Metaculus

Forecasting AI

Weekly forecasting sessions

We’re running weekly online forecasting hangouts, open to Metaculus AI members. These remote sessions are similar to the team meetings run in the Good Judgement Project (which was one of the key ways of boosting performance, in addition to training and aggregation).


Note: this page is not up to date. You can find the latest details about our workshops on our Discord channel.



If you’re keen on joining, please sign-up so that we can plan in accordance with interest.

Session 1

Time: Sunday February 10, at 10 AM PST/6 PM GMT

Agenda: $10,000 Starcraft question

We improved a Starcraft Guesstimate model to incorporate algorithmic progress and factorised pathways to superhuman, sub-$10k Starcraft. You can find the notes from the session here.

Session 2

Time: Sunday February 24, at 10 AM PST/6 PM GMT

Platform: Google hangouts

Agenda:

Due to the high demand, we split the session into two groups.

Group 1 focused on the questions on the policy implications of OpenAI's GPT-2. Summary notes, transcript and recording can be found here.

Group 2 tried to make progress on the modularity of future architecture search. Summary notes can be found here.

Session 3

Time: Sunday March 10, at 10 AM PST/6 PM GMT

Platform: Google hangouts

Agenda: Forecasting MIRI's research progress in 2019, and the general problem of forecasting organisational research progress

Notes and recording can be found here.

Session 4

Time: Sunday March 24, at 11 AM PST/6 PM GMT

Platform: Google Hangouts

Agenda: The implications of OpenAI LP, in particular whether there massive compute increases will be used for safety work

Notes and recording can be found here.

Session 4

Time: Sunday April 7, at 11 AM PST/6 PM GMT

Platform: Google Hangouts

Agenda: OpenAI Five finals, and the implication of Dota 2 and Starcraft 2 performance for AGI timelines

Notes can be found here.

Session 5

Time: Sunday April 21, at 11 AM PST/6 PM GMT

Platform: Google Hangouts

Agenda: Forecasting algorithmic progress and reduction in years of experience required to train OpenAI Five

Sign-up here


For questions, comments, and more, you can comment below or just reach out to Jacob (jacobklagerros@gmail.com) or Ben (goldhaber.ben@gmail.com).