As a counterpoint, I also tried writing something with Claude last weekend: a Google docs clone[1]. I spent $170 on Anthropic API credits, and got something that did mostly what I asked but was basically useless. It seems that for simple interfaces for which there is an exact specification, like the recent compiler and web browser examples, it's possible to write bigger projects that "work" as a demo although not in a way they'd be viable alternatives. For anything that requires taste and judgment, we've still got a long way to go. There are lots of great demos out there but few if any real examples of vibe coded (or whatever you want to call it) software standing alone as an alternative to project people wrote.
So we are now at the stage that AI coding agents don’t work because you can’t create a good google docs alternative from scratch without dependencies with $170 in one weekend.
I think the conclusion is that Vibe oding is shit, and I agree with this, but having an AI assistant that can do some specific tasks that you can review is a good strategy.
I also have a story about Vibe Coding, I had the AI make for me a mqarkdown editor with some extra features, it worked fine but the problem is I have no idea how it works, I do not know if I can add feature X easily or it needs a rewrite from scratch, I have no ideas for improvements or new features, if there is a problem I have no clue what causes it since I I never looked at the code.
So I concluded that I (others can do whatever they want, they are free people with their own standards ) I will only use Vibe Coding for throw away personal shit, that I will not make public, like for example I made some bash scripts, some python scripts to automate some stuff.
Does "vibe coding" now mean the agent needs to produce a working product in one shot? Or not even just working, but meeting unspecified requirements? I thought in the original phrasing by Karpathy it meant that you don't care about the code, yet you may still iterate because you might care about the product. I tried out the "Impeccable" skill for Claude Code and found it to be very useful if you care about visual aspects and good UI/UX. (I vibe coded a web app for tracking personal finances where I used this skill, see: https://github.com/AdrianVollmer/Solvency. Not one-shotted, and not perfect, but personally I'm impressed with what is possible in a few weekends with a $90 subscription to Claude.)
> For anything that requires taste and judgment, we've still got a long way to go. There are lots of great demos out there but few if any real examples of vibe coded (or whatever you want to call it) software standing alone as an alternative to project people wrote
Yeah, we still have exactly the same problem as before LLMs/agents, namely that we lack people with "Good Taste". I've outlined how I feel about before (https://emsh.cat/good-taste/) but the TLDR is basically that while LLMs can help you move faster, they won't suddenly mean you'll make better choices, probably the reverse is true, you'll move faster and make worse choices.
Having good taste and knowing how things actually should function is 80% of the work of building good software, and so far all these tools that are trying to replace human choices lead to worse software, and we need more tooling that puts the human and the LLM working together, instead of just outsourcing from the human to the LLM.
> The result is OK. It has all the features I asked for, and includes document sharing, collaborative editing in real time, support for fonts and line spacing, etc. etc. I could not have paid a developer $170 and got this. The problem, of course, is that, while abstractly impressive, this is completely useless
Well, what would you expect from a few hours of running in a loop with these constraints?
> This project exists to build a document editor from the ground up. Violating these constraints defeats the entire purpose.
> FORBIDDEN dependencies (do NOT install or use these):
> Rich text editor frameworks: ProseMirror, Slate, Quill, TipTap, Draft.js, CKEditor, TinyMCE, Lexical, or any similar library
> CRDT/OT libraries: Yjs, Automerge, ShareDB, OT.js, or any similar library
> Full CSS frameworks: Bootstrap, Tailwind, Material UI (small utility libs for specific needs are OK)
> ORMs: Prisma, TypeORM, Sequelize (use raw SQL or a thin query builder)
I can't help but wonder what you thought you would achieve, and how getting "mostly what you asked for" is still disappointing to you.
> there is no taste being applied.
There are 0 lines in AGENT_PROMPT.md about "taste". You have instructed something/someone on how to build more than what to build.
Your goals are (from a quick skim):
- The goal of this project is ultimately to generate a working alternative to Google Docs with the same functionality.
- You are an autonomous software engineer building AltDocs, a from-scratch alternative to Google Docs.
I see a FEATURES.md file, but not clear if this is from you or expanded by the model. It seems pretty slim.
All in all, I don't get the "disappointment". It seems, from your blog post, that the "model" did most of the things you asked for. The disappointment might come from what you asked for, more than from the "model" being bad... To paraphrase a line from a sitcom: "Damn, Andrew, I can't control the weather!" :)
This blog post is a sophisticated piece of content marketing for a company called JUXT and their proprietary tool, "Allium." While the technical achievement is plausible, the framing is heavily distorted to sell a product.
Here is the breakdown of the flaws and the "BS" in the narrative.
1. The "I Didn't Write Code" Lie
The author claims, "I didn't write a line of implementation code." The Flaw: He wrote 3,000 lines of "Allium behavioural specification." The BS: Writing 3,000 lines of a formal specification language is coding. It’s just coding in a proprietary, high-level language instead of Kotlin.
The Ratio is Terrible: The post admits the output was ~5,500 lines of Kotlin. That means for every 1 line of spec, he got roughly 1.8 lines of code.
Why this matters: True "low-code" or "no-code" leverage is usually 1:10 or 1:100. If you have to write 3,000 lines of strict logic to get a 5,000-line program, you haven't saved much effort—you've just swapped languages.
2. The "Weekend Project" Myth
The post frames this as a casual project done "between board games and time with my kids." The Flaw: This timeline ignores the massive "pre-computation" done by the human. The BS: To write 3,000 lines of coherent, bug-free specifications for a Byzantine Fault Tolerant (BFT) system, you need to have the entire architecture fully resolved in your head before you start typing. The author is an expert (CTO level) who likely spent weeks or years thinking about these problems. The "48 hours" only counts the typing time, not the engineering time.
3. The "Byzantine Fault Tolerance" (BFT) Bait-and-Switch
The headline claims "Byzantine fault tolerance," which implies a system that continues to operate correctly even if nodes lie or act maliciously (extremely hard to build). The Flaw: A "Resolved Question" block in the text admits: "The system's goal is Byzantine fault detection, not classical BFT consensus." The BS: Real BFT (like PBFT or Raft with signatures) is mathematically rigorous and keeps the system running. "Fault Detection" just means "if the two copies don't match, stop." That is significantly easier to build. Calling it "BFT" in the intro is a massive overstatement of the system's resilience.
4. The "Maintenance Nightmare" (The Vendor Lock-in Trap)
The post glosses over how this system is maintained. The Flaw: You now have 5,500 lines of Kotlin that no human wrote. The BS: This is the "Model Driven Architecture" (MDA) trap from the early 2000s.
Scenario: You find a bug in the Kotlin code.
Option A: You fix the Kotlin. Result: Your code is now out of sync with the Spec. You can never regenerate from Spec again without losing your fix.
Option B: You fix the Spec. Result: You hope the AI generates the exact Kotlin fix you need without breaking 10 other things.
The Reality: You are now 100% dependent on the "Allium" tool and Claude. If you stop paying for Allium, you have a pile of unmaintainable machine-generated code.
5. The Performance "Turning Point"
The dramatic story about 10,000 Requests Per Second (RPS) has a hole in it. The Flaw: The "bottleneck" wasn't the code; it was a Docker proxy setting (gvproxy). The BS: This is a standard "gotcha" for anyone using Docker on Mac. Framing this as a triumph of AI debugging is a stretch—any senior engineer would check network topology when seeing high latency but low CPU usage. 10k RPS is also not "ambitious" for a modern distributed system; a single well-optimized Node.js or Go process can handle that easily.
I find it interesting that you request me to discuss subject matter when your post is intellectually equivalent to: "I generated this sequence of numbers using my pseudo-random number generator"
Stop telling people what to do (they will never listen because you told them). Or at least have the courtesy to point out why it's "slop" or what's unfounded about it.
I don't think people should be rude to you, but the comment was AI-generated, right? Lots of people dislike that as it feels kind of wasteful and disrespectful of our time; it can literally take you less time to generate the comment than for us to read it, and the only information you added is whatever was in the (presumably much shorter) prompt. If you'd written it yourself, it may or may not be interesting and correct, but I'd at least know that someone cared enough to write it and all of it made sense from that person's perspective. Sometimes I am interested in an LLM's take on a topic, but not when browsing a forum for humans.
I'm sorry but if you're blaming some text on a website to be "disrespectful of our time", I don't know what to say to you.
I stand behind everything on my comment and I have engaged in good faith with every single reply to it here (even though none of them talk about anything in the comment itself).
Go through my profile, see how I engage with people and tell me again I'm AI.
If you do not have anything to say to the subject matter of a comment and just have personal snide remarks, I do think it's a waste of your time but do not blame me for it or tell me to leave the platform.
Typing this comment right now is a waste of time for me but I do not feel the need to grandstand over it as if there's a massive opportunity cost to it. I'm a human writing/interacting in a "forum for humans."
I didn't say or think that your account was AI-run, and I didn't tell you to leave the platform. I just tried to explain why your comment might have annoyed people and triggered negative responses (while agreeing that the rude ones were inappropriate).
Sure, cheers then. I don't care about negative responses if they're negative just because they think it's AI-generated, without having to say anything substantial on the actual comment or the article. I have demonstrated my willingness to engage in good faith but those comments have not.
If negative responses have no substance behind it, it makes no sense to care about them or take them seriously.
Also, the fact that you assume it takes more time reading that comment than it took for me to write is pretty weird (I still don't get what was so wrong about the comment that simply reading it is a waste of time to people).
Is there even a sync to be had? The same prompt to the same LLM at different times will yield different artifacts, even if you were to save and re-use the seed.
Is there a link to the specification and the resulting generated code ? I skimmed through the article and the author's github profile, but couldn't find anything related.
Seems like a serious oversight if this is your selling point.
Sorry, I meant the spec and code of the toy project described in the post.
I don't necessarily expect their secret sauce to be openly available. But given the grandiose claims made there, I do expect something to back it up other than a corpospeak "trust me bro".
This post appears to essentially be an ad for Allium, so I’m gonna focus on that.
I don’t fully understand how allium solves the inherent issues with markdown, i.e. free-form language specs. It appears to have some loose syntactical rules, but outside of that relies on comments which are, again, free-form. Moreso, the “Resolved” portions remind of me how LLM agents like Claude code just go and edit existing doc files and make a mess of repetitive and overlapping thoughts.
Over a weekend, between board games and time with my kids, Claude and I built a distributed system with Byzantine fault tolerance, strong consistency and crash recovery under arbitrary failures. I described the behaviour I wanted in Allium, worked through the bugs conversationally and didn't write a line of implementation code.
I don't see why you need to bring your kids into this, and as a parent any suggestion of being distracted by tech during time with the kids raises my suspicions.
We have a strict no laptops and no phones rule when the kids are around (unless we're specifically doing something with them using those tools - looking at the weather forecast, or looking up some information).
"I can prompt AI while playing with the kids" is not a future I want.
They said "between", not "during"; I think the point was that they didn't spend a full weekend with Claude.
edit: anyone care to explain why this is a bad comment, rather than just downvoting? The GP comment says "Over a weekend, between board games and time with my kids,", and the parent comment lectures them based on an obvious strawman: "I can prompt AI while playing with the kids"
"between time with the kids" is another person's "during time with the kids".
For me, "between time with the kids" means my kids are engaged in another activity that does not require my input until they are done with it. Whatever I am doing during this time also is typically very interruptible, so I am ready to help the kids along to their next "thing" (the joys of being a parent!). On a typical weekend (my oldest is 7), I'll get maybe 2 hours of this time during the time my kids are awake.
That all makes sense, but I don't see how it supports your uncharitable read of the original comment. Both because they could very easily have meant something a bit different (we have no idea how old the kids are or whether they were even around all weekend; maybe there were times when they were at friends' houses or similar), and because vibe coding could be the interruptible activity they do while the kids are busy. Maybe you feel strongly that it's not a legitimate use of that kind of semi-downtime, but we don't know if that's what they were talking about in the first place.
"Hours spent on the project" would be a much more useful metric, with no confounding variables. As it is, the mere mention of interleaving time with your kids and time engrossed in tech hits a nerve at time when IMO too many parents are doing this already, and lends unwarranted validity to the idea.
As a counterpoint, I also tried writing something with Claude last weekend: a Google docs clone[1]. I spent $170 on Anthropic API credits, and got something that did mostly what I asked but was basically useless. It seems that for simple interfaces for which there is an exact specification, like the recent compiler and web browser examples, it's possible to write bigger projects that "work" as a demo although not in a way they'd be viable alternatives. For anything that requires taste and judgment, we've still got a long way to go. There are lots of great demos out there but few if any real examples of vibe coded (or whatever you want to call it) software standing alone as an alternative to project people wrote.
[1] https://www.marble.onl/posts/this_cost_170.html
So we are now at the stage that AI coding agents don’t work because you can’t create a good google docs alternative from scratch without dependencies with $170 in one weekend.
I think the conclusion is that Vibe oding is shit, and I agree with this, but having an AI assistant that can do some specific tasks that you can review is a good strategy.
I also have a story about Vibe Coding, I had the AI make for me a mqarkdown editor with some extra features, it worked fine but the problem is I have no idea how it works, I do not know if I can add feature X easily or it needs a rewrite from scratch, I have no ideas for improvements or new features, if there is a problem I have no clue what causes it since I I never looked at the code.
So I concluded that I (others can do whatever they want, they are free people with their own standards ) I will only use Vibe Coding for throw away personal shit, that I will not make public, like for example I made some bash scripts, some python scripts to automate some stuff.
Does "vibe coding" now mean the agent needs to produce a working product in one shot? Or not even just working, but meeting unspecified requirements? I thought in the original phrasing by Karpathy it meant that you don't care about the code, yet you may still iterate because you might care about the product. I tried out the "Impeccable" skill for Claude Code and found it to be very useful if you care about visual aspects and good UI/UX. (I vibe coded a web app for tracking personal finances where I used this skill, see: https://github.com/AdrianVollmer/Solvency. Not one-shotted, and not perfect, but personally I'm impressed with what is possible in a few weekends with a $90 subscription to Claude.)
> For anything that requires taste and judgment, we've still got a long way to go. There are lots of great demos out there but few if any real examples of vibe coded (or whatever you want to call it) software standing alone as an alternative to project people wrote
Yeah, we still have exactly the same problem as before LLMs/agents, namely that we lack people with "Good Taste". I've outlined how I feel about before (https://emsh.cat/good-taste/) but the TLDR is basically that while LLMs can help you move faster, they won't suddenly mean you'll make better choices, probably the reverse is true, you'll move faster and make worse choices.
Having good taste and knowing how things actually should function is 80% of the work of building good software, and so far all these tools that are trying to replace human choices lead to worse software, and we need more tooling that puts the human and the LLM working together, instead of just outsourcing from the human to the LLM.
> The result is OK. It has all the features I asked for, and includes document sharing, collaborative editing in real time, support for fonts and line spacing, etc. etc. I could not have paid a developer $170 and got this. The problem, of course, is that, while abstractly impressive, this is completely useless
Well, what would you expect from a few hours of running in a loop with these constraints?
> This project exists to build a document editor from the ground up. Violating these constraints defeats the entire purpose.
> FORBIDDEN dependencies (do NOT install or use these):
> Rich text editor frameworks: ProseMirror, Slate, Quill, TipTap, Draft.js, CKEditor, TinyMCE, Lexical, or any similar library
> CRDT/OT libraries: Yjs, Automerge, ShareDB, OT.js, or any similar library
> Full CSS frameworks: Bootstrap, Tailwind, Material UI (small utility libs for specific needs are OK)
> ORMs: Prisma, TypeORM, Sequelize (use raw SQL or a thin query builder)
I can't help but wonder what you thought you would achieve, and how getting "mostly what you asked for" is still disappointing to you.
> there is no taste being applied.
There are 0 lines in AGENT_PROMPT.md about "taste". You have instructed something/someone on how to build more than what to build.
Your goals are (from a quick skim):
- The goal of this project is ultimately to generate a working alternative to Google Docs with the same functionality.
- You are an autonomous software engineer building AltDocs, a from-scratch alternative to Google Docs.
I see a FEATURES.md file, but not clear if this is from you or expanded by the model. It seems pretty slim.
All in all, I don't get the "disappointment". It seems, from your blog post, that the "model" did most of the things you asked for. The disappointment might come from what you asked for, more than from the "model" being bad... To paraphrase a line from a sitcom: "Damn, Andrew, I can't control the weather!" :)
This blog post is a sophisticated piece of content marketing for a company called JUXT and their proprietary tool, "Allium." While the technical achievement is plausible, the framing is heavily distorted to sell a product.
Here is the breakdown of the flaws and the "BS" in the narrative.
1. The "I Didn't Write Code" Lie The author claims, "I didn't write a line of implementation code." The Flaw: He wrote 3,000 lines of "Allium behavioural specification." The BS: Writing 3,000 lines of a formal specification language is coding. It’s just coding in a proprietary, high-level language instead of Kotlin.
The Ratio is Terrible: The post admits the output was ~5,500 lines of Kotlin. That means for every 1 line of spec, he got roughly 1.8 lines of code.
Why this matters: True "low-code" or "no-code" leverage is usually 1:10 or 1:100. If you have to write 3,000 lines of strict logic to get a 5,000-line program, you haven't saved much effort—you've just swapped languages.
2. The "Weekend Project" Myth The post frames this as a casual project done "between board games and time with my kids." The Flaw: This timeline ignores the massive "pre-computation" done by the human. The BS: To write 3,000 lines of coherent, bug-free specifications for a Byzantine Fault Tolerant (BFT) system, you need to have the entire architecture fully resolved in your head before you start typing. The author is an expert (CTO level) who likely spent weeks or years thinking about these problems. The "48 hours" only counts the typing time, not the engineering time.
3. The "Byzantine Fault Tolerance" (BFT) Bait-and-Switch The headline claims "Byzantine fault tolerance," which implies a system that continues to operate correctly even if nodes lie or act maliciously (extremely hard to build). The Flaw: A "Resolved Question" block in the text admits: "The system's goal is Byzantine fault detection, not classical BFT consensus." The BS: Real BFT (like PBFT or Raft with signatures) is mathematically rigorous and keeps the system running. "Fault Detection" just means "if the two copies don't match, stop." That is significantly easier to build. Calling it "BFT" in the intro is a massive overstatement of the system's resilience.
4. The "Maintenance Nightmare" (The Vendor Lock-in Trap) The post glosses over how this system is maintained. The Flaw: You now have 5,500 lines of Kotlin that no human wrote. The BS: This is the "Model Driven Architecture" (MDA) trap from the early 2000s.
Scenario: You find a bug in the Kotlin code.
Option A: You fix the Kotlin. Result: Your code is now out of sync with the Spec. You can never regenerate from Spec again without losing your fix.
Option B: You fix the Spec. Result: You hope the AI generates the exact Kotlin fix you need without breaking 10 other things.
The Reality: You are now 100% dependent on the "Allium" tool and Claude. If you stop paying for Allium, you have a pile of unmaintainable machine-generated code.
5. The Performance "Turning Point" The dramatic story about 10,000 Requests Per Second (RPS) has a hole in it. The Flaw: The "bottleneck" wasn't the code; it was a Docker proxy setting (gvproxy). The BS: This is a standard "gotcha" for anyone using Docker on Mac. Framing this as a triumph of AI debugging is a stretch—any senior engineer would check network topology when seeing high latency but low CPU usage. 10k RPS is also not "ambitious" for a modern distributed system; a single well-optimized Node.js or Go process can handle that easily.
Hello, could you please put your over-sensationalized, overly-long, AI-generated comments somewhere else? Thank you.
Kindly, the HN Community.
Could you please stop being the messiah of the HN community and talk to the subject matter instead of making personal snide remarks? Thank you.
I find it interesting that you request me to discuss subject matter when your post is intellectually equivalent to: "I generated this sequence of numbers using my pseudo-random number generator"
Stop posting slop.
Stop telling people what to do (they will never listen because you told them). Or at least have the courtesy to point out why it's "slop" or what's unfounded about it.
I don't think people should be rude to you, but the comment was AI-generated, right? Lots of people dislike that as it feels kind of wasteful and disrespectful of our time; it can literally take you less time to generate the comment than for us to read it, and the only information you added is whatever was in the (presumably much shorter) prompt. If you'd written it yourself, it may or may not be interesting and correct, but I'd at least know that someone cared enough to write it and all of it made sense from that person's perspective. Sometimes I am interested in an LLM's take on a topic, but not when browsing a forum for humans.
I'm sorry but if you're blaming some text on a website to be "disrespectful of our time", I don't know what to say to you.
I stand behind everything on my comment and I have engaged in good faith with every single reply to it here (even though none of them talk about anything in the comment itself).
Go through my profile, see how I engage with people and tell me again I'm AI.
If you do not have anything to say to the subject matter of a comment and just have personal snide remarks, I do think it's a waste of your time but do not blame me for it or tell me to leave the platform.
Typing this comment right now is a waste of time for me but I do not feel the need to grandstand over it as if there's a massive opportunity cost to it. I'm a human writing/interacting in a "forum for humans."
I didn't say or think that your account was AI-run, and I didn't tell you to leave the platform. I just tried to explain why your comment might have annoyed people and triggered negative responses (while agreeing that the rude ones were inappropriate).
Sure, cheers then. I don't care about negative responses if they're negative just because they think it's AI-generated, without having to say anything substantial on the actual comment or the article. I have demonstrated my willingness to engage in good faith but those comments have not.
If negative responses have no substance behind it, it makes no sense to care about them or take them seriously.
Also, the fact that you assume it takes more time reading that comment than it took for me to write is pretty weird (I still don't get what was so wrong about the comment that simply reading it is a waste of time to people).
> Your code is now out of sync with the Spec
Is there even a sync to be had? The same prompt to the same LLM at different times will yield different artifacts, even if you were to save and re-use the seed.
Is there a link to the specification and the resulting generated code ? I skimmed through the article and the author's github profile, but couldn't find anything related.
Seems like a serious oversight if this is your selling point.
Their docs say:
> Allium has no compiler and no runtime. It is purely descriptive, defined entirely by its documentation
From what I understand that means there is no spec, no parser, it's just vibe evaluated by LLMs.
Sorry, I meant the spec and code of the toy project described in the post.
I don't necessarily expect their secret sauce to be openly available. But given the grandiose claims made there, I do expect something to back it up other than a corpospeak "trust me bro".
This post appears to essentially be an ad for Allium, so I’m gonna focus on that.
I don’t fully understand how allium solves the inherent issues with markdown, i.e. free-form language specs. It appears to have some loose syntactical rules, but outside of that relies on comments which are, again, free-form. Moreso, the “Resolved” portions remind of me how LLM agents like Claude code just go and edit existing doc files and make a mess of repetitive and overlapping thoughts.
> Allium has no compiler and no runtime. It is purely descriptive, defined entirely by its documentation
That seems ill advised at best.
Over a weekend, between board games and time with my kids, Claude and I built a distributed system with Byzantine fault tolerance, strong consistency and crash recovery under arbitrary failures. I described the behaviour I wanted in Allium, worked through the bugs conversationally and didn't write a line of implementation code.
I don't see why you need to bring your kids into this, and as a parent any suggestion of being distracted by tech during time with the kids raises my suspicions.
We have a strict no laptops and no phones rule when the kids are around (unless we're specifically doing something with them using those tools - looking at the weather forecast, or looking up some information).
"I can prompt AI while playing with the kids" is not a future I want.
They said "between", not "during"; I think the point was that they didn't spend a full weekend with Claude.
edit: anyone care to explain why this is a bad comment, rather than just downvoting? The GP comment says "Over a weekend, between board games and time with my kids,", and the parent comment lectures them based on an obvious strawman: "I can prompt AI while playing with the kids"
I'll bite.
"between time with the kids" is another person's "during time with the kids".
For me, "between time with the kids" means my kids are engaged in another activity that does not require my input until they are done with it. Whatever I am doing during this time also is typically very interruptible, so I am ready to help the kids along to their next "thing" (the joys of being a parent!). On a typical weekend (my oldest is 7), I'll get maybe 2 hours of this time during the time my kids are awake.
That all makes sense, but I don't see how it supports your uncharitable read of the original comment. Both because they could very easily have meant something a bit different (we have no idea how old the kids are or whether they were even around all weekend; maybe there were times when they were at friends' houses or similar), and because vibe coding could be the interruptible activity they do while the kids are busy. Maybe you feel strongly that it's not a legitimate use of that kind of semi-downtime, but we don't know if that's what they were talking about in the first place.
"Hours spent on the project" would be a much more useful metric, with no confounding variables. As it is, the mere mention of interleaving time with your kids and time engrossed in tech hits a nerve at time when IMO too many parents are doing this already, and lends unwarranted validity to the idea.