Home

Mosaic is a research project with the primary purpose to test whether difficult problems can be solved using HCH-like systems. Difficult problems include questions that require:

• Creative insight (e.g., some math problems)
• Extended learning over time (e.g., questions in textbooks)
• Incremental reasoning about evidence (e.g., fact checking)
• Interaction with the external world (e.g., dialog)
• Meta-reasoning (e.g. reasoning about what cognitive strategies to apply to solve a problem)

See more in the list of proposed tasks.

Targeted Userbase

Testing will eventually be done by having a pool of people do many (>100+ hours worth) tasks across bounded domains. Besides gathering data on whether problems can be solved in a distributed fashion, we will also use the resulting data to see what fraction of human behavior in these tasks can be automated using ML.

The users of Mosaic are expected to be a relatively small / narrow people of intelligent people, probably ones with knowledge of computer science. This means that the user interface can be optimized for people with a general understanding of computer science, modern web browsers, and who enjoy keyboard shortcuts.

Participants will either be paid to use the service, or will otherwise be already motivated to participated. As such, the goal isn't to particularly to optimize for user enjoyment. The most important factors are the ability to get experimental data for reasonable costs (meaning the process cant' be incredibly inefficient) and to be able to do that experimentation relatively soon (within 6 months). That said, we do care about some strategies to make the user's task easier, especially if they might result in a qualitative change in what problems can be solved with users who have (say) 15 minutes per task.