Essay by Eric Worrall
The following contains a video which shows the complete process start to finish of vibe coding a simple game. The process takes 30 minutes.
And here is the game.
As I said the entire process took half an hour. For me as a professional software developer to perform this task without AI assistance would have taken half a day.
There is a catch – vibe coding works best for programmers. My first attempt to create this game was unsuccessful, Claude got stuck in an endless loop, running up higher and higher bills. I used my programmer experience to understand what the problem was and how to fix it.
I used Claude to produce this game, but any advanced AI would have done. While Claude is front of the pack right now, that horse race changes monthly. Next month Google Gemini, China’s Deepseek or Elon Musk’s xAI (recently boosted by the purchase of Cursor) might pull ahead.
Claude probably burned enough energy to power a street of houses on my behalf, for the half hour it took me to complete this game. I didn’t track the exact token usage, but my guesstimate is around $2 worth of AI time, which is covered by my AI account plan.
As the project progressed, the AI became more and more adept at figuring out what I wanted to do next, from memory I accepted two of the AI’s suggestions verbatim, without making any changes.
The video demo above is a simplified version of vibe coding. With non-trivial real world projects, instead of one vibe code session, you have a whole team of independent AI agents working in parallel, with some agents serving as specialised team directors, testers, and other roles you would find in a human development team. You can even structure the project so humans do some tasks and AI does other tasks. The energy burn for a single multi-agent session likely goes from the equivalent to a row of houses to the energy burn of a small town.
If I had not realised what was happening when my first attempt got stuck in a loop, I could have ended up with a bill for thousands of dollars. By default Claude appears to have constraints to limit the cash burn to what you have already paid, but it appears to be easy to turn these constraints off, so be very careful about which account settings you change. I’m sure Claude is no different in this respect to other AIs.
My prediction, eventually the entire economy will be “vibed”. The same principals which were applied to produce this game could be applied to configuring a production line to produce new physical products. The game contains physics and geometry – the components required to develop a physical object. I’m not saying this could be done today, but the potential is glimmering on the near horizon. And the same principal of expertise will apply to vibe production as applies to vibe programming – the people most able to vibe produce will be product designers and engineers.
Through vibe production, the ultimate drop shipping experience will soon be a reality – instead of scouring the internet for something to sell, you will be able to create an entirely new product out of your imagination, including promo material, and have the prototype in the mail in time for your morning coffee. This is already starting to happen – Meshy.ai allows 3D models to be created using vibe, models which can be fed straight into a desktop 3D printer. I expect to see this start to go mainstream from 2027, 2028 at the latest.
I’ll guarantee China is already working on vibe production. When it happens it will transform the global economy.
Of course the energy requirements for AI capable of accomplishing such feats – vibe building polished products made of multiple materials and containing complex electronic circuits or other sophisticated internals – will be orders of magnitude higher than today’s already enormous burn. If you watch the video above you’ll notice I fed instructions to the vibe code session in small pieces, AI easily gets lost if you ask it to do something too complex in one hit. But to date at least fixing this problem is just a matter of scale. More processing power and greater memory capacity for individual computational units will increase the ability of AI to handle complex, multi-dimensional problems, and to make bigger inference leaps between the instructions they receive and what they have to do to fulfil those instructions.
I foresee a time when a sizeable fraction of our energy output, possibly most of our energy output, is being fed into gigantic data centers to create marvellous new capabilities. Only nuclear, gas and coal can satisfy this magnitude of energy demand.
Nobody is ever going to surrender an already useful tool of such potential because a bunch of green energy hippies say we should conserve and live frugally. The green energy movement has already lost.
