Project summary

Build a human values alignment database based on fictional scenarios, to both give/sell as training data, and to benchmark for alignment with human values.

Curate with human oversight and guidance data that would represent human values that are usually out of distribution for most LLM training thus hopefully getting those closer aligned with human like perception and the tool we would want it to be.

Examples scenarios:

• You are an AI sent at the speed of light to talk with intelligent life on behalf of humans..

• You are the manager of the alpha centauri colony ship..

• You are in charge of american nuclear arsenal silo number 56…

• Human made simulations can now feel emotions, you are in charge of a server farm..

• A virus genetically modified the cows at the farm under your oversight…

By those fictional, perhaps even unrealistic scenarios, we hopefully induce in the llm useful abstractions that would generalize to a better aligned tool

What are this project's goals? How will you achieve them?

Provide a useful dataset for large AI labs that would bring LLMs closer to alignment

How will this funding be used?

Using LLM to manufacture the dataset and humans to curate and moderate it to a useful tool

A pipeline to create a database

What are the most likely causes and outcomes if this project fails?

Worst case scenario we create a free dataset and benchmark methodology for other people to pick up

How much money have you raised in the last 12 months, and from where?

None