5 ms·

$85,000 in tokens later: What I learned from scaling agentic coding at Lovable

21 points by aliclark 3mo ago

aliclark 3mo ago
I am not the author but Lovable tells me I'm in the top 10% of users. It's a great tool. Not affiliated in any way.
justinclift 3mo ago
> Reviewing AI-written code line-by-line isn't practical or a good use of anyone's time. And the usual answer to problems created by the use of AI is to use more AI, so you switch to AI reviews by default. Ugh. Sure, for non-critical stuff that might be acceptable, but for anyone working on core banking or infrastructure PLEASE don't be doing this.
- alunchbox 3mo ago
  It depends I imagine. Core bankend bank system, database, transactions and such sure, but what about the web and mobile client? I'm down to get better customer experience if they can ship faster without breaking the bank. What about fraud detection and prevention? I hope to think I'm not alone in that perspective that it's acceptable in those cases.
  - hoppp 3mo ago
    Its fine for web and mobile as long as you can't get sued if data leaks or you get hacked.
    - kelseyfrog 3mo ago
      Fortunately AI lawyers are much cheaper than real lawyers.
      - qmarchi 3mo ago
        Except that any lawyer worth their salt won't touch AI for fear of getting disbarred.
      - justinclift 3mo ago
        For good reason. :) This is a website collecting the AI mishaps in the law field around the world: https://www.damiencharlotin.com/hallucinations/ https://www.damiencharlotin.com/hallucinations/ It seems pretty extensive (1696 cases identified so far), and appears to be updated regularly. :)
      - mhitza 3mo ago
        And AI layers are so good that even OpenAI complained that it would be unable to navigate compliance if every US state can have it's own AI legislation. And they have infinite access to AI that passed the bar exam and is brandished with "PhD-level intelligence."
  - justinclift 3mo ago
    > What about fraud detection and prevention? It's easy to see both pros and cons of that. I guess, like most stuff, it depends upon the appetite for risk vs the downsides of false positives/negatives and incorrect analysis.
  - fragmede 3mo ago
    Fraud detection and prevention have been using AI to determine things long before ChatGPT came on the scene. Of course, they call it machine learning, but it's all linear algebra under the hood. Just that ChatGPT has a really big hood.
- altmanaltman 3mo ago
  For critical stuff, there is already a ton of regulations in place that prevent this kind of thing from happening. Why do people think programmers working in such industries have no oversight and can ship whatever they want and however they want? It simply doesn't work that way in those industries.
  - justinclift 3mo ago
    > Why do people think programmers working in such industries have no oversight and can ship whatever they want and however they want? I've worked in those industries, and there's a large gap between the theory of what you're saying vs the practical reality when the CxO suite are pushing AI on everyone. :(
    - altmanaltman 3mo ago
      :( indeed if that's the case
      - justinclift 3mo ago
        Unfortunately, it is the case. And I fully agree with you.
hoppp 3mo ago
[dead]
willtemperley 3mo ago
Instead of employing an engineer for a year we burned an obscene amount of resources to generate code which will enable vibe coders to burn more resources. But we’re lovable! Cute smile. Heart emoji.
- nylonstrung 3mo ago
  50M projects built on Lovable and somehow I've never seen any of them
  - realusername 3mo ago
    Because these things do not scale beyond a landing page, it's the modern Frontpage
StrLght 3mo ago
> During the first week of June I merged 293 PRs, and have found no production defects tracing back to those changes so far. The latter part is a bit of good luck — I think 2-3 minor and 1 major defect would be acceptable for this volume. At this point, articles about LLMs paired with meaningless metrics have become a classic combo. I get that it's typical corporate BS, but publishing this widely is just weird. "Look, my productivity is skyrocketing according to a chart that only my manager cares about!" Mitchell Hashimoto put it well: https://xcancel.com/mitchellh/status/2071971627748020409 https://xcancel.com/mitchellh/status/2071971627748020409
- tangenter 3mo ago
  “Spotify ships 4,500 production deploys a day” LMAO
altmanaltman 3mo ago
I had an idea for a video long time ago. The idea was to try building lovable using lovable and then building another lovable within that using it. They market to normal people as if anything is possible but its very disingenous to offer a no code platform with the type of marketing they do. They are not alone, selling people the dream of "just use this to make a million dollar app!" while they make money off them even if the app works out or not. These platforms should do a revenue-share pricing where all these amazing apps created on the platform should only pay it if the app actually generates revenue if they really buy their marketing.
gershy 3mo ago
I think this article may have been a lot more insightful if it gave some insight into what product goals were achieved.
- tempodox 3mo ago
  You’d have to ask the agent, but it looks like they didn’t prompt for that.
vivzkestrel 3mo ago
- i swear to god it has been 3 years since LLMs came out - i still have no idea what people are running for more than 5 mins - if you are sitting and writing 20000 page requirement documents for your next project and having agentic AI agents create the whole damn project from scratch, you are doing it all wrong - you ll end up eroding all your skills that translate requirements to code and worse you are dependent on these so called couple of frontier labs - in about 5-10 years you are going to see absolutely horrible effects of this LLM stuff at scale when most engineers wont be able to write a script tag in html without an LLM - mark my words
- jdjdkrjdjjdj 3mo ago
  [flagged]
  - vivzkestrel 3mo ago
    - wrong comparison - ever heard the word called threshold? there is always a threshold for everything in this universe - when you say i dont need to manually do math anymore because i use a calculator, the threshold for offloading your thinking is very very low - when you say i dont need assembly language anymore because i learnt python, once again the threshold for offloading your thinking to an external source is very very slow because you are still thinking how that program works in python - but now you are saying here is a document that describes how the system works and let the agent generate the code - you are not thinking at all here are you? - even if you were to claim that writing requirement documents is a form of thought, you are still offloading the parts of your brain that were capable of translating requirement documents to code - i can still add and subtract 6 digit numbers effortlessly and rapidly without a calculator because i stopped using one long time back - can you say the same about the current generation of engineers 10 yrs from now?
    - jdjdkrjdjjdj 3mo ago
      It was a yes or no question. You didn't answer it.
siren2026 3mo ago
This article is very long with almost no content. Except "spending more tokens is good". Obviously the author got something to sell us.
- sevenseacat 3mo ago
  They went from 20-30 merged PRs a week to 150! What's in them? Who knows, no-one has read any of them
paradox242 3mo ago
What was the end result of all of this exactly? Seems like 85k and all you have is a masturbatory blog post.
ImPostingOnHN 3mo ago
> Human review is an exception I see quite a few bloggers saying this. Not super surprising coming from a vibe coding company trying to encourage vibe coding. The problem with applying this in other companies is that human review by the vibe coder is the exception. It still gets human reviewed, just by a tired human who is jaded at daily getting numerous 20 page PR review requests with 20 distinct unrelated changes because the "author" figured as long as they're making a change with AI they might as well say yes to every suggested follow-on prompt and as long as they're making a PR why not include all these changes in it and the code is moved all around for no reason and every line shows changes because of formatting changes so we'll have to add more rules to the linter now which we never needed before and the functions are renamed and reimplemented and replaced and removed because of course it's just better that way and the documentation is full of wrong assumptions and emojis and oh god the comments have emojis too and it's ok they can implement those fixes just give them a moment so they can add another commit to the 20 fix commits in the PR and the commit they push doesn't have anything to do with the thing it's supposed to fix did they even look at this ok just talk to them and oh they've been away on chat for hours and oh they're an innovator look at all those commits made and PRs opened and oh geez if those old geezers would stop holding up the PRs with slow human brains and being blockers maybe they'd be the ones getting the promotions and the fast-tracks and the innovator awards
- zacharylee 3mo ago
  [flagged]
nylonstrung 3mo ago
When I go to a "dev tools" site and landing page is just a prompt that's when I reach for my revolver
esailija 3mo ago
If you don't human review anything and no human reads the codebase guess what I could do hundreds of PRs in an hour and appear the most productive while not doing any real work at all. These kind of idiotic companies sound like great places to work for as overemployed .
NewsaHackO 3mo ago
I wonder how much he burned on this article.
Stitch4223 3mo ago
Let’s screw some numbers to make your Codex subscription look as impressive. By spending on the $200 plan, you get 20x the amount of tokens. So effectively $4,000 if you’d buy it at the worst way possible. Now do so for a year: $48,000 on tokens. To reach the conclusion of the article, you could just also have used the free tier. “Was it worth it? My current productivity is simply beyond the reach of old unassisted development techniques. I feel there's no going back in that sense.” Money spent versus value gained seems pretty low.
tempodox 3mo ago
Was that an education that was worth $85k?

21 points by aliclark 3mo ago

I am not the author but Lovable tells me I'm in the top 10% of users. It's a great tool. Not affiliated in any way.

> Reviewing AI-written code line-by-line isn't practical or a good use of anyone's time. And the usual answer to problems created by the use of AI is to use more AI, so you switch to AI reviews by default. Ugh. Sure, for non-critical stuff that might be acceptable, but for anyone working on core banking or infrastructure PLEASE don't be doing this.

alunchbox 3mo ago
It depends I imagine. Core bankend bank system, database, transactions and such sure, but what about the web and mobile client? I'm down to get better customer experience if they can ship faster without breaking the bank. What about fraud detection and prevention? I hope to think I'm not alone in that perspective that it's acceptable in those cases.
- hoppp 3mo ago
  Its fine for web and mobile as long as you can't get sued if data leaks or you get hacked.
  - kelseyfrog 3mo ago
    Fortunately AI lawyers are much cheaper than real lawyers.
    - qmarchi 3mo ago
      Except that any lawyer worth their salt won't touch AI for fear of getting disbarred.
    - justinclift 3mo ago
      For good reason. :) This is a website collecting the AI mishaps in the law field around the world: https://www.damiencharlotin.com/hallucinations/ https://www.damiencharlotin.com/hallucinations/ It seems pretty extensive (1696 cases identified so far), and appears to be updated regularly. :)
    - mhitza 3mo ago
      And AI layers are so good that even OpenAI complained that it would be unable to navigate compliance if every US state can have it's own AI legislation. And they have infinite access to AI that passed the bar exam and is brandished with "PhD-level intelligence."
- justinclift 3mo ago
  > What about fraud detection and prevention? It's easy to see both pros and cons of that. I guess, like most stuff, it depends upon the appetite for risk vs the downsides of false positives/negatives and incorrect analysis.
- fragmede 3mo ago
  Fraud detection and prevention have been using AI to determine things long before ChatGPT came on the scene. Of course, they call it machine learning, but it's all linear algebra under the hood. Just that ChatGPT has a really big hood.
altmanaltman 3mo ago
For critical stuff, there is already a ton of regulations in place that prevent this kind of thing from happening. Why do people think programmers working in such industries have no oversight and can ship whatever they want and however they want? It simply doesn't work that way in those industries.
- justinclift 3mo ago
  > Why do people think programmers working in such industries have no oversight and can ship whatever they want and however they want? I've worked in those industries, and there's a large gap between the theory of what you're saying vs the practical reality when the CxO suite are pushing AI on everyone. :(
  - altmanaltman 3mo ago
    :( indeed if that's the case
    - justinclift 3mo ago
      Unfortunately, it is the case. And I fully agree with you.

hoppp 3mo ago

[dead]

willtemperley 3mo ago

Instead of employing an engineer for a year we burned an obscene amount of resources to generate code which will enable vibe coders to burn more resources. But we’re lovable! Cute smile. Heart emoji.

nylonstrung 3mo ago
50M projects built on Lovable and somehow I've never seen any of them
- realusername 3mo ago
  Because these things do not scale beyond a landing page, it's the modern Frontpage

StrLght 3mo ago

> During the first week of June I merged 293 PRs, and have found no production defects tracing back to those changes so far. The latter part is a bit of good luck — I think 2-3 minor and 1 major defect would be acceptable for this volume. At this point, articles about LLMs paired with meaningless metrics have become a classic combo. I get that it's typical corporate BS, but publishing this widely is just weird. "Look, my productivity is skyrocketing according to a chart that only my manager cares about!" Mitchell Hashimoto put it well: https://xcancel.com/mitchellh/status/2071971627748020409 https://xcancel.com/mitchellh/status/2071971627748020409

tangenter 3mo ago
“Spotify ships 4,500 production deploys a day” LMAO

altmanaltman 3mo ago

I had an idea for a video long time ago. The idea was to try building lovable using lovable and then building another lovable within that using it. They market to normal people as if anything is possible but its very disingenous to offer a no code platform with the type of marketing they do. They are not alone, selling people the dream of "just use this to make a million dollar app!" while they make money off them even if the app works out or not. These platforms should do a revenue-share pricing where all these amazing apps created on the platform should only pay it if the app actually generates revenue if they really buy their marketing.

gershy 3mo ago

I think this article may have been a lot more insightful if it gave some insight into what product goals were achieved.

tempodox 3mo ago
You’d have to ask the agent, but it looks like they didn’t prompt for that.

vivzkestrel 3mo ago

- i swear to god it has been 3 years since LLMs came out - i still have no idea what people are running for more than 5 mins - if you are sitting and writing 20000 page requirement documents for your next project and having agentic AI agents create the whole damn project from scratch, you are doing it all wrong - you ll end up eroding all your skills that translate requirements to code and worse you are dependent on these so called couple of frontier labs - in about 5-10 years you are going to see absolutely horrible effects of this LLM stuff at scale when most engineers wont be able to write a script tag in html without an LLM - mark my words

jdjdkrjdjjdj 3mo ago
[flagged]
- vivzkestrel 3mo ago
  - wrong comparison - ever heard the word called threshold? there is always a threshold for everything in this universe - when you say i dont need to manually do math anymore because i use a calculator, the threshold for offloading your thinking is very very low - when you say i dont need assembly language anymore because i learnt python, once again the threshold for offloading your thinking to an external source is very very slow because you are still thinking how that program works in python - but now you are saying here is a document that describes how the system works and let the agent generate the code - you are not thinking at all here are you? - even if you were to claim that writing requirement documents is a form of thought, you are still offloading the parts of your brain that were capable of translating requirement documents to code - i can still add and subtract 6 digit numbers effortlessly and rapidly without a calculator because i stopped using one long time back - can you say the same about the current generation of engineers 10 yrs from now?
  - jdjdkrjdjjdj 3mo ago
    It was a yes or no question. You didn't answer it.

siren2026 3mo ago

This article is very long with almost no content. Except "spending more tokens is good". Obviously the author got something to sell us.

sevenseacat 3mo ago
They went from 20-30 merged PRs a week to 150! What's in them? Who knows, no-one has read any of them

paradox242 3mo ago

What was the end result of all of this exactly? Seems like 85k and all you have is a masturbatory blog post.

ImPostingOnHN 3mo ago

> Human review is an exception I see quite a few bloggers saying this. Not super surprising coming from a vibe coding company trying to encourage vibe coding. The problem with applying this in other companies is that human review by the vibe coder is the exception. It still gets human reviewed, just by a tired human who is jaded at daily getting numerous 20 page PR review requests with 20 distinct unrelated changes because the "author" figured as long as they're making a change with AI they might as well say yes to every suggested follow-on prompt and as long as they're making a PR why not include all these changes in it and the code is moved all around for no reason and every line shows changes because of formatting changes so we'll have to add more rules to the linter now which we never needed before and the functions are renamed and reimplemented and replaced and removed because of course it's just better that way and the documentation is full of wrong assumptions and emojis and oh god the comments have emojis too and it's ok they can implement those fixes just give them a moment so they can add another commit to the 20 fix commits in the PR and the commit they push doesn't have anything to do with the thing it's supposed to fix did they even look at this ok just talk to them and oh they've been away on chat for hours and oh they're an innovator look at all those commits made and PRs opened and oh geez if those old geezers would stop holding up the PRs with slow human brains and being blockers maybe they'd be the ones getting the promotions and the fast-tracks and the innovator awards

zacharylee 3mo ago
[flagged]

nylonstrung 3mo ago

When I go to a "dev tools" site and landing page is just a prompt that's when I reach for my revolver

esailija 3mo ago

If you don't human review anything and no human reads the codebase guess what I could do hundreds of PRs in an hour and appear the most productive while not doing any real work at all. These kind of idiotic companies sound like great places to work for as overemployed .

NewsaHackO 3mo ago

I wonder how much he burned on this article.

Stitch4223 3mo ago

Let’s screw some numbers to make your Codex subscription look as impressive. By spending on the $200 plan, you get 20x the amount of tokens. So effectively $4,000 if you’d buy it at the worst way possible. Now do so for a year: $48,000 on tokens. To reach the conclusion of the article, you could just also have used the free tier. “Was it worth it? My current productivity is simply beyond the reach of old unassisted development techniques. I feel there's no going back in that sense.” Money spent versus value gained seems pretty low.

tempodox 3mo ago

Was that an education that was worth $85k?