A few things to point out after reading and thinking about this:
- Another AI firm building products focused on Fortune 500 scale problems. If you're not at a F500, this tool isn't necessarily a good fit for you, so YMMV.
- static analysis tools that produce flowcharts and diagrams like this have existed since antiquity, and I'm not seeing any new real innovation other than "letting the LLM produce it".
They say it's ZDR, so maybe I don't fully understand what problem they're trying to solve, but in general I don't see the value add for a system like this. Also onboarding isn't necessarily just presenting flow charts and diagrams: one of the biggest things you can do to onboard somebody is level-set and provide them with problem context. You COULD go into a 30 minute diatribe about how "this is the X service, which talks to the Y service, and ..." and cover a whiteboard in a sprawling design diagram, or you could just explain to them "this is the problem we're working on", using simple, compact analogies where/when applicable. If the codebase is primarily boilerplate patterns, like CRUD, MVC, or Router/Controller/Service/DB, why talk about them? Focus on the deviant patterns your team uses. Focus on the constraints your team faces, and how you take the unbeaten path to navigate those constraints.
> static analysis tools that produce flowcharts and diagrams like this have existed since antiquity, and I'm not seeing any new real innovation other than "letting the LLM produce it".
Inherent limitation of static analysis-only visualization tools is lack of flexibility/judgement on what should and should not be surfaced in the final visualization.
The produced visualizations look like machine code themselves. Advantage of having LLMs produce code visualizations is the judgement/common sense on the resolution things should be presented at, so they are intuitive and useful.
Although I haven't personally experienced the feeling of "produced visualizations looking like machine code", I can appreciate the argument you're making wrt judgment-based resolution scaling.
> static analysis tools that produce flowcharts and diagrams like this have existed since antiquity
Today I am apparently on of xkcd's "Lucky 10,000".
Does anyone have any recommendations for such tools? Ideally open source, but that's not a hard requirement. (Although "Enterprise - if you have to ask the price you can't afford it" options will not work for me.)
I'm particularity interested tools that work with Python, Java, and Javascript (Angular flavoured Javascript, it if matters)?
We are building https://noesis.vision/ a similar tool to extract the system architecture from the source code according to the patterns. We are now in beta for .NET.
After working with the topic for multiple months I can tell you that introductions for new-joiners are not the only use case for this kind of extracted knowledge. Many ppl in the organizations need insights into the software structure as they either impact decisions shaping this structure (e.g. analysts) or depend on the decisions about this structure (e.g. testers, or support agents)
It's all the matter of giving access to reliable architecture knowledge structured by a consistent ontology. Garbage in / garbage out - the higher the knowledge quality, the better the output - both for human and agentic knowledge consumers.
https://www.ensoftcorp.com/products/atlas is the Java/c oriented flavor I'm most familiar with. I've used them for Javascript before previously, although I'd have to do some digging to find the particular package I used. I am confident that you could find one with an npm/pypi search.
I'm currently building one based on tree-sitter: https://github.com/CRJFisher/code-charter. It's still in development but will be released soon (this year). Will support Python, JS/TS, Rust to begin with others (like Java) to follow.
It's been a few years, but I remember jetbrains IDEs had these, though I'm not totally sure it was built-in rather than a plugin. Personally I find automated diagrams are not that useful. You generally need to have some understanding of what's happening to know what to hide and collapse to tame the spaghetti. So it needs someone in the know to be explanatory, and as a tool it requires a similar amount of fiddling to just drawing some boxes in slideware.
I wanted to try this out, so I opened Windsurf for the first time in ages and clicked the "Upgrade Available" button, which sent me to: https://windsurf.com/editor/update-linux
Did you install using apt or apt-get? If so...
1. Update package lists
sudo apt-get update
2. Upgrade Windsurf
sudo apt-get upgrade windsurf
Whle `apt-get upgrade windsurf` will technically upgrade Windsurf, instructing users to run a command that will attempt to upgrade all packages on their system is nuts when the command is provided in a context that strongly implies it will only upgrade Windsurf and has no warnings or footnotes to the contrary. Good thing I didn't ask Windsurf's agent to ugprade itself for me, I guess.
EDIT - I don't want to detract from the topic at hand, however - after upgrading (with `sudo apt-get install --only-upgrade windsurf` :)) and playing around a bit, the Codemaps feature indeed seems very nifty and worth checking out. Good job!
So `apt-get upgrade $PACKAGE` has ridiculous semantics that no one would expect, and the actual syntax for upgrading a package is in neither the man page nor the command help.
I have been using Debian for literally decades and I didn't even know "apt-get upgrade $PACKAGE" existed. It is weird, it doesn't appear in the documentation, it doesn't work with the "apt" command, it means it is probably a relic of the past left there for compatibility reasons and you probably shouldn't use it.
My guess is that someone or some LLM hallucinated this command, "apt-get upgrade" is for upgrading your system, not for upgrading a single package, and it takes no extra argument.
For upgrading a single package, just do "apt install $PACKAGE". It is the same command as for installing. The semantics is rather clear to me, upgrading is like installing the new version on top of the old version. It also makes no sense to install a package you already have or to upgrade a package you don't have, but if you want to be sure, for example because you don't know if you already have the package installed or not, there are the --no-upgrade and --only-upgrade options.
--no-upgrade
Do not upgrade packages; when used in conjunction with install, no-upgrade will prevent packages on the command line from being upgraded if they are already installed. Configuration Item: APT::Get::Upgrade.
The canonical way to do the thing you want via apt-get is `apt-get install`. And if you read the man page from start to finish, it'd be clear to you... but it is tucked away there in the most obtuse, indirect, ungreppable way. :'D
That would be a great addendum to an EXAMPLES section! In the meantime, this is documented well and clearly in the tldr page for apt-get².
Fwiw, apt-get not only sucks, but has been known to suck for many, many years (more than a decade at least). Its interface sticks around because it's basically plumbing at this point. But you, as a user, should never use it (or `apt-cache` or `apt-*`, if you can avoid it.
Aptitude is preferable for a whole host of reasons, not least of which being that its upgrade commands have the semantics you'd intuitively expect³. They take packages as an optional list of positional args, and upgrade everything only if you don't pass any. (Aptitude also has a ton of other nice features and I highly recommend it.)
There's also an official new porcelain in APT itself, aptly called "apt". It preserves⁴ the semantics of apt-get's `upgrade` command, but its usage message actually matches that syntactically— hopefully it'll barf if you tell it `apt upgrade windsurf` or whatever.
But automation needs to rely on the ugly, old, disparate APT commands that have been around forever and can't really change. That probably goes, too, for things guides want you to copy and paste, or instructions handed over to LLMs.
(This is one reason that if you only learn to use APT from guides/tutorials whose primary concern is something other than documenting or teaching how to use Debian-based systems, you'll probably never learn to use the correct tools (the nicer, newer ones).)
apt-get has a more stable interface and is more suitable for scripts and instructions intended to be followed to the letter.
apt is better for interactive use and by people who are not just blindly following instructions.
Here there are arguments for both. As commands intended to be copy-pasted in a terminal, using apt-get makes sense as it is the safest choice. But it is also intended for humans, it is not a script, so maybe apt would be better. To me, both ways make sense.
I really love this comment, it's got a very "tree-falling-in-the-woods" vibe to it.
On the direct face of it, no, it turns out it doesn't matter: plant cellulose is not toxic to humans, a certain level of it is in many processed foods, and that information isn't secret.
By the time it matters to people, it's at the level where you can tell it's happened: large, pointy chunks, eg, or so much the flavour or texture is ruined. Or toxic contaminants, albeit at the significant risk that one might only be able to tell at the point of suffering from the consequences.
But if we modify the proposition a little, we get a statement about the possibility of a vegan's metaphorical sawdust being cut with ground beef. Now, it's more likely to matter. By and large, dietary choices like that are based on some belief structure, so the presence of the unwanted ingredient could be considered as an attack on the belief system.
When we move the metaphor back to AI generated code, does this reveal a belief system at play? If the resulting program is not poor quality, but the use of AI is objectionable nevertheless, does that make a "no AI in software" stance a sort of veganism for code? (And can we coin a good term for that stance? I vote for hominism, because while I quite like anthropism that leads to anthropic which is obviously not going to work.)
Given there's a regulatory number on acceptable bug parts per million for confectionary, is there a hypothetical acceptable bytes per million for AI-generated code that can still be called hoministic?
The HN guidelines explicitly ask you to steel man arguments you reply to. It is obvious that the point of the comment is not sawdust specifically;
they could have used anything else, like cyanide, and the point would stand. Spending multiple paragraphs of rebuttal on a nitpick which fails to address the crux of the argument is precisely the kind of bad argument the HN guidelines aim to avoid.
Seems like you haven’t understood my comment, but I’m unsure how to clarify it for you. Perhaps start by not assuming that expressing disagreement means taking offence? Not everything needs to be emotionally charged. Again, steel man.
Just checked again to give you the benefit of the doubt, and I still see the same thing. I read the long post as a thoroughly steelmanned response. Nobody has yet engaged with the philosophical content of that post. You cried foul for reasons I still can't understand. Would you tell us what you thought about the post on an intellectual level?
I eat meat but I'm one of those people who is ethically opposed to consuming AI content. An AI-vegan you might say.
I've had a shouting fight with someone who tried to spoon feed an AI summary to me in a regular human conversation.
But. I know that people are going to sneak AI content into what I consume even if I do everything within my power to avoid it.
The question is straightforward if immensely complex. Do I have a right to not be fed AI content? Is that even a practical goal? What if I can't tell?
> I read the long post as a thoroughly steelmanned response.
Steel manning means engaging with the strongest interpretation of the argument. The original comment clearly used sawdust not as sawdust specifically but as a substitute for something harmful or inappropriate. It’s not even about eating. So spending half a comment on “ackchyually, sawdust is good for you” (this is a caricature for brevity) is nitpicking something which doesn’t matter and derails the rest of the comment which is based on it.
Steel manning would’ve meant engaging in good faith, understanding “eating sawdust” isn’t meant literally but as a random choice for “something bad”, and replying to the latter, not the former.
In other words (I’m explaining it three times to drive the point home), steel manning means not nitpicking the exact words of someone’s argument but making the effort to respond to their meaning. It’s addressing the spirit of the comment above its letter (https://en.wikipedia.org/wiki/Letter_and_spirit_of_the_law). Sometimes the difference between those isn’t obvious, but I’m arguing that in this case it is.
> I eat meat but I'm one of those people who is ethically opposed to consuming AI content.
Eating meat or being vegan has nothing to do with the original comment. Again, it’s not even about eating, that was clearly a random example which could be substituted by a myriad other things. When you describe your eating habits you’re already engaging with a derailed, straw manned version of the argument instead of the original point the person was making.
I do apologise if my response came across as deliberately nitpicking on the specific item; my intent was to highlight that there are many cases where things we might broadly find unpalatable actually do happen all the time, with no harm except to our belief structures; from that perspective, sawdust or any other non-toxic contaminant in food is a pretty good analogy for AI in content, because in very small dilutions the only possible harm it can carry is to a belief structure.
On the flip side, it does seem to me like you have deliberately chosen the worst possible interpretation of what I wrote, so ... pot, kettle?
I really think more people should give Windsurf a go. It's really good. I'm a senior engineer and do a mix of agentic and regular coding and I really think people are looking past Windsurf.
As the conversation shifted towards Cursor vs Claude code vs Codex people seem to have stopped mentioning it which is a shame.
Source: user for 12 months - not a shill.
Codemaps was a very pleasant surprise when it showed up.
I co-sign this as a similarly-credentialed person. I use windsurf at work and recently started enjoying Claude Code, but the UX of Windsurf is actually a legit value add. Codemaps especially - been using them for weeks and they're excellent. Ask me again in a year maybe; churn in code could make maintaining codemaps annoying, but even that seems solvable.
appreciate the feedback! just a reminder codemaps are based on snapshots of your code when you run them; technically there's nothing to maintain, because you just rerun them if you need to.
Yep it's low friction, but is it easy to discover that I need to? I guess it's the "needing to know that I need to rerun" that I'm less enthused about.
you missed the point. Zed develops and pioneers ACP (Agent Client Protocol), which I can also use in other editors and with other agents. at the moment, only Neovim is available as an alternative editor, but nothing stops, say, JetBrains from implementing it. I can plug Codex, Gemini, Claude Code, and Goose directly into my editor of choice.
This is enough for me to give this a go. I've tried a few different tools; abacus.ai (and their IDE), claude CLI, crush-cli. My workflows are still mostly on the command line, and a little in VS Code. I haven't found a flow that works "right", yet.
first mention i've heard of abacus.ai and IDE. what do you think stands out about them?
you might struggle with Windsurf since you're so command line heavy. but pro tip - ask for command line work to be done inside of Windsurf's Cascade agent. they were first to the terminal-inside-aichat pattern and i really like how it's much better at command line stuff than i am (or can do the legwork to specify command line commands based on a few english descriptions)
> first mention i've heard of abacus.ai and IDE. what do you think stands out about them?
Their reasoning agent is better than anything else I've used, tbh. The inability to use it in a CLI environment is why I stopped using it. They have a router that they hook into that "intelligently" chooses models for you in a normal "chat" setting. The power comes with their DeepThink (or whatever) mode that has a VM hooked up to it, as well as many, many well designed agents and internal prompts that handle all sorts of interesting things, from planning to front-end dev, to reasoning about requirements and requirements fulfillment.
I'm surprised that people still aren't discussing Github Copilot baked into VSCode. I pair agent mode + Sonnet 4 + Sequential Thinking + Tavily MCP servers and it works wonders. I recently prototyped the first version of our SaaS with this setup in a minimal amount of time. Also worth nothing, the pricing is extremely reasonable. Free credits + pay per use. I frequently max out the free tier and have never spent more than $40 per month.
I second this. As a previous AI skeptic. VS Code with Agent Mode is amazing. Perfect amount of control over what the AI is doing. For me its been a game-changer. I would describe it as letting the AI do lots of the work and me being the guiding hand.
I will say, it is extremely important to have a good AGENTS.md file and other .md files that the agent can refer to when it needs to. Also having tests is more important than ever.
And when you notice common hiccups, document it in the AGENTS.md.
+1 this approach. One recurring issue I've ran into is the agents producing a lot of unused methods and files. Have you ran into this as well? We're working on a pure Typescript system, and found the package Knip helpful. Knip will report these deadfiles/methods. I ask Sonnet to run Knip before opening a merge request and have it clean up it's own mess.
Agreed. I'm back and forth about whether I want to spend the time with an agentic coding editor yet, because it's sitting right on the cusp of distraction/enhancement.
I've also tried the 3 C's, and it still feels like Windsurf has the net best user experience.
I as a big Windsurf advocate, miles ahead of Cursor IMO but I've fully switched to Codex these days. The cloud environments are just such a nice feature.
Still like Windsurf though their pricing is what drove me to not roll it out across my company.
I have. I tried Devin after the initial release and then again when they did their 2.0 release. Found it to be the worst of the tools I've tried.
More of a tangent but an underappreciated part of Codex is their PR review bot. Just miles ahead of all the competitors we've tried (greptile, charlie, cubic)
>2) how about the new claude code teleport feature from web to cli?
I have not. I rarely use Claude Code these days but I will give it a spin because you just told me it existed.
I just tried out windsurf yesterday, The only thing I hate for now is that when there are changes and I accept one of them, then trying to accept the others gives an error saying the file was changed
I've used it, and I thought it was absolute trash. Goes crazy doing shit I don't want. I spend more time deleting crap I didn't want and reviewing and changing its code than I do just writing it myself.
I know what you're going to say: I need to learn to use this groundbreaking technology that is so easy to use that my product manager will soon be doing my job but also is too hard for me a senior engineer, to find value in.
Kindly: no, I trust my judgement, and the data backs me up.
Have you taken measurements of how many features and bugs you've shipped over the last twelve months or are you just like the engineers in the METR study who self reported an improvement but when measured, had been impaired? What evidence do you have that your attitude is not simply informed by the sunk cost of your subscription?
Nothing my friends that heavily use AI for is groundbreaking at all. It's stuff they already entirely know how to do, describe in full detail what they want implemented, then double-check all of the results. I'm not convinced at all that they're doing architectural and long-term design thinking in this process. They're just "making the thing". I don't think they really care enough to do any of that hard thinking either. Not that they should be, considering the state of the industry and the lack of loyalty companies have to developers.
A feature like this isn't useful because knowing what connects to what, dependencies, etc. means nothing without business context. AI will never know the why behind the architecture, it will only take it at face value. I think technical design docs which have some context and reading the code is more than enough. This sits in the middle ground where it lacks the context of a doc and is less detailed than the code.
To add to that, a lot of business context is stuck in people‘s heads. To reach the level of a human engineer, the coding agent would have to autonomously reach out and ask them directed questions.
> AI will never know the why behind the architecture…
That's true only if you don't provide that context. The answer is: Do provide that context. My experience is that LLM output will be influenced and improved by the why's you provide.
it takes longer to explain the context to the model than it does to just write the code based on the context I already understand, especially since code is more terse than natural language
Definitely, iff you have to provide the context with every task. If agent memory worked better and across your whole team, then providing context might be much easier
agree that AI can kinda infer business context sometimes. in my experience, it doesn't work that well.
a lot of the time, debugging isn't a logic issue, but more of a state exploration issue. hence you need to add logging to see the inputs of what's going on, just seeing a control flow isn't super useful. maybe codemaps could simulate some inputs in a flow which would be super cool, but probably quite hard to do.
> AI will never know the why behind the architecture, it will only take it at face value.
There is no reason to believe that at some point in the future AI will know the business context for apps when they are vibecoded (because the prompts should contain the business context).
Well, interesting idea, but can you trust that it generates it properly? Because if it doesn't, then your understanding of the code will be incorrect, even worse than lack of knowledge; and if you do need to check all the things it has generated for you, as a description - doesn't it defeat the purpose of the tool?
Arguably, if you vibe code you don't really care about the code, thus even less about the diagram. So you'll vibe something, get a diagram of something to show your boss, and you can move on.
Do you have any examples/demos for very large codebases? Because it's easy to get decent quality output using LLMs on small codebases. But large, messy codebases is where you really need help with understanding.
Less a question and more a strong suggestion: codemaps should be viewable in the main pain. The sidebar is FAR too small. Either default or a button or something to open it like an editor tab.
Right but you're an IDE and I want to view code (and especially the links) in the IDE.
Hopefully we can get codemaps in the main IDE panel sooner rather than later. Feels like the very impressive thing (codemaps) is being held back by a trivial thing (reading them in a 200px panel) to the point it's impractical to use themm
I created a side project ~3 years ago based on a similar idea.
It was before LLMs were a big thing, and AI could render the code relationships for you.
I started with go and java (the two languages I was using in my job) and built AST parsers that would extract the relationships between the classes and functions and draw them with graphviz. Then I created a filter syntax based on regex to filter the graphs.
I never followed through on the idea to a production ready version, but it helped massively as a personal tool when working on unfamiliar codebases.
The site is still here, but lots of it is probably broken by now..
I did something similar but for non-classes based language (Go) and in 3D [1]
But I saw it as next step towards shifting programming from sitting and scanning texts into something more tangible, where developer has broad overview of software, and can work on smaller parts while seeing context of how these parts are connected. Ended up concluding that this stuff should work in VR.
I really like the idea of visualizing code in any other way than text and have given it some thought from time to time. However, I think the problem here can quickly become that you tend to fall in love with a bad idea. No one, and I don't mean that in bad faith, wants to look at these diagrams. Usually they do not communicate the meaning they intend to do, and I find that I have to spend some time understanding what exactly is meant by any type of box or arrow.
What people might want to look at is their mental visualization of the code or math they are working on (or their LLM made for them). At least to me, that is much more tied to what the data will look like at runtime and how different parts from different data-structures will interact with each other.
If you were to visualize a flutter app, and nowhere in that visualization the tree-like structure of the widget-tree would appear as such, that would collide with my mental model of how such an app functions.
This visualization will be induced by reading the code, much like reading a novel will produce pictures in your head. I'm not sure LLMs are the technology that will produce code-movies you would rather watch.
Feature request: add a Github Action so I can generate a codemap for my repo and throw it on my README. Then update it when major PRs change the codemap.
This was my conclusion, too! Over time as agentic coders get better at handling higher-complexity tasks, this kind of bracing will become less and less necessary.
Out of nowhere Cognition with a banging product. Probably not 100% yet but the idea is so good I'll be surprised if within 6 months all the other IDEs aren't copying.
Looks cool! I've been doing this a lot recently - a workflow I set up to create diagrams for code using the Claude cli and mermaid. It works pretty well but it's just a diagram - no links to the actual code. The latter is a neat addition that will likely get me to try codemaps.
we obviously agree but one of the problems i had with onboarding as a key proposition is that onboarding "seems" like a onetime problem. i lacked the datapoints or anecdotes to convincingly pitch "onboarding = context switching", a more recurring problem as the size of your team and size of your codebase and size of your tenure grows, even if its technically the same codebase, you're "always" onboarding to wahtever it is you're working on or maintaining or putting out fires.
Depends on how big the codebase is and how many people work on it and how often you need to switch to an unfamiliar context, but yes agreed, it's context switching, and is a regular thing. It is part of the job of software engineering, learning piles of code you were previously unfamiliar with.
Great idea. I always end up having to tag the relevant files/abstractions anyways to avoid having the LLM produce duplicated slop, and something like this makes collecting this info much easier.
This looks awesome. I’m a very heavy Claude Code user (and Codex) in both the CLI and VS Code (and now in the web too!) and it’s quite infuriating when the agent just gets lost after context compaction and I have to point it to read CLAUDE/AGENTS.md (and update it if a lot of changes have been made)
I tried Windsurf a while back but I’ll definitely come back ASAP just to play with this and see how it does in a somewhat complex project I’m working on.
Is this like an entire astro-turfing for AI thread? Can mods like ban these topics? This is clearly just like monetarily incentivized amoral behavior.
As far as I understand still working with this stuff regularly. None of the actual problems have yet to be solved at all. AI still produces garbage for anything complex. And if it doesn't, it's because you specified in full detail how it should do everything and heavily hand-held it and reviewed the results, taking more time than just doing it yourself.
And it's either that or they are flying by the seat of their pants with the thing and free-balling their way to a broken system.
A few things to point out after reading and thinking about this:
- Another AI firm building products focused on Fortune 500 scale problems. If you're not at a F500, this tool isn't necessarily a good fit for you, so YMMV.
- static analysis tools that produce flowcharts and diagrams like this have existed since antiquity, and I'm not seeing any new real innovation other than "letting the LLM produce it".
They say it's ZDR, so maybe I don't fully understand what problem they're trying to solve, but in general I don't see the value add for a system like this. Also onboarding isn't necessarily just presenting flow charts and diagrams: one of the biggest things you can do to onboard somebody is level-set and provide them with problem context. You COULD go into a 30 minute diatribe about how "this is the X service, which talks to the Y service, and ..." and cover a whiteboard in a sprawling design diagram, or you could just explain to them "this is the problem we're working on", using simple, compact analogies where/when applicable. If the codebase is primarily boilerplate patterns, like CRUD, MVC, or Router/Controller/Service/DB, why talk about them? Focus on the deviant patterns your team uses. Focus on the constraints your team faces, and how you take the unbeaten path to navigate those constraints.
> static analysis tools that produce flowcharts and diagrams like this have existed since antiquity, and I'm not seeing any new real innovation other than "letting the LLM produce it".
Inherent limitation of static analysis-only visualization tools is lack of flexibility/judgement on what should and should not be surfaced in the final visualization.
The produced visualizations look like machine code themselves. Advantage of having LLMs produce code visualizations is the judgement/common sense on the resolution things should be presented at, so they are intuitive and useful.
Although I haven't personally experienced the feeling of "produced visualizations looking like machine code", I can appreciate the argument you're making wrt judgment-based resolution scaling.
> static analysis tools that produce flowcharts and diagrams like this have existed since antiquity
Today I am apparently on of xkcd's "Lucky 10,000".
Does anyone have any recommendations for such tools? Ideally open source, but that's not a hard requirement. (Although "Enterprise - if you have to ask the price you can't afford it" options will not work for me.)
I'm particularity interested tools that work with Python, Java, and Javascript (Angular flavoured Javascript, it if matters)?
We are building https://noesis.vision/ a similar tool to extract the system architecture from the source code according to the patterns. We are now in beta for .NET.
After working with the topic for multiple months I can tell you that introductions for new-joiners are not the only use case for this kind of extracted knowledge. Many ppl in the organizations need insights into the software structure as they either impact decisions shaping this structure (e.g. analysts) or depend on the decisions about this structure (e.g. testers, or support agents)
It's all the matter of giving access to reliable architecture knowledge structured by a consistent ontology. Garbage in / garbage out - the higher the knowledge quality, the better the output - both for human and agentic knowledge consumers.
Love what you are trying to do here.
Will evaluate it on a couple of PHP codebases I maintain at work.
Is there a way to contribute to the project?
Thanks, please join our Discord here: https://discord.gg/QF5PMX4Dqg It would be easiest to discuss all the options there :)
https://www.ensoftcorp.com/products/atlas is the Java/c oriented flavor I'm most familiar with. I've used them for Javascript before previously, although I'd have to do some digging to find the particular package I used. I am confident that you could find one with an npm/pypi search.
I'm currently building one based on tree-sitter: https://github.com/CRJFisher/code-charter. It's still in development but will be released soon (this year). Will support Python, JS/TS, Rust to begin with others (like Java) to follow.
It's been a few years, but I remember jetbrains IDEs had these, though I'm not totally sure it was built-in rather than a plugin. Personally I find automated diagrams are not that useful. You generally need to have some understanding of what's happening to know what to hide and collapse to tame the spaghetti. So it needs someone in the know to be explanatory, and as a tool it requires a similar amount of fiddling to just drawing some boxes in slideware.
one of my earliest blogposts at Netlify https://www.netlify.com/blog/2018/08/23/how-to-easily-visual...
Sounds very cool.
I wanted to try this out, so I opened Windsurf for the first time in ages and clicked the "Upgrade Available" button, which sent me to: https://windsurf.com/editor/update-linux
Whle `apt-get upgrade windsurf` will technically upgrade Windsurf, instructing users to run a command that will attempt to upgrade all packages on their system is nuts when the command is provided in a context that strongly implies it will only upgrade Windsurf and has no warnings or footnotes to the contrary. Good thing I didn't ask Windsurf's agent to ugprade itself for me, I guess.EDIT - I don't want to detract from the topic at hand, however - after upgrading (with `sudo apt-get install --only-upgrade windsurf` :)) and playing around a bit, the Codemaps feature indeed seems very nifty and worth checking out. Good job!
So `apt-get upgrade $PACKAGE` has ridiculous semantics that no one would expect, and the actual syntax for upgrading a package is in neither the man page nor the command help.
I have been using Debian for literally decades and I didn't even know "apt-get upgrade $PACKAGE" existed. It is weird, it doesn't appear in the documentation, it doesn't work with the "apt" command, it means it is probably a relic of the past left there for compatibility reasons and you probably shouldn't use it.
My guess is that someone or some LLM hallucinated this command, "apt-get upgrade" is for upgrading your system, not for upgrading a single package, and it takes no extra argument.
For upgrading a single package, just do "apt install $PACKAGE". It is the same command as for installing. The semantics is rather clear to me, upgrading is like installing the new version on top of the old version. It also makes no sense to install a package you already have or to upgrade a package you don't have, but if you want to be sure, for example because you don't know if you already have the package installed or not, there are the --no-upgrade and --only-upgrade options.
> So `apt-get upgrade $PACKAGE` has ridiculous semantics that no one would expect
Especially not an LLM!
Sure it is¹ (kinda):
The canonical way to do the thing you want via apt-get is `apt-get install`. And if you read the man page from start to finish, it'd be clear to you... but it is tucked away there in the most obtuse, indirect, ungreppable way. :'DThat would be a great addendum to an EXAMPLES section! In the meantime, this is documented well and clearly in the tldr page for apt-get².
Fwiw, apt-get not only sucks, but has been known to suck for many, many years (more than a decade at least). Its interface sticks around because it's basically plumbing at this point. But you, as a user, should never use it (or `apt-cache` or `apt-*`, if you can avoid it.
Aptitude is preferable for a whole host of reasons, not least of which being that its upgrade commands have the semantics you'd intuitively expect³. They take packages as an optional list of positional args, and upgrade everything only if you don't pass any. (Aptitude also has a ton of other nice features and I highly recommend it.)
There's also an official new porcelain in APT itself, aptly called "apt". It preserves⁴ the semantics of apt-get's `upgrade` command, but its usage message actually matches that syntactically— hopefully it'll barf if you tell it `apt upgrade windsurf` or whatever.
But automation needs to rely on the ugly, old, disparate APT commands that have been around forever and can't really change. That probably goes, too, for things guides want you to copy and paste, or instructions handed over to LLMs.
(This is one reason that if you only learn to use APT from guides/tutorials whose primary concern is something other than documenting or teaching how to use Debian-based systems, you'll probably never learn to use the correct tools (the nicer, newer ones).)
--
1: https://manpages.debian.org/trixie/apt/apt-get.8.en.html
2: https://tldr.inbrowser.app/pages/linux/apt-get
3: https://manpages.debian.org/trixie/aptitude/aptitude.8.en.ht...
4: https://manpages.debian.org/trixie/apt/apt.8.en.html
hiya! team noticed your comment and agreed - and it is fixed.
Why not use apt?
apt-get has a more stable interface and is more suitable for scripts and instructions intended to be followed to the letter.
apt is better for interactive use and by people who are not just blindly following instructions.
Here there are arguments for both. As commands intended to be copy-pasted in a terminal, using apt-get makes sense as it is the safest choice. But it is also intended for humans, it is not a script, so maybe apt would be better. To me, both ways make sense.
Did you also generate this with “AI”?
https://tenor.com/view/westworld-if-you-cant-tell-does-it-ma...
My reading is that GP can tell, and they're trying to highlight it by asking a question.
If you couldn't tell your food had been cut with sawdust would it matter to you if you found out?
I really love this comment, it's got a very "tree-falling-in-the-woods" vibe to it.
On the direct face of it, no, it turns out it doesn't matter: plant cellulose is not toxic to humans, a certain level of it is in many processed foods, and that information isn't secret.
By the time it matters to people, it's at the level where you can tell it's happened: large, pointy chunks, eg, or so much the flavour or texture is ruined. Or toxic contaminants, albeit at the significant risk that one might only be able to tell at the point of suffering from the consequences.
But if we modify the proposition a little, we get a statement about the possibility of a vegan's metaphorical sawdust being cut with ground beef. Now, it's more likely to matter. By and large, dietary choices like that are based on some belief structure, so the presence of the unwanted ingredient could be considered as an attack on the belief system.
When we move the metaphor back to AI generated code, does this reveal a belief system at play? If the resulting program is not poor quality, but the use of AI is objectionable nevertheless, does that make a "no AI in software" stance a sort of veganism for code? (And can we coin a good term for that stance? I vote for hominism, because while I quite like anthropism that leads to anthropic which is obviously not going to work.)
Given there's a regulatory number on acceptable bug parts per million for confectionary, is there a hypothetical acceptable bytes per million for AI-generated code that can still be called hoministic?
The HN guidelines explicitly ask you to steel man arguments you reply to. It is obvious that the point of the comment is not sawdust specifically; they could have used anything else, like cyanide, and the point would stand. Spending multiple paragraphs of rebuttal on a nitpick which fails to address the crux of the argument is precisely the kind of bad argument the HN guidelines aim to avoid.
You read the same response I did, right? And you... thought it was... literally about sawdust? ...and you took offense? I'm so confused...
Seems like you haven’t understood my comment, but I’m unsure how to clarify it for you. Perhaps start by not assuming that expressing disagreement means taking offence? Not everything needs to be emotionally charged. Again, steel man.
Just checked again to give you the benefit of the doubt, and I still see the same thing. I read the long post as a thoroughly steelmanned response. Nobody has yet engaged with the philosophical content of that post. You cried foul for reasons I still can't understand. Would you tell us what you thought about the post on an intellectual level?
I eat meat but I'm one of those people who is ethically opposed to consuming AI content. An AI-vegan you might say.
I've had a shouting fight with someone who tried to spoon feed an AI summary to me in a regular human conversation.
But. I know that people are going to sneak AI content into what I consume even if I do everything within my power to avoid it.
The question is straightforward if immensely complex. Do I have a right to not be fed AI content? Is that even a practical goal? What if I can't tell?
> I read the long post as a thoroughly steelmanned response.
Steel manning means engaging with the strongest interpretation of the argument. The original comment clearly used sawdust not as sawdust specifically but as a substitute for something harmful or inappropriate. It’s not even about eating. So spending half a comment on “ackchyually, sawdust is good for you” (this is a caricature for brevity) is nitpicking something which doesn’t matter and derails the rest of the comment which is based on it.
Steel manning would’ve meant engaging in good faith, understanding “eating sawdust” isn’t meant literally but as a random choice for “something bad”, and replying to the latter, not the former.
In other words (I’m explaining it three times to drive the point home), steel manning means not nitpicking the exact words of someone’s argument but making the effort to respond to their meaning. It’s addressing the spirit of the comment above its letter (https://en.wikipedia.org/wiki/Letter_and_spirit_of_the_law). Sometimes the difference between those isn’t obvious, but I’m arguing that in this case it is.
> I eat meat but I'm one of those people who is ethically opposed to consuming AI content.
Eating meat or being vegan has nothing to do with the original comment. Again, it’s not even about eating, that was clearly a random example which could be substituted by a myriad other things. When you describe your eating habits you’re already engaging with a derailed, straw manned version of the argument instead of the original point the person was making.
I do apologise if my response came across as deliberately nitpicking on the specific item; my intent was to highlight that there are many cases where things we might broadly find unpalatable actually do happen all the time, with no harm except to our belief structures; from that perspective, sawdust or any other non-toxic contaminant in food is a pretty good analogy for AI in content, because in very small dilutions the only possible harm it can carry is to a belief structure.
On the flip side, it does seem to me like you have deliberately chosen the worst possible interpretation of what I wrote, so ... pot, kettle?
Is the result is molecular identical? If so, no.
Yes, it does matter.
Wow, what's the upside to that syntax? I never would have guessed.
I really think more people should give Windsurf a go. It's really good. I'm a senior engineer and do a mix of agentic and regular coding and I really think people are looking past Windsurf.
As the conversation shifted towards Cursor vs Claude code vs Codex people seem to have stopped mentioning it which is a shame.
Source: user for 12 months - not a shill.
Codemaps was a very pleasant surprise when it showed up.
I co-sign this as a similarly-credentialed person. I use windsurf at work and recently started enjoying Claude Code, but the UX of Windsurf is actually a legit value add. Codemaps especially - been using them for weeks and they're excellent. Ask me again in a year maybe; churn in code could make maintaining codemaps annoying, but even that seems solvable.
appreciate the feedback! just a reminder codemaps are based on snapshots of your code when you run them; technically there's nothing to maintain, because you just rerun them if you need to.
Yep it's low friction, but is it easy to discover that I need to? I guess it's the "needing to know that I need to rerun" that I'm less enthused about.
I prefer IDEs like Zed that don't lock me in to their ecosystem and force me to "log in" to use them.
`codeium` which is now `windsurf` started out as a vscode fork IIRC
Codeium started as VS Code extension (after their pivot). The whole Windsurf rebrand fork happened years later
Not to be confused with `vscodium` which is an open source build of vscode
Yeah definitely confused it with vscodium, thx
you missed the point. Zed develops and pioneers ACP (Agent Client Protocol), which I can also use in other editors and with other agents. at the moment, only Neovim is available as an alternative editor, but nothing stops, say, JetBrains from implementing it. I can plug Codex, Gemini, Claude Code, and Goose directly into my editor of choice.
This is enough for me to give this a go. I've tried a few different tools; abacus.ai (and their IDE), claude CLI, crush-cli. My workflows are still mostly on the command line, and a little in VS Code. I haven't found a flow that works "right", yet.
first mention i've heard of abacus.ai and IDE. what do you think stands out about them?
you might struggle with Windsurf since you're so command line heavy. but pro tip - ask for command line work to be done inside of Windsurf's Cascade agent. they were first to the terminal-inside-aichat pattern and i really like how it's much better at command line stuff than i am (or can do the legwork to specify command line commands based on a few english descriptions)
> first mention i've heard of abacus.ai and IDE. what do you think stands out about them?
Their reasoning agent is better than anything else I've used, tbh. The inability to use it in a CLI environment is why I stopped using it. They have a router that they hook into that "intelligently" chooses models for you in a normal "chat" setting. The power comes with their DeepThink (or whatever) mode that has a VM hooked up to it, as well as many, many well designed agents and internal prompts that handle all sorts of interesting things, from planning to front-end dev, to reasoning about requirements and requirements fulfillment.
ah yeah i have heard about their router. i wonder if GPT5 doing a router hurt it a bit.
I'm surprised that people still aren't discussing Github Copilot baked into VSCode. I pair agent mode + Sonnet 4 + Sequential Thinking + Tavily MCP servers and it works wonders. I recently prototyped the first version of our SaaS with this setup in a minimal amount of time. Also worth nothing, the pricing is extremely reasonable. Free credits + pay per use. I frequently max out the free tier and have never spent more than $40 per month.
I second this. As a previous AI skeptic. VS Code with Agent Mode is amazing. Perfect amount of control over what the AI is doing. For me its been a game-changer. I would describe it as letting the AI do lots of the work and me being the guiding hand.
I will say, it is extremely important to have a good AGENTS.md file and other .md files that the agent can refer to when it needs to. Also having tests is more important than ever.
And when you notice common hiccups, document it in the AGENTS.md.
+1 this approach. One recurring issue I've ran into is the agents producing a lot of unused methods and files. Have you ran into this as well? We're working on a pure Typescript system, and found the package Knip helpful. Knip will report these deadfiles/methods. I ask Sonnet to run Knip before opening a merge request and have it clean up it's own mess.
Agreed. I'm back and forth about whether I want to spend the time with an agentic coding editor yet, because it's sitting right on the cusp of distraction/enhancement.
I've also tried the 3 C's, and it still feels like Windsurf has the net best user experience.
I as a big Windsurf advocate, miles ahead of Cursor IMO but I've fully switched to Codex these days. The cloud environments are just such a nice feature.
Still like Windsurf though their pricing is what drove me to not roll it out across my company.
cloud envs are good :)
1) did you compare codex cloud with devin?
2) how about the new claude code teleport feature from web to cli?
just wanted to pry for more opinions on what matters to you
>1) did you compare codex cloud with devin?
I have. I tried Devin after the initial release and then again when they did their 2.0 release. Found it to be the worst of the tools I've tried.
More of a tangent but an underappreciated part of Codex is their PR review bot. Just miles ahead of all the competitors we've tried (greptile, charlie, cubic)
>2) how about the new claude code teleport feature from web to cli?
I have not. I rarely use Claude Code these days but I will give it a spin because you just told me it existed.
I just tried out windsurf yesterday, The only thing I hate for now is that when there are changes and I accept one of them, then trying to accept the others gives an error saying the file was changed
I’m sorry to hear about that. What version are you on? Looking to fix / repro this asap
(also you can hit "share conversation" or "view response statistics" and then "copy request id" and send to support!)
I'm on v 1.12.28
I've used it, and I thought it was absolute trash. Goes crazy doing shit I don't want. I spend more time deleting crap I didn't want and reviewing and changing its code than I do just writing it myself.
I know what you're going to say: I need to learn to use this groundbreaking technology that is so easy to use that my product manager will soon be doing my job but also is too hard for me a senior engineer, to find value in.
Kindly: no, I trust my judgement, and the data backs me up.
Have you taken measurements of how many features and bugs you've shipped over the last twelve months or are you just like the engineers in the METR study who self reported an improvement but when measured, had been impaired? What evidence do you have that your attitude is not simply informed by the sunk cost of your subscription?
Please share your data below
Wholeheartedly agree.
Nothing my friends that heavily use AI for is groundbreaking at all. It's stuff they already entirely know how to do, describe in full detail what they want implemented, then double-check all of the results. I'm not convinced at all that they're doing architectural and long-term design thinking in this process. They're just "making the thing". I don't think they really care enough to do any of that hard thinking either. Not that they should be, considering the state of the industry and the lack of loyalty companies have to developers.
[flagged]
https://knowyourmeme.com/memes/sir-this-is-a-wendys
Stop, I bruised a rib laughing at this
This isn't really a comment fit for Hacker News.
A feature like this isn't useful because knowing what connects to what, dependencies, etc. means nothing without business context. AI will never know the why behind the architecture, it will only take it at face value. I think technical design docs which have some context and reading the code is more than enough. This sits in the middle ground where it lacks the context of a doc and is less detailed than the code.
To add to that, a lot of business context is stuck in people‘s heads. To reach the level of a human engineer, the coding agent would have to autonomously reach out and ask them directed questions.
> AI will never know the why behind the architecture…
That's true only if you don't provide that context. The answer is: Do provide that context. My experience is that LLM output will be influenced and improved by the why's you provide.
if you know that context, you don't need a codemap
As you just said, codemaps don't include the "why" behind the architecture. That's context you need to add.
this is possible if you have a couple two-pizza teams. beyond that, good luck.
it takes longer to explain the context to the model than it does to just write the code based on the context I already understand, especially since code is more terse than natural language
Definitely, iff you have to provide the context with every task. If agent memory worked better and across your whole team, then providing context might be much easier
But wouldn’t the context also be useful, in written form, to colleagues?
It absolutely is, yes.
you might be surprised how much business context leaks into a codebase and that's plenty to work on :)
https://deepwiki.com/search/vimfnfnname-lets-you-call-neov_e...
but also how much you kinda dont need it when you're just debugging code
https://windsurf.com/codemaps/87532afd-092d-401d-aa3f-0121c7...
agree that AI can kinda infer business context sometimes. in my experience, it doesn't work that well.
a lot of the time, debugging isn't a logic issue, but more of a state exploration issue. hence you need to add logging to see the inputs of what's going on, just seeing a control flow isn't super useful. maybe codemaps could simulate some inputs in a flow which would be super cool, but probably quite hard to do.
[dead]
> AI will never know the why behind the architecture, it will only take it at face value.
There is no reason to believe that at some point in the future AI will know the business context for apps when they are vibecoded (because the prompts should contain the business context).
> because knowing what connects to what, dependencies, etc. means nothing without business context. AI will never know the why behind the architecture
since when has coding become so trivial that things are only useful if it helps with the "why" and "business context"?
Well, interesting idea, but can you trust that it generates it properly? Because if it doesn't, then your understanding of the code will be incorrect, even worse than lack of knowledge; and if you do need to check all the things it has generated for you, as a description - doesn't it defeat the purpose of the tool?
Arguably, if you vibe code you don't really care about the code, thus even less about the diagram. So you'll vibe something, get a diagram of something to show your boss, and you can move on.
Yeah but there are a lot of vibe engineers, and we care about the code, because we have to own it
this is the right way to try and tackle this problem imo. too much focus in AI dev tooling has been on building "products" that only half work.
making codebases understandable to humans, and LLMs etc, is a better approach
self documenting, interpretable systems would actually solve a lot of dev churn in big companies
plus it's not like artifacts have to be limited to code once that's figured out
I don't think it's a choice; I use both. Code understanding is especially useful in new code bases, but once that's over you need to get work done.
(coauthor) happy to take any questions! see 1 min demo video here https://x.com/cognition/status/1985755284527010167
this is brainchild of cognition cto steven who doesn't like the spotlight but he deserves it for this one https://x.com/stevenkplus1/status/1985767277376241827
if you leave qtns here he'll see it
Do you have any examples/demos for very large codebases? Because it's easy to get decent quality output using LLMs on small codebases. But large, messy codebases is where you really need help with understanding.
we dogfooded this with some v large customers i cant name before launch so .. yeah reasonably confident, but ofc no guarantees overall.
Less a question and more a strong suggestion: codemaps should be viewable in the main pain. The sidebar is FAR too small. Either default or a button or something to open it like an editor tab.
acked. you can also open it in a browser window for now https://windsurf.com/codemaps/9e2791c4-0b14-4757-b4be-a71488...
Right but you're an IDE and I want to view code (and especially the links) in the IDE.
Hopefully we can get codemaps in the main IDE panel sooner rather than later. Feels like the very impressive thing (codemaps) is being held back by a trivial thing (reading them in a 200px panel) to the point it's impractical to use themm
just cahtted internally
turns out our dev actually had a PR for it but wasnt sure it was valuable
you just helped tip it over the line. see it in next release
I created a side project ~3 years ago based on a similar idea. It was before LLMs were a big thing, and AI could render the code relationships for you.
I started with go and java (the two languages I was using in my job) and built AST parsers that would extract the relationships between the classes and functions and draw them with graphviz. Then I created a filter syntax based on regex to filter the graphs.
I never followed through on the idea to a production ready version, but it helped massively as a personal tool when working on unfamiliar codebases.
The site is still here, but lots of it is probably broken by now..
https://packagemap.co
I did something similar but for non-classes based language (Go) and in 3D [1]
But I saw it as next step towards shifting programming from sitting and scanning texts into something more tangible, where developer has broad overview of software, and can work on smaller parts while seeing context of how these parts are connected. Ended up concluding that this stuff should work in VR.
[1] https://divan.dev/posts/visual_programming_go/
for people too lazy to download windsurf to try it, codemaps is also in deepwiki
example: https://deepwiki.com/search/how-do-react-hooks-work-under_7a...
this does a pretty good job of going in the weeds of how the useState hook works in react
I really like the idea of visualizing code in any other way than text and have given it some thought from time to time. However, I think the problem here can quickly become that you tend to fall in love with a bad idea. No one, and I don't mean that in bad faith, wants to look at these diagrams. Usually they do not communicate the meaning they intend to do, and I find that I have to spend some time understanding what exactly is meant by any type of box or arrow. What people might want to look at is their mental visualization of the code or math they are working on (or their LLM made for them). At least to me, that is much more tied to what the data will look like at runtime and how different parts from different data-structures will interact with each other. If you were to visualize a flutter app, and nowhere in that visualization the tree-like structure of the widget-tree would appear as such, that would collide with my mental model of how such an app functions. This visualization will be induced by reading the code, much like reading a novel will produce pictures in your head. I'm not sure LLMs are the technology that will produce code-movies you would rather watch.
Feature request: add a Github Action so I can generate a codemap for my repo and throw it on my README. Then update it when major PRs change the codemap.
So it is the same thing when I ask Claude to build me mermaid charts of code flows? So no point in this tool?
Btw, claude code is a lot better at graphviz than mermaid! I have been using it a lot for architecture designs.
This was my conclusion, too! Over time as agentic coders get better at handling higher-complexity tasks, this kind of bracing will become less and less necessary.
Out of nowhere Cognition with a banging product. Probably not 100% yet but the idea is so good I'll be surprised if within 6 months all the other IDEs aren't copying.
hey we been shipping!!
https://cognition.ai/blog/swe-1-5
https://cognition.ai/blog/swe-grep
https://cognition.ai/blog/devin-agent-preview-sonnet-4-5
Actually another piece of feedback, the "waves" release notes/videos were pretty cool. Might want to consider bringing them back.
theyre not gone, just havent had a waves type thing to do in a while
Looks an interesting enough feature to give Windsurf a try!
Looks cool! I've been doing this a lot recently - a workflow I set up to create diagrams for code using the Claude cli and mermaid. It works pretty well but it's just a diagram - no links to the actual code. The latter is a neat addition that will likely get me to try codemaps.
Figuring out new codebases is definitely one of the most challenging and time-intensive things I have had to do in my jobs.
we obviously agree but one of the problems i had with onboarding as a key proposition is that onboarding "seems" like a onetime problem. i lacked the datapoints or anecdotes to convincingly pitch "onboarding = context switching", a more recurring problem as the size of your team and size of your codebase and size of your tenure grows, even if its technically the same codebase, you're "always" onboarding to wahtever it is you're working on or maintaining or putting out fires.
Depends on how big the codebase is and how many people work on it and how often you need to switch to an unfamiliar context, but yes agreed, it's context switching, and is a regular thing. It is part of the job of software engineering, learning piles of code you were previously unfamiliar with.
>whereas people get into trouble when the code they generate and maintain starts to outstrip their ability to understand it.
If you are trying to understand code, then you are not vibe coding by definition.
I feel like understanding the codebase should be the default thing the AIs would need to do first. This sounds promising though.
I really like this kind of applied statistical data infrastructure approach, feels much more natural than just raw text + immediate HIL
huh? whats statistical about it? i hope i didnt give the wrong impression in the post.
but yes, keeping the human in the loop, in charge, on top of the code, is the way to prevent ai slop code
This is extremely useful and helpful ! How great is this use case for AI !!!
So we're back to UML?
Great idea. I always end up having to tag the relevant files/abstractions anyways to avoid having the LLM produce duplicated slop, and something like this makes collecting this info much easier.
Aggressively ironic for a pitch deck that has no respect for the readers critical thinking skills
This looks awesome. I’m a very heavy Claude Code user (and Codex) in both the CLI and VS Code (and now in the web too!) and it’s quite infuriating when the agent just gets lost after context compaction and I have to point it to read CLAUDE/AGENTS.md (and update it if a lot of changes have been made)
I tried Windsurf a while back but I’ll definitely come back ASAP just to play with this and see how it does in a somewhat complex project I’m working on.
Kudos to the team!
this is amazing!
Is this like an entire astro-turfing for AI thread? Can mods like ban these topics? This is clearly just like monetarily incentivized amoral behavior.
As far as I understand still working with this stuff regularly. None of the actual problems have yet to be solved at all. AI still produces garbage for anything complex. And if it doesn't, it's because you specified in full detail how it should do everything and heavily hand-held it and reviewed the results, taking more time than just doing it yourself.
And it's either that or they are flying by the seat of their pants with the thing and free-balling their way to a broken system.