is this like Obsidian's graph view? Looks pretty/makes cool screenshots but has no actual value and is just cumbersome to use? (btw, this isn't meant to be a mean comment, just a question after looking at the output.)
What evidence is there that this makes any difference at all? There are a gazillion (and one) codebase understanding solutions using knowledge graphs. How do I know if it's any good compared to just using Codex or Claude Code?
Did anyone actually use this on a complex codebase and have any kind of intuition for it ?
Like, having looked at the demo, it feels less intuitive and extra complex than going through the codebase myself with tmux + codex + reading it myself.
I think for you to understand the codebase, it should be easier to interact without. This seems to introduce way too many steps to interact with the codebase
Interesting approach. I built something similar https://github.com/nilbuild/diffity to understand unknown codebases. The difference is it gives you the interactive walk-through with mermaid diagrams, guiding you through the feature or part of the codebase that you're looking at.
Are those 9.7k real users? I mean, maybe I am too old fashioned, but whenever I tried to use such tools long before AI, it actually didn't help much. It was much easier to read the codebase and find the needed connection on my own.
It reminds me on NX graphs, which are helpful to find the circular dependencies, but other than that, doesn't provide a lot of value as I can see the same kind of structure just looking at the codebase.
I don't quite get this stars and users connection. The stuff I use, I use, I don't need to star it. The reference is saved somewhere else. I bookmark, which is what the star is, stuff that seems interesting. So basically star for me is that I don't actually use the project, most of the time.
And yes, I glance at stars for a popularity cutoff, but forks, PRs, issues are much more telling.
For me, practical knowledge comes from trying to figure things out. The more polished and "ELI5" the material is, the less I retain. I've played with quite a few LLM tools that promised to help me "understand anything", but I don't think they help with intuition all that much. For what it's worth, it's not an LLM-specific problem. I like YouTube content like 3Blue1Brown, but I don't think that I retained anything useful from any of it.
I don't question that LLMs are useful for answering questions about codebases, but this is closer to "turn a codebase into a curriculum", and... does that actually work?
As I understand it, in teaching there's an idea of the "Zone of Proximal Development" (ZPD). Some things you can do without help, other things you can only do if others do it for you, and then in between there's all the stuff that you can do with some amount of assistance. Being in this zone is important for learning, at least in theory.
I suspect that's kind of happening here. If you're trying to learn something too abstract or distant from what you currently know, you'll probably use more polished or eli5-y sorts of material, because you don't yet have the skills to understand a more complex version. You're probably not in the ZPD. But if you can figure some things out by yourself, possibly with some amount of help, then you're in the learning zone and can meaningfully progress.
I have similar experiences to you with 3B1B - it's interesting, but I rarely retain anything meaningful after I've finished - and I think it's because he has to explain every part for me to understand what's going on. I'm not in the zone of proximal development because I can't do enough of the work myself. So the end result is an interesting video where someone explains a cool concept to me, but it's not learning because it's not also doing all the foundation work that gets me to the point where I can understand the video for myself.
You are the first one I know who said that. Thank you for saying that!
I think his videos are amazing but they are NOT meant to teach you the material.
They are providing a high level intuition which I haven’t found the use for yet.
Perhaps it’s just me, but I do NOT learn from intuition and analogies at all. I need to get lost in the details and rigor first, and then develop my own intuition second, and maybe look at someone else’s intuition third, maybe.
You're conflating two points: (1) learning math requires work, not just watching videos, and (2) learning math requires proofs, not just intuition and analogies.
Both are true but only the first is pertinent, because 3b1b videos typically contain complete mathematical arguments, not just intuition.
I don't agree. Many 3b1b videos provide a rigorous mathematical argument in the usual sense (i.e., they don't contain every conceivable detail — math papers don't do that either — but the gap from what is explained to rigorous-as-you-like is routine).
As a representative example, consider his video on the Basel problem. The video features geometric intuition, but it is intuition that drives a proof that is being presented. There's nothing hand-wavy about it.
I actually find some 3b1b content really useful. When I was learning linear algebra his course provided a visual representation for the fundamental operations and properties, which helped solidify and consolidate the formal stuff. The LLM videos are great too (ref https://xkcd.com/1838/).
To be fair, I think visual representations are particularly useful for reasoning about vectors and transformations, and I think I retain less from the videos that focus more on a specific advanced problem rather than the fundamentals of a topic. Those videos feel a bit more clickbaity too.
A big az graph with 100s of spagatti nodes is the kind of learning I try to avoid. Is better to just ask directly, "where do i start?", "teach me about...". This is over engineered education.
The phrase going around the interwebs is "You can outsource your thinking but not your understanding". A phrase that can at times seem like this weird human<>llm endless loop; depending on what you think you understand and what the llm "thinks" to help you understand, it can seem like an LLM also understand. But it does not.
Its clear one can't really think about anything without building a basic understanding about it. Worth stating that these are distinct from learning. But, I would argue that it is important to know what you *have* to understand now and why is that important. An LLM can help you understand a great many things, you just need to know what you are looking for and that is something no artificial intelligence can really *do* for you. Trial and error, building a sense of self awareness, and talking to people is a better way to know what this is especially for fairly open ended problems.
I was talking to a teacher and she was explaining how everyone is reaching for AI to have everything explained to them. "I'm too dumb to understand things," is the basic assumption people are now growing up with, reaching for AI summaries all the time without trying to understand anything themselves.
Instead of trying to understand things, people are reaching for better tools to have the thinking done for them. We are losing something huge.
I'm exhausted by these shiny vibe coded projects that overpromise and underdeliver.
Knowledge comes from doing the hard work, not from being spoon fed information. All these fancy graphs represent a tentative mental model produced as a result of research and learning. Everyone's model is different based on their own experience and focus, so trying to present it as a unique map will more than likely not be conducive to understanding at all. Besides the fact that it will almost certainly miss important details or be hallucinated.
HN users: stop upvoting and promoting this garbage. HN mods: please give us tools to label and filter this content.
I know I'm bandwagoning but just to make sure the signal beats the noise:
- This looks like vibe-code + fake github stars
- There's a difference between "Summarize this verbose report from my the PM so I can get the gist" and "ELI5 this complex subject so I feel like I understand it".
One thing I realized while working on large repos is that most “code graph” tools are still fundamentally navigation tools.
You can see structure, dependencies, call graphs, etc., but you still spend a lot of time manually building a mental model of why things exist and how concepts connect across the codebase.
What I’m trying to explore with Understand Anything is whether LLMs + structured graphs can help generate higher-level semantic understanding instead of only visualization.
For example:
1. tracing how a business concept propagates through services/modules
2. mapping requirements ↔ implementation ↔ data flow
3. surfacing architectural patterns automatically
4. helping new contributors build a mental model faster
Still very early obviously, but that’s the direction I’m interested in exploring.
They lost me at the first gif. Scrolling around a large graph that looks like mostly empty space.. seems you can make the same info in a compact screen-sized text with nested <ul>
32 comments
[ 3.0 ms ] story [ 62.4 ms ] threadLike, having looked at the demo, it feels less intuitive and extra complex than going through the codebase myself with tmux + codex + reading it myself. I think for you to understand the codebase, it should be easier to interact without. This seems to introduce way too many steps to interact with the codebase
It reminds me on NX graphs, which are helpful to find the circular dependencies, but other than that, doesn't provide a lot of value as I can see the same kind of structure just looking at the codebase.
Am I doing something wrong with these tools?
I don't quite get this stars and users connection. The stuff I use, I use, I don't need to star it. The reference is saved somewhere else. I bookmark, which is what the star is, stuff that seems interesting. So basically star for me is that I don't actually use the project, most of the time.
And yes, I glance at stars for a popularity cutoff, but forks, PRs, issues are much more telling.
I don't question that LLMs are useful for answering questions about codebases, but this is closer to "turn a codebase into a curriculum", and... does that actually work?
I suspect that's kind of happening here. If you're trying to learn something too abstract or distant from what you currently know, you'll probably use more polished or eli5-y sorts of material, because you don't yet have the skills to understand a more complex version. You're probably not in the ZPD. But if you can figure some things out by yourself, possibly with some amount of help, then you're in the learning zone and can meaningfully progress.
I have similar experiences to you with 3B1B - it's interesting, but I rarely retain anything meaningful after I've finished - and I think it's because he has to explain every part for me to understand what's going on. I'm not in the zone of proximal development because I can't do enough of the work myself. So the end result is an interesting video where someone explains a cool concept to me, but it's not learning because it's not also doing all the foundation work that gets me to the point where I can understand the video for myself.
https://youtu.be/dcolM6W5Odc?si=W-OL-ek5scYSvz2C
You are the first one I know who said that. Thank you for saying that!
I think his videos are amazing but they are NOT meant to teach you the material.
They are providing a high level intuition which I haven’t found the use for yet.
Perhaps it’s just me, but I do NOT learn from intuition and analogies at all. I need to get lost in the details and rigor first, and then develop my own intuition second, and maybe look at someone else’s intuition third, maybe.
Both are true but only the first is pertinent, because 3b1b videos typically contain complete mathematical arguments, not just intuition.
And they don't claim to either.
As a representative example, consider his video on the Basel problem. The video features geometric intuition, but it is intuition that drives a proof that is being presented. There's nothing hand-wavy about it.
To be fair, I think visual representations are particularly useful for reasoning about vectors and transformations, and I think I retain less from the videos that focus more on a specific advanced problem rather than the fundamentals of a topic. Those videos feel a bit more clickbaity too.
Its clear one can't really think about anything without building a basic understanding about it. Worth stating that these are distinct from learning. But, I would argue that it is important to know what you *have* to understand now and why is that important. An LLM can help you understand a great many things, you just need to know what you are looking for and that is something no artificial intelligence can really *do* for you. Trial and error, building a sense of self awareness, and talking to people is a better way to know what this is especially for fairly open ended problems.
Instead of trying to understand things, people are reaching for better tools to have the thinking done for them. We are losing something huge.
Knowledge comes from doing the hard work, not from being spoon fed information. All these fancy graphs represent a tentative mental model produced as a result of research and learning. Everyone's model is different based on their own experience and focus, so trying to present it as a unique map will more than likely not be conducive to understanding at all. Besides the fact that it will almost certainly miss important details or be hallucinated.
HN users: stop upvoting and promoting this garbage. HN mods: please give us tools to label and filter this content.
- This looks like vibe-code + fake github stars - There's a difference between "Summarize this verbose report from my the PM so I can get the gist" and "ELI5 this complex subject so I feel like I understand it".
Similar idea, but much simpler and focused!
You can see structure, dependencies, call graphs, etc., but you still spend a lot of time manually building a mental model of why things exist and how concepts connect across the codebase.
What I’m trying to explore with Understand Anything is whether LLMs + structured graphs can help generate higher-level semantic understanding instead of only visualization.
For example:
1. tracing how a business concept propagates through services/modules 2. mapping requirements ↔ implementation ↔ data flow 3. surfacing architectural patterns automatically 4. helping new contributors build a mental model faster
Still very early obviously, but that’s the direction I’m interested in exploring.