This weird trend reached an apex in a Feb 2026 OpenAI blog post [1], recently on the front page [2], which describes the process for building... something... written 100% by agents.
There is no description of what the thing is, no indication of what value it provides its users. The closest it gets is "the product has been used by hundreds of users internally, including daily internal power users".
But the fact that the thing has a million lines of code is repeated twice in the first few hundred words.
The reality distortion field is strong around anthropic. Anthropic posts tons of equally bullshit blog posts, written entirely by AI, saying absolutely nothing, to the front page and they consistently average many hundreds of upvotes.
Yeah it was very disappointing that so few details were provided. One of the reasons I think it's going to be an open source project or effort that is going to show, sooner or later, how effective these things actually are.
In their podcast interview, they mention that it's an Electron app that users download, and so they periodically create a new build. See section "Autonomous Merging Flow" here: https://www.latent.space/p/harness-eng
I'm constantly thinking about that Microsoft guy who posted something like "we want 1 million LoC per engineer per month", which basically read as satire to most engineers I talked to, except apparently it was not satire at all, and indeed seemed to reflect the position of many CEOs etc when it comes to LLM code generation.
I do think that over the past few months, it feels like the hype around producing unmaintainable amounts of LoC has started dying down. More pragmatic and realistic takes are seemingly shared more openly, and are maybe even getting through to top leadership at some tech companies. Maybe not all is lost yet.
> I do think that over the past few months, it feels like the hype around producing unmaintainable amounts of LoC has started dying down.
I wonder if a small part of this is more and more business and product people actually trying to incorporate AI into their daily workflows. I have seen this in both small companies I work for. People were very excited about getting Claude Cowork a couple of months ago, and while they use it daily, I would say they are rather underwhelmed compared to the magic they were expecting.
Complaints include the output being mediocre and verbose, it getting the most basic things wrong, hitting token limits all the time, and people going back to doing things themselves because it is faster.
Sure, there is some degree of holding it wrong in the beginning, but people are realizing that maybe, just maybe, there is still somewhat of a gap between what AI CEOs, LinkedIn grifters, and YouTube AI supplement peddlers claim and reality.
I once worked in a company where there was an 80% code coverage requirement. Some enterprising contractor had a script that generated a single file with its own covering test suite the size of which could be tuned to achieve 80% over the whole codebase. Mostly the code was untested.
It has been incredibly hilarious to watch the C-suites sudden realization that tokens COST MONEY and immediately revise their guidelines for how employees should use AI.
Like maybe having every engineer generate 1 million lines of code per month every month…with no thought to how those lines of code would make the company money…or how many tokens would be burned to accomplish this at what cost…wasn’t fully thought through.
When I read recent news on HN, I feel it is a fable about Goodhart's Law. The law says: 'When a measure becomes a target, it ceases to be a good measure.' The dog should wag its tail. But the tail is wagging the dog.
Because it's fun. And why shouldn't we be into incremental automation?
I still write code manually to keep my trad-coding skills from withering away, but using AI without a doubt has allowed me to better test my existing apps. Create playwright automations I would've never had the time for. Allowed me to search through docs many times faster. And it just making programming more fun when I do use it for more challenging problems, and I actually get something working at the end of the day.
Deciding what to build. Reviewing Code. And testing code. Are the new bottleneck.
So of course we don't see massive productivity gains. Because these parts of the SCLC were always bottlenecked but their capacity matched the throughout. We fired all the dedicated QAs years ago. Sr+ engineers that do all the code review are limited.
Teams have not re-organized to match the new code-input velocity.
Engineers don't want to do QA because it's "beneath them".. and most engineers don't like performing or are not Sr enough to do extensive or high quality code review.
AI is the new cloud. There's no market for people or companies who aren't committed to it. If you're a dev who refuses to use AI, no company will hire you; and should a company decide not to use AI they will have a hard time retaining devs (and they will need more devs). Their investors and big-ticket customers will also think twice before signing off on major commitments.
So yes, use AI. Don't nitpick the costs and benefits. The world is headed this way; if you want to develop software for a living and afford to eat, you need to be too.
This is already changing again now that CEOs have wised up to the fact that they're paying for code by the line but these lines don't translate to profit.
If your A+ senior developer spends 8 months working on a feature that ultimately doesn’t get shipped or a MVP that gets killed, then you wasted that A+ senior developer and their productivity was the same as the other two B+ engineers that also worked on the project. This is actually a very common issue and usually ignored when it comes to things like hiring or assigning resources to a project. AI won’t change that in a meaningful way, your team may just finish their tasks a lot faster but the bureaucratic layer above will likely remain the same, which will make any AI coding gains negligible. Companies would have to be rebuilt from the top down for AI and that’s very unlikely to happen.
I think engineers tend to over index on this kind of thing being "waste". You didn't waste that investment, you paid for the option to ship that feature or MVP and the research into the question of whether it made sense to ship it.
Are you sure those 8 months is not being spent on just “coding”? There’s design, product team input and iteration, etc. Where did you see that an A+ engineer goes into a cubicle and come X months after with an MVP in isolation?
It’s worth looking at sectors where LLM code generation hasn’t been very visible, such as certification-accredited flight-control, braking, train-control, medical, or nuclear-control source code involving real-time embedded operating systems. This sector relies on assurance: deterministic scheduling requirements, detailed commit traceability, tool qualification, configuration management, independent verification, etc.
Since this is an area where failure can lead not to Instagram accounts getting hacked, but planes falling out of the sky and nuclear reactors spewing radioactive elements, it’s worth a close look. Some of the most visible companies in this sector include: QNX, Wind River, SYSGO, Lynx, Green Hills, Siemens Embedded, etc. None of them seem to have much if any adoption of LLMs for source code generation based on public statements.
Research in this area agrees with this view:
“In this paper, I have conducted a comparative analysis of the C++ code generated by popular LLMs including: OpenAI ChatGPT, Google Gemini, DeepSeek, Meta AI, and Microsoft Copilot for compliance with MISRA C++. The study revealed that none of the evaluated LLMs generated MISRA-compliant code despite clear prompts, with DeepSeek showing the fewest violations and Meta AI the most.”
This study showed that people writing computer games had little interest in productivity tools (because they are producing something that is really used). But people who produce things that not really used are obsessed with productivity:
> the perennially unprofitable venture-backed startup, for which faux productivity is connected to the generally immaterial nature of its high valuations, versus the game studio that lives and dies by the profitability of its products.
> In a sector of the economy where "it's not about how much you earn, but about how much you're worth," the labors of the companies whose workflows are built on the kinds of productivity apps that today comprise nearly 40 percent of Product Hunt's output are not actually directed at the creation of a thing, but at the appearance of the creation of a thing.
Maybe this is why Silicon Valley seems to have become obsessed with productivity and AI whereas the people in the industries you mention don't seem as excited. It's because they are actually making real things so they don't have to 'look busy' in order to justify themselves.
More that LoC is a simple metric that has always been a problem.
Non-Functional requirements is a vestigial term from ‘function point analysis’ which is from the late 70s, and which also ended up being a proxy for LoC.
The entire industry is so focused on measuring now, and incentives are so skewed to short term that lagging indicators like maintainability are a non starter in many organizations that it will be challenging to fix this time.
The kloc fallacy never actually disappeared. Project and engineering managers got wise to the fact that it was only loosely correlated with shipping features, and stopped emphasizing it. Most everyone else has carried on silently believing it without really thinking about it. And of course engineers themselves have always believed it. How many times have you heard some guy talk about how he wrote 10kloc over the weekend as a brag?
> But! Hold my beer… in February 2026 METR effectively walked it back : their follow-up estimates flipped to a speedup (with error bars wide enough to ride a Moto Guzzi, with panniers, through!), and they abandoned the study design entirely - because developers now refuse to work without AI, and can’t reliably self-report time on agentic work. Their latest position: AI probably speeds developers up in 2026, and we can no longer cleanly measure by how much.
This may be true, but they followed in May with this [0]:
> Importantly, survey results are not necessarily grounded in reality. There are reasons to be skeptical of people’s responses to counterfactual questions such as about AI’s effect on productivity — for instance, our study in early 2025 found that people overestimated AI’s effect on their time spent on tasks by 40 percentage points on average.
Author here. Thanks, this is a useful addition I'd missed. The 40-point overestimate from their early-2025 work is exactly why I read METR's current position as "we can no longer cleanly measure this" rather than "AI definitely speeds you up now"; self-reports are doing a lot of load-bearing work in every direction, including in my own sense of my productivity.
The thing LOC measures best is how much code someone now has to read, understand, and keep alive. That number going up is a cost, not an output.
I spend a lot of my time taking over codebases other people left behind, and the AI-heavy ones have a recognizable shape: lots of plausible-looking code, thin tests, and nobody who can tell you why a given abstraction exists. Writing was never the hard part. Deciding what not to build, and being able to delete it confidently later, is the part that does not get faster with a model.
What did get faster for me is reading and reverse-engineering unfamiliar code - which is a little ironic, since the same tools are now producing more of the unfamiliar code that needs reverse-engineering in the first place.
A) a newly-receptive audience - engineers who have discovered that they very much enjoy and appreciate the tradeoff of proximity to the code for amplified velocity and impact, now that it's possible to achieve without being a manager of messy human teams.
B) an ecosystem in which it's grown nearly impossible to connect a functional description of something to how much bespoke construction and effort was involved, partially because of marketing and partially because of how much software already exists to be built on top of. It's impossible to tell from a few paragraphs of functional description whether something was built in a weekend or took a team 4 years to ship, so volume of code is the natural fallback for describing complexity.
94 comments
[ 3.2 ms ] story [ 82.8 ms ] threadUgh. Just imagine the following on a normal curve:
Pre-AI: The goal is to make more money.
With-AI: The goal is to ship more code.
Post-AI: The goal is to make more money.
Can't wait to see how we get there...
I mean, if you give 219 people a free text box and ask them to explain anything, you're extremely unlikely to get the exact same answer twice...
There is no description of what the thing is, no indication of what value it provides its users. The closest it gets is "the product has been used by hundreds of users internally, including daily internal power users".
But the fact that the thing has a million lines of code is repeated twice in the first few hundred words.
[1] https://openai.com/index/harness-engineering/
[2] https://news.ycombinator.com/item?id=48416264
>We intentionally chose this constraint so we would build what was necessary to increase engineering velocity by orders of magnitude
What kind of wanky bs is "engineering velocity". Maybe the post was written by AI?
https://github.com/openai/symphony
In their podcast interview, they mention that it's an Electron app that users download, and so they periodically create a new build. See section "Autonomous Merging Flow" here: https://www.latent.space/p/harness-eng
I do think that over the past few months, it feels like the hype around producing unmaintainable amounts of LoC has started dying down. More pragmatic and realistic takes are seemingly shared more openly, and are maybe even getting through to top leadership at some tech companies. Maybe not all is lost yet.
I wonder if a small part of this is more and more business and product people actually trying to incorporate AI into their daily workflows. I have seen this in both small companies I work for. People were very excited about getting Claude Cowork a couple of months ago, and while they use it daily, I would say they are rather underwhelmed compared to the magic they were expecting. Complaints include the output being mediocre and verbose, it getting the most basic things wrong, hitting token limits all the time, and people going back to doing things themselves because it is faster.
Sure, there is some degree of holding it wrong in the beginning, but people are realizing that maybe, just maybe, there is still somewhat of a gap between what AI CEOs, LinkedIn grifters, and YouTube AI supplement peddlers claim and reality.
I wish I were joking.
(The had never been an engineer.)
100000 LOC per month /25 days per month /8 hours per day / 60 minutes per hour
That seems...problematic for anyone doing code reviews.
No, it's incentive to let LLMs do the reviews, supporting your tokenmaxxing efforts.
Like maybe having every engineer generate 1 million lines of code per month every month…with no thought to how those lines of code would make the company money…or how many tokens would be burned to accomplish this at what cost…wasn’t fully thought through.
Why?
I still write code manually to keep my trad-coding skills from withering away, but using AI without a doubt has allowed me to better test my existing apps. Create playwright automations I would've never had the time for. Allowed me to search through docs many times faster. And it just making programming more fun when I do use it for more challenging problems, and I actually get something working at the end of the day.
Deciding what to build. Reviewing Code. And testing code. Are the new bottleneck.
So of course we don't see massive productivity gains. Because these parts of the SCLC were always bottlenecked but their capacity matched the throughout. We fired all the dedicated QAs years ago. Sr+ engineers that do all the code review are limited.
Teams have not re-organized to match the new code-input velocity.
Engineers don't want to do QA because it's "beneath them".. and most engineers don't like performing or are not Sr enough to do extensive or high quality code review.
It's not the first article I've read recently that is an ad for AI after a short context pretending to criticize it, with nothing connecting them.
So yes, use AI. Don't nitpick the costs and benefits. The world is headed this way; if you want to develop software for a living and afford to eat, you need to be too.
Are you sure those 8 months is not being spent on just “coding”? There’s design, product team input and iteration, etc. Where did you see that an A+ engineer goes into a cubicle and come X months after with an MVP in isolation?
Since this is an area where failure can lead not to Instagram accounts getting hacked, but planes falling out of the sky and nuclear reactors spewing radioactive elements, it’s worth a close look. Some of the most visible companies in this sector include: QNX, Wind River, SYSGO, Lynx, Green Hills, Siemens Embedded, etc. None of them seem to have much if any adoption of LLMs for source code generation based on public statements.
Research in this area agrees with this view:
“In this paper, I have conducted a comparative analysis of the C++ code generated by popular LLMs including: OpenAI ChatGPT, Google Gemini, DeepSeek, Meta AI, and Microsoft Copilot for compliance with MISRA C++. The study revealed that none of the evaluated LLMs generated MISRA-compliant code despite clear prompts, with DeepSeek showing the fewest violations and Meta AI the most.”
https://arxiv.org/abs/2506.23535
> the perennially unprofitable venture-backed startup, for which faux productivity is connected to the generally immaterial nature of its high valuations, versus the game studio that lives and dies by the profitability of its products.
> In a sector of the economy where "it's not about how much you earn, but about how much you're worth," the labors of the companies whose workflows are built on the kinds of productivity apps that today comprise nearly 40 percent of Product Hunt's output are not actually directed at the creation of a thing, but at the appearance of the creation of a thing.
Maybe this is why Silicon Valley seems to have become obsessed with productivity and AI whereas the people in the industries you mention don't seem as excited. It's because they are actually making real things so they don't have to 'look busy' in order to justify themselves.
https://components.news/the-gamer-and-the-nihilist/
https://news.ycombinator.com/item?id=47235774
Non-Functional requirements is a vestigial term from ‘function point analysis’ which is from the late 70s, and which also ended up being a proxy for LoC.
The entire industry is so focused on measuring now, and incentives are so skewed to short term that lagging indicators like maintainability are a non starter in many organizations that it will be challenging to fix this time.
That is why I have created one (Open Honest Slop Audit).
This may be true, but they followed in May with this [0]:
> Importantly, survey results are not necessarily grounded in reality. There are reasons to be skeptical of people’s responses to counterfactual questions such as about AI’s effect on productivity — for instance, our study in early 2025 found that people overestimated AI’s effect on their time spent on tasks by 40 percentage points on average.
[0] https://metr.org/blog/2026-05-11-ai-usage-survey/#productivi...
I spend a lot of my time taking over codebases other people left behind, and the AI-heavy ones have a recognizable shape: lots of plausible-looking code, thin tests, and nobody who can tell you why a given abstraction exists. Writing was never the hard part. Deciding what not to build, and being able to delete it confidently later, is the part that does not get faster with a model.
What did get faster for me is reading and reverse-engineering unfamiliar code - which is a little ironic, since the same tools are now producing more of the unfamiliar code that needs reverse-engineering in the first place.
A) a newly-receptive audience - engineers who have discovered that they very much enjoy and appreciate the tradeoff of proximity to the code for amplified velocity and impact, now that it's possible to achieve without being a manager of messy human teams.
B) an ecosystem in which it's grown nearly impossible to connect a functional description of something to how much bespoke construction and effort was involved, partially because of marketing and partially because of how much software already exists to be built on top of. It's impossible to tell from a few paragraphs of functional description whether something was built in a weekend or took a team 4 years to ship, so volume of code is the natural fallback for describing complexity.