Jorge Zaccaro · 张豪

LLM Knowledge Bases Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. In this way, a large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge (stored as markdown and images). The latest LLMs are quite good at it. So: Data ingest: I index source documents (articles, papers, repos, datasets, images, etc.) into a raw/ directory, then I use an LLM to incrementally "compile" a wiki, which is just a collection of .md files in a directory structure. The wiki includes summaries of all the data in raw/, backlinks, and then it categorizes data into concepts, writes articles for them, and links them all. To convert web articles into .md files I like to use the Obsidian Web Clipper extension, and then I also use a hotkey to download all the related images to local so that my LLM can easily reference them. IDE: I use Obsidian as the IDE "frontend" where I can view the raw data, the the compiled wiki, and the derived visualizations. Important to note that the LLM writes and maintains all of the data of the wiki, I rarely touch it directly. I've played with a few Obsidian plugins to render and view data in other ways (e.g. Marp for slides). Q&A: Where things get interesting is that once your wiki is big enough (e.g. mine on some recent research is ~100 articles and ~400K words), you can ask your LLM agent all kinds of complex questions against the wiki, and it will go off, research the answers, etc. I thought I had to reach for fancy RAG, but the LLM has been pretty good about auto-maintaining index files and brief summaries of all the documents and it reads all the important related data fairly easily at this ~small scale. Output: Instead of getting answers in text/terminal, I like to have it render markdown files for me, or slide shows (Marp format), or matplotlib images, all of which I then view again in Obsidian. You can imagine many other visual output formats depending on the query. Often, I end up "filing" the outputs back into the wiki to enhance it for further queries. So my own explorations and queries always "add up" in the knowledge base. Linting: I've run some LLM "health checks" over the wiki to e.g. find inconsistent data, impute missing data (with web searchers), find interesting connections for new article candidates, etc., to incrementally clean up the wiki and enhance its overall data integrity. The LLMs are quite good at suggesting further questions to ask and look into. Extra tools: I find myself developing additional tools to process the data, e.g. I vibe coded a small and naive search engine over the wiki, which I both use directly (in a web ui), but more often I want to hand it off to an LLM via CLI as a tool for larger queries. Further explorations: As the repo grows, the natural desire is to also think about synthetic data generation + finetuning to have your LLM "know" the data in its weights instead of just context windows. TLDR: raw data from a given number of sources is collected, then compiled by an LLM into a .md wiki, then operated on by various CLIs by the LLM to do Q&A and to incrementally enhance the wiki, and all of it viewable in Obsidian. You rarely ever write or edit the wiki manually, it's the domain of the LLM. I think there is room here for an incredible new product instead of a hacky collection of scripts.

来たよ。

一个人怎么能和汉字熟成这样，好想知道正在学中文的外国朋友看完是什么心情？

"The buying of more books than one can read is nothing less than the soul reaching toward infinity." A. Edward Newton

Beijing taxi drivers are famous for chatting up passengers about pretty much anything—especially international politics. On my way to the airport, my driver got really curious and started talking. “I think the Americans were just trying to pull the same playbook they used in Venezuela on Iran—figuring if they took out the leader, the whole country would just give in.” “Do you think that actually worked?” “No way, definitely not. The U.S. just wants to control Iran and install someone who’ll do whatever they’re told. But Israel? They want to wipe Iran out once and for all to feel ‘secure’ in the region. Their goals were never aligned to begin with. Trump’s pretty simple, but Bibi was evil. I reckon Trump got played and must be regretting it now.” Then I stopped at a little street restaurant for breakfast, and right nearby two older ladies were chatting about the U.S. midterms. “If Trump loses this fight in Iran, he’s going to tank in the midterms,” one of them said. “If he can’t keep oil prices down and the American economy starts hurting, he’s done for.” Her friend nodded. “Exactly. That’s why China pushing so hard on electric vehicles and cutting energy dependence is such a smart move.” It’s always fascinating hearing what everyday people think about global affairs.

Prototyping UIs has always been a good fit for vibe coding, because code quality matters less than when shipping to prod. But with the latest models, things have gotten kinda ridiculous… Opus 4.6 Fast can ask me 50 interview questions about a spec in rapid succession. That process reaches such clarity that it can then one-shot a big feature roughly aligned with the vision in my head, at an adequate quality level to feel out the concept and share the idea. Further iterations happen in seconds. Sometimes the integration tests for thousands of LOC pass on the first try which makes me chuckle—that’s not human level performance! In the past few days I’ve made two prototypes in a large codebase at work. Each one took a few hours from the initial seed of the idea to working demo, in total flow the entire time. I predict they would have taken days without AI (partially due to my unfamiliarity with the large codebase). In fact, without AI I would have chosen a different medium at this early stage. With modern tools, I find prototyping in prod code is often the fastest way for me to feel something out. Surprisingly the upfront interview is one of the most valuable parts — it feels amazing to have design decisions and judgment pulled out of me, without needing to stumble into the questions as I build; it feels like having a super sharp dev at the project kickoff. The faster model also promotes single-tasking focus which I love. For creative prototyping work (where figuring out what to build is the goal), I’m not a big fan of slow models and parallel multitasking; flow matters. Overall, production engineering has a ways to go with LLMs, but it feels like this problem of “UI prototyping assistance” is close to solved. The main bottleneck is my own decision making and judgment. While my main feeling is one of tremendous excitement and relief that I can validate the ideas in my head so quickly now, I do always worry a bit about the unintended consequences of such dramatic process change. Prototyping is a delicate art of working with a material and having reactions to it. There are no shortcuts; spending time is necessary to have good ideas. So I’m trying to keep an eye on that: what are the moments in my personal prototyping process that matter and must be preserved, and what are the parts that can be fast-forwarded? Tentatively things feel OK to me—using the draft UI and reacting to it is where the magic happens, I think, and the faster I can iterate on that UI the faster I can build intuition, without getting stuck in the mud of broken code. But it’s hard to know for sure, and as things speed up further I expect I may need to add more speed bumps to the process to ensure the same level of depth.

"AI code review" should be about teaching people how the code works, not just verifying that it's correct We can now generate essays, diagrams, and explorable explanations on demand!! Why are we still reading raw code diffs as the primary UI?

秒懂了哈哈

A lot of critiques of vibe coding actually don't cut deep enough, imo Sure, "the AI wrote bad code" may be true...but that is changing fast! The deeper issue is that when you're designing a complex system, you need good mental models of how it works in order to invent great ideas for the next iteration. Until now, if you built it by hand, that would guarantee some amount of understanding. But now the two have been decoupled, so we need to avoid the temptation of skipping understanding Reminiscent of the debate around whether memory matters in the modern era. It turns out you can't "just Google facts" -- you actually need to have ideas represented in your head to work with them fluently. The good news is AI can actually help us develop understanding faster and better, if used correctly! So the takeaway isn't to stop using AI for coding. It's to find ways to use it to develop understanding and move faster