Prior work scales the training side: more data, more compute, more parameters, a lower loss. These methods are effective and well funded. We study the other side of the model: what it is fed after training. Our setup uses no GPUs, no weight updates and no loss curve. None of the three was available to us.
We present a vending machine in which every snack encodes exactly one prompt. The user selects a snack, presses VEND and pastes the result into ChatGPT, Claude or Gemini. Feeding completes in a single paste. No fine-tuning, plug-in or sign-up is required.
One prompt per wrapper. One paste and it's eaten.
Evaluated in ChatGPT, Claude and Gemini before release.
Cost to the user: $0. Shelf space is funded by brands.
I grab a fistful of popcorn. The lights die. One kernel hits the floor and the speakers answer it.
We arrange the house menu along a single axis, from tweak to transform. At the tweak end, a snack adjusts how the model answers: one straight answer, bold ideas, a sparring partner, a fed-up genius. At the transform end, it replaces the chat with an artefact: a fortune that spoils your future, your encyclopedia page from 2050, a life sim in which you raise yourself, a game of your life, the trailer for the film of it. The slider traverses the axis.
Fun is plotted against price on a log scale. Fed models (★) run on the same model and the same plan the user already pays for, plus one snack. The snack costs $0, and log(0) is undefined, so the result cannot be placed on the chart. We report it above the chart instead.
Hunger was observed in every model, in every session. No model reached satiety. The machine stayed fully stocked throughout. We leave zero hunger to future work.
Per snack, in any model. Compare the rates above.
To feed. The effect lasts one chat.
From, to put your brand on a snack. The tag shows what you paid.
Abstract. One can of energy drink moves the model from a hedged list of options to the boldest idea that could actually work: concrete, with its real risk in one line. Bold, not reckless.
Read the snackAbstract: Your AI gives you one straight answer. First line is the answer. No intro, no recap.
Read the snackAbstract: Your encyclopedia page as it reads in 2050.
Read the snackA prompt, packaged as a product. It either changes how your AI answers or turns the chat into something else. The full range is on the menu above.
Pick a snack and press VEND. It copies the snack: its picture and its prompt. Paste both into ChatGPT, Claude or Gemini and send. That's it.
ChatGPT, Claude and Gemini.
For you, yes. Brands pay for shelf space.
No. One chat. Every house snack tells your AI not to save any of it to memory. New chat, normal AI.
Yes. For AI.
The house makes the eleven on the shelf. Anyone can add a snack, free. Brands put their name on one – or order a CUSTOM made just for them.
Open the snack form. Pick a form, name it, say what it does, write the prompt. We read every one before it goes up. If we change anything, we ask you first. It goes on the shelf with your name and your link.
Pick a snack, put your name, colour and logo on it, choose a slot, pay – from $10. It's on the shelf the moment payment confirms, with a link to your site. The price tag shows what you paid.
Yes. Any slot is open, even a taken one. Pay its price plus $1 or plus $5 for #1 – and it's yours. The brand that held it moves down one, nothing is removed. We email them once, and they can raise any time to win it back.
A snack made for your brand: your own form, your own prompt. We design it, you approve the preview, then it goes on the shelf. $49 for the first 5 brands, then $299.