If you are interested in this subject, Levar Burton in Roots (1977) is a great series to watch. I think I watched it in 6th grade and then rewatched it in my 20's. It holds up to factual accounts and might help you understand that what you are discovering isn't a revelation in the historical analysis.
I really wish that this was true. I just don’t have the same faith in the general public.
Safety AI engineering is partly done in the thought of service. Service to people who don’t necessarily know what they are being protected from. Also, it is invisible by design and a thankless job because if it is done well it is seamless.
The closest example to this idea that I can think of is when the US federal government cyclically furloughs its employees for policy negotiations between house/senate/president. The federal US government is made up of people who have ... (read more)
A solution could be a council. Councils love making rules and setting boundaries. They also love giving a voice to everyone and can be very productive… or not (if a slowdown is needed). An established council of peers this magnitude would be seen in the telecomm community where standards are set at intervals. The telecom council can’t really stop devices being sold that are non-compliant, but they can make things dysfunctional. The FCC works with the telecom council for enforcement, but the burden of decision is moved away from the government.
To continue w... (read more)
Most of the proposed fixes in this comment thread accrue a privatized cost, available only to those who can pay it in money, infrastructure, or skill. A $10 gripper, running hot water requires electricity and water paid by the user, retained knowledge and space to store a gripper. It’s infuriating when you get a drink from a gas station on the go and you can’t open it.
This post shows a known concept of an implicit gendered tax. It can also be shown with office hvac systems being designed for male metabolic rates…making women supply an extra sweater or spac... (read more)
Thank you for the work, it's great! Appendix G is informative.
I had a question about early-layer divergence. Fig 17 shows the expert jacobians disagreeing in direction at earlier layers and agreeing later on. Figure 16 shows lower effective rank early. This could be what you would see if the Jacobians are partly rotated relative to each other and averaging would cause a loss of information (incoherent combining, in signal processing).
Did you try aligning the experts before averaging? If I were going to do it, I would try polar-decomposing each expert Jaco... (read more)