If AGI is the future, vibe coding is what we should all be doing
...plus the changing definitions of work
So, the consensus in many parts of the AI world is that we'll all be living in luxury fully automated space communism by the time my kids finish high school. Not only will certain tasks be automated, entire jobs would be, and not just jobs we have but every job we could ever think of. “Entire data centers filled with Nobel prize winners” Dario Amodei, co-founder of Anthropic, said.
Now look, you might not believe it, most people don't, after all it sounds kind of crazy. Even most investors who invest in this space don't believe in it, as you can easily see by their investments. Nobody is really betting on employees becoming almost too cheap to meter or thinking beyond what if the future's like today but with more automation.
But you don't need to believe that the entire thing will necessarily happen, all or nothing, for you to think that the world of work will change quite a bit in the coming years.
If you believe this is coming, or even if it doesn't quite hit the metric of automating “every job that can be done with a computer”, then surely doing things like writing code by hand is as archaic as when people used to calculate ballistic missile trajectories by hand. Computers used to be a job, but now it's just a machine.
And soon, we would be able to merely ask for something, and the entire apparatus would click into gear and things would just get made. In fact that’s what many have been doing since ChatGPT came out a couple years ago. Ask for pieces of code, copy paste it into your favourite IDE, and press run. Voila, you have yourself a running python script or a website or an app.
Andrej Karpathy recently named this phenomenon of asking the models to do things and accepting it immediately as “vibe coding”. As all fantasy novels have taught us, naming is incredibly powerful, so this term has completely taken off. Vibe is of course the word of the century at this point.
Most of the arguments against doing this basically were different versions of saying to do this he will get burned because the models do not yet know how to do things very well. People leaked login credentials and got laughed at.
If you want to know what the world will look like once this is the norm, we recently had a glimpse. It got particularly popular when Pieter created a simple, very simple, flying game, which became massively popular. How popular? It's barely a month and he's already making $100,000 a month from it. Not just because the game is really fun, after all virtual dog fighting games have been dime-a-dozen over the years. But because he created it almost entirely with vibe coding. Just him, Claude, and Cursor.
This made game developers, especially professional game developers, really angry. Understandably so because while they had to work really really hard to make something a thousand times as good and not make a fraction of the money he's making. He developed the game in public, each new feature added either the same day or next day, and the features are as simple as “hey you can now pilot a triangle around instead of a plane”, and “do you want your name on the side of a blimp”.
Right now Pieter needs to be smart enough to fix things quite a lot but presumably the idea is that very soon he would not need to be smart at all, or at least not conversant with code.
(I also think this should be an existing benchmark we all see regularly to see how well new models do arbitrarily complex things. The measure is “how good a game can a non-coder make with this”.)
And people who are very conversant with code, they get leverage. Some are writing thousands of lines of code a day by managing multiple Claude code agents, each of which (whom?) is working on one particular feature, each of whom will submit PR for the author to review, and effectively automating large chunks of software development. This isn't tomorrow, it's happening now. Today.
Yeah yeah it's still slow and expensive and error prone and hallucinates and sometimes tries to change the test to pass it and can't do long code bases and … But still. You get to type in a thing you want or point to a JIRA ticket and bam! Almost working code!
It's got vast limitations still. It cannot work on super large code bases, makes hallucinations or mistakes, it's sometimes so eager to pass a test that it will try to hard code the answer or find an end run around it. But still, man a couple years ago it could barely write a python script that was right…
Whether you want to or not, you're gonna change from being an Individual Contributor to a Manager. The only question is what you manage and how tiresome they are to manage.
What this means for work is chaos. Almost everyone will have “part time AI trainer” as their job description, that's for sure. Every company will have its AI employees outnumber its human employees. There will be a dislocation, arguably it’s started happening now, where hiring for PMs and sales and marketing and engineers have already subsided. And guess what. This mostly won't matter much because humans are already outnumbered by any number of things and we just grow the organisations or make more work to compensate.
It also means individual productivity will be a function of how much inference you can “suck” from the models. There's a capex component if you want to host it or train it. But there's also an opex component. I racked up $6-15 per hour when I was messing with Claude code. I'm not the most productive engineer, so the actual would have to be higher than that. And even higher if you have multiple agents running at a time, as you should, and some already are.
Since I wrote this Steve Yegge wrote an excellent post that said something similar.
Running N agents at once multiplies your devs' daily agent spend of $10/day by N, not counting cloud costs, just token burn. If your developers are each on average running, say, five agents at once – a very conservative number, since agents will work mostly independently, leaving the dev free to do other stuff – then those devs are each now spending $50/hr, or roughly $100k/year.
It’s not really a steal anymore, so much as a heist. We’re talking about each developer gradually boosting their productivity by a multiplier of ~5x by Q4 2025 (allowing for ramp-up time), for an additional amortized cost of only maybe $50k/year the first year. Who wouldn’t go for that deal?
Unfortunately, you almost certainly didn't include $50k/year per developer of LLM spend in your 2026 operating budget.
And this goes so much further once you have Agents which can run the Claude Code coding agents. To do your PR review and check it they ran the unit tests properly and if it makes sense and other things you'd end up doing otherwise. This too won't be perfect but will slowly get better. When I say slowly I mean week by week not year by year.
This is not just true of coding, it's true of an extraordinarily large percentage of white collar work. Coding just happens to be the obsession for the models right now, mostly because people developing them are obsessed with coding. Sort of like how silicon valley cannot stop creating new devops startups.
Any job which has sufficient data to train, which is almost every job, and has decent ways of telling if something is right or wrong, which is also a large enough number of those jobs, can't help but transform. Can you imagine the same principle being applied to finance? Literature reviw? Report writing? PRDs? Contract writing?
It's already started. Every forward leaning tech company is already doing this. From Stripe to the big tech to Linear to every startup especially from the YC group, it's absolutely dominating every hiring decision. Not just coding, though that's the leading one, but also marketing and sales and operations and product managers, but also lawyers and compliance.
Or even infographic creators.
This is the future that’s being built right now. The world of work has already transformed. We're much wealthier than three decades ago, programmers do vastly different tasks helped along by automation and the congealed remnants of decisions made firm from the eras gone past. Consultants today, for instance, do on their first day what would've taken an entire three month project in the 90s. And yet, consultants have grown as an industry. Most jobs are like this.
Even if the AI train comes to a crashing halt for some reason, this trend will only slow and not stop. If this is where it stops the world of work will be transformed. Especially if you believe that AI will get better, which seems true whether that's 2x or 200x, the only way to do it well is to use it.
I started coding on machines with 1k of memory. By the end of the 80s - I was writing code that was directly manipulating disk sectors (not saving files, but actual blocks of data at specific locations on a spinning disk). Programming languages had relatively little in way of standard libraries or components, and even for the same language, code wasn’t particularly portable between different computer platforms.
By the early 90s - things had progressed a lot - you could open a socket on an ISDN line to a remote server, and only had to code everything else on top (the handshaking and message exchange). The desktop OS war had a winner, and it wasn’t IBM OS/2. Unix was standardising, with relatively high portability of C code between different Unix platforms.
Then we got Java, HTTP, XML, SSL- and connecting two systems would take days rather than weeks.
15 years ago - it took me minutes to write some JavaScript to make an AJAX call to an external service and display the results as a data grid.
And so it keeps accelerating - authentication is standardised, API documentation has standardised around OpenAPI in a way that WSDL never quite achieved.
Where I’m going with this - at each stage, the barriers to software development have been reduced - it requires less technical knowledge.
What can be done for $10,000 has massively changed, and that in turn has created a massive market for software development.
If you go back even further, to the first uses of commercial computing, only the very largest enterprises could see any point in investing in computer based automation.
But when adding an automatic telephony agent is something that you pay for as a service, even a mom-and-pop store can have one as a plugin on their Shopify store.
Easier cheaper development creates a larger market for development. It results in businesses refreshing front end design for reasons of fashion, because it’s not going to cost $20 million.
What I’ve not seen yet, is ‘the business’ actually wanting to do the testing and fault analysis.
If we get to the point where humans write no code - and we already write incredibly little code to produce megabytes of machine language - then ‘developer’ is still the role of the person who builds and QAs the system. It requires a kind of system thinking that is different to marketing or finance. In a small business, one person may wear all those hats - but in any larger business, there will always be specialisation in roles.
(And in some ways, this might be a return to earlier stages of IT. The notion of the full time tester is relatively new, as is the division between frontend and backend engineer. Or the backend engineer who doesn’t ’do databases’. That kind of hyper-specialisation is under threat)
So good Rohit. And the key to the future is the consumer, not the vibe producer. Humans are so picky and fickle. We like this today that tomorrow. Wait, maybe I don't even like your post any more. Oh yes, I do. So holding attention matters. Don't buy slop is a great motto for the late 2020s. Run marathons. Read Voltaire. Tend your garden.