The archive · Consumer Apps · Product decision · 2019
AI Dungeon 2 bets a fine-tuned GPT-2 can make infinite text adventures
A developer launches a free text adventure whose every twist is generated by a 1.5B-param GPT-2; the Show HN post draws 584 points and 220 comments.
AI Dungeon
What the business is
AI Dungeon 2 is a free browser text adventure where a 1.5B-parameter GPT-2 model, fine-tuned on scraped choose-your-own-adventure stories, generates the world and narrates the result of whatever the player types.
How it started
Nick Walton spent several months on the project before the December 2019 launch. He fine-tuned OpenAI's 1.5B-parameter GPT-2 on roughly 30 MB of second-person stories scraped from chooseyourstory.com, training on a DGX-1 through a professor's lab for about 12–16 hours. The idea grew out of an earlier experiment by Nathan using the smallest GPT model on a small dataset of Zork adventures; Walton's first version was already playable at ai-adventure.appspot.com.
What happened
The Show HN on 2019-12-05 hit 584 points and 220 comments as players tested the limits — marrying dragons, installing Linux, becoming president — alongside sharp critiques about story coherence and memory. Demand was so heavy that the free Google Colab notebooks threw 403 errors mid-thread; the author patched them live, open-sourced the code for people with 12GB+ GPUs, and said he had no idea how he would afford to host the model himself.
No ending yet — it is still running.
Background
AI Dungeon 2 is a free browser text adventure that replaces a scripted story with a language model: a 1.5B-parameter GPT-2, fine-tuned on scraped choose-your-own-adventure stories, generates every next passage from whatever the player types. In the builder's words, in any other game you are constrained by what developers programmed, while here you can do anything you can imagine and the game plays off of it.
Nick Walton spent several months on the project. He fine-tuned the model on about 30 MB of second-person stories scraped from chooseyourstory.com, using a DGX-1 cluster through his professor's lab for roughly 12–16 hours, and the idea grew out of a friend's earlier experiment with the smallest GPT model on Zork adventures.
The Show HN on 2019-12-05 drew 584 points and 220 comments as players tested the limits — marrying dragons, installing Linux, becoming president — alongside critiques of the game's dream-logic coherence. Demand was so heavy that the free Google Colab notebooks threw 403 errors mid-thread; the author patched them live, open-sourced the code for local play, and said he could not afford to host the model himself.
What has to be true
- The launch signal was immediate and large: 584 points and 220 comments on Hacker News on 2019-12-05, with the thread filling with playthroughs within hours.
- The demo was the product: no roadmap or screenshots, just a link that let thousands type anything and watch GPT-2 react, and demand overwhelmed the free Colab backend.
- It was open by necessity: the author could not afford the GPUs to host a 1.5B model, so the game ran on free Colab notebooks and the code was public for anyone with a 12GB+ GPU.
- The reactions marked the trade-off later AI storytelling would face: delight at genuinely open-ended play, next to critiques of coherence and memory that the first version could not answer.
What can be applied
Show the impossible thing working, not a roadmap: letting thousands play a flawed but magical GPT-2 game created demand that no feature list could have.
Aftermath
As of 2019-12-05 AI Dungeon 2 was live and free, served through Google Colab notebooks that could not keep up with demand; the author was fixing 403 errors from the thread, and the code was open-sourced for local play. The material records no revenue, company or funding at this point.
Sources
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card