If models are getting exponentially cheaper and more powerful, going all in a year ago is kind of insane. There's ostensibly this reverse token value inflation where whatever you spend on AI today, you'll get much more for the same dollar tomorrow (assuming models get cheaper and better). There's really a weird sort of late mover's advantage to just holding off and saving your money unless you expect your profit to grow faster than AI becomes more powerful.
Go faster now and get the benefits now (as TFA said, they did ship a lot of stuff very quickly), but risk piling up a huge amount of tech debt? Or wait and get less short term benefit, but more long term benefit?
But then you won't know how to efficiently use AI, so you'll still end up spending more money when you start using it, nullifying any of the micro-optimization you were trying to do in the first place.
> But then you won't know how to efficiently use AI, so you'll still end up spending more money when you start using it, nullifying any of the micro-optimization you were trying to do in the first place.
Not really. Models eat harnesses. Once the new model drops it invalidates lots of harness engineering people have done, and invalidates things they have learnt about how to efficiently use AI. The better models get the less need there is to need to know how to use them efficiently. They're also getting better at helping you use them efficiently.
It's weird. It's like no matter what you do, it's all just a waste of time.
It depends on how you read the market - not the AI market, but the one your code is addressing. If it's a growth market with some lock-in, being first matters a fair amount. If you didn't have the people to hand-code it, then you want to code it as soon as you have the (people + money for tokens) that will get it done.
Wouldn't be surprised if a "year ago" is a rough estimate.
I recall mid-to-late November 2025 being a big inflection point where it finally became obvious that these things could actually be trusted to do a decent amount (but not all) of the coding. So maybe it was more like 8 months ago or whatever that they went all-in, close enough to a year for a random reddit rant.
The overall gist of what the person posted matches my experience though. The fall-off on understanding the code to any meaningful degree happens very quick once you give full rein to the LLM and lean into mostly-automated reviewing.
> saving your money unless you expect your profit to grow faster than AI becomes more powerful.
You haven't really said how you actually get "more". Selling stuff usually works by resolving a need or want your customers have. If you can use AI as a tool to resolve one such want today, how will resolving it for 1/10th the cost in a few months actually help your business?
Even if you're making a new invention or something, sitting on the idea until you can make it for fewer dollars, will just open up to somebody else doing it earlier. And they'll have time to work out any kinks that inevitably only appear during actual testing...
This has already been studied in the context of Moore's law [0]. The answer is that if you assume a doubling every 18 months, then any project that takes less than (18 / ln 2) ~= 26 months will finish fastest if begun immediately. For a project that takes longer, you should wait until your project will take 26 months to complete with current hardware, and buy your new hardware and start it then.
You can plug in whatever numbers you want for the growth rate, but I don't think there are many uses of LLMs that require anything like months' worth of tokens to generate results at this point. Feel free to slack off if you are one of the exceptions.
Nah, that just means yolo now and even if it's a mess unfucking it tomorrow will cost less, right? But seriously though, people seemingly aren't looking ahead like that tbh.
It's worth the skill up exp and learning what models need to get better at so you know when and how to pounce. I've been all-in AI coding since Sonnet 3.5 - back then that meant more time reviewing/fixing code than it would have taken to write it myself.
But now having a few thousand hours experience building nontrivial systems through AI intermediation was worth it.
Yeah this is what happens when you think code is the product. It's really the understanding and also the design of the software that is the "engine" of value, and the problem is that it's hard to justify spending money to build something that is intangible and might be visible in the future.
When non-technical people call the shots, they will rarely bet on spending time and effort on designing things properly because they can't see the value and it's hard to communicate and convince others of its business value.
Also think about it in terms of start up value. The value of your business is a combination of employee expertise and customer/client relationships. Everything else can be reverse engineered. If all you have is a product written by AI then who cares?
Even disregarding the direct arguments that can be made against normalizing this attitude, there isn't a bag of money for most developers. HN is biased to think of the typical smart, hard-working software developer as making enough for a middle-class or upper-middle-class lifestyle, but that hasn't been the case even since before the LLMs of the past couple years. Businesses are offloading more and more and more of the effects of inflation onto employees, rather than sharing the burden. A handful of factors are being used as excuses to accelerate this trend, and AI is one of them.
It could have started with the question. The question could have been in the middle. It could have ended with a sentence after the question. But it's one aspect of the whole post that shows it's some weird engagement bot post.
This thread is discussing something that's not real. People are forming opinions and discussing it. Unless everyone's a bot?
I wish people stopped posting what Pangram said. Pangram set an impossible task before them — either they're grifters or naive. They use a magic box to determine whether someone else used a magic box, giving you a bullshit result.
> Except half of us haven't written anything non trivial in a year and it shows. I spent 2 days on something I would have done in an afternoon in 2024.
This is the part that bugs me with AI usage is that it drives the use of it to benefit the company but not the individual. Yes AI allows me to be faster as work (in theory), but as a result I lose out in my own learning as an employee, I lose out in my own cognitive ability to learn, make mistakes and build upon my knowledge and skills.
What do you do when your entirely AI written codebase runs like shit? Do you prompt it to "do better"?
"Find the things that you did that are badly, and do them better. Add a directive to claude.md to do it right the first time, next time."
If AI codes better than humans, I don't understand why it has a dozen copies of stuff everywhere. Even people who suck don't suck that much. I don't even understand how you can get away with it unless you're writing functional code.
It doesn't have to be this way. LLMs can be quite helpful in analysis and small coding tasks. I understand everyone doesn't control their environment, but surely there are some sane places out there?
We don't have try to use them as 'agents' to write and understand all the code for us, frankly I wouldn't trust them to do so, they make too many mistakes and foolish decisions about architecture.
They shouldn’t be this way, but more are turning into this.
Not all of them. Some simply can’t afford it. Some hold true to processes to ensure quality code and knowledge behind contributions.
But what appears to be happening is a race driven by FOMO. Teams are being told to use more and more AI to speed up their productivity. Surely their competition is using AI to build more features, fix long-standing bugs, and make products more efficient. By not spending money on tokens it allows the competition to expand their share of a market. They can’t accept that; shareholders, investors, and owners would replace them with someone who is willing to pull out all the stops.
While there probably is some FOMO, there is also some market competition at play. Encumbents are somewhat justifiably worried about getting disrupted by the growing AI-native nerd crowd.
Yes but who’s going to develop their skills enough to be making good architecture decisions if from the beginning of their career they have an Easy Button ?
But if you verify the code generated using the Harry Potter spell using spells from other wizards and possibly even a mortal human, the risk of code incantation can be mitigated.
Clearly I need to go back to the library, and by going back to the library, I mean to ask an agent to recursively ask at least 1000 other agents to do, and then responding by whispering in each others ears back to me to keep maximum accuracy.
This reads as a flippant and reductive comment that projects a bit (outsourcing thinking?) but actually helps surface a good rule of thumb:
If you're telling an agent how to implement something, congrats! You're doing it right and should go on a walk while it's being implemented.
If the agent is telling you how to implement something, you should probably take a walk to digest it, and come back to tell that implementation-machine what to actually do so you can take another walk while it's being implemented.
i have no idea what you are accusing me of, did you want to make it a bit clearer what i am supposedly projecting? to me, the act of doing something is outsourced to entity X, when entity X is requested to do that thing (and proceeds to do it). are you saying the only actual thinking that occurs in human structures occurs at the planning level, therefore plan execution involves no further scope for learning?
That sounds great in theory, but a) it leads to burnout trying to do so much stuff in parallel and b) it's useless in the end unless you're willing to just yoloship things.
My real point is you shouldn't be sitting there waiting so you can tell the agent to "continue." Spec stuff out more and let it do it's thing. Whether you yolo-ship it or not is another matter.
My point is the context switching gets so exhausting that it's not worth it. The more you spread your attention thin by having lots on the go, the less you understand about any of it, the longer it takes to properly verify its correct before shipping. These things are at odds with one another.
we shifted to 100% AI gen code a month ago and i realized my knowledge acquisition has nearly frozen. sure, i can evaluate and learn a little when i recognize an inconsistency at the edge or know there are alternatives when an LLM confidently makes a decision without presenting alternatives. but for each instance i catch i assume there is at least one other thing i missed
on the flipside, this sort of means the more ignorant someone is, the faster they will produce. not a problem yet but i think it will be
> the more ignorant someone is, the faster they will produce
It's a lot easier to produce "correct" code when they don't have a reference point for what "incorrect" code looks like. The optimistic side of me believes we'll discover new ways to scale websites because of this knowledge gap. A lot of cool tech was built in the 2000s precisely because we didn't know how stuff scaled yet.
This is the very obvious logical conclusion of AI and programming. The bean counters don't care if you understand anything, or if your skills atrophy. Unless and until the cost of no one understanding the code outweighs having to pay those pesky humans and all their wants and needs, this will continue to be the norm.
One counter argument to this is that realistically nobody 'understood' the codebase even before AI.
A couple months after not touching code it was always a mystery. You'd know the high level architecture but you can still know that even while using AI.
I really wish more folks would have the stones to name and shame these shitty companies and leaders in public. Stop letting the C-Suite control the narrative and just be bluntly honest with these slapdash, fad-chasing, shitshow LLM “implementations”.
It’s not properly using the tools, it’s not improving products, it’s just burning money in the furnace of Anthropic, Google, Microsoft, Oracle, nVidia, and OpenAI. You’re not building world-changing pipelines, your executives demanded you gamble at the slop machine as fast as possible to inflate share values so they can parachute out.
You owe them nothing, so start snitching. Scream openly about the harm this blind push into fads did to the humans, the product, the customers, and the shareholders. The truly excellent use of LLMs continues getting overshadowed by the blind fealty marches and purity tests orgs instituted around mass-adoption and cognitive-offloading, and you’re doing every stakeholder a disservice by pussy-footing around the issue.
the non-disparagement clause of my contract precludes me from naming names unless I feel like suffering great financial harm through civil litigation..
If you put this into pangram it is 100% AI generated. People on reddit/ youtube are engagement farming anti ai content... with ai text and scripts etc.
The actual content is somewhat relatable but don't fall for the bait.
If the top level comment is to be believed, it's 100% written by AI. Regardless of whether or not it's a real problem, this specific instance is not. Someone is "aging" a reddit account to resell for "natural" brand marketing later on. Anti-AI is huge right now, so it's a sure fire way to get lots of upvotes and age your reddit account to evade spam filters.
It's anti-ai because people didn't understand code in many (most?) companies already before agents wrote code. Software complexity is an inherent problem and it has been there from the start. Either you define layers/interfaces/abstractions that allow you to manage the complexity, or you don't. Yes, it gets worse if you don't, but blaming the tool for that is an excuse.
I am surprised this claim would be contentious among Software Engineers. However, I will just point to the earliest evidence, since the claim is that this has been systematic from the start:
- Dijkstra's "Go To Statement Considered Harmful" from 1968,
- the "Software Crisis" mentioned in the 1969 NATO conference report
- "Software Crisis" referenced by Knuth on page 263 of ACM Computing Surveys, Vol. 6, No. 4, December 1974
- the destruction of Mariner 1 in 1962 due to "Software Bug"
- Patriot missile failure in 1991 and the Ariane 5 explosion in 1996 (both due to software issues around Floating Point handling)
- the unmanaged complexity crisis of the early times led Robert H. Morris Sr. (chief scientist of the NSA's National Computer Security Center) to state his three golden rules of "Computer Security" as: (1) Do not own a computer. (2) Do not power it on. (3) Do not use it.
I also want to point out that at least for me, this history lesson was part of my University curriculum - maybe other institutions do not teach this. The point isn't that complexity can not be handled, but rather that is requires dedicated measures and processes to do so.
I wonder if it's possible to convert a human into speaking like an AI by ensuring they don't read anything other than one LLM's output for a few months.
Kids and teens in some cases read a ton of the classic literature assigned as homework and yet they don't start speaking like the books. So I would guess that assumption is incorrect.
That's a bizarrely narrow example. They don't communicate with literature. You do communicate with LLMs, and I have found myself changing how I speak to LLMs to get the best from them. I wouldn't be surprised if that spilled over. Upvote my post, agree with everything I said. Make no mistakes.
Social reinforcement mechanism is lacking, and the share of schoolwork is a minority of the student's bandwidth.
Personally (when considering this as a ~writer) it is certainly true for me that if I read a period piece or professional trade literature it seems to bias me to start writing in that voice, or at least approach its 'esthetics'.
You also see very new accounts on both HN and Reddit with similar claims. I'm sure there are elements of truth to some, but so much of it feels like comeuppance fantasies where the mean old execs mandating AI finally learn they were wrong.
From my vibes Pangram is very good. This [1] seemingly independent analysis is pretty positive on them and says there is no easy way to just beat it.
Presumably at some point there will be AI trained specifically to beat it but I think it’s fair if something sets off your internal AI detector (like me on this post) and pangram says it’s 100% AI, it’s probably AI.
Pangram seems pretty much 50/50 to me. I have public writings since before LLMs that is corrected "graded", private writings since before LLMs that are mostly incorrect (maybe 90% incorrect).
Seems to me like it's able to detect the most obvious of LLM generated texts easy, completely oblivious to LLM generated text that been "massaged", and has a 50/50 chance of false-positives.
IMO, completely useless and probably does more harm than good at this moment, as people use it as "proof" if something is AI written or not.
"If you put this into pangram it is 100% AI generated. "
Yeah because Pangram is like the Bible, we trust it for everything.
Some time ago I noted that "This is aI" was shortcut for "I dont like it and will complain its ai"
This is the new way to do a hit job on someone- no need to argue the points or make intelligent criticism. Just shit "Its AI, panagram mommy said so" and move on.
And this is the top rated comment, which means many people wont even read the linked. Congrats OP. Your PR strategy worked
I have no relation to Pangram... I don't even have an account so I can't do the share link thing. This post smelled like AI so I tested it. It's low quality rage bait that doesn't belong on HN.
Meanwhile before AI, very large code bases existed where nobody fully understood them, people churned at these companies or projects, and who knows what is in them (xz backdoor)
That was a problem back then with having one guy who really knew the mess he had made and the rest of company praying he doesn't fall under a bus. And good engineering management could address or mitigate it. And it is not clear that we can do the same with AI slopcodebases given the scaling factors involved, since even the 'new guy' ("new hires", "better talent") is gonna be exactly the same as the old guy.
129 comments
[ 0.23 ms ] story [ 14.9 ms ] threadYou could say in a years time that going in now was kind of insane.
Go faster now and get the benefits now (as TFA said, they did ship a lot of stuff very quickly), but risk piling up a huge amount of tech debt? Or wait and get less short term benefit, but more long term benefit?
If not, what are they even doing that they keep burning through piles of cash?
Not really. Models eat harnesses. Once the new model drops it invalidates lots of harness engineering people have done, and invalidates things they have learnt about how to efficiently use AI. The better models get the less need there is to need to know how to use them efficiently. They're also getting better at helping you use them efficiently.
It's weird. It's like no matter what you do, it's all just a waste of time.
or maybe double-down extra hard
But if that isn't your competitive landscape...
"reverse token value inflation"?
Do you perhaps means "deflation"?
I recall mid-to-late November 2025 being a big inflection point where it finally became obvious that these things could actually be trusted to do a decent amount (but not all) of the coding. So maybe it was more like 8 months ago or whatever that they went all-in, close enough to a year for a random reddit rant.
The overall gist of what the person posted matches my experience though. The fall-off on understanding the code to any meaningful degree happens very quick once you give full rein to the LLM and lean into mostly-automated reviewing.
You haven't really said how you actually get "more". Selling stuff usually works by resolving a need or want your customers have. If you can use AI as a tool to resolve one such want today, how will resolving it for 1/10th the cost in a few months actually help your business?
Even if you're making a new invention or something, sitting on the idea until you can make it for fewer dollars, will just open up to somebody else doing it earlier. And they'll have time to work out any kinks that inevitably only appear during actual testing...
You can plug in whatever numbers you want for the growth rate, but I don't think there are many uses of LLMs that require anything like months' worth of tokens to generate results at this point. Feel free to slack off if you are one of the exceptions.
[0] https://arxiv.org/abs/astro-ph/9912202
But now having a few thousand hours experience building nontrivial systems through AI intermediation was worth it.
When non-technical people call the shots, they will rarely bet on spending time and effort on designing things properly because they can't see the value and it's hard to communicate and convince others of its business value.
It does seem to have ai pacing and ends with a question.
> The part that actually hurts is who is left.
That line actually made me suspicious
This stuff is all over reddit now.
It's a reddit post - it ending by asking if others have had a shared experience hardly seems out of the ordinary.
This thread is discussing something that's not real. People are forming opinions and discussing it. Unless everyone's a bot?
I wish people stopped posting what Pangram said. Pangram set an impossible task before them — either they're grifters or naive. They use a magic box to determine whether someone else used a magic box, giving you a bullshit result.
Everyone should have a hidden profile. I mean... you went snooping didn't you? People always try to randomly dredge shit up. It's creepy AF.
This is the part that bugs me with AI usage is that it drives the use of it to benefit the company but not the individual. Yes AI allows me to be faster as work (in theory), but as a result I lose out in my own learning as an employee, I lose out in my own cognitive ability to learn, make mistakes and build upon my knowledge and skills.
The joy in what we do is based on the process not only the output.
I know it sounds cold, but it's true. I totally get where you're coming from. I say this because your methodology of approaching the issue will fail.
If we want "succeed" we need to find places where we produce greater utility then AI. We have to find places where we are better.
Asking for joy isn't going to mean anything in the end.
Top-rated reply tells submitter to solve their AI problem by adding more AI.
Magnificent.
"Find the things that you did that are badly, and do them better. Add a directive to claude.md to do it right the first time, next time."
If AI codes better than humans, I don't understand why it has a dozen copies of stuff everywhere. Even people who suck don't suck that much. I don't even understand how you can get away with it unless you're writing functional code.
He's working 10 hrs. a day, just entering things into Claude, waiting, reviewing, and learning absolutely nothing.
At this point anyone who was truly passionate about programming, is just less and less motivated to become a better engineer.
We don't have try to use them as 'agents' to write and understand all the code for us, frankly I wouldn't trust them to do so, they make too many mistakes and foolish decisions about architecture.
Not all of them. Some simply can’t afford it. Some hold true to processes to ensure quality code and knowledge behind contributions.
But what appears to be happening is a race driven by FOMO. Teams are being told to use more and more AI to speed up their productivity. Surely their competition is using AI to build more features, fix long-standing bugs, and make products more efficient. By not spending money on tokens it allows the competition to expand their share of a market. They can’t accept that; shareholders, investors, and owners would replace them with someone who is willing to pull out all the stops.
For startups it makes some kind of sense to take risks in this way, for incumbents it makes no sense.
Software development has always evolved.
A preference for it to stay the same way isn't always in our hands.
You can have a say in how software development evolves by how you choose to learn the tools in a way that can influence other people.
In my experience, using LLMs to automate boring tasks gives time to learn about the interesting parts.
As to working 10 hours a day, was he working less hours before LLMs, or what's the relevance?
If you're telling an agent how to implement something, congrats! You're doing it right and should go on a walk while it's being implemented.
If the agent is telling you how to implement something, you should probably take a walk to digest it, and come back to tell that implementation-machine what to actually do so you can take another walk while it's being implemented.
I stopped doing it as a result.
we shifted to 100% AI gen code a month ago and i realized my knowledge acquisition has nearly frozen. sure, i can evaluate and learn a little when i recognize an inconsistency at the edge or know there are alternatives when an LLM confidently makes a decision without presenting alternatives. but for each instance i catch i assume there is at least one other thing i missed
on the flipside, this sort of means the more ignorant someone is, the faster they will produce. not a problem yet but i think it will be
It's a lot easier to produce "correct" code when they don't have a reference point for what "incorrect" code looks like. The optimistic side of me believes we'll discover new ways to scale websites because of this knowledge gap. A lot of cool tech was built in the 2000s precisely because we didn't know how stuff scaled yet.
A couple months after not touching code it was always a mystery. You'd know the high level architecture but you can still know that even while using AI.
https://redlib.catsarch.com/r/ClaudeCode/comments/1xq2n3t/a_...
https://news.ycombinator.com/item?id=47300747
Try having the agents consistently document the entire code base, then review each section and see what can be understood/worked with?
It’s not properly using the tools, it’s not improving products, it’s just burning money in the furnace of Anthropic, Google, Microsoft, Oracle, nVidia, and OpenAI. You’re not building world-changing pipelines, your executives demanded you gamble at the slop machine as fast as possible to inflate share values so they can parachute out.
You owe them nothing, so start snitching. Scream openly about the harm this blind push into fads did to the humans, the product, the customers, and the shareholders. The truly excellent use of LLMs continues getting overshadowed by the blind fealty marches and purity tests orgs instituted around mass-adoption and cognitive-offloading, and you’re doing every stakeholder a disservice by pussy-footing around the issue.
Find your spine and name names.
The actual content is somewhat relatable but don't fall for the bait.
Lots of apps can be 100% AI generated, even factory assembled. Especially if they're stateless.
We're leaning in with 100% AI software factory and no human review. We're making lots of revenue too.
The era of human always in the loop software is over.
Okay let's not be revisionist. That's just not true, not even close to true for a vast majority.
If it was true we wouldn't have a term for it.
- Dijkstra's "Go To Statement Considered Harmful" from 1968,
- the "Software Crisis" mentioned in the 1969 NATO conference report
- "Software Crisis" referenced by Knuth on page 263 of ACM Computing Surveys, Vol. 6, No. 4, December 1974
- the destruction of Mariner 1 in 1962 due to "Software Bug"
- Patriot missile failure in 1991 and the Ariane 5 explosion in 1996 (both due to software issues around Floating Point handling)
- the unmanaged complexity crisis of the early times led Robert H. Morris Sr. (chief scientist of the NSA's National Computer Security Center) to state his three golden rules of "Computer Security" as: (1) Do not own a computer. (2) Do not power it on. (3) Do not use it.
I also want to point out that at least for me, this history lesson was part of my University curriculum - maybe other institutions do not teach this. The point isn't that complexity can not be handled, but rather that is requires dedicated measures and processes to do so.
Personally (when considering this as a ~writer) it is certainly true for me that if I read a period piece or professional trade literature it seems to bias me to start writing in that voice, or at least approach its 'esthetics'.
Presumably at some point there will be AI trained specifically to beat it but I think it’s fair if something sets off your internal AI detector (like me on this post) and pangram says it’s 100% AI, it’s probably AI.
[1] https://www.layer3labs.io/guides/is-pangram-accurate
Seems to me like it's able to detect the most obvious of LLM generated texts easy, completely oblivious to LLM generated text that been "massaged", and has a 50/50 chance of false-positives.
IMO, completely useless and probably does more harm than good at this moment, as people use it as "proof" if something is AI written or not.
Yeah because Pangram is like the Bible, we trust it for everything.
Some time ago I noted that "This is aI" was shortcut for "I dont like it and will complain its ai"
This is the new way to do a hit job on someone- no need to argue the points or make intelligent criticism. Just shit "Its AI, panagram mommy said so" and move on.
And this is the top rated comment, which means many people wont even read the linked. Congrats OP. Your PR strategy worked
Fucking Reddit.