing else for you than be a kind of screenshot vacuum cleaner,
00:52:06and use a local model to break these things down, the way you know it from the Knowledge
00:52:12Distiller and the SkillSafe, and it tries to get everything possible out of the image,
00:52:18basically.
00:52:19So if there's a photo, it tries to understand with Apple's Vision framework,
00:52:24aha, there's a tree and a laughing child.
00:52:26It tries to extract the texts, it has data, selectors, what it picks up,
00:52:30are addresses, dates, links in it, reads QR codes and all that stuff.
00:52:36The app is not yet available to everyone, right?
00:52:41Well, it's currently in review.
00:52:45And that's really funny, because I started the thing on my phone.
00:52:51Honestly, the phone got a bit warm, because it had to somehow.
00:52:55Well, the first number was a four, and then a few more numbers came,
00:52:59because it's the number of screenshots it should import from the last twelve months.
00:53:03And when it was finished
00:53:05and you then also in the app on the iPhone, I then had it with chat and with this day view,
00:53:12and you navigate through it, but there's also a Mac version for that, and I have MCP support,
00:53:18so I can then go at it with Claude or whatever, and suddenly all
00:53:23these screenshots that I've taken over the last 15 months, I don't want
00:53:29to sound pathetic, but living knowledge, because I took the screenshots
00:53:34and I've forgotten most of it, yes. The ADHD comes to mind, like when I recently
00:53:38found my company badge at the coal machine, my bag in the room where I started the
00:53:43meeting of the day, only that in the meantime it was six p.m. And at that point it helps
00:53:49me a lot, because I can now just take a screenshot, and I know, if
00:53:57there's an appointment on it that I want to be reminded of, if there's a link on it
00:54:01that I don't want to forget, if there's anything else, that this system collects it and then sends me a push message and so on.
00:54:09And now you might ask, hey, a push message, that's where the circle closes to the context,
00:54:14yes. I wrote this small skill for myself, which basically tries not to store knowledge for you,
00:54:21but to determine the relevance of knowledge for you. Until recently, for example, I was
00:54:28responsible for everything at our company that was mobile app development. That means my professional interest is
00:54:33very strongly mobile. iOS, Android, yes, for the most varied projects that
00:54:38you're standing there. That means ContextMe didn't exist back then, but I could try it out on this
00:54:44example. If you had run it back then, it would have said, look out
00:54:49for iOS, Apple and Google and Samsung and so on, very important, very important. Today I'm interested
00:54:55in none of that anymore. iOS only privately, and professionally more the whole topic of AI. So
00:55:03above all, higher in importance than mobile. And ContextMe picks that up over
00:55:10time, because I tell it, please look over a period X, what mail I have,
00:55:15what Teams messages I have, which projects I'm in, in which
00:55:20role I work, and from that the system learns, roughly, what the market is interested in,
00:55:24right now, how it's changing, professionally and privately, and that then feeds back into
00:55:33the SkillSafe and into the app that's on its way right now, just briefly...
00:55:36Yes, wait, wait, just briefly, so, quite
00:55:39concretely, you've installed ContextMe and then really use it as a slash
00:55:47command.
00:55:48No, no, no, it's different, I've set it on a timer, so it goes once
00:55:54a day over all mails, all projects, all Teams messages, and updates. That's
00:56:02once a day a good rhythm, maybe once a week or once a month would be enough, so not that erratic
00:56:07I am, that might be enough, so it's simply set up that way for now.
00:56:11So that it looks at what interests Mark.
00:56:16I have it running as jobs on my inbox, you just mentioned the Inbox directory.
00:56:22On the Mac mini at home I have an import job running, and it uses
00:56:27context, like, on the different import directories, and I have a daily briefing, where my
00:56:37day is summarized, yesterday's day, all the information, that is relevant today,
00:56:43which appointments I have, what I have to keep in mind. And that's also where
00:56:49context is included. It's then used read-only, basically. So I take it
00:56:57to work out what interests me, and then I take this
00:57:01ContextMe result and bring it, for example, into this app that
00:57:08manages the screens, these screenshots. Along the lines of, I don't want to
00:57:13get a push message for every bit of stuff, like, tomorrow at 12:00 there's
00:57:17this, which I maybe screenshotted because my daughter is into horses.
00:57:21I personally have nothing against horses, but full stop.
00:57:26But I don't want it to immediately go, oh, very important tomorrow,
00:57:30drive to the stable at 1 p.m. tomorrow, right away.
00:57:32But if, for example, it's a screenshot of some cool
00:57:36gadget, then I do, possibly, want
00:57:40it ranked higher for me, and that I can then
00:57:44basically consume it faster, have it prepared for me, and I use it there
00:57:50just like for my personal second brain: I tell it, look out,
00:57:55throw the things in there, and you know, along the lines of, okay, the iPhone
00:58:02I've always liked and still do, but when I use it
00:58:05professionally, it knows, look out, that now the
00:58:09iPhone Duo is coming out, which the market finds totally great, but which is now
00:58:13secondary in its professional context. There, maybe it's more important right now that
00:58:19Maestro 6.1 has done something. And that's how ContextMe helps my second brain and me keep an overview in
00:58:30a world, and bring order to the chaos I mentioned at the start.
00:58:38We opened with knowledge management being complex, and that fits too.
00:58:45I think anyone who's listened to us until now deserves congratulations too.
00:58:49They're entitled to basically a SkillSafe...
00:58:52Exactly, everyone who has listened so far
00:58:55is entitled to a special edition of SkillWorkshop instead of a SkillSafe download.
00:59:00Because I was just about to say,
00:59:02because so far we've only talked about SkillSafe,
00:59:06the history, and above all about you or me doing something for
00:59:13ourselves or for others, but not yet with others. And that's something we both
00:59:20have already done, that we said, I want to maintain such a SkillSafe with others.
00:59:26And for that I used, for example, the SkillSafe workshop, which
00:59:32at least for me answered the question, okay, I just said, in the family there are
00:59:41situations where you say, well, I have things I discuss with my partner, and
00:59:46no idea, we somehow have to make decisions on certain topics,
00:59:49but there we definitely don't want to involve Grandma, but for some other things we want
00:59:55maybe to keep Grandma informed too, and you have that kind of situation in a company context too,
01:00:00where you say, in a project team you work out something, and one thing is directed
01:00:08stakeholder communication, where you say, well, that one I have to, there I have to regularly
01:00:14somehow align expectations. But there's also undirected stakeholder communication,
01:00:19where you say, well, in general other colleagues may also be interested
01:00:24in finding out what the project is actually about. What are they doing right now? What
01:00:28learnings could they share with me, and want to, especially when it's about the AI field,
01:00:34there are a lot of people who are really interested in that. And for exactly that kind of thing, from my
01:00:40point of view you have quite a good solution, with which the SkillSafe workshop was created, because on the one hand
01:00:45we can basically have internal knowledge, for example in a SharePoint directory or another
01:00:52synchronized directory. And on the other hand, you've already said,
01:00:58versioned knowledge, so to say, okay, now we have this knowledge sitting in a directory.
01:01:05Greetings to all the department directories of this world, well, information goes to die.
01:01:09Now it would be very good if we could curate that and then make it available to others too,
01:01:16in a way where they know, okay, this is current, and that is, that is,
01:01:22the other is no longer current, that was last week's state. This knowledge
01:01:26is outdated now, though. And for that you've, you've created a way
01:01:31to automate it via Git and something like CI-CD, that is pipelines, workflows,
01:01:40to automate releases. Shall we talk about that briefly?
01:01:43Well, I think at this point I've tried to lead over a bit, because these changes, along the lines of ContextMe, I did say help me with this connection to the screenshots, but it actually helps me mainly with the workshop, because the SkillSafe created a skill with knowledge.
01:02:03And then you distributed it, the workshop says, okay, I'm not alone in the world.
01:02:09I have people, agents, whatever, who consume it together with me, but also maintain it.
01:02:16And that's where the topic comes in, damn it, I can't lock it away in that sense,
01:02:23instead I need it, or want it, jointly consumable, jointly maintainable.
01:02:30And the idea behind the SkillSafe workshop was to say, okay, if you don't lock it up,
01:02:37then you have additional challenges, something like temporal assignment.
01:02:43Yes, yes.
01:02:44Yes, that gets complicated, but temporal assignment, I mean, what does temporal
01:02:47assignment mean?
01:02:48It's the question of how you deal with newer information.
01:02:53So does only the new information apply, or does historical knowledge also need to be
01:02:58kept, like, what Mark has, he has an annual salary of two million, it has to
01:03:04be said out loud, yes. And now he's had since the first... But the whole
01:03:08thing yesterday, he got a salary cut. Exactly. And then it's already
01:03:16important to know over what period Mark had the two million, or whether
01:03:21it's really only to know that Mark is now down to 20,000. So, and
01:03:26those are the things. The second thing is that the knowledge isn't locked up and distributed in a SkillSafe, but is alive.
01:03:36Do you have a certain curation obligation? Is it all still consistent in itself, check routines you want to run.
01:03:48Above all when several people work on the system, so you don't end up running into the pre-dule,
01:03:54where, due to some error, branches of knowledge get cut off, and that the system also
01:04:01is able to restructure itself. So this workshop has the topic that
01:04:08you can also talk with it about how the branches look. In the system, what
01:04:15are the clusters the SkillWorkshop arranges itself by, and if you now say, okay,
01:04:21the classes have changed, then you can also ask the workshop why.
01:04:26The workshop can also export it, and then we're back at the
01:04:31SkillSafe.
01:04:32And the workshop has something where I thought, damn it, because that's
01:04:36the next problem when you work with several people: either you invite them
01:04:40all over for coffee and cake, to your computer, and avoid competing
01:04:44write access by letting them all at your keyboard, or you try
01:04:49to come up with something that helps, so that when people sit at distributed systems,
01:04:56they can still work on the system. But what do they have access to then? And there you
01:05:02very quickly end up with things like iCloud or Dropbox or OneDrive, or I also wanted a few
01:05:10other tools that aren't from Microsoft, so to speak, like the Rediff of iCloud.
01:05:15Like Transferner, although Wrenswehr has no live file access, it only has download...
01:05:21Google Drive.
01:05:23Exactly, and how do you deal with people being able to do that?
01:05:27There's another sound too, next.
01:05:28Super.
01:05:29Yes, thanks.
01:05:31Sovereignty comes first.
01:05:33And how do you manage it so that people, on their own devices, are able
01:05:41to consume the knowledge, to maintain the knowledge. Not because it's a backup of files,
01:05:47but because it's the representation of connected knowledge in it. And also approaches
01:05:54to sync conflicts, written at the same time with locks, so that you prevent writing at the same time
01:06:01in this value, so you give the system something that, when the
01:06:06agents do the maintenance, because you're still interacting with the agent,
01:06:09you tell your Claude and your Gemini, maintain this here, here's the budget for it, go, go, go, go.
01:06:15That the system says, I can't right now, because another agent is on it, I'll check something else first,
01:06:21I'll restructure it so that you, I don't know, the physical limit of simultaneous humans is,
01:06:31but that you get a bit of order into it, and that you also have the chance here.
01:06:35To put it bluntly, to have a collaboration platform for cheap money,
01:06:40where you can know, edit together and share, because for that money, that's fun, you
01:06:45can also think in the process about how to do it well.
01:06:47Yes, but the interesting thing I found, and there I wondered how much
01:06:56of that you've actually built in?
01:06:58You have a few basic rules in SkillSafe workshop, along the lines of, that
01:07:02you have to agree on as a group working together on the vault, so that's something like a
01:07:08maintainer list, and what I found totally fascinating, again a question for the
01:07:18listeners: who of you has ever worked with decision records, that is, basically
01:07:24decision templates, and I was constantly confronted with those as SkillSafe
01:07:31workshop imported them, basically. We have, in OneDrive, in our case we have
01:07:38a Vault-Inbox directory, and we said, okay, everyone just throws their stuff in there,
01:07:43and then you can have the SkillSafe workshop import it regularly. So
01:07:49we throw PowerPoints in there, we throw meeting transcripts in there.
01:07:54The SkillSafe workshop imported the stuff and then said, in the meeting you
01:07:59by the way made the following decisions.
01:08:01Oh, wait, only one of you was in that meeting, and by the way the following
01:08:05things were discussed.
01:08:06There's again a contradiction to the decision you made last week in that and that
01:08:10meeting.
01:08:11And now you have to please carry out a conflict resolution here.
01:08:14I expect votes from each one of you, and I've set up the template here
01:08:19already.
01:08:20Please always use that for such decisions from now on, and now everybody
01:08:24vote, please.
01:08:25I've also set up task management for everyone in the vault, and if now, well,
01:08:32I have no idea, I do that, then it works out, that's completely absurd, it looks
01:08:37it up, when I say, what are my tasks, it looks up, who is my, who's actually talking
01:08:42to me here, hello, who are you, it then looks at the machine name somehow,
01:08:46some cryptic corporate computer, then next it looks at, ah,
01:08:52Ah, there, he has a Git directory in his home directory.
01:08:56Ah, look, there's a name and an email address in it.
01:08:59Look at you.
01:09:03I know, Schiere, I know where you live and where your car is parked.
01:09:06Yes, yes, yes, of course.
01:09:08I found that super fascinating, and then next up is Mr. Gagau, and he said,
01:09:12ah, wait a moment, in the meeting you talked about
01:09:15this project, where two of you are working on it,
01:09:18there's also this Jira project. I'm syncing the tasks that influence this other vault, the ones the group works on together, and I'm bringing them over here, because the others in the group should weigh in too and should know about it, so I'm putting the information in here, but keeping the reference to Jira.
01:09:40And so, Jira is a task management system from the company.
01:09:45I think anyone who's listened this far will get Jira sorted out.
01:09:49Okay.
01:09:50Because I think we've probably already lost some of the listeners,
01:09:53but the topic, I mean, the longer we talk about it and the more small examples and anecdotes we give,
01:10:00they basically show one thing, right:
01:10:06it's not that Mark uploaded a cool GitHub project, but that the systems are able
01:10:12to do many things you wouldn't have trusted them with.
01:10:16And when you introduced that.
01:10:20That's not so cool for me.
01:10:22The thing looks at it.
01:10:24What needs do we have as a group, and it makes suggestions about them?
01:10:28I hope what I wanted to get at, before I get to the sentence,
01:10:33is that it's basically an intrinsic enrichment for every further use.
01:10:37So, the more SkillSafes and knowledge spaces and workshops you have,
01:10:43or whatever you call them, the more intelligently, let's say, they can support you.
01:10:49Because at some point these systems know more and more about what matters to you,
01:10:52what matters to the group, what matters to the project, what they pay attention to.
01:10:56And that way you also manage not to be the smartest person in the room
01:11:00to set up such a thing, but the thing is basically the sum of all the decisions
01:11:06you've made in the past. Probably not all of them, and some things also get lost there,
01:11:09but it complements itself. And I already have the feeling
01:11:14that, like back in the day with "let me google that for you", there were people who
01:11:18entered a search term and got hits, while others thought of the same search term
01:11:21but never found anything, so that at some point, on the one hand,
01:11:25you have a knack for it, and on the other hand so does the machine. A knack for you,
01:11:28your needs, your concerns, whether it's you as an individual person or you
01:11:32in professional contexts. But what I also wanted to get at in this context
01:11:37and I find this is also something that has nothing to do with Second Brain and workshop and
01:11:42so on, but it's a nice example for explaining it. I know someone
01:11:47who once said to me, Mark, you know, I don't want to quote it, because
01:11:52it really came out wrong, let's say, it drove me crazy. I sat
01:11:56for three months thinking, how do I build myself a second brain?
01:12:00And then you come along, well, me, after a weekend, and say, hey,
01:12:04I've got something, and it works, and it has somehow 80, 90 percent of his thoughts in it.
01:12:09And then he says, honestly, I'm making the months-long head, how do you do that?
01:12:13And there it is again, you know, I've got the knack, I said, blond.
01:12:17Yes, no, that too, but I think the art simply lies in you,
01:12:21when you don't know how things will turn out at that point,
01:12:24what might not have been thought of, or what you have to think about, that you approach this
01:12:30with, let's call it, a kind of grounded attitude, combined with, yes, I've now also been
01:12:37over 20 years in the corporate world, so, let's call it life experience, professional
01:12:43life experience, we don't want to talk about private life, simply professional
01:12:46life experience, you bring in professional life experience, you have a tool
01:12:50at hand that empowers you rather than hinders you, because it's also gotten used to, and you have
01:12:55not seen it.
01:12:56And then you try things out, and based on your experience you notice, okay, there one should
01:12:59maybe pay attention to version conflicts, there one should pay attention to binding commitments,
01:13:04one should pay attention to, to, to, and so on, and so you come closer to this solution and
01:13:07don't lose yourself in months of conceptual work.
01:13:12Exactly, because you do it, and that's what I have to say brought me the
01:13:17greatest added value. Those who may have stumbled across my blog already
01:13:22will know it's called Overthinker's Thoughts, and that was really my hobbyhorse. The
01:13:30whole time thinking about things, and seriously, no joke, five years ahead,
01:13:35thinking about what will all happen and what the influencing factors are, and with all
01:13:40that thinking not managing to take the first step. And that's why
01:13:45my plea from the beginning is also: install it, and create an environment in which
01:13:51I once said, reduced blast radius. So where you ask, what's the worst that can happen?
01:13:57Okay, I can handle that. Once you've reached that situation, then you have
01:14:03a safe space, an area created for you, where you can run small experiments,
01:14:11where you can use such a thing and simply try it out.
01:14:15And honestly, the moment seems to come, greetings to my colleagues, where you
01:14:21feel overwhelmed, where you see that Alex made this thing,
01:14:26which is really great.
01:14:27And now let's say, my Copilot should show me the things I
01:14:35have to do, and Copilot buries you in 80 decision records, where it says, yes,
01:14:43by the way, Alex has imported your meeting minutes from the last three months from the
01:14:48whole projects and initiatives you're running, and by the way, these are all contradictory
01:14:53things and decisions you wanted to make. And sometimes the reaction of
01:15:00the colleagues is just to close it again and ask whether I've
01:15:03been drinking varnish. So I'm now running to Copilot's rescue, well, you can do all of that with Copilot too. What's nice about the non-Copilot things,
01:15:18classic Copilot, that we can use the skills, not because we have a SKILL.md file for it, a maximum of 13, sorry, the teasing,
01:15:30but that we can have real skills in the deterministic, not the generative data, you didn't see that, everything finds a home.
01:15:38But if you take something away from this, look at the blog and try the skills out.
01:15:47The skills don't break anything for you, except that they pull the rug out from under you a bit,
01:15:54namely understanding what I actually have hidden in your data. I said you can use it for small things, for 70 Apple talks or 28 gigabytes of data.
01:16:04And I think.
01:16:06As you also said, we're not the experts here, we're just trying things out and seeing what comes of it.
01:16:12This exploratory thing, I tinker with it and make it usable, and if it's good I talk about it, and if it isn't, I explain that it isn't.
01:16:22That helps. That's a hard thing to carry. But you report it, you pass it on right away.
01:16:27One tip I definitely want to share, because I've
01:16:32already had feedback on it several times: it really was a game changer, you
01:16:36didn't have it. It's not only the command lines, the CLIs of Claude, of Codex, etc.,
01:16:46that they can now see each other's sessions. So if I have one chat window
01:16:52open, the other chat window can see it. That means, if you say, you've
01:16:59tried something and you're stuck and say, damn it, I'm
01:17:05not getting anywhere, open a new window and say to Claude or ChatGPT, hey, I
01:17:12have a separate session open with you and I'm not getting anywhere.
01:17:16Please look at this and give me tips.
01:17:19Tell me what I can do.
01:17:21How do I get further?
01:17:22That's my goal.
01:17:23So you can step up to a meta level at any point and talk with another instance,
01:17:30so to speak, with another window of Claude or of ChatGPT,
01:17:33about how to get further in that original window.
01:17:38And that's a real advantage.
01:17:41I'd sometimes wish for that with people: you step out of your body, take Holger out of your body, talk to him about it, and the problem is putting him back in.
01:17:57What's wrong with Holger? Why doesn't he understand me?
01:18:03I've got much more trouble engaging with my new project, but unfortunately it's already late, and no matter how long we've been doing this, we've set a new record.
01:18:16I'm pretty sure that if we keep going like this, I'd like to quote a few rain.
01:18:22He had a child the first time, twins the second time, and a third time he didn't want to risk it, because he didn't know how the Wattmarsch rule goes.
01:18:30goes on. At this point, I'll definitely invite you again. For everyone who's listened so far,
01:18:36of course, we'll put it all in the show notes, so maybe also a
01:18:42small note there. All the episodes we have are also published on GitHub as a
01:18:48transcript, things like that. If someone says, oh, they talked for so long, I switched off,
01:18:53or I'll ask the system. I fell asleep, which,
01:18:56at what point did they talk about what, and will you tell me that?
01:18:59I always find that very exciting too, for us, when Jens
01:19:06then goes over the podcast episodes again and looks at what
01:19:10we talked about in there.
01:19:11It's always nice to have the material on hand for a guest too.
01:19:16On that note, a second brain won't run out on us. We may still have
01:19:21more second brains, and then the term second brain probably won't be
01:19:24quite right anymore.
01:19:25I once had a colleague who always called his dog Unudor Stresquattro and so on.
01:19:32But I don't know what his dog is called today, or whether he still kept to that rule.
01:19:36But I digress. Next time, you were here.
01:19:39Yes, maybe.
01:19:41Thanks for being here.
01:19:42And greetings to Jens too, because I think we haven't greeted him yet.
01:19:46But maybe next time with Jens again.
01:19:49Yes, I'm looking forward to next time with Jens too.
01:19:52We'll see, maybe I'll want to, sure, Jens, then... Oh no, I actually recorded with him then. Well,
01:19:56I have it in my calendar, my second brain knows about it too,
01:20:00maybe it's in the eight houses and unread emails. And at this
01:20:04point, if you liked the episode, leave a thumbs up and a like, if
01:20:07you didn't like it, please write us a comment, because then too
01:20:12we know you've listened until now, and wish you a pleasant
01:20:17night, a pleasant day with lots of people, especially those for whom a
01:20:22second brain would also have done others good. And with that, until next time. Ciao.
01:20:28Thanks. Ciao.
01:20:32Welcome to Think Different. Think AI., the podcast by Mark and Jens.
01:20:38Two tech-loving minds who don't just talk about artificial intelligence, but live it.
01:20:44Here you get clear classifications, real practical insights and a fresh look at what's possible.
01:20:51Down to earth, critical, and always with a wink.
01:20:55Something to think about, to smile about, and above all to join in on.