Hover over a name to see the position and date range. This table only
includes positions where at least the start date is known. The positions count
can count the same person multiple times if they held different positions;
similarly, the list of staff may include the same person multiple times if
they held more than one position during a single year.
For each year, a person is included if they were at the organization for any
part of the year; this means the actual staff count at any point during
the year can be lower (or higher, if some staff held multiple positions in a
single year).
Year
Positions count
Researchers
General staff
Associates
Board members
Advisors
Number of full-time staff at the beginning each year
The following table lists some dates and people who were at the organization
on the given date (namely, the start of the year). The table may not list every
person who worked for the organization (e.g. they could have joined and left in
the middle of a single year). This table excludes associates, interns,
advisors, and board members.
Date
Staff count
Staff
Full history of additions and subtractions
This table shows the full change history of positions. Each row corresponds
to at least one addition or removal of a position. Additions are in green and
subtractions are in red. If a position name changed,
it is listed simultaneously as an addition (of the new name) and removal (of
the old name) and colored yellow. Additionally there are faded variants of each
color for visited links.
Robert Wiblin interviews Dario Amodei for the 80,000 Hours podcast about working at OpenAI and about the domains of AI and AI safety. The latter half of the podcast includes advice for people training to work in AI organizations such as OpenAI and DeepMind
Blog post on LessWrong announcing the recursive reward modeling agenda. Some comments in the discussion thread clarify various aspects of the agenda, including its relation to Paul Christiano’s iterated amplification agenda, whether the DeepMind safety team is thinking about the problem of whether the human user is a safe agent, and more details about alternating quantifiers in the analogy to complexity theory. Jan Leike is listed as an affected person for this document because he is the lead author and is mentioned in the blog post, and also because he responds to several questions raised in the comments.
This paper introduces the (recursive) reward modeling agenda, discussing its basic outline, challenges, and ways to overcome those challenges. The paper also discusses alternative agendas and their relation to reward modeling.