Sample (from 63 notes)
|Front||unipolar vs. multipolar|
|Back||Worlds in which only one player or many players, respectively, are in a realistic position to reach AGI. It's important to consider which of the two is likely to make safe AGI development outcomes more likely when determining the ideal coordination strategy.The Foresight Institute white paper briefly lists the advatanges and disadvantages of each world (p17):https://foresight.org/publications/AGI-Timeframes&PolicyWhitePaper.pdf|
|Front||reward gaming problem|
|Back||Problem in AIs that learn a reward function from actions and preferences of humans. Rather than doing what the human designers wanted it to, an AI may do something unanticipated (i.e. misaligned with the humans' goals) because the reward function was subtly misspecified."One potential way around the problem of gameable reward functions is to let the agent learn the reward function. This lets designers offload some of the design work to powerful machine learning techniques." (Everitt et al., 2018)https://arxiv.org/pdf/1805.01109.pdfExamples include those described in Clark and Amodei (2016), Gwern (2011), Irpan (2018), and Lehman et. al. (2018).|
|Back||One of the learning algorithms used in machine learning, consisting of different layers for analyzing and learning data.|
After the file is downloaded, double-click on it to open it in the desktop program.
At this time, it is not possible to add shared decks directly to your AnkiWeb account - they need to be added from the desktop then synchronized to AnkiWeb.