If you are interested in AI Safety, come visit the AI Safety Reading Group.
The AI Safety reading group meets on Skype Wednesdays at 18:45 UTC, discussing new and old articles on different aspects of AI Safety. We start with a presentation round, then a summary of the article is presented, followed by discussion both on the article and in general.
Sometimes we have guests. On Wednesday the 14th, Stuart Armstrong will be giving a presentation on his research agenda in the reading group:
https://www.alignmentforum.org/posts/CSEdLLEkap2pubjof/research-agenda-v0-9-synthesising-a-human-s-preferences-into
Join us by Skype, by adding ‘soeren.elverlin’.
Previous guests include Eric Drexler, Rohin Shah, Matthijs Maas, Scott Garrabrant, Robin Hanson, Roman Yampolskiy, Vadim Kosoy, Abram Demski and Paul Christiano. A full list of articles read can be found at https://aisafety.com/reading-group/
* The question addressing Gwern's post about Tool AIs wanting to be Agent AIs.
When Søren posed the question, he identified the agent / tool contrast with the contrast between centralized and distributed processing, and Eric denied they are the same contrast. He then went on to discuss the centralized / distributed contrast. He regards it as of no particular significance. In any system, even within a neural network, different processes are conditionally activated according to the task in hand and don't use the whole network. These different processes within the system can be construed as different services.
Although there is mixing and overlapping of processes within the human brain, this is a design flaw rather than a desirable feature.
I thought there was some mutual misunderstanding here. I didn't find the tool / agent distinction being addressed in our discussion.
* The question addressing his optimism about progress without theoretical breakthroughs (related to NNs/DL).
Regarding breakthroughs versus incremental progress: Eric reiterated his belief that we are likely to see improvements in doing particular tasks but a system that – in his examples – is good at counting leaves on a tree is not going to be good at navigating a Mars rover, even if both are produced by the same advanced learning algorithm. I couldn't identify any crisp arguments to support this.