"I met a founder who writes 10k lines of code a day thanks to AI"
twitter.comHe says that this is not 10,000 lines of bug-filled crap… but, unless you are explicitly verifying and testing the code, how would you know? And, what kind of solution needs 10,000 lines of code added to it daily? Then again, these could be separate solutions. but, still… this 10,000-lines-of-code-a-day thing seems strange.
Exactly. And I wonder why Paul Graham is jumping on such a statement. Of everything I've read by him so far, this has the least substance.
An instant to ruin reputation that took decades to build. )':
Social media is, paraphrasing Jon Stewart, ultra-processed shit that's not good for most people.
I suppose VC love it. They want to lower the bar to building products, and make them cheaper. They want a steady line of investable companies with cheap hiring down the line.
More generally, it's weird. 10 years ago, the technical aspects would legitimately be seen as a major, perhaps main obstacle to building a product. It's questionable whether that is really gone (can you really, really build a software product now with just LLMs? With deployment, security etc).
But for sure the technological bar is now much lower, it might disappear altogether, and what remains is actually finding a useful product, marketing, finding clients etc. Which I'd argue were actually always the biggest step to overcome.
> 10 years ago, the technical aspects would legitimately be seen as a major, perhaps main obstacle to building a product.
Yeah I think that this was really more the advancement of frameworks, tooling and SaaS software. I see it all the time with non-technical people at work who demo something they built with LLMs and its always Next.js, Supabase and Tailwaind that are actually doing the heavy lifting, with the LLM invoking some commands for them. This all gets attributed to LLMs though since this is often the first exposure to those tools for non-technical people. Most engineers also knew how to scaffold a SaaS with Rails, Django or Next.js, Stripe etc. without writing much code.
> Most engineers also knew how to scaffold a SaaS with Rails, Django or Next.js, Stripe etc. without writing much code.
Most web engineers do, certainly. But I work in data, so I didn't. And for me, LLM based tools have made it much, much easier to build relatively simple frontends for my data stuff that make my life a lot easier.
I'm partial to the notion that LLMs don't really raise the ceiling, rather they raise the floor by making web n00bs like me more effective at shipping stuff that can be used by others.
To be fair though, I'm much handier at Django now than I used to be, and am slowly learning JS/browser stuff as a result of these projects.
It might also mean overproduction which is bad for economy.
"Shame about the recession, here's a portfolio of innovative companies ready to 10x while the stock market tanks"
> VC love it. They want to lower the bar to building products, and make them cheaper.
Yes! VCs want to make it so cheap and easy to build that you don’t need VCs anymore!
VCs fund companies. For them, employee compensation is a cost against their profits. In tech, this is the largest cost by far. The owners of capital deeply deeply deeply want a future where they can earn their billions without paying the rabble anything. AI is selling this idea to them.
and the rabble deeply wants a future where CEOs are deprived of billions because AI has taken their job.
Hey, who wants to start a project for CEO.ai?
Unfortunately, it isn't actually about being able to replace somebody from a technical perspective. CEOs are picked by capital and selected from amongst capital. An effective "CEO.ai" wouldn't be used for all the same reasons that an ineffective "Employee.ai" will be used.
huh?
he would be extremely pleased to pay for (and have to listen to the complaints of) far fewer programmers.
HTML/CSS? Anything UI is unnecessary verbose, and full of boilerplate code.
[edit] also any interaction with storage / database isn't fun either, unless you use some ORM frameworks, and then you have some even bigger problems.
So the output of this company is websites? It's not 1995 any more.
History doesn't repeat itself,
but it definitely rhymes.
I can't wait for SQL injections to be a thing again!
And full of XSS and weird bugs if you don’t know what you’re doing.
I would say, if you're fully embracing AI altering code frequently, and want to do it safely...
Then a lot of things about the way code should be structured can change. If you're not editing by hand, and want to limit the scope of changes, and have AI tools to make future edits faster... THEN you can have a lot of redundancy in your code. You can just not have complex code that shares responsibility and instead have redundant repeated code that solves one problem differently.
It's not what I would advocate for every single part of every program, but many things now make sense to be structured in this way. Separate out concerns way more, and not try to have complex solutions that solve many problems at once, and favour repetitive code slightly more.
Much as you would structure a project if you had some junior programmers, you might have sections that only seniors should be touching - complex cores and difficult parts where you have way more oversight. And use more AI in the areas where you'd let juniors do scut work.
Paul Graham became thoroughly Thomas Friedman-ized about 15 years ago. He's a 30,000-foot cheerleader for "innovation" and "free markets" these days.
To be fair. After having done this for the last 6 months (sabbaticals are fantastic things). About 5k lines of code corresponds to 1-2 "decent" features. And though an ai skeptic originally I've turned around - so 10k is not hard if:
A. you are doing this "full time"
B. Know what to build
C. More importantly you know "how" you want it built and you push the llm to refactoring regularly in your style.
I don't buy into the "let the agent code for 6 hours straight" agentic nonsense. I review every (big) change and just saving me stackoverflow/Google visits has been a huge productivity booster. Heck I wrote css, proto plugins, dsl parsers, simple games and more because now I needed much less "onboarding" times.
Maybe they create 10k lines per day, but only 10% actually make it into production?
> unless you are explicitly verifying and testing the code, how would you know?
Ask your AI to write tests and to review the code, of course /s
> And, what kind of solution needs 10,000 lines of code added to it daily?
In the first days, adding 10,000 useful lines a day can be very helpful. It means you get your MVP out earlier and/or makes it way easier to experiment with different ways to solve a problem.
"Measuring programming progress by lines of code is like measuring aircraft building progress by weight." Bill Gates
"I met a founder today who said he ..."
I meet people all the time who say the most incredible things. Sometimes, rarely, these things turn out to be true.
Especially in Silicon Valley where there is an increased density of counterproductive/self-defeating crazy people and smooth, Cheshire-cat-like liars who charm almost everyone like cult leaders.
Sigh. (Take away pg's weed for today.) That's 100-200 defects/day and unknown vulnerabilities. Building one's own coffin by voluminous tech debt to be repaid later with interest. It's wiser to reuse what already exists and reduce kLoC for maximum delivered utility. And it's also an absolutely dubious claim to assert because there's absolutely no evidence... it's purely a vague, testimonial vibe. (Sorry, pg.)
This comes to mind: https://www.folklore.org/Negative_2000_Lines_Of_Code.html
Even if I wrote 10k lines of code a day I will be out of things to implement in day two. Maybe I am working on wrong codebases, but I never felt like I am behind because of not enough lines got writen. It was always finding and forming correct perspective.
Just for fun I checked the codebase stats for platform I am working on and it is 70k lines of code. We are talking about enterprise saas processing payments in billions per year.
This is like praising an "author" who self-publishes hundreds of AI-ghostwritten novels a day and then preemptively defending against criticism by going "He's not naive. These aren't books of typo-filled, grammatically incorrect crap".
Why would anyone care who Paul Graham hangs out with?
many people do care because he provides more signals the guy PG hangs out with is more likely to be legit than a random person
...or just somebody who plays the SV networking game well enough to be in the room. Any number of Adam Neumanns and Sam Bankman-Fried-adjacent personality types could get through that door.
But...would Paul write that if he hadn't read (some of) the code?
...you know he can read code, don't you?
Is that a joke?
I bet he'd be pumped to know he created a community that's so counterculture it either doesn't know who he is, or disdains him.
Actual lol.
Mark Zuckerberg created much bigger and more useful communities. Must be the better man. Actual lol.
Huh?
Counterculture? Do you think you have to be counterculture not to know who Paul Graham is? What the fuck lol. Ask anyone outside of your little bubble and they won't have a fucking clue.
Edit: The above was needlessly confrontational. I don't know you and it was uncalled for. I just fucking despise Silicon Valley.
You know where you're posting, right?
Not know who he is funny, not counterculture.
Having disdain for him is getting there.
I honestly never heard of this person before now. I presume some startupper.
Is that the guy with the purple Porsche? Who or what were we talking about again?
How would I know which car someone I don't know has?
It's really sad to see what big tech has become these days. At least some folks are ready to stop drinking the kool-aid.
Which folks are you talking about?
"Say more"
I met a company that has a huge technical debt.
Unpaid technical debt who will eventually want others to pay it.
"I met a guy who takes on $10,000 debt a day thanks to an unlimited line of credit."
What could go wrong?
If it was _truly_ unlimited, that might not be an unreasonable thing to do. Of course, nothing ever is.
VERY important to note that Paul Graham is a dumb guy.
That statement definetively shows him out of touch. Is such a bad take.
Better link, at least for me:
https://xcancel.com/paulg/status/1953289830982664236
I've done that, occasionally, though through deterministic means, i.e. code generation tools. If you know what you're getting it's no big deal. If you don't, it probably is, but not in a good way.
If it's actually true and the result reasonable, I expect ~7-8 KLOC out of those 10 to be adequate and necessary tests, and roughly 4 out of those 12 h a day to be waiting for tests to finish. I also expect static analysis tooling, linters, ast-grep and so on get a lot of use.
> I also expect static analysis tooling, linters, ast-grep and so on get a lot of use.
Better yet, feedback from these kind of tools is exactly what AI is great at in iteratively resolving so you don't have to review code below your standard of quality.
It is literally impossible for someone to understand what they're "coding" in a day if 10,000 lines (!!) just got farted out of some codegen tools. I cannot believe anyone would say this responsibly with a straight face.
They're not coding, they're not vibe coding, they're asking for the code to an LLM.
When people say "I did this" showing something they did with AI, the reality is "I asked an AI for this". Okay, not every time, but something on this scale would, they don't even know what those 10K lines of code do, they asked and received.
10k lines added daily or generating 10k lines, of which only a fraction makes it?
Either way, extraordinary claims demand extraordinary evidence and from that one visible tweet without an account I remain unconvinced.
But is he a true Scotsman? https://idiallo.com/byte-size/ai-scotsman
Every time something like this is said, the author turns out (in this case, obviously is given the yc’s recent investment profile) to be a peddler of AI coding tools.
PG has become a parody of himself.
He was some 20..10 years ago a proponent of succint code and even authored the Arc language which is used to run HN.
Is HN any more sophisticated than a PHPbb forum? The value is in the community and moderation, not the back end.
Given that is has soldiered on almost two decades as a single-server system, I'd give a bit more credit to the back end.
Maybe it's an inside baseball joke taken out of context that I don't understand.
So
do
I
As*
Great but I don't want meet myself letting AI write 10k lines of "code" for me
"I met your wife, she was happy I did not say a word about AI".
AI is a too and can be useful at times. But amount of hype is unjustified in my opinion. Ergo, I am getting a feeling that there is some kind of elite conspiracy going on.
pg somewhat backs off with the claim in a subsequent tweet "The fact that you can write x number of LOC in 12 hours doesn't imply that you work 12 hour days every day." The first tweet says "lines of code a day".
He’s an indie developer then, not a founder?
Will artificial intelligence steal my ideas?
Only if you get the BCI implants or talk in your sleep.
No, but it might alter the world in ways which make it futile to accomplish some goals. Your goals in particular if you do not account for basilisk-compliance.
> But he's not naive. This is not 10,000 lines of bug-filled crap.
> As it turns out I met another startup today that's using AI to do ongoing testing of changing codebases. (He's not using them yet though.)
This is a satire, right?
When someone becomes sufficiently and necessarily old and grizzled, our existence collapses into schadenfreudian enjoyment of confusing the uninitiated.
Or: too much weed tweeting.
Twitter is such a superficial format that obliterates all context, reason, and sense of humanity. I'm still 100% unsure what the intent of that conveyance was. Brain fart? That's why I don't tweet. Also, I'm too old.
show me their github stats then pg
Well, Paul, show us the code and let us judge ourselves instead of using the old "trust me, bro" rhetoric that doesn't work any more.
What a load of crap.
Literally, for anyone thinking about investing in this company.
how are you even meant to review 10k lines of code a day?
One key process advantage Meta has is that all commits (diffs) require review by someone other than you, preferably supported by added tests and tests passing (whether locally or by automated test infrastructure). It gets more complicated when testing and reducing risk of deploying infrastructure and client platform code rather than pure service-oriented software running as part of an app.
yes, everyone who cares about code quality should do that, that was what I'm talking about - even after automated tests, and automated linting, 10k is a huge amount of code for one person to coherently review each day.
That's absolutely true. Not arguing with that. It's absolutely absurd.
The point of requiring others to review code is it essentially eliminates cowboy coding and incomprehensible crap because another person/people puts their credibility on the line when they sign-off. It doesn't devolve into "trust me, bro, just rubber stamp it" because of the seriousness, professionalism tech culture.
Code is an answer. What is the question?
I stopped looking at my code about 2 months back. I am launching "Lets Order", a micro SaaS, entirely generated by Claude Code, Gemini CLI, Google Jules, etc. I follow a structured approach, using issues on GitHub, Git branches, etc. - as I have done as an engineer for more than a decade.
Rust (backend) and TypeScript - Admin app (for managers), Menu app (for guests) and marketing site. SQLite database with LiteStream for backups. Basic tests, CI/CD.
All this - solo, in 10 days. Plus I market a lot more than I have ever had time.
- Source code and GitHub issues: https://github.com/brainless/letsorder
- Site: https://letsorder.app
I am sure there are parts that will break, but what I am getting out of a Claude Pro $20/month in value is a dream. I am generating a couple other products in parallel too, all sources available.
Thanks for sharing this, very brave.
Looking at the code, there is a good chance this codebase is vulnerable to SQL injection.
I will not be surprised if there is but that is not a problem that cannot be fixed with some effort. The point is if we can produce deploy-able, full-stack apps, which are manageable, this changes what it means for software and for startups.
I live in a remote village in Himalayas, WB, India that I am sure no one on HN has heard of. I got 5G based broadband that is flaky just a few weeks back. By the end of this year, I am sure I will be able to attempt 4-5 products and market them more than I have ever done in my 16 years of professional life.
Software development has changed, forever.
The design is fundamentally flawed, with queries in close to hundred .js files.
Perhaps you marker and sell a few, but it looks insecure and would be hard to refactor.
Are you sure of queries in `.js` files? Or do you mean `.rs` files?
grep -r -i --include="*.jsx" --include="*.tsx" "SELECT\|INSERT\|UPDATE\|DELETE" ./adminapp grep -r -i --include="*.jsx" --include="*.tsx" "SELECT\|INSERT\|UPDATE\|DELETE" ./menuappApologies, they are in JSON files. GitHub truncated the last two characters on my mobile.
https://github.com/brainless/letsorder/blob/main/backend/.sq...
I had the same doubt early on, I had asked Claude and checked this:
https://docs.rs/sqlx/latest/sqlx/macro.query.html#offline-mo...
Looks like we are supposed to check them into version control.
Neither of your demos actually launch when you press the buttons (Chrome on Android). Does this have users?
No, it is not live yet. I am launching this hopefully in a week.
Good luck! Great to see people having a go. Save a screenshot of my cynicism to send me when you make bank.
Thank you. Making bank means I will have to do something I never did - make a sale or two. I am an engineer and trying founder for more than a decade. Always struggled to keep things in balance.
Things changed dramatically in the last few months. I can enjoy my super cheap nomad life of $200 / month while I build and market all my ideas, without working over 40 hours a week.
I can see that.
Last week I have been working on remote controlling Borunte welding robots, I wrote 10k lines of code per day, creating all kinds of interfaces, weird ik strategies, camera calibrations, weird joypad controls, cad mouse controls, simulations etc.
None of it was prod ready, I was just experimenting, but tbh I would've made 1/10th of the progress if it wasn't for gemini and claude.
Now I know what to write.
PS: the default ik engine on those robots have a bug that can easily kill you when it gets singular, so if you are working on them, keep your distance :)
You vibe-coded safety-critical software? Jesus.
nop, I was just trying out stuff, also the factory ik engine was not vibecoded and I would say it tried to kill me :)