Skip to content

With artificial intelligence trained on real startup pitches and conversations with entrepreneurs, Recuria creates surprisingly plausible random startup ideas.

Notifications You must be signed in to change notification settings

winsonluk/Recuria

Repository files navigation

Recuria

Winson Luk, wluk@ucsd.edu

Abstract

Every startup claims to be disrupting an industry or changing the world. Most startup ideas are destined to fail, but some truly change the world. By training on thousands of startup taglines, articles, and interviews, this project aims to generate a lot of bad startup ideas, and a few good ones.

Most startup ideas can be summarized in just one paragraph. The tagline describes the overarching concept (e.g., "Uber is finding you better ways to move, work, and succeed"), and the next few sentences can provide a more detailed description of the product, as well as context on the startup's history, people, and industry.

The tagline can be created by finetuning GPT-2 with a dataset of startup taglines, and the subsequent sentences can be generated by feeding this tagline as a prefix into a second GPT-2 model finetuned with company analyses (e.g., most TechCrunch articles).

The ideas generated have been fairly realistic (most are bad, some are good), so there are plans to incorporate these results into a faux startup website similar to https://tiffzhang.com/startup, with a few million permutations of ideas. And with that many ideas, there might be one in there that could actually change the world.

Model/Data

Code

Results

Technical Notes

  • The multi-gpu fork of gpt-2-simple needs to be installed to train with the 774M model.
  • I used 4 x Tesla V100 GPUs and 16 GB of RAM on Vast.ai to train the models. Training will fail with single GPUs or less than 16 GB of RAM. After training, generation can be performed with a single GPU, though 16 GB of RAM is still necessary.
  • The startup tagline model is finetuned to a loss of 0.05, while the larger TechCrunch model is finetuned to a loss of 1.8.
  • I sampled both models with temperature ranges from 0.2 to 2.0 and top-p from 0.1 to 1.0 (higher values translate to more "creativity" in the text) to find the optimal parameters for realistic text generation.

Examples

  • 1
  • 2
  • 4
  • 5
  • 7
  • 8
  • 9
  • 10
  • 11

Bloopers

Lowering unemployment

Lowering unemployment

Strategic arms sales

Strategic arms sales

Internet of things

Internet of things

Workers of the world, unite!

Workers of the world, unite!

10,000 hours

10,000 hours

Reference

References to any papers, techniques, repositories you used:

About

With artificial intelligence trained on real startup pitches and conversations with entrepreneurs, Recuria creates surprisingly plausible random startup ideas.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published