do you have plans to expand this to include more advanced topics like architecture-level reasoning, refactoring patterns, or onboarding workflows for large-scale repositories?
Yes! This is an initial prototype. Good to see the interest, and I'm considering digging deeper by creating more tailored tutorials for different types of projects. E.g., if we know it's web dev, we could generate tutorials based more on request flows, API endpoints, database interactions, etc. If we know it's a more long-term maintained projects, we can focus on identifying refactoring patterns.
Have you ever seen komment.ai? Is so did you have any issues with the limitation of the product?
I haven't used it, but it looks like it's in the same space and I've been curious about it for a while.
I've tried my own homebrew solutions, creating embedding databases by having something like aider or simonw's llm make an ingests json from every function, then using it as a rag in qdrant to do an architecture document, then using that to do contextual inline function commenting and make a doxygen then using all of that once again as an mcp with playwright to hook that up through roo.
It's a weird pipeline and it's been ok, not great but ok.
I'm looking into perplexica as part of the chain, mostly as a negation tool
One thing to note is that the tutorial generation depends largely on Gemini 2.5 Pro. Its code understanding ability is very good, combined with its large 1M context window for a holistic understanding of the code. This leads to very satisfactory tutorial results.
However, Gemini 2.5 Pro was released just late last month. Since Komment.ai launched earlier this year, I don't think models at that time could generate results of that quality.
I've been using llama 4 Maverick through openrouter. Gemini was my go to but I switched basically the day it came out to try it out.
I haven't switched back. At least for my use cases it's been meeting my expectations.
I haven't tried Microsoft's new 1.58 bit model but it may be a great swap out for sentencellm, the legendary all-MiniLM-L6-v2.
I found that if I'm unfamiliar with the knowledge domain I'm mostly using AI but then as I dive in the ratio of AI to human changes to the point where it's AI at 0 and it's all human.
Basically AI wins at day 1 but isn't any better at day 50. If this can change then it's the next step
Yeah, I'd recommend trying Gemini 2.5 Pro. I know early Gemini weren't great, but the recent one is really impressive in terms of coding ability. This project is kind of designed around the recent breakthrough.
The Linux repository has ~50M tokens, which goes beyond the 1M token limit for Gemini 2.5 Pro.
I think there are two paths forward: (1) decompose the repository into smaller parts (e.g., kernel, shell, file system, etc.), or (2) wait for larger-context models with a 50M+ input limit.
Some huge percentage of that is just drivers. The kernel is likely what would be of interest to someone in this regard; moreover, much of that is architecture specific. IIRC the x86 kernel is <1M lines, though probably not <1M tokens.
The AMDGPU driver alone is 5 million lines - out of about 37 million lines total. Over 10% of the codebase is a driver for a single vendor, although most of it is auto generated per-product headers.
I would have preferred to see what would have been generated without my guidance, but since you asked:
* Explanations of how each sub-component is organized and works would be useful.
* Explanations of the modern disk format (an updated ZFS disk format specification) would be useful.
* Explanations of how the more complex features are implemented (e.g. encryption, raid-z expansion, draid) would be interesting.
Basically, making guides that aid development by avoiding a need to read everything line by line would be useful (the ZFS disk format specification, while old, is an excellent example of this). I have spent years doing ZFS development, and there are parts of ZFS codebase that I do not yet understand. This is true for practically all contributors. Having guides that avoid the need for developers to learn the hard way would be useful. Certain historical bugs might have been avoided had we had such guides.
As for the others, LLVM could use improved documentation on how to make plugins. A guide to the various optimization passes would also be useful. Then there is the architecture in general which would be nice to have documented. Documentation for various esoteric features of both FreeBSD and Linux would be useful. I could continue, but I the whole point of having a LLM do this sort of work is to avoid needing myself or someone else to spend time thinking about these things.
I guess you already know what a “Functional API” is and feel patronized. Also possibly you dislike the “cute analogy” factor.
I think this could be solved with an “assume the reader knows …” part of the prompt.
Definitely looks like ELI5 writing there, but many technical documents assume too much knowledge (especially implicit knowledge of the context) so even though I’m not a fan of this section either, I’m not so quick to dismiss it as having no value.
(The images aren't working. I believe those were auto-generated class inheritance or dependency diagrams.)
* The first paragraph is pretty good.
* The second paragraph is incorrect to call pw_rpc the "core" of Pigweed. That implies that you must always use pw_rpc and that all other modules depend on it, which is not true.
* The subsequent descriptions of modules all seemed decent, IIRC.
* The big issue is that the wiki is just a grab bag summary of different parts of the codebase. It doesn't feel coherent. And it doesn't mention the other 100+ modules that the Pigweed codebase contains.
When working on a big codebase, I imagine that tools like mutable.ai and Pocket Flow will need specific instruction on what aspects of the codebase to document.
Woah, this is really neat.
My first step for many new libraries is to clone the repo, launch Claude code, and ask it to write good documentation for me. This would save a lot of steps for me!
I suppose I'm just a little bit bothered by your saying you "built an AI" when all the heavy lifting is done by a pretrained LLM. Saying you made an AI-based program or hell, even saying you made an AI agent, would be more genuine than saying you "built an AI" which is such an all-encompassing thing that I don't even know what it means. At the very least it should imply use of some sort of training via gradient descent though.
At the top are some neat high-level stuffs, but, below that, it quickly turns into code-written-in-human-language.
I think it should be possible to extract some more useful usage patterns by poking into related unit tests. How to use should be what matters to most tutorial readers.
Exactly, kudos to the author because AI didn’t came up with that.
But that’s what they sell, that AI could do what the author did with AI.
The question is, is it worth to put all that money and energy in AI. MS sacrificed its CO2 goals for email summaries and better autocomplete not to mention all the useless things we do with AI
Some people already said it’s useless to learn to program because AI will do, that‘s the hype of AI not that AI isn’t useful as such like parent comment suggested.
They push AI into everything like it’s the ultimate solution but it is not instead is has serious limitations.
The AI companies sell it like the AI could do it by itself and developers are obsolete but in reality it‘s a tool that still needs developers to make something useful
Doesn’t claim it isn’t useful just it’s not as useful as they thought.
For instance to me AI is useful because I don’t have to write boilerplate code but that’s rarely the case. For other things it still useful to write code but I am not faster because the time I save writing the code I need to fix the prompt, audit and fix the code.
Really nice work and thank you for sharing. These are great demonstrations of the value of LLMs which help to go against the negative view on the impacts to junior engineers.
This helps bridge the gap of most projects lacking updated documentation.
This is actually really cool. I just tried it out using an AI studio API key and was pretty impressed. One issue I noticed was that the output was a little too much "for dummies". Spending paragraphs to explain what an API is through restaurant analogies is a little unnecessary. And then followed up with more paragraphs on what GraphQL is. Every chapter seems to suffer from this. The generated documentation seems more suited for a slightly technical PM moreso than a software engineer. This can probably be mitigated by refining the prompt.
The prompt would also maybe be better if it encouraged variety in diagrams. For somethings, a flow chart would fit better than a sequence diagram (e.g., a durable state machine workflow written using AWS Step Functions).
exactly it is. I'd rather impressive but at the same time the audience is always going to be engineers, so perhaps it can be curated to still be technical to a degree? I can't imagine a scenario where I have to explain to the VP my ETL pipeline
Ensure the tone is welcoming and easy for a newcomer to understand{tone_note}.
- Output only the Markdown content for this chapter.
Now, directly provide a super beginner-friendly Markdown output (DON'T need ```markdown``` tags)
So just a change here might do the trick if you’re interested.
But I wonder how Gemini would manage different levels.
From my take (mostly edtech and not in English) it’s really hard to tone the answer properly and not just have a black and white (5 year old vs expert talk) answer.
Anyone has advice on that?
"Write simple, rigorous statements, starting from first principles, and making sure to take things to their logical conclusion. Write in straightforward prose, no bullet points and summaries. Avoid truisms and overly high-level statements. (Optionally) Assume that the reader {now put your original prompt whatever you had e.g 5 yo}"
Sometimes I write a few more lines with the same meaning as above, and sometimes less, they all work more or less OK. Randomly I get better results sometimes with small tweaks but nothing to make a pattern out of -- a useless endeavour anyway since these models change in minute ways every release, and in neural nets the blast radius of a small change is huge.
Answers like this are sort of what makes me wonder what most engineers are smoking when they think AI isn’t valuable.
I don’t think the outright dismissal of AI is smart. (And, OP, I don’t mean to imply that you are doing that. I mean this generally.)
I also suspect people who level these criticisms have never really used a frontier LLM.
Feeding in a whole codebase that I’m familiar with, and hearing the LLM give good answers about its purpose and implementation from a completely cold read is very impressive.
Even if the LLM never writes a line of code - this is still valuable, because helping humans understand software faster means you can help humans write software faster.
Many devs still think their job is to write code not build products their business needs. I use LLMs extensively and it’s helped me work better faster.
LLMs excel at some things and work very poorly at others. People working on different problems have had different experiences, sometimes opposite ends of the spectrum.
I think the people who claim 10x-100x productivity improvements are working on tasks where LLMs work really well. There is a lot of development work out there that is relatively simple CRUD and LLMs are very good at it. On the complete opposite end we have designing new algorithms/data structures or extending them in a novel way. Or implementing drivers for new hardware from incomplete specs. LLMs do not do well on these tasks or even slow down developers 10x.
So, I think the claims of improvement in productivity and regression in productivity can be true at the same time (and it's not just that people who don't find using LLMs productive are just prompting them wrong).
I think most can be gained by learning in which areas LLMs can give large productivity boosts and where it's better to avoid using them. Of course, this is a continuous process, given that LLMs are still getting better.
Personally, I am quite happy with LLMs. They cannot replace me, but they can do a chunk of the boring/repetitive work (e.g. boilerplate), so as a result I can focus on the interesting problems. As long as we don't have human-like performance (and I don't feel like we are close yet), LLMs make programming more interesting.
They are also a great learning aid. E.g., this morning I wanted to make a 3D model for something I needed, but I don't know OpenSCAD. I iteratively made the design with Claude. At some point the problem becomes too difficult for Claude, but with the code generated at that point, I have learned enough about OpenSCAD that I can fix the more difficult parts of the project. The project would have taken me a few hours (to learn the language, etc.), but now I was done in 30 minutes and learned some OpenSCAD in a pleasant way.
Your OpenSCAD experience is an important point in the productivity debates that is often not discussed. A lot of projects that were previously impossible are now feasible. 10 years ago, you might have searched the OpenSCAD docs, watched videos, felt like it was impossible to find the info you needed, and given up. Claude and similar tools have gotten me past that initial blocker many times. Finding a way to unblock 0 to 1 productivity is perhaps as important (or maybe even more important than) as enabling 1 to 10 or 1 to 100.
You don't even need such fancy examples. There are plenty of codebases where people are working with code that is over a decade old and has several paradigms all intermixed with a lot of tribal knowledge that isn't documented in code or wiki. That is where AI sucks. It will not be able to make meaningfully change in that environment.
There is also the frontend and tnpse code bases don't need to be very old at all before AI falls down. NPM packages and clashing styles in a codebase and AI has been not very helpful to me at all.
Generally speaking, which AI is a fine enhancement to autocomplete, I haven't seen it be able to do anything more serious in a mature codebase. The moment business rules and tech debt sneak in in any capacity, AI becomes so unreliable that it's faster to just write it yourself. If I can't trust the AI to automatically generate a list of exports in an index.ts file. What can I trust it for?
When is the last time you tried using LLMs against a large, old, crufty undocumented codebase?
Things have changed a lot in the past six weeks.
Gemini 2.5 Pro accepts a million tokens and can "reason" with them, which means you can feed it hundreds of thousands of lines of code and it has a surprisingly good chance of figuring things out.
OpenAI released their first million token models with the GPT 4.1 series.
OpenAI o3 and o4-mini are both very strong reasoning code models with 200,000 token input limits.
These models are all new within the last six weeks. They're very, very good at working with large amounts of crufty undocumented code.
Ultimately LLMs don’t really understand what the code does at runtime. Sure, just parsing out the codebase can help make a good guess but in some cases it’s hard to trust LLMs with changes because the consequences are unknown in complex codebases that have weird warts nobody documented.
Maybe in a generation or two codebases will become more uniform and predictible if fewer humans do it by hand. Same with self driving cars, if there were no human drivers out there the problem would become trivial to conquer.
That's a lot less true today than it was six weeks ago. The "reasoning" models are spookily good at answering questions about how code runs, and identifying the source of bugs.
They still make mistakes, and yeah they're still (mostly) next token predicting machines under the hood, but if your mental model is
"they can't actually predict through how some code will execute" you may need to update that.
> Answers like this are sort of what makes me wonder what most engineers are smoking when they think AI isn’t valuable.
Honestly, I wonder if I'm living in some parallel universe, because my experience is that "most engineers" are far from that position. The reactions I'm seeing are either "AI is the future" or "I have serious objections to and/or problems with AI".
If you're calling the latter group "the outright dismissal of AI", I would disagree. If I had to call it the outright dismissal of anything, it would be of AI hype.
> I also suspect people who level these criticisms have never really used a frontier LLM.
It's possible. At my workplace, we did a trial of an LLM-based bot that would generate summaries for our GitHub PRs. I have no idea whether it's a "frontier" LLM or not, but I came out of that trial equally impressed, disappointed, and terrified.
Impressed, because its summaries got so many details right. I could immediately see the use for a tool like that: even when the PR author provides a summary of the PR, it's often hard to figure out where to start looking at the PR and in which order to go through changes. The bulleted list of changes from the bot's summary was incredibly useful, especially because it was almost always correct.
Disappointed, because it would often get the most important thing wrong. For the very first PR that I made, it got the whole list of changes right, but the explanation of what the PR did was the opposite of the truth. I made a change to make certain behavior disabled by default and added an option to enable it for testing purposes, and the bot claimed that the behavior was impossible before this change and the PR made it possible if you used this option.
Terrified, because I can see how alluring it is for people to think that they can replace critical thinking with AI. Maybe it's my borderline burnout speaking, but I can easily imagine the future where the pressure from above to be more "efficient" and to reduce costs brings us to the point where we start trusting faulty AI and the small mistakes start accumulating to the point where great damage is done to millions of people.
> Even if the LLM never writes a line of code - this is still valuable, because helping humans understand software faster means you can help humans write software faster.
I have my doubts about this. Yes, if we get an AI that is reliable and doesn't make these mistakes, it can help us understand software faster, as long as we're willing to make the effort to actually understand it, rather than delegating to the AI's understanding.
What I mean by that is that there are different levels of understanding. How deep do you dive before you decide it's "deep enough" and trust what the AI said? This is even more important if you start also using the AI to write the code and not just read it. Now you have even less motivation to understand the code, because you don't have to learn something that you will use to write your own code.
I'll keep learning how to use LLMs, because it's necessary, but I'm very worried about what we seem to want from them. I can't think of any previous technological advance that aimed to replace human critical thinking and creativity. Why are we even pursuing efficiency if it isn't to give us more time and freedom to be creative?
The value is that it got the details correct as you admit. That alone is worth the price of admission. Even if I need to rewrite or edit parts it has saved me time, and has raised the quality of PRs being submitted across the board. The key point with these tools is *Accountability*. As an engineer you are still accountable for your work. Using any tool doesn't take that away. If the PR tool gets it wrong, and you still submit it, that on the engineer. If you have a culture of accountability, then there is nothing to be terrified of. Any by the way the most recent tools are really, really good at PRs and commit messages.
> Even if the LLM never writes a line of code - this is still valuable, because helping humans understand software faster means you can help humans write software faster.
IMHO, Ai text additions are generally not valuable and I assume, until proven wrong, that Ai text provides little to no value.
I have seen so many startups fold after they made some ai product that on the surface level appeared impressive but provided no substantial value.
Now, I will be impressed by the ai that can remove code without affecting the product.
> hearing the LLM give good answers about its purpose and implementation from a completely cold read
Cold read ability for this particular tool is still an open question. As others have mentioned, a lot of the example tutorials are for very popular codebases that are probably well-represented in the language model's training data. I'm personally going to test it on my private, undocumented repos.
> Spending paragraphs to explain what an API is through restaurant analogies is a little unnecessary. And then followed up with more paragraphs on what GraphQL is.
It sounds like the tool (as it's currently set up) may not actually be that effective at writing tutorial-style content in particular. Tutorials [1] are usually heavily action-oriented and take you from a specific start point to a specific end point to help you get hands-on experience in some skill. Some technical writers argue that there should be no theory whatsoever in tutorials. However, it's probably easy to tweak the prompts to get more action-oriented content with less conceptual explanation (and exclamation marks).
>> This can probably be mitigated by refining the prompt
Sometimes it explains things like I am a child and sometimes it doesn't explain things well enough. I think fixing this just by a simple prompt change won't work - it may fix it in one part and make things worse in the other part. This is a problem which I have with LLM: you can fine-tune the prompt for a specific case but I find it difficult to write a universally-working prompt. The problem seems to be LLM "does not understand my intents", like it can't deduce what I need and "proactively" help. It follows requirements from the prompt but the prompt has to (and can't) handle all situations. I am getting tired of LLM.
Isn't that overly optimistic? The postgres source code is really complex, and reading a dummy tutorial isn't going to make you a database engine ninja. If a simple tutorial can, imagine what a book on the topic could do.
No, I am not that optimistic about LLMs. I just think that something is better then nothing.
The burden of understanding still is with the engineers. All you would get is some (partially inaccurate at places) good overview of where to look for.
I built browser use. Dayum, the results for our lib are really impressive, you didn’t touch outputs at all?
One problem we have is maintaining the docs with current codebase (code examples break sometimes). Wonder if I could use parts of Pocket to help with that.
As a maintainer of a different library, I think there’s something here. A revised version of this tool that also gets fed the docs and asked to find inaccuracies could be great. Even if false positives and false negatives are let’s say 20% each, it would still be better than before as final decisions are made by a human.
Thank you! And correct, I didn't modify the outputs. For small changes, you can just feed the commit history and ask an LLM to modify the docs. If there are lots of architecture-level changes, it would be easier to just feed the old docs and rewrite - it usually takes <10 minutes.
180 comments
[ 3.9 ms ] story [ 243 ms ] threadI haven't used it, but it looks like it's in the same space and I've been curious about it for a while.
I've tried my own homebrew solutions, creating embedding databases by having something like aider or simonw's llm make an ingests json from every function, then using it as a rag in qdrant to do an architecture document, then using that to do contextual inline function commenting and make a doxygen then using all of that once again as an mcp with playwright to hook that up through roo.
It's a weird pipeline and it's been ok, not great but ok.
I'm looking into perplexica as part of the chain, mostly as a negation tool
One thing to note is that the tutorial generation depends largely on Gemini 2.5 Pro. Its code understanding ability is very good, combined with its large 1M context window for a holistic understanding of the code. This leads to very satisfactory tutorial results.
However, Gemini 2.5 Pro was released just late last month. Since Komment.ai launched earlier this year, I don't think models at that time could generate results of that quality.
I haven't switched back. At least for my use cases it's been meeting my expectations.
I haven't tried Microsoft's new 1.58 bit model but it may be a great swap out for sentencellm, the legendary all-MiniLM-L6-v2.
I found that if I'm unfamiliar with the knowledge domain I'm mostly using AI but then as I dive in the ratio of AI to human changes to the point where it's AI at 0 and it's all human.
Basically AI wins at day 1 but isn't any better at day 50. If this can change then it's the next step
Do you have anything in mind? Are you familiar enough with any of those codebases to suggest something useful?
The task will be much more interesting if there is not a good existing tutorial that the LLM may have trained on.
OS kernel: tutorial on how to write a driver?
OpenZFS: ?
https://github.com/openzfs/zfs/graphs/contributors
I would have preferred to see what would have been generated without my guidance, but since you asked:
* Explanations of how each sub-component is organized and works would be useful.
* Explanations of the modern disk format (an updated ZFS disk format specification) would be useful.
* Explanations of how the more complex features are implemented (e.g. encryption, raid-z expansion, draid) would be interesting.
Basically, making guides that aid development by avoiding a need to read everything line by line would be useful (the ZFS disk format specification, while old, is an excellent example of this). I have spent years doing ZFS development, and there are parts of ZFS codebase that I do not yet understand. This is true for practically all contributors. Having guides that avoid the need for developers to learn the hard way would be useful. Certain historical bugs might have been avoided had we had such guides.
As for the others, LLVM could use improved documentation on how to make plugins. A guide to the various optimization passes would also be useful. Then there is the architecture in general which would be nice to have documented. Documentation for various esoteric features of both FreeBSD and Linux would be useful. I could continue, but I the whole point of having a LLM do this sort of work is to avoid needing myself or someone else to spend time thinking about these things.
I think this could be solved with an “assume the reader knows …” part of the prompt.
Definitely looks like ELI5 writing there, but many technical documents assume too much knowledge (especially implicit knowledge of the context) so even though I’m not a fan of this section either, I’m not so quick to dismiss it as having no value.
https://news.ycombinator.com/item?id=42542512
The latter is what this thread claims ^
I did however archive the wiki that it generated for the project I work on: https://web.archive.org/web/20240815184418/wiki.mutable.ai/g...
(The images aren't working. I believe those were auto-generated class inheritance or dependency diagrams.)
* The first paragraph is pretty good.
* The second paragraph is incorrect to call pw_rpc the "core" of Pigweed. That implies that you must always use pw_rpc and that all other modules depend on it, which is not true.
* The subsequent descriptions of modules all seemed decent, IIRC.
* The big issue is that the wiki is just a grab bag summary of different parts of the codebase. It doesn't feel coherent. And it doesn't mention the other 100+ modules that the Pigweed codebase contains.
When working on a big codebase, I imagine that tools like mutable.ai and Pocket Flow will need specific instruction on what aspects of the codebase to document.
Like at least one other person in the comments mentioned, I would like a slightly different tone.
Perhaps good feature would be a "style template", that can be chosen to match your preferred writing style.
I may submit a PR though not if it takes a lot of time.
I bet in the next few months we'll be getting dynamic, personalized documentation for every library!! Good times
I think it should be possible to extract some more useful usage patterns by poking into related unit tests. How to use should be what matters to most tutorial readers.
The hype is that AI isn’t a tool but the developer.
But that’s what they sell, that AI could do what the author did with AI.
The question is, is it worth to put all that money and energy in AI. MS sacrificed its CO2 goals for email summaries and better autocomplete not to mention all the useless things we do with AI
Can you give an example of what you meant here? The author did use AI. What does "AI coming up with that" mean?
In the few years we will see complaints that it’s not AI that built a power station and a datacenter, so it doesn’t count as well.
They push AI into everything like it’s the ultimate solution but it is not instead is has serious limitations.
The AI companies sell it like the AI could do it by itself and developers are obsolete but in reality it‘s a tool that still needs developers to make something useful
https://news.ycombinator.com/item?id=41542497
For instance to me AI is useful because I don’t have to write boilerplate code but that’s rarely the case. For other things it still useful to write code but I am not faster because the time I save writing the code I need to fix the prompt, audit and fix the code.
The prompt would also maybe be better if it encouraged variety in diagrams. For somethings, a flow chart would fit better than a sequence diagram (e.g., a durable state machine workflow written using AWS Step Functions).
Ensure the tone is welcoming and easy for a newcomer to understand{tone_note}.
- Output only the Markdown content for this chapter.
Now, directly provide a super beginner-friendly Markdown output (DON'T need ```markdown``` tags)
So just a change here might do the trick if you’re interested.
But I wonder how Gemini would manage different levels. From my take (mostly edtech and not in English) it’s really hard to tone the answer properly and not just have a black and white (5 year old vs expert talk) answer. Anyone has advice on that?
"Write simple, rigorous statements, starting from first principles, and making sure to take things to their logical conclusion. Write in straightforward prose, no bullet points and summaries. Avoid truisms and overly high-level statements. (Optionally) Assume that the reader {now put your original prompt whatever you had e.g 5 yo}"
Sometimes I write a few more lines with the same meaning as above, and sometimes less, they all work more or less OK. Randomly I get better results sometimes with small tweaks but nothing to make a pattern out of -- a useless endeavour anyway since these models change in minute ways every release, and in neural nets the blast radius of a small change is huge.
I don’t think the outright dismissal of AI is smart. (And, OP, I don’t mean to imply that you are doing that. I mean this generally.)
I also suspect people who level these criticisms have never really used a frontier LLM.
Feeding in a whole codebase that I’m familiar with, and hearing the LLM give good answers about its purpose and implementation from a completely cold read is very impressive.
Even if the LLM never writes a line of code - this is still valuable, because helping humans understand software faster means you can help humans write software faster.
So, I think the claims of improvement in productivity and regression in productivity can be true at the same time (and it's not just that people who don't find using LLMs productive are just prompting them wrong).
I think most can be gained by learning in which areas LLMs can give large productivity boosts and where it's better to avoid using them. Of course, this is a continuous process, given that LLMs are still getting better.
Personally, I am quite happy with LLMs. They cannot replace me, but they can do a chunk of the boring/repetitive work (e.g. boilerplate), so as a result I can focus on the interesting problems. As long as we don't have human-like performance (and I don't feel like we are close yet), LLMs make programming more interesting.
They are also a great learning aid. E.g., this morning I wanted to make a 3D model for something I needed, but I don't know OpenSCAD. I iteratively made the design with Claude. At some point the problem becomes too difficult for Claude, but with the code generated at that point, I have learned enough about OpenSCAD that I can fix the more difficult parts of the project. The project would have taken me a few hours (to learn the language, etc.), but now I was done in 30 minutes and learned some OpenSCAD in a pleasant way.
There is also the frontend and tnpse code bases don't need to be very old at all before AI falls down. NPM packages and clashing styles in a codebase and AI has been not very helpful to me at all.
Generally speaking, which AI is a fine enhancement to autocomplete, I haven't seen it be able to do anything more serious in a mature codebase. The moment business rules and tech debt sneak in in any capacity, AI becomes so unreliable that it's faster to just write it yourself. If I can't trust the AI to automatically generate a list of exports in an index.ts file. What can I trust it for?
Things have changed a lot in the past six weeks.
Gemini 2.5 Pro accepts a million tokens and can "reason" with them, which means you can feed it hundreds of thousands of lines of code and it has a surprisingly good chance of figuring things out.
OpenAI released their first million token models with the GPT 4.1 series.
OpenAI o3 and o4-mini are both very strong reasoning code models with 200,000 token input limits.
These models are all new within the last six weeks. They're very, very good at working with large amounts of crufty undocumented code.
Maybe in a generation or two codebases will become more uniform and predictible if fewer humans do it by hand. Same with self driving cars, if there were no human drivers out there the problem would become trivial to conquer.
They still make mistakes, and yeah they're still (mostly) next token predicting machines under the hood, but if your mental model is "they can't actually predict through how some code will execute" you may need to update that.
Honestly, I wonder if I'm living in some parallel universe, because my experience is that "most engineers" are far from that position. The reactions I'm seeing are either "AI is the future" or "I have serious objections to and/or problems with AI".
If you're calling the latter group "the outright dismissal of AI", I would disagree. If I had to call it the outright dismissal of anything, it would be of AI hype.
> I also suspect people who level these criticisms have never really used a frontier LLM.
It's possible. At my workplace, we did a trial of an LLM-based bot that would generate summaries for our GitHub PRs. I have no idea whether it's a "frontier" LLM or not, but I came out of that trial equally impressed, disappointed, and terrified.
Impressed, because its summaries got so many details right. I could immediately see the use for a tool like that: even when the PR author provides a summary of the PR, it's often hard to figure out where to start looking at the PR and in which order to go through changes. The bulleted list of changes from the bot's summary was incredibly useful, especially because it was almost always correct.
Disappointed, because it would often get the most important thing wrong. For the very first PR that I made, it got the whole list of changes right, but the explanation of what the PR did was the opposite of the truth. I made a change to make certain behavior disabled by default and added an option to enable it for testing purposes, and the bot claimed that the behavior was impossible before this change and the PR made it possible if you used this option.
Terrified, because I can see how alluring it is for people to think that they can replace critical thinking with AI. Maybe it's my borderline burnout speaking, but I can easily imagine the future where the pressure from above to be more "efficient" and to reduce costs brings us to the point where we start trusting faulty AI and the small mistakes start accumulating to the point where great damage is done to millions of people.
> Even if the LLM never writes a line of code - this is still valuable, because helping humans understand software faster means you can help humans write software faster.
I have my doubts about this. Yes, if we get an AI that is reliable and doesn't make these mistakes, it can help us understand software faster, as long as we're willing to make the effort to actually understand it, rather than delegating to the AI's understanding.
What I mean by that is that there are different levels of understanding. How deep do you dive before you decide it's "deep enough" and trust what the AI said? This is even more important if you start also using the AI to write the code and not just read it. Now you have even less motivation to understand the code, because you don't have to learn something that you will use to write your own code.
I'll keep learning how to use LLMs, because it's necessary, but I'm very worried about what we seem to want from them. I can't think of any previous technological advance that aimed to replace human critical thinking and creativity. Why are we even pursuing efficiency if it isn't to give us more time and freedom to be creative?
I'll just wait for a winner to shake out and learn that one. I've gotten tired of trying AIs only to get slop.
IMHO, Ai text additions are generally not valuable and I assume, until proven wrong, that Ai text provides little to no value.
I have seen so many startups fold after they made some ai product that on the surface level appeared impressive but provided no substantial value.
Now, I will be impressed by the ai that can remove code without affecting the product.
Current AIs can already do this decently. With the usual caveats about possible mistakes/oversight.
Cold read ability for this particular tool is still an open question. As others have mentioned, a lot of the example tutorials are for very popular codebases that are probably well-represented in the language model's training data. I'm personally going to test it on my private, undocumented repos.
Gemini 2.5 Flash Preview 04-17 another powerful model has 10 and 500.
OpenAI also allows you to use their API for free if you agree to share the tokens.
It sounds like the tool (as it's currently set up) may not actually be that effective at writing tutorial-style content in particular. Tutorials [1] are usually heavily action-oriented and take you from a specific start point to a specific end point to help you get hands-on experience in some skill. Some technical writers argue that there should be no theory whatsoever in tutorials. However, it's probably easy to tweak the prompts to get more action-oriented content with less conceptual explanation (and exclamation marks).
[1] https://diataxis.fr/tutorials/
Sometimes it explains things like I am a child and sometimes it doesn't explain things well enough. I think fixing this just by a simple prompt change won't work - it may fix it in one part and make things worse in the other part. This is a problem which I have with LLM: you can fine-tune the prompt for a specific case but I find it difficult to write a universally-working prompt. The problem seems to be LLM "does not understand my intents", like it can't deduce what I need and "proactively" help. It follows requirements from the prompt but the prompt has to (and can't) handle all situations. I am getting tired of LLM.
Put in postgres or redis codebase, get a good understanding and get going to contribute.
The burden of understanding still is with the engineers. All you would get is some (partially inaccurate at places) good overview of where to look for.
With the rise of AI understanding software will become relatively easy
DEFAULT_INCLUDE_PATTERNS = { ".py", ".js", ".jsx", ".ts", ".tsx", ".go", ".java", ".pyi", ".pyx", ".c", ".cc", ".cpp", ".h", ".md", ".rst", "Dockerfile", "Makefile", ".yaml", ".yml", } DEFAULT_EXCLUDE_PATTERNS = { "test", "tests/", "docs/", "examples/", "v1/", "dist/", "build/", "experimental/", "deprecated/", "legacy/", ".git/", ".github/", ".next/", ".vscode/", "obj/", "bin/", "node_modules/", ".log" }
Can see some finetuning after generation being required, but assuming you know your own codebase that's not an issue anyway.
You might get two or three tutorials built for yourself inside the free 25/day limit, depending on how many chapters it needs.