The self-authoring wiki, beating brain fry, and Obsidian as memory is a trap

0:05Andrew:So Andrew, are you as excited as I am about this news that Google is going to start punishing websites that hijack your back button?

0:14Ben:I mean, for me, it's a little too late. It's like, I'm not going to the websites anymore. My agents are. So did it take my agents complaining for them to finally do something about the back button not working on bad websites?

0:28Andrew:Oh, yeah. So you're saying it's Google's token costs were getting too high because their agent's back button is getting hijacked. So now they're going to take action on it.

0:37Ben:Or maybe I didn't take the thesis that far, but now that you've said it, that's very compelling that these agents were getting caught in these loops. But I do love the movement that Google is, you know, continuing to fight spam. Especially during a time when, like, bad practices on the web are at an all-time high in terms of spam and vibe-coded websites that don't act like you maybe expect. And so any kind of pulse check from Google of like, hey, we still care about spam on the Internet is or, you know, bad Internet website practices is a win in my book. What do you think about it?

1:08Andrew:Yeah, I mean, it makes me just wonder, like, how many other things out there do Google see that are like, wow, it's our agents are having a really hard time consuming websites that do this. So we should go punish those websites. Like they're probably like the only company in the space that actually has the authority and power to do that. Right. They truly are.

1:28Ben:They have a unique moat in terms of, you know, kind of owning how people were consuming the web in the first place.

1:35Andrew:Yeah. Yeah. Well, awesome. Welcome to the Friday Deploy from Linear B and DevInterrupted. I'm your host, Ben Lloyd Pearson. And I'm your host, Andrew Ziegler. And this week we're covering Google's offline AI breakthrough, agent swarms for data teams, Obsidian isn't AI memory, or is it when we talk about Carpathia in his self-writing wiki? And lastly, we'll close out with the AI brain fry epidemic. Andrew, let's just start right at the top with this new Google Gemma news. What do we have here?

2:05Ben:Yes. So continuing the story we've been covering about small language models, open source models, alternatives to the foundation models like Anthropic and Claude. This is an article about highlighting Gemma 4's ability to run natively on the iPhone, which is something that we have called out here on the show before when we first talked about it after its unveiling about two weeks ago. So it's a great reminder about all the different variants of this model and how they're optimized for mobile devices, devices on the edge, devices and Internet as bad as mine. And this represents like a major shift towards Internet in places where Internet or rather AI in places where Internet connectivity is spotty.

2:43Ben:There's like a real danger in the world of a lot of populations getting, frankly, left behind in the AI revolution. And personally, I see Gemma as a really great step towards making AI more accessible to the rest of the world.

2:56Andrew:Yeah, a lot of a lot of people and use cases, I think. And I think it's it's interesting to know that I think it feels like Google is really trying to position this as something for developers and power users to treat as like a foundation for for future capabilities rather than like a feature that they're rolling out to users, which totally makes sense to me. I mean, the average AI user isn't going to know like which tasks are appropriate to hand off to a local model, nor would they even know how to do that in the first place often. You know, and like you said, we've been covering a lot of these stories of how local models are becoming more efficient, higher quality, easier to delegate subagent tasks to.

3:34Andrew:And I really do think this type of stuff is going to open up a lot of efficiency gains, but it's also going to open up a lot of new use cases where you need either offline AI completely or some sort of edge AI capability where it's just too expensive to go back to a central service for API services versus just trying to do it locally. So, yeah, you know, a lot of companies are competing in this space more and more. It's really exciting to see Google being a part of it as well, particularly because, you know, as we covered in our opening, they have a lot of power in their incumbency that lets them, you know, push this stuff out in a way that other companies may not be able to do.

4:16Andrew:So, yeah, I think this is we'll definitely be following this a lot more because I think local models are increasingly going to become like the story of the rest of this year, maybe so.

4:26Ben:I completely agree. It's an opportunity for everyone to get involved and to reduce their costs. Like we even covered the Shopify story, I think like two weeks ago. How they not only cut their inference costs from OpenAI by 75 times, but they also got the multi-agent architecture then for free, I mean cheaper even, out of the box because they were able to leverage these really well -tuned models. So lost a couple in this story. Don't be sleeping on small language models. They are here to stay.

4:55Andrew:All right, let's move on from Google. We've given them enough attention. And let's talk about Watertown, the agent swarm data stack. So what do we have here, Andrew?

5:04Ben:Yes, this is a really fun article from Jordan Tagani. It introduces Watertown. And Jordan, he works at Mother Duck. It's a database service. It's largely used for AI inference, vector stores, but it's also just, you know, a database. And he works with large amounts of data. In this framework, it takes Yege's Gastown idea and it maps it into eight stage progressions for data teams that are using AI automation. Starting from the very beginning of like automating SQL assistants to having cell peeling pipelines around querying and manipulating data. And it uses structured communication in really smart ways.

5:42Ben:Borrowing lots of techniques that Steve has used to create his software factory. Now they have a data analysis factory. is a really great glimpse at how people can adapt to this kind of methodology for their own use cases.

5:56Andrew:Yeah, there was a line in this article that like really stood out to me and made me chuckle a little bit. And that is agents are like violence. The only solution to the problems they cause is to use more of them. And now I may not have used violence as the example, but that line resonates with me extremely well, you know, because I feel like the challenges that we solve with AI only create additional challenges that can also be solved with AI. And hopefully those new challenges are bigger and better challenges rather than just like dealing with the toil and the slop that AI, like the poorly optimized AI can create.

6:33Andrew:So yeah, there's this eight step maturity model for agentic use within data pipelines that is outlined in this article. And I really think this is an exercise that we should be applying to most of knowledge work today. So, you know, we do a ton of content production here at Dev Interrupted. And I haven't really thought about a system for like leveling where we are agentically within that. You know, we are definitely progressing deeper and deeper into agentic workflows. But I think what these models, what makes these models particularly important is that they give you this maturity roadmap that shows you where you should be focusing next.

7:08Andrew:So you can see where you are now, but then also what the next step looks like. The types of things that you would need to build to achieve that step. There's also another really great metaphor in this article, borrowing again from Stevie Ige's ideas. It's based on the book Master and Commander, I think, or something like that. It's about basically how an agentic data pipeline operates sort of like a ship at sea. So you have a log of observations. You have orders that agents receive to do something. There are flags that show when there's human feedback that's needed. There's captains. There's other roles within this ship that help maintain and build data pipelines.

7:48Andrew:There's also this metaphor about a scribe, which I also particularly love because I've used that exact metaphor myself in the past when building agentic content pipelines. So, yeah, you know, Andrew, I think you and I are both firmly in the camp of metaphors as being a really great way of conveying meaning to AI. And it can be really powerful, like particularly if you can just really tie everything to that central metaphor. So, you know, I love seeing these concepts getting applied to new fields. And I think, you know, like other stories we're covering today, we're just going to see more and more of this as time goes on.

8:21Ben:So the idea of agents needing to get brought in to solve problems that agents create, that's so true. And it just, it really speaks actually to not the whack-a-mole situation of using them as a solution, but actually rather just like the next order scale that you end up working on. And what it makes me think of is like, if you're going to go build a huge highway bridge over like a bay or a river, you're probably not going to use like hammers and nails and, you know, things by head. You're probably going to use power tools and trucks and like massive cranes and equipment to do it. And you know what?

8:56When you bring those in, that is then going to require more heavy machinery, more specialized people, more logistics. fixed. And you might think like, oh, gosh, like, why couldn't they have just built it with, you know, hammer and nails? And we all know the answer to that to build like the sturdy thing that's going to be resonating and staying in the future. We have to use the tool smartly. And that means acknowledging that way. Sometimes you got to put tools on tools. But I'm strongly with you here on the idea of like Gastown is coming to lots of different disciplines. I think this is a great example of using it to create a semi-autonomous data working system.

9:34And particularly what was really interesting to me is how it can flag these states for users and for humans to come in, almost like the human in the loop, but more like a human tool call that the agents can do to kind of get more information. I think we're going to start seeing this pattern more. I first learned about this, I think it was last week when I was at HumanX and I hosted a panel. One of the panelists, Angela McNeil of Thread AI, they're the team that created Lemma. She made this really smart point about exactly that, that humans are evolving from being in the loop to being on the loop, or they are the loop.

10:11Ben:They're the tool call itself. It's a really fascinating way of looking at it and it's really great to see it make its way into this example. I love how people are adapting Gastown and I hope to see more of these.

10:21Andrew:All right, let's move on to our next one. Stop calling it memory. The problem with every AI plus Obsidian tutorial. And I want to cover this one because I feel like this is almost like a direct attack on me as somebody who recently became an Obsidian convert because of agentic workflows. But in this article, the author argues that Obsidian Markdown files as an AI memory, in quotes, is a fundamentally flawed architecture that doesn't scale. So, you know, while Obsidian works really well for like personal notes. It fails as a data store, according to the author, because you can't query, filter or handle relationships of that data at scale.

11:01Andrew:And in this article, the author argues that engineering teams need to use actual databases like SQLite for structured data instead of flat files that get dumped into context windows. You should use proper infrastructure like Kuzu, OpenBrain, Supabase, you know, all the tools that help build these agentic systems rather than just sort of hacking it with flat files in Obsidian. And, you know, like I said, I feel a little bit attacked by this, but I also am not going to argue that a flat file system is with like Markdown and JSON is the most scalable system for when you need to do a whole bunch of agentic works.

11:38Andrew:I'm not going to make that claim. But for personal use, you know, I feel like, or even for a small team, it is really often the quickest and easiest way to start building better context for your AI. And it's funny, Andrew, because I feel like you and I are sort of taking opposite directions on this often. Like, you've gone very cloud-centric where all of your agentic workflows are triggered with stuff in the clouds that you can do it when you're like on your phone in the cab or working out in the morning or something. Versus me, I'm going like pure local only with everything. Like everything is now just a markdown or a JSON document that gets saved locally so I can just feed it into claw desktop or something like that.

12:20Andrew:So what did you think about this article, Andrew? As a heavy Obsidian user, I think that they really hit the nail on the head on how it can be used and what Obsidian really means for AI. Because Obsidian existed, obviously, before AI came into the scene, note-taking in this kind of way, locally on your machine with markdown files, with tags and links to each other in like a semi-wiki way.

12:43Ben:This is a really established way of building your own personal knowledge, management system, something that I've been a big fan of even before AI. I really love how this calls out the cargo cult thinking and like the AI influencer space thinking that like, oh, wow, when OpenClaw came out, you know, they just had this like memory.markdown file. And this was the secret sauce. This was why it was able to like have this persistence to its memory and its capabilities. But when you look closer, that's not actually the truth. Like one of the things being that not long after OpenClaw hit the scene and became viral, they quietly bolted on SQLite because of this exact reason.

13:23You need the ability to query and fetch only what you need. If you just dump a file into your context window, that's just brute force. It's not retrieval. There's no schema. No matter how beautiful your front matter is on your markdown, there's no joins or traversals. And your ability to scale that, it actually comes a lot faster than you think. and the wiki links and stuff that link between them are great for human discovery, but they are not queryable. You can, of course, create sorts of tools and stuff for an LLM to query it, but then, you know, I'm just going to call it out. Then you're just writing a database.

14:01And so you should then just use a database. Now, all of this said, there are very powerful ways to use Obsidian as a memory tool. One of them being that if you use, for example, a AI-focused plugin, there are a few. There's like a co-pilot plugin for Obsidian. You connect it with a subscriber token that you have. It could be like from your favorite foundation model provider. And the key here is taking those notes and embedding them. These tools can actually take your notes and turn them into embedded semantic lookup in like a local, basically like a vector database for your LLM to use. This is the connector that you really need to make it more like a memory store.

14:42However, I will say as someone who's used Obsidian every single day of my life for probably the last four years, I don't use Obsidian at all in any combination with agents. Like you called out, like I work in the cloud. I like to work in an ability where I can access everything from everywhere, talk to my phone or my terminal from any kind of place. And that's what fits me. Like right now I'm sitting in a hotel room. I'm about to go spend the whole day at a conference. Yesterday I was in a keynote and I shipped two things. It's like the ability to work that way is only because I've taken the opposite things.

15:16I don't want to be locked in my local machine. The one other thing I will call out here is that like the unlock of taking things that matter to you and distilling them into a place where you can collect them over time and you and the agent can access them is absolutely critical. And why I don't use Obsidian for this,

15:35Ben:I certainly have tools and mechanisms for doing that. So this will get you like 50 % of the way there. Just be mindful of how retrieval actually works. And I don't just think that it's working under the hood. You'd be surprised at how much you lean into the hallucinations of your agent when you think it's sitting on top of all your thoughts.

15:54Andrew:Yeah, you know, in your approach, I'm not going to, you know, try to sugarcoat this at all. Your approach is way more scalable than mine. If you're like talking about scaling it out to a team or to your organization or putting it in production or something like that. But I also don't really view what I'm doing with Obsidian as building memory when I record all of my artifacts into it. Instead, I view it as more of a record of historical context, right? So instead of losing all that information to the ether as I used to do, it all now gets captured in a single location that I can push to a Git repo.

16:28Andrew:I can feed directly into Claw Desktop. I can copy and paste files into whatever GPT service I might want to use instead. And, you know, I think we don't really need to be focused as much on the tech that we use to solve this, especially in this day and age. Like it's really easy to use your agents to just migrate to new tech, particularly if it's something that you're just doing for your personal use. So, you know, instead we should focus more on the process behind it all. So if Obsidian is the thing that solves it for you today, then you should use Obsidian because it's a great tool. But if you need to be more scalable, this article has some really great insights on how you can take it beyond just flat file architecture.

17:07Andrew:But I would actually bet that most people right now are not yet at a state where they need that level of scale. And again, the most important thing that you can do right now is start these practices rather than focusing on which tools you're going to adopt. Bingo. Yeah. With all that said, I'm sticking with flat files on my local machine. For now, I was not convinced to change from this. They just work so well for me and I can move really quickly and it's great. But I also understand why this article exists and why some people out there might have issues with it as well. All right, well, let's move on from an article that says Obsidian is the wrong answer to all this to the opposite, an article that argues or that makes an argument for how you can use Obsidian in really great ways for this type of work.

17:52Andrew:And so this article focuses on some recent learnings from Andre Carpate, who recently shifted from using LLMs for coding to building self-maintaining knowledge bases. So he's spending more time manipulating tokens for knowledge rather than manipulating code itself. And this article does a really great job outlining a system that uses a two-tier architecture with raw sources in an AI compiled wiki that synthesizes and cross links information automatically. So talking about some of the relational challenges that we were just discussing. And this approach is a little different from traditional note-taking practices because AI is doing all of the synthesis work and basically just removing a lot of the cognitive friction around understanding of things.

18:43Andrew:For listeners that don't know, Andrea Karpathy was one of the founding figures of the modern AI era. and coins the term vibe coding. And this article kind of breaks down a lot of the practices that he's been following recently. It really resonates with how I feel. Like, you know, I don't do a whole lot of code writing these days, but I do a lot of content, which is, has a lot of similar challenges. And for the last four months, I feel like most of my time at Dev Interrupted and Linear B has been really just focused on the things that I want to convey rather than how I convey those things. The AI does all of the structuring and turning it into clear and actionable information, but I'm the one that's there just giving guidance along the whole way and just making sure that the AI stays on the right track and has all the information it needs to make the right decisions.

19:32Andrew:So what'd you think about this article, Andrew?