- 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
- 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?
Look my point here is not to troll. My point here is to make you think. I learnt assembly and I never want to do it again, I learnt to code in many languages (compiled and interpreted) and you know what, there has not been an issue i found that required me to dig into the assembly of a compiled language (there are edge cases if there was a compiler bug, but I have not seen them myself).
Now the change from assembly to compiled languages was a step change. The same is true now. What was once hard is now easy. And again we must adapt.
For you I think your identity has been tied up in being good at what you do ... I was the same... The issue is now, that we need to find something harder to fill our minds...
Finally, I want to remind you that at one point in our history a computer was a job. I think programmer/software engineer is going the same way.
In 10 years I hope we are not doing the same mundane crap as what we were .. There are so many other interesting things we could be filling our brains with!!!
> there has not been an issue i found that required me to dig into the assembly of a compiled language (there are edge cases if there was a compiler bug, but I have not seen them myself).
The compiler did exactly what you said because the programming language forced you to be exact. There was no room for misunderstanding.
This will never be the case with an LLM even if it becomes infallible (it won't). It's you who is fallible and sloppy until you force yourself to be precise. A programming language can help with that.
Eh this is also not true. Non determinism doesn’t disqualify something from being useful and necessary. If you are a manager, you employ people (who are non deterministic) to get stuff done.
In fact entirety of humanity has been delegating work to non deterministic specialists. Like how I delegate making food to non deterministic supply chains and capitalism. No one complains about skill atrophy here I’m sure.
“But humans can learn..” well sure agents can learn too, adapt the markdown files.
“But humans are accountable..” distinction without a difference
My argument was not that LLMs aren't useful but that a compiler is a completely misleading anamlogy.
You compared them to humans. Humans live in the world, have a world model, meet people for lunch and exchange ideas. Agents read markdown files and code.
Simanwords, you spend a lot of time on HN promoting LLMs and being dismissive of humans in comparison. I have to ask:
- to answer your question, i did not start with assembly
- i started in 2010s with c++ directly but I did go back and learn how to do 8086 opcodes and while i am not good at it, i am still learning it slowly
- mundane stuff? algorithms, business logic, critical thinking, architectural understanding is being outsourced to LLMs at an unprecedented rate and there is nothing mundane about any of this
- you can sit and talk about how 10 yrs down the line everyone has to sit and write english paragraphs of how they want their website to look like and nobody needs to write code but I am going to have to back out of this one
- Writing code from requirements is an art as much as a critical skill and bringing an LLM to do so doesnt bode well for the entire industry
- Remember that concept of smart terminals and dumb terminals that used to be discussed about long time back. How dumb terminals merely offloaded everything to the server while they did nothing.
- Frontier labs are turning humans into dumb terminals while they wield data, power and critical thinking from the masses at scale
- This kinda power and leverage should not be handed out especially to frontier labs controlled by a few people
- Even if you are one of those few insightful people out there who will sit and actually discuss what the function should do and then review the generated code, vast majority of programmers are literally copy pasting vibe coding stuff day in and day out now
- This ll create a special type of deficiency a few years down the line in every major corporation out there while concentrating even more power into the hands of a few frontier labs
- Remember that meme image of an ape slowly evolving into a human and the in the last step, it says "oops we need to turn back one step". This honestly is that one step. I think the software industry has taken a messy step here
Very respectfully this all sounds incoherent. You have also untastefully mixed in some concern over frontier labs having too much power - this is orthogonal to the discussion.
Your reckoning for when we will look back to this time and regret using LLMs won’t happen, as much as you crave for it.
- Here’s what will happen: agents allow you to understand less and less of the code (directionally) and allow you to focus on high level design.
- this won’t make it such that no one will look at code, just like some still look at assembly
- LLMs won’t allow you to never look at code (yet) but it pushes you in such a way that you may need to read less and less of your code
Also the fact that you have written about how writing code from requirements is an art. This is just not gonna be true anymore with LLMs. Very few in enterprise is writing artful code - they are getting stuff done. The art part is understanding the requirements and high level design. If you’ve ever worked in enterprise, there’s always a few people who nitpick on code reviews to artify code - these sorts eventually learn and forget this tic.
They are right though. The entire history of human civilisation involves a thing called “specialisation”. The primary reason we were able to prosper as humans is because we all dont have to know how to procure our food. Did humanity lose the skills for making food as people specialised? No.
same thing with memory management - as Java became popular, should the concern have been towards losing skills on memory management? That’s stupid.
LLMs don’t allow you (yet) to not understand the code completely. But it allows you to understand less and less directionality. We are getting there. I don’t read all code anymore, just high level design.
This kind of performative concern over skill atrophy is annoying.
Non determinism doesn’t discount something from still being useful. Companies outsource projects to consultancies all the time. Imagine you wanted to rebrand your logo - you hire a freelance guy to make it. He’s non deterministic. Does this stop you from hiring him? Why or why not?
Would you call him or her "the next step in the evolution of the brush" or something like that? Would you call an assistant who reads from books to or for you the "next level of literacy"? Does objecting to such statements -- because they're false and insulting the intelligence of the reader on the surface, and outright sinister in their long term implications -- mean I am "anti" either of these or other things?
No x3. It also doesn't mean that I think if you have taste, even maybe some artistic skill, you will instantly lose it all when you hire that one person for that one logo, and so on. It just means what it says, that this comparison is nonsense. I don't have to justify myself for recognizing that, it should be the other way around if anything.
I actually ended up learning a little assembly (Knuth's new one MMIX) and it's actually been a really useful skill, as it gives me a (lot) more sympathy for the machine.
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.
> 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.
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.
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."
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.
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.
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.
> 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. :(
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.”
> 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!"
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 .
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.
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
- 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
Can you write assembly?
- 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?
It was a yes or no question. You didn't answer it.
Look my point here is not to troll. My point here is to make you think. I learnt assembly and I never want to do it again, I learnt to code in many languages (compiled and interpreted) and you know what, there has not been an issue i found that required me to dig into the assembly of a compiled language (there are edge cases if there was a compiler bug, but I have not seen them myself).
Now the change from assembly to compiled languages was a step change. The same is true now. What was once hard is now easy. And again we must adapt.
For you I think your identity has been tied up in being good at what you do ... I was the same... The issue is now, that we need to find something harder to fill our minds...
Finally, I want to remind you that at one point in our history a computer was a job. I think programmer/software engineer is going the same way.
In 10 years I hope we are not doing the same mundane crap as what we were .. There are so many other interesting things we could be filling our brains with!!!
> there has not been an issue i found that required me to dig into the assembly of a compiled language (there are edge cases if there was a compiler bug, but I have not seen them myself).
The compiler did exactly what you said because the programming language forced you to be exact. There was no room for misunderstanding.
This will never be the case with an LLM even if it becomes infallible (it won't). It's you who is fallible and sloppy until you force yourself to be precise. A programming language can help with that.
Eh this is also not true. Non determinism doesn’t disqualify something from being useful and necessary. If you are a manager, you employ people (who are non deterministic) to get stuff done.
In fact entirety of humanity has been delegating work to non deterministic specialists. Like how I delegate making food to non deterministic supply chains and capitalism. No one complains about skill atrophy here I’m sure.
“But humans can learn..” well sure agents can learn too, adapt the markdown files.
“But humans are accountable..” distinction without a difference
My argument was not that LLMs aren't useful but that a compiler is a completely misleading anamlogy.
You compared them to humans. Humans live in the world, have a world model, meet people for lunch and exchange ideas. Agents read markdown files and code.
Simanwords, you spend a lot of time on HN promoting LLMs and being dismissive of humans in comparison. I have to ask:
Whose side are you on?
- to answer your question, i did not start with assembly
- i started in 2010s with c++ directly but I did go back and learn how to do 8086 opcodes and while i am not good at it, i am still learning it slowly
- mundane stuff? algorithms, business logic, critical thinking, architectural understanding is being outsourced to LLMs at an unprecedented rate and there is nothing mundane about any of this
- you can sit and talk about how 10 yrs down the line everyone has to sit and write english paragraphs of how they want their website to look like and nobody needs to write code but I am going to have to back out of this one
- Writing code from requirements is an art as much as a critical skill and bringing an LLM to do so doesnt bode well for the entire industry
- Remember that concept of smart terminals and dumb terminals that used to be discussed about long time back. How dumb terminals merely offloaded everything to the server while they did nothing.
- Frontier labs are turning humans into dumb terminals while they wield data, power and critical thinking from the masses at scale
- This kinda power and leverage should not be handed out especially to frontier labs controlled by a few people
- Even if you are one of those few insightful people out there who will sit and actually discuss what the function should do and then review the generated code, vast majority of programmers are literally copy pasting vibe coding stuff day in and day out now
- This ll create a special type of deficiency a few years down the line in every major corporation out there while concentrating even more power into the hands of a few frontier labs
- Remember that meme image of an ape slowly evolving into a human and the in the last step, it says "oops we need to turn back one step". This honestly is that one step. I think the software industry has taken a messy step here
Very respectfully this all sounds incoherent. You have also untastefully mixed in some concern over frontier labs having too much power - this is orthogonal to the discussion.
Your reckoning for when we will look back to this time and regret using LLMs won’t happen, as much as you crave for it.
- Here’s what will happen: agents allow you to understand less and less of the code (directionally) and allow you to focus on high level design.
- this won’t make it such that no one will look at code, just like some still look at assembly
- LLMs won’t allow you to never look at code (yet) but it pushes you in such a way that you may need to read less and less of your code
Also the fact that you have written about how writing code from requirements is an art. This is just not gonna be true anymore with LLMs. Very few in enterprise is writing artful code - they are getting stuff done. The art part is understanding the requirements and high level design. If you’ve ever worked in enterprise, there’s always a few people who nitpick on code reviews to artify code - these sorts eventually learn and forget this tic.
You created a throwaway to “not troll” and post the same three tired tropes every tokenmaxxing vibe coder trots out
They are right though. The entire history of human civilisation involves a thing called “specialisation”. The primary reason we were able to prosper as humans is because we all dont have to know how to procure our food. Did humanity lose the skills for making food as people specialised? No.
same thing with memory management - as Java became popular, should the concern have been towards losing skills on memory management? That’s stupid. LLMs don’t allow you (yet) to not understand the code completely. But it allows you to understand less and less directionality. We are getting there. I don’t read all code anymore, just high level design.
This kind of performative concern over skill atrophy is annoying.
Apples and oranges, since there is not determinism. That this argument gets trotted out so often is part of the dumbing down IMO.
FreeBSD took some 30 years to get deterministic builds
Non determinism doesn’t discount something from still being useful. Companies outsource projects to consultancies all the time. Imagine you wanted to rebrand your logo - you hire a freelance guy to make it. He’s non deterministic. Does this stop you from hiring him? Why or why not?
Would you call him or her "the next step in the evolution of the brush" or something like that? Would you call an assistant who reads from books to or for you the "next level of literacy"? Does objecting to such statements -- because they're false and insulting the intelligence of the reader on the surface, and outright sinister in their long term implications -- mean I am "anti" either of these or other things?
No x3. It also doesn't mean that I think if you have taste, even maybe some artistic skill, you will instantly lose it all when you hire that one person for that one logo, and so on. It just means what it says, that this comparison is nonsense. I don't have to justify myself for recognizing that, it should be the other way around if anything.
I actually ended up learning a little assembly (Knuth's new one MMIX) and it's actually been a really useful skill, as it gives me a (lot) more sympathy for the machine.
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.
50M projects built on Lovable and somehow I've never seen any of them
Because these things do not scale beyond a landing page, it's the modern Frontpage
> 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.
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.
Its fine for web and mobile as long as you can't get sued if data leaks or you get hacked.
Fortunately AI lawyers are much cheaper than real lawyers.
For good reason. :)
This is a website collecting the AI mishaps in the law field around the world:
https://www.damiencharlotin.com/hallucinations/
It seems pretty extensive (1696 cases identified so far), and appears to be updated regularly. :)
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."
Except that any lawyer worth their salt won't touch AI for fear of getting disbarred.
> 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.
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.
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.
> 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. :(
:( indeed if that's the case
Unfortunately, it is the case. And I fully agree with you.
When I go to a "dev tools" site and landing page is just a prompt that's when I reach for my revolver
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.
I wonder how much he burned on this article.
I think this article may have been a lot more insightful if it gave some insight into what product goals were achieved.
This article is very long with almost no content. Except "spending more tokens is good". Obviously the author got something to sell us.
They went from 20-30 merged PRs a week to 150! What's in them? Who knows, no-one has read any of them
> 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
“Spotify ships 4,500 production deploys a day” LMAO
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 .
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.
> 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
[dead]
[dead]
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.