We previously launched Auto Wiki (https://news.ycombinator.com/item?id=38915999) in Jan 2024 and broke the ground for AI generated wikis that explain your code. Now this product has been rebuilt by the same team, as well as others and launched as a part of Google. Hope you enjoy.
Although, I've recently moved on to working on Gemini and AI research, I'm still involved as an advisor and founder emeritus of sorts. This team moves extremely fast and while we don't have full availability yet, we're working hard on addressing some early feedback before we make it more widely available including for private repos. Personally, I think the NotebookLM integration is a nice touch and distinguishing factor that we could only do as Google.
There is no data carried over from Mutable if that's what you're referring to. Please refer to the final email sent to users which I don't have access to.
If they choose to publish no-truth-value garbage about my life's work I will f**ing shred them with words.
This project and the whole philosophy behind it is just dripping with disrespect so you won't find me in line to be polite to the people who made it. If they're going to walk over and chunder their lukewarm slop onto me they can expect a verbal fistfight.
Ok so it was acquired and merged into that google offering? Does this mean we lost the open source nature and ability to preserve and protect our data?
I don’t understand how this can be more optimal than just reading the documentation that a human already wrote. All “normal” uses can be answered by reading the docs, everything advanced you can just read the code. I’m not sure when I would ever use this?
Reading a ton of docs and interpreting them is a tedious activity. If you can get deducible or inferable answers with AI, that's a huge win. I have faced issues with kubernetes that needed me to wade through the code for days to find it was a missed case or unsupported. AI would help me in minutes. That's the claim here, if it works that way.
I’ve never had AI understand a problem of that depth though. It can maybe surface the right part of the docs to reference but in the times I have used it it leads me down the wrong path half of the time.
Generating and maintaining docs is a massive cost. Presumably the point of this is to reduce that cost. And for projects where the existing documentation is poor or nonexistent, this might be far better than what’s available today.
How well this actually works, though, I have no idea.
Well "system that calls functions in a loop" doesn't sound like it is from a science fiction story. Obviously, that is no good if you are trying to sell science fiction fantasies to people.
Anectodally, I've found Gemini to provide useful, but made-up solutions to issues I've encountered with Google products. It showed reasonable understanding of my problem, but then gave me solutions which the actual documentation essentially made clear were not possible. YMMV
Seems great: I looked at two largish code bases I'm familiar with, and learned something each time.
But is this just a summary for the impatient, or can it reduce the effort for developers writing docs?
Docs have always been the mirror of code, and thus hard to get and keep right. Can we do without the mirror, or parts of it?
Does it work when you haven't written documentation for your code? Let's say one is fanatical about writing such clear code that names are sufficient to convey what's happening (i.e., no documentation and no comments). Does it work?
If not, does it work when there are only (clear) comments?
Does it tell you when documentation, comments, or code is unclear or missing?
I.e., I'd like it to go beyond summarization to fill easy gaps and point developers to the hard ones.
If this is being regenerated every commit, I’d be interested to see the version history and/or being able to see the CodeWiki diff inside a pull request.
Maybe it’s too noisy, if the LLM isn’t stable about the way it’s wording things, or maybe it’s only useful for commits that make significant changes to architecture. However, I do think it’d be interesting to see how the documentation changes over time, as well as seeing how any specific PR changes it.
Also, I looked at golang, and I was definitely expecting a multi-page architecture with lots of cross references, not just one long scrolling field of content.
Interesting. I've found Claude Code very useful for answering questions about codebases. I think that kind of functionality is on the whole more useful than static AI-generated documentation, but maybe there's a place for an always-ready and google-able starting point.
It badly needs to split up the pages for different parts of large codebases. The golang/go page is way too long and the table of contents sidebar makes you watch a scrolling animation to expand subsections.
I like the idea behind this and think it could be useful for agentic applications and reduce hallucinations. However, I usually just read the source code for the most up to date understanding and it's my best approach. After a quick look at CodeWiki, it could be useful to cross reference while analyzing the code and I look forward to testing it out next time it could apply.
Not sure how this is supposed to show. Does it only show stuff about existing repositories?
I put in the URL for a project of mine in Codeberg and didn't seem to be able to do anything - i expected that it'd go clone the repository, parse the code and attempt to tell stuff about it, but all i got was an error :-P.
The whole thing is automatically generated? Does anything persist? If I could be in the middle of reading it, and the next day it's completely different, that's a huge waste of my time.
Hello HN,
We previously launched Auto Wiki (https://news.ycombinator.com/item?id=38915999) in Jan 2024 and broke the ground for AI generated wikis that explain your code. Now this product has been rebuilt by the same team, as well as others and launched as a part of Google. Hope you enjoy.
Although, I've recently moved on to working on Gemini and AI research, I'm still involved as an advisor and founder emeritus of sorts. This team moves extremely fast and while we don't have full availability yet, we're working hard on addressing some early feedback before we make it more widely available including for private repos. Personally, I think the NotebookLM integration is a nice touch and distinguishing factor that we could only do as Google.
I hope you enjoy.
Thank You, Omar (Formerly Founder/CEO MutableAI)
How do I opt out?
There is no data carried over from Mutable if that's what you're referring to. Please refer to the final email sent to users which I don't have access to.
Can you email me at the email in my HN profile so I can clarify what I mean by opt-out in more detail?
Without snark, delete all instances of your code from github or similar places.
If they choose to publish no-truth-value garbage about my life's work I will f**ing shred them with words.
This project and the whole philosophy behind it is just dripping with disrespect so you won't find me in line to be polite to the people who made it. If they're going to walk over and chunder their lukewarm slop onto me they can expect a verbal fistfight.
I opt out
Ok so it was acquired and merged into that google offering? Does this mean we lost the open source nature and ability to preserve and protect our data?
So this isn’t a one time thing; you have a pattern of using the word “wiki” to describe products that have nothing to do with wikis?
Seems fitting that someone who doesn’t care about correct meaning would work on misinformation slop generators.
You’ve truly found your calling in life.
I don’t understand how this can be more optimal than just reading the documentation that a human already wrote. All “normal” uses can be answered by reading the docs, everything advanced you can just read the code. I’m not sure when I would ever use this?
Reading a ton of docs and interpreting them is a tedious activity. If you can get deducible or inferable answers with AI, that's a huge win. I have faced issues with kubernetes that needed me to wade through the code for days to find it was a missed case or unsupported. AI would help me in minutes. That's the claim here, if it works that way.
I’ve never had AI understand a problem of that depth though. It can maybe surface the right part of the docs to reference but in the times I have used it it leads me down the wrong path half of the time.
I encourage you to put your preference to the test on the NixOS and nixpkgs documentation.
Generating and maintaining docs is a massive cost. Presumably the point of this is to reduce that cost. And for projects where the existing documentation is poor or nonexistent, this might be far better than what’s available today.
How well this actually works, though, I have no idea.
This doesn't replace docs, nor technical writers.
Anyone else get an eye twitch when they read or hear "Agentic"?
Full on cringe shivers
aye, shiver me timbers, argh
Well "system that calls functions in a loop" doesn't sound like it is from a science fiction story. Obviously, that is no good if you are trying to sell science fiction fantasies to people.
what word do you prefer?
Six seven
Anectodally, I've found Gemini to provide useful, but made-up solutions to issues I've encountered with Google products. It showed reasonable understanding of my problem, but then gave me solutions which the actual documentation essentially made clear were not possible. YMMV
Will it be possible to document private repositories? And how can we prevent Google from using the code to train its AI?
Does anyone know of any trustworthy, usable alternatives? Perhaps even ones that run 100% on-premises?
[dead]
Seems great: I looked at two largish code bases I'm familiar with, and learned something each time.
But is this just a summary for the impatient, or can it reduce the effort for developers writing docs?
Docs have always been the mirror of code, and thus hard to get and keep right. Can we do without the mirror, or parts of it?
Does it work when you haven't written documentation for your code? Let's say one is fanatical about writing such clear code that names are sufficient to convey what's happening (i.e., no documentation and no comments). Does it work?
If not, does it work when there are only (clear) comments?
Does it tell you when documentation, comments, or code is unclear or missing?
I.e., I'd like it to go beyond summarization to fill easy gaps and point developers to the hard ones.
The marketing blurbs on point are not helpful.
Blog post: https://developers.googleblog.com/en/introducing-code-wiki-a...
When will it be deprecated?
Check out deepwiki.com, which is quite similar and works well
Thank you!
If this is being regenerated every commit, I’d be interested to see the version history and/or being able to see the CodeWiki diff inside a pull request.
Maybe it’s too noisy, if the LLM isn’t stable about the way it’s wording things, or maybe it’s only useful for commits that make significant changes to architecture. However, I do think it’d be interesting to see how the documentation changes over time, as well as seeing how any specific PR changes it.
Also, I looked at golang, and I was definitely expecting a multi-page architecture with lots of cross references, not just one long scrolling field of content.
Interesting. I've found Claude Code very useful for answering questions about codebases. I think that kind of functionality is on the whole more useful than static AI-generated documentation, but maybe there's a place for an always-ready and google-able starting point.
It badly needs to split up the pages for different parts of large codebases. The golang/go page is way too long and the table of contents sidebar makes you watch a scrolling animation to expand subsections.
I like the idea behind this and think it could be useful for agentic applications and reduce hallucinations. However, I usually just read the source code for the most up to date understanding and it's my best approach. After a quick look at CodeWiki, it could be useful to cross reference while analyzing the code and I look forward to testing it out next time it could apply.
I have been doing some reading about Kafka lately so decided to check the repo and it looks like a great reference: https://codewiki.google/github.com/apache/kafka
Mmmmh for this one, I'd say 1.5 before joining the cemetery.
Not sure how this is supposed to show. Does it only show stuff about existing repositories?
I put in the URL for a project of mine in Codeberg and didn't seem to be able to do anything - i expected that it'd go clone the repository, parse the code and attempt to tell stuff about it, but all i got was an error :-P.
Apparently gotta request a new repo. Kinda lame.
The whole thing is automatically generated? Does anything persist? If I could be in the middle of reading it, and the next day it's completely different, that's a huge waste of my time.
Good idea, low quality execution. The landing page looks awesome the wikis just a wall of text, no visual appealing.
cool idea, but I tried searching for ffmpeg, linux, llvm-project - how does it not have the top 10 open source projects in it?
How will this purely AI generated content do from SEO perspective, I wonder.
deepwiki is far far better than this, sorry google
Correct me if I'm wrong. You're also generating a decent YouTube video from a code base? Pretty cool.