> I still review every line, shape the architecture, and carry the responsibility for how it runs in production. But the sheer volume of what I now let an agent generate would have been unthinkable even six months ago.
>That said, none of this removes the need to actually be a good engineer. If you let the AI take over without judgment, you’ll end up with brittle systems and painful surprises (data loss, security holes, unscalable software). The tools are powerful, but they don’t absolve you of responsibility.
I feel the same. AI is "the code monkey". Like very inexperienced that works hard and fast, has learned a lot but can't put it into practice. They need constant supervision and review.
This will be very challenging for inexperienced programmers. Normally learn by coding. You write code for fun or for money, get review from more experienced, ask questions and improve.
Now a new programmer is expected to review AI generated code and learn programming and managing AI.
>This will be very challenging for inexperienced programmers.
I'm feeling this for art. AI generated art still has obvious issues, but it produces works much better than a beginner. You also can't meaningfully reprompt or beat the model in to fixing the issues. So seemingly the only way to get actually good art is to grind through hundreds of hours of getting worse results than you could generate in seconds until you beat the models.
How many people are going to push through the painful unsatisfying work to eventually become experienced now.
> So seemingly the only way to get actually good art is to grind through hundreds of hours of getting worse results than you could generate in seconds until you beat the models.
Even more frustrating: if those hundreds of hours of practice are 8 hours per week, by the time you've gotten OK the SOTA GenAI will have become noticeably better.
> This will be very challenging for inexperienced programmers
It doesn't look like the undergrad-to-industry pipeline works anymore for programmers.
Those who teach themselves out of passion will always have a place. But for most coders, a graduate or PhD may become necessary. There just isn't may not be profitable niche for a CS grad out of undergrad in the private sector with AI.
I suspect this already applies more broadly than just programming.
I was already reading similar stories something like a decade ago in law, with the claim then being that the first thing most law graduates did to learn the ropes and get practical experience was being automated by simple file search and almost everything being digitised (might be "digital discovery", but also I might be conflating terms).
Last I heard (may be out of date already), robotics and computer vision is currently a dichotomy of either {not good enough} and/or {not fast enough} to be a "junior gardener" or a "junior hairdresser", so this (probably) isn't yet true for all roles, but I suspect it may be true for most* desk jobs.
* not all, most: if you need a human face somewhere, the GenAI real-time conversation agents do still sometimes mess up and TTS out the | and < that come from the LLM, but other than that…
> The service is written in Go with few dependencies and an OpenAPI-compatible REST API. At its core, it sends and receives emails. I also generated SDKs for Python and TypeScript with a custom SDK generator. In total: about 40,000 lines, including Go, YAML, Pulumi, and some custom SDK glue.
40000 lines of code for sending and receiving emails? Hmmh...
What? All they can talk about is how many lines of code they generated. The other day someone was bragging about having built a 500k SLOC behemoth completely AI generated. Good effing luck maintaining that, and I wonder what the hell were they building that needs so much code.
If it were a solo coder writing a half a million lines of code application in the span of a few months by themselves, they would either be a savant or a complete lunatic.
What for? You responded in hostile and dismissive ways to others that have done just that in other comments - what is there to think you’ll do any different this time?
Thanks for your invaluable contribution. I can’t speak to what this other person has built with AI but on my side I’ve built two live startups with AI so far and working on more (I lead a venture studio doing exactly this).
If you’re actually curious (instead of just trolling or pretending everyone who’s using ai to code is having mass hallucinations), you can join my talk at Google Brussels next week. Sign up here: https://gdg.community.dev/events/details/google-gdg-cloud-be...
Since the title isn't very informative, here's a tldr:
> Is 90% of code going to be written by AI? I don’t know. What I do know is, that for me, on this project, the answer is already yes. [...] At the same time, for me, AI doesn’t own the code. I still review every line, shape the architecture, and carry the responsibility for how it runs in production. But the sheer volume of what I now let an agent generate would have been unthinkable even six months ago.
Written by Armin Ronacher of Flask, Jinja, and general Python fame.
Assuming the AI coding tools progress at their current rate (which is probable but not guaranteed), 90% of the code generated by AI will become commonplace for certain types of projects.
An LLM is essentially a statistical text engine that can produce convincing code for any problem for which there are already similar solutions. Most projects have many such problems, and some projects involve 100% solved problems that just need to be packaged into a new solution.
However, there is a certain class of problems that are too technically innovative and novel. It is often difficult to even describe these problems in human language. AI will mostly hallucinate for such a class of problems, which will actually slow down a competent programmer because the necessary training data is missing.
> Some things that would have taken me a day or two to figure out now take 10 to 15 minutes.
The advantage of spending a day or two of figuring something out is that is (mostly) a one-time process, after that you've learnt something you can apply later again and again. Taken to its extreme, you spend a few years learning programming when your agent can do it for you in far less time, but as this post states, this kind of work wouldn't have been possible (or at least reliable) had the author not taken the time earlier to learn programming, systems architecture, etc... themselves.
I'm not saying that AI can't help you learn something, but I think when you measure its success in time saving, learning gets unknowingly pushed to the back as a waste of time.
IMO this take is naive when it comes to actual senior developers (not people calling themselves senior after 3y).
Most of these things that would take me a day or two to figure out, take this time because of the process. Not because of any learning happening. I have enough experience that learning can happen in a very, very small amount of time and explanations.
AI increases the amount of things I learn because it decreases the amount of time I need to spend to get to the signal. But this doesn’t apply to everyone; less experienced developers need the practice and for this your comment does apply I think.
I think when the human effort is eliminated in making of a product, the value of the product goes down as well. The value comes down to the running costs of the machine and cost of raw materials.
Food remains valuable even though production is increasingly automated; there's not much difference in value of hand-picked {insert name of soft fruit whose harvesting has yet to be automated here} over a machine that sticks a net around the base of a tree and shakes it until the fruit falls off, but there sure is a difference in price.
For e.g. software: when I'm buying, I don't care in the slightest if it was made by one single human, a team of humans of my nationality, contracted and outsourced to the lowest bidder in a developing nation, a sentient typewriter, or an evolutionary algorithm — I care that it solves the problem that I had and for which I entered the market looking to buy.
Maybe I don't understand your equation. For me, value and price are not too different. Value of your car is the price that you can put on it today. Price of goods must come down to the cost of making it plus margins. If making costs are low, price goes low. As simple as that for me.
Ah, a person whose technical assessment I just can't believe at all. I notice their blog also posts all their socials, including a Bluesky. What do I see in their latest 10 posts? Stereotypical sociopolitical hate farming about Israel, immigration, capitalism, etc. Yup .. all checks out - person with crazy technical beliefs is pretty off the rails across the board.
AI cannot do anything meaningful beyond basic functions and boilerplatey stuff with moderate difficulty. Don't believe me? Ask your AI model to generate a production grade typescript 5.x application with a tsconfig.json file that uses ts-node instead of tsx with path aliases configured and uses biome 2.x for linting formatting. Add vitest for testing library with graceful shutdowns that handle SIGTERM and SIGINT events from your express server wih an ioredis connection that can be shut down gracefully along with using pg-promise to open a postgresql connection. Make it use @dotenvx/dotenvx for managing development, testing, staging and production environments. Add commitlint to ensure all git commit messages adhere to a specific format. Ask it to add lefthook along with a configuration file for running, lint format and test jobs on every commit. Now add pino and pino-http with custom serializers to redact passwords, tokens and apiKeys from the logging output. Development version of the logs should be pretty printed, production version of the logs should be optimized JSON sent to a cloudwatch log transport. Add CORS to handle frontends running on localhost:5173 for development, https://example.com for production and add helmet for managing security loopholes. Add CSRF protection to all POST, PATCH and PUT routes in the application. Generate 4 different docker files, one for development, one for staging, one for testing and one for production. The development and testing versions should run a separate postgres and redis container with whom we have seamless networking setup. Development versions use self signed SSL certificates at every level. The staging and production environments should connect with an actual elasticache and RDS instance supplied. Both staging and production use actual SSL certificates. We also need Github actions for CI / CD with templates for pull requests, feature requests and tasks to run tests on PR with dependabot updates completely disabled. We need NGINX running as forward proxy to this endpoint with SSL certificates generated by letsencrypt that are auto rotated. Talk to me when your AI model can generate 10% of this setup
I sincerely hope that isn't your prompt. These boilerplate tasks are easy for an AI. You just can't expect to it to get it right all in one shot. You certainly wouldn't do all that in a single commit or even PR.
if you are going to prompt this over the next 50 hours with 300 different prompts, you might as well build it without using AI. that was the point I was trying to make
15 years ago I made, and even sold, games that I spent less than 50 hours making. But that's with the advantage of starting with a generic template that already held all the game menu structure (achievements, preferences, etc.), and even then the games weren't very good.
If you can make a "production grade … application" (irregardless of language details or anything else) in 50 hours, by hand and without AI assistance, you're an unusual person. Most "agile" places I've been, even a single sprint is longer than that.
I'm pretty much doing that in a containerized deployment for project I'm looking to open source soon called webbin:
Technology Stack
Frontend: React 18 + TypeScript + Vite + TailwindCSS
Backend: Node.js + Express + Clustering
Database: PostgreSQL 16 with performance optimization
Cache: Redis 7 with active defragmentation
Security: HTTPS/TLS with container-to-container encryption
Orchestration: Docker Compose with health checks
Monitoring: Built-in APM and performance tracking
Services
webbin-frontend - React TypeScript frontend with HTTPS
webbin-backend - Node.js API with clustering support
webbin-postgres - PostgreSQL 16 with performance tuning
webbin-redis - Redis 7 with advanced caching
webbin-certbot - SSL Certificate management, openssl for dev, LE for production
webbin-testrunner - Testbot
webbin-nginx - Proxy
Access Points
Frontend: https://localhost:5173
Backend API: https://localhost:3001
Health Check: https://localhost:3001/api/health
I'm around 15k LOC, all built in ~80 hours of interactive prompting mostly with Claude 4.0 Sonnet, then some Gemini 2.5 Pro for more devops activities.
> I still review every line, shape the architecture, and carry the responsibility for how it runs in production. But the sheer volume of what I now let an agent generate would have been unthinkable even six months ago.
>That said, none of this removes the need to actually be a good engineer. If you let the AI take over without judgment, you’ll end up with brittle systems and painful surprises (data loss, security holes, unscalable software). The tools are powerful, but they don’t absolve you of responsibility.
I feel the same. AI is "the code monkey". Like very inexperienced that works hard and fast, has learned a lot but can't put it into practice. They need constant supervision and review.
This will be very challenging for inexperienced programmers. Normally learn by coding. You write code for fun or for money, get review from more experienced, ask questions and improve. Now a new programmer is expected to review AI generated code and learn programming and managing AI.
>This will be very challenging for inexperienced programmers.
I'm feeling this for art. AI generated art still has obvious issues, but it produces works much better than a beginner. You also can't meaningfully reprompt or beat the model in to fixing the issues. So seemingly the only way to get actually good art is to grind through hundreds of hours of getting worse results than you could generate in seconds until you beat the models.
How many people are going to push through the painful unsatisfying work to eventually become experienced now.
> So seemingly the only way to get actually good art is to grind through hundreds of hours of getting worse results than you could generate in seconds until you beat the models.
Even more frustrating: if those hundreds of hours of practice are 8 hours per week, by the time you've gotten OK the SOTA GenAI will have become noticeably better.
> This will be very challenging for inexperienced programmers
It doesn't look like the undergrad-to-industry pipeline works anymore for programmers.
Those who teach themselves out of passion will always have a place. But for most coders, a graduate or PhD may become necessary. There just isn't may not be profitable niche for a CS grad out of undergrad in the private sector with AI.
I suspect this already applies more broadly than just programming.
I was already reading similar stories something like a decade ago in law, with the claim then being that the first thing most law graduates did to learn the ropes and get practical experience was being automated by simple file search and almost everything being digitised (might be "digital discovery", but also I might be conflating terms).
Last I heard (may be out of date already), robotics and computer vision is currently a dichotomy of either {not good enough} and/or {not fast enough} to be a "junior gardener" or a "junior hairdresser", so this (probably) isn't yet true for all roles, but I suspect it may be true for most* desk jobs.
* not all, most: if you need a human face somewhere, the GenAI real-time conversation agents do still sometimes mess up and TTS out the | and < that come from the LLM, but other than that…
> The service is written in Go with few dependencies and an OpenAPI-compatible REST API. At its core, it sends and receives emails. I also generated SDKs for Python and TypeScript with a custom SDK generator. In total: about 40,000 lines, including Go, YAML, Pulumi, and some custom SDK glue.
40000 lines of code for sending and receiving emails? Hmmh...
Surprisingly no AI lover counts SLOCs. I've seen formatters expand readable 1-liners into 4.
Yeah, verbosity is the big thing LLMs do.
Lots of the work after the initial code generation is cajoling it into writing something shorter then going back and just doing that.
What? All they can talk about is how many lines of code they generated. The other day someone was bragging about having built a 500k SLOC behemoth completely AI generated. Good effing luck maintaining that, and I wonder what the hell were they building that needs so much code.
If it were a solo coder writing a half a million lines of code application in the span of a few months by themselves, they would either be a savant or a complete lunatic.
thats go for you my man
So which startups are near 100% AI code and what has he built with AI?
None of these blog posts would be needed hyping up vibe coding if people actually built something.
Useless take. There’s plenty of examples, nobody needs to prove anything to you in particular.
Plenty of examples? Why not provide some?
What for? You responded in hostile and dismissive ways to others that have done just that in other comments - what is there to think you’ll do any different this time?
I simply asked for evidence to back up an unfounded statement. Which part of that was hostile and dismissive?
they won't provide examples because they cant.
Thanks for your invaluable contribution. I can’t speak to what this other person has built with AI but on my side I’ve built two live startups with AI so far and working on more (I lead a venture studio doing exactly this).
If you’re actually curious (instead of just trolling or pretending everyone who’s using ai to code is having mass hallucinations), you can join my talk at Google Brussels next week. Sign up here: https://gdg.community.dev/events/details/google-gdg-cloud-be...
Still haven't provided any examples.
I'm sceptical that there are no tells like the useless comments that are word for word what the code will do.
Since the title isn't very informative, here's a tldr:
> Is 90% of code going to be written by AI? I don’t know. What I do know is, that for me, on this project, the answer is already yes. [...] At the same time, for me, AI doesn’t own the code. I still review every line, shape the architecture, and carry the responsibility for how it runs in production. But the sheer volume of what I now let an agent generate would have been unthinkable even six months ago.
Written by Armin Ronacher of Flask, Jinja, and general Python fame.
Assuming the AI coding tools progress at their current rate (which is probable but not guaranteed), 90% of the code generated by AI will become commonplace for certain types of projects.
An LLM is essentially a statistical text engine that can produce convincing code for any problem for which there are already similar solutions. Most projects have many such problems, and some projects involve 100% solved problems that just need to be packaged into a new solution.
However, there is a certain class of problems that are too technically innovative and novel. It is often difficult to even describe these problems in human language. AI will mostly hallucinate for such a class of problems, which will actually slow down a competent programmer because the necessary training data is missing.
> Some things that would have taken me a day or two to figure out now take 10 to 15 minutes.
The advantage of spending a day or two of figuring something out is that is (mostly) a one-time process, after that you've learnt something you can apply later again and again. Taken to its extreme, you spend a few years learning programming when your agent can do it for you in far less time, but as this post states, this kind of work wouldn't have been possible (or at least reliable) had the author not taken the time earlier to learn programming, systems architecture, etc... themselves.
I'm not saying that AI can't help you learn something, but I think when you measure its success in time saving, learning gets unknowingly pushed to the back as a waste of time.
IMO this take is naive when it comes to actual senior developers (not people calling themselves senior after 3y).
Most of these things that would take me a day or two to figure out, take this time because of the process. Not because of any learning happening. I have enough experience that learning can happen in a very, very small amount of time and explanations.
AI increases the amount of things I learn because it decreases the amount of time I need to spend to get to the signal. But this doesn’t apply to everyone; less experienced developers need the practice and for this your comment does apply I think.
Saying that AI is good because it produces 90% of code is like saying that newsletters are good because they produce 90% of e-mails.
I have a hard time following all the AI news. How is Claude Code different from Cursor? Is there any reason to switch over?
I think when the human effort is eliminated in making of a product, the value of the product goes down as well. The value comes down to the running costs of the machine and cost of raw materials.
Value != price
Food remains valuable even though production is increasingly automated; there's not much difference in value of hand-picked {insert name of soft fruit whose harvesting has yet to be automated here} over a machine that sticks a net around the base of a tree and shakes it until the fruit falls off, but there sure is a difference in price.
For e.g. software: when I'm buying, I don't care in the slightest if it was made by one single human, a team of humans of my nationality, contracted and outsourced to the lowest bidder in a developing nation, a sentient typewriter, or an evolutionary algorithm — I care that it solves the problem that I had and for which I entered the market looking to buy.
Maybe I don't understand your equation. For me, value and price are not too different. Value of your car is the price that you can put on it today. Price of goods must come down to the cost of making it plus margins. If making costs are low, price goes low. As simple as that for me.
Price is what it costs, value is brings you.
If the only thing you can imagine doing with a thing is selling it, as in you cannot imagine using it, price and value become the same.
> Price of goods must come down to the cost of making it plus margins
The difference is "profit" or "loss" depending which is higher.
Buyers will (in general) only buy when they value a thing more than the price charged.
[flagged]
Ah, a person whose technical assessment I just can't believe at all. I notice their blog also posts all their socials, including a Bluesky. What do I see in their latest 10 posts? Stereotypical sociopolitical hate farming about Israel, immigration, capitalism, etc. Yup .. all checks out - person with crazy technical beliefs is pretty off the rails across the board.
AI cannot do anything meaningful beyond basic functions and boilerplatey stuff with moderate difficulty. Don't believe me? Ask your AI model to generate a production grade typescript 5.x application with a tsconfig.json file that uses ts-node instead of tsx with path aliases configured and uses biome 2.x for linting formatting. Add vitest for testing library with graceful shutdowns that handle SIGTERM and SIGINT events from your express server wih an ioredis connection that can be shut down gracefully along with using pg-promise to open a postgresql connection. Make it use @dotenvx/dotenvx for managing development, testing, staging and production environments. Add commitlint to ensure all git commit messages adhere to a specific format. Ask it to add lefthook along with a configuration file for running, lint format and test jobs on every commit. Now add pino and pino-http with custom serializers to redact passwords, tokens and apiKeys from the logging output. Development version of the logs should be pretty printed, production version of the logs should be optimized JSON sent to a cloudwatch log transport. Add CORS to handle frontends running on localhost:5173 for development, https://example.com for production and add helmet for managing security loopholes. Add CSRF protection to all POST, PATCH and PUT routes in the application. Generate 4 different docker files, one for development, one for staging, one for testing and one for production. The development and testing versions should run a separate postgres and redis container with whom we have seamless networking setup. Development versions use self signed SSL certificates at every level. The staging and production environments should connect with an actual elasticache and RDS instance supplied. Both staging and production use actual SSL certificates. We also need Github actions for CI / CD with templates for pull requests, feature requests and tasks to run tests on PR with dependabot updates completely disabled. We need NGINX running as forward proxy to this endpoint with SSL certificates generated by letsencrypt that are auto rotated. Talk to me when your AI model can generate 10% of this setup
I sincerely hope that isn't your prompt. These boilerplate tasks are easy for an AI. You just can't expect to it to get it right all in one shot. You certainly wouldn't do all that in a single commit or even PR.
if you are going to prompt this over the next 50 hours with 300 different prompts, you might as well build it without using AI. that was the point I was trying to make
15 years ago I made, and even sold, games that I spent less than 50 hours making. But that's with the advantage of starting with a generic template that already held all the game menu structure (achievements, preferences, etc.), and even then the games weren't very good.
If you can make a "production grade … application" (irregardless of language details or anything else) in 50 hours, by hand and without AI assistance, you're an unusual person. Most "agile" places I've been, even a single sprint is longer than that.
I'm pretty much doing that in a containerized deployment for project I'm looking to open source soon called webbin:
I'm around 15k LOC, all built in ~80 hours of interactive prompting mostly with Claude 4.0 Sonnet, then some Gemini 2.5 Pro for more devops activities.