Nobody is writing 90% of their code using AI. Here's the uncomfortable truth: The real productivity gain from using AI to write code is closer to 10%, nowhere near the 90% people claim. Sundar Pichai said in 2024 that 30% of the new code at Google was AI-generated. However,... he went on to admit, during Lex Friedman's podcast, that engineering velocity had only increased by about 10%. AI-generated code isn't free code. It still has to be reviewed, tested, and made production-ready. Optimizing a single step (code generation) doesn't boost output if bottlenecks shift elsewhere (code reviews). It doesn't matter how much code you generate if you can't keep up the review process. The solution: Automate as much as you can the review and verification of your code. I'm working with Sonar, who is sponsoring this post, and they will take care of the code quality and security analysis of your code: • They review over 300B lines of code every single day • They cover reliability, security, and maintainability for your code • You can integrate them into your CI/CD pipeline • You can install them in your IDE (I use their VSCode extension) • Support for more than 30 languages Here is a link so you can check them out:show more

AI writes 90% of my current code. Code which works, and which is productive, and which was completed 5X faster than I could do it myself. Maybe you're doing it wrong?

Not true. I might have believed this 6 months ago but not today. I might take an hour to build a moderately complex feature. review it, ask questions, tell it to make changes, etc. I ask it to write comprehensive tests in minutes that might take me a whole day. I review that, too. When I feel it's ready, I put it out for others to review. Lately all the comments I get are minor. I tell AI to do those as well. at the end of the day I've completed a feature that might have taken me a week. but it doesn't stop there. This code is going to production - and my dumb human error bugs have gone to ZERO, forgot that null check? Never happens, If you're not seeing this gain - you're making life harder for yourself and by the time you realize it, people like myself are already getting 100X returns.

Are you getting 100x returns, or is that just a number you just came up with?

@mk_fsd_xai Why don't people who make these claims demonstrate what they have done?

@svpino Actually, I'm 100% serious - ask me this question in 2-3 months and I promise you I'll show you. My side project 8 months ago has turned into my startup. I'll be launching before the year's end.

@JadeCole2112 @svpino Mike are you a bot - what's your GitHub show me AI generated code

dude i'm not a bot. bots don't talk like me. I will prove it when I launch my 60K lines of AI generated code as my startup. my github is private - I mean it's my startup. perhaps one day if I am successful and my story gets out there I will be happy to be the one to prove to the world "yes, you can launch a legit startup with 100% AI generated code" and then I'll be happy to. bookmark this thread. u don't even need to follow me - if you want to evaluate what I've built and have a real conversation about whether what I did was garbage or ahead of the times - I'd love to. hint - it's in the retail trading industry focusing on options trading.

The uncomfortable truth is that the greatest productivity gains are not for engineers but for product people with some coding knowledge and experience. They can now instruct an AI agent about a feature, provide details, and end with, “Ask any clarification questions before starting.” This works remarkably well for me. As an engineer doing product work for the past 12 years, highly proficient in Laravel, I’ve used this approach for a year. I can confirm it would have taken me 5-10 times longer to write the code, including reviews and tests. Engineers are not the primary audience for coding AIs. People like me are.

I don’t agree, there has been a massive shift in coding agent capability since GPT5 and sonnet 4.5 It’s a different world to two months ago. Good devs ARE writing so called “vibe code” and it’s good, if you’re careful.

90% is probably the bare minimum. I jump in and write a line or two here and there, but the AI is doing pretty much all of the work. I architect solutions, research things, give it locations of docs and function definitions. AI implements my vision and fills in my skill gaps.

100% written by ai

This is false and I am living proof.

There are PLENTY of people doing serious things having 90% of their code written by AI. You not wanting to believe that doesn't make it not true.

It's the ultimate hiring filter. Anyone saying more than a third of their code is AI generated is not someone I'd want on my team. I think you identified the problem but are off on the solution. Code where a third or more is AI generated tends to be crap, mostly not because individual chunks of AI code are actually all that bad, but because the whole AI generated part of the codebase lacks in application of DRY principles and thus should have been maybe a quarter or even a tenth of the total size. For now,we need to keep engineers away from our teams who don't realize that what AI generated code most desperately needs is a few rounds of DRY refactoring, and claims of one third or more being AI generated are actually real because these guys just don't see the problem. I've reduced the LOC count for one AI-gen heavy sub project by a little over 80% because the researchers (engineers tend to know better, though there are exceptions) did manually exactly what you are describing without "any" DRY efforts. Don't get me wrong, automated quality improvement is great, but you still need human eye balls. So what we really need is the inverse of the chart generative AI of LLMs : Reductive AI. AI that can take the bloat generated by an LLM and DRY the hell out of it. Reduce the code base first so static analysis and human eyes can polish it into something usable, secure, and maintainable. I think all of that is probably a few years away though,possibly more, so for now our strongest weapon is the use of AI-gen percentages as a hiring filter, keep the DRY-blind folks away from the projects where quality and efficient use of human time matters.

AI helps a lot. But, you still need to have technical knowledge about what you're doing because AI makes mistakes. I built a mobile app entirely using AI. I didn't write a single line of code but had to fix a lot of things by prompting.

You are wrong. Many orgs are generating 90% of new code. This is happening now.

AI writes 100% of my code. I create plans and guide it and review etc but the writing part is 100% AI. The planning and specs are also 100% “written” by ai I just chat with it to research and come up together with what needs to be done. I would say. 80% human architected

I believe you are right. But you lost me at "sponsoring this post". Everything tagged with "ad" gets uninteresting.

lol imagine being so dense that you're wrong i've created three profit-producing products with ai coding literally every bit of it.

I wrote 99% of code through ai. tbf spent time writing the instructions first

Maybe you aren’t. Because you’re retarded. Everyone else though…

The “90% AI code” meme hides the real constraint: code isn’t done until it’s reviewed, tested, and shipped. AI removes friction from typing, but the choke point is trust. That’s why velocity gains are single-digit, not 90%. The real unlock? Automating review, verification, and security. Until then, AI is a typing accelerator, not an engineering multiplier.

In the last six months, I have written > 50% of my code using AI. Six months before that it was 30. This is not to say I “vibe code” without supervision. I still verify every line generated so the productivity gain is not minimal, but still debatable.

I know experienced coders who are writing almost 100% of their code with AI.

I concur. I use it to write helper functions, and basic logic. And at that it does well and saves me time. However, for anything novel it generally does not do well, even when given very explicit instructions.

Was the speed of writing code ever the bottleneck?

90% of the code could be true, coz I use it as autocomplete. I setup structure and everything else myself. Its like driving a Tesla

Not true at all. An AI agent now writes 90% of my code at 2-3x the speed it would have taken me by hand. Its first attempt may be slop, but it’s easy to give it feedback to come up with nice clean results.

Obviously varies on what you’re trying to accomplish but your claim is objectively false that it’s only 10%

Interesting take. I actually generate 100% of my code with AI. To do that I had to redefine my workflow—focusing more on planning, design & reviews than raw coding. The result? ~5x faster or more. Here’s my process:

In 10 years there will be an existential panic just like happened with the offshoring craze of the 2000s. We will then spend the next 10+ years after that hiring humans to fix the AI slop, or at least keep it running, as systems age.

Ai hater whose late to the party and has claims that were rock solid in feb 2025 but sound like a joke now. This post should be ratio to bring this dude back to reality #aicancodenow

I think currently a factor 2 or 3 is possible provided a structured approach and discipline.

It’s not 90% (yet) but definitely not 10%. Biggest mistake ppl are making is looking at what AI is capable of today not realizing how fast its iteration and growth is happening. 1 year from now anyone who’s not using it is toast.

It's a sponsored post. Dont listen to this guy

Skill issue

Prior to using AI, easily 90% of my coding was done by pulling up various sources online and copying/pasting then tweaking.

The only REASON people aren't writing 90% of code via AI is INERTIA. People tend to just carry on in old ways or refuse to embrace the new. Also, MOST people have not yet learned how to effectively "converse" with AI to make their goals clear.

99% is written in ai , production ready code (and used by thousands)

The point is not writing but reviewing. No one should blindly trust AI-generated code; review it to ensure project maintainability.