AMA | i taught an AI to solve the trolley problem
New video went live today. Director's commentary is uploading. Do y'all have any questions about this video - from research, content, to technical details - ask me anything in the comments and I'll do my best to answer :)
Comments (15)
Question on how the scripting vs. shooting vs. re-scripting vs. editing vs. animation workflow happens? How much of the B-Roll and animations are decided before you even start? Curious about your workflow generally.
Are you using a non-standard color palette in Jupyter? I wasn't sure because it was so quick on-screen. If so, what colors are you using?
I used Spyder and Google Colab in the video.
How did you first get introduced to ML? And are there any other areas of ML (like reinforcement learning, etc) you're interested in learning more about?
Also, my partner does RL, and at least from what she's told me it is very very hard. So you dodged a bullet by not going that route lol
Also off topic but really nice job in the graphics this episode! They were amazing
Thanks! They were directly inspired by the animations in this video: https://www.youtube.com/watch?v=ir1YXtdhRPg
When working on a video, how do you determine if/when to pivot and what to pivot to? I can imagine a lot of the time you don't know whether your hypothesis will work until well into the research process
We split research into 2ish stages. The first stage is checking the idea for appetite, feasibility, and contingency. Would people be interested in this subject? Could we get a pretty good shot at an answer, even if we don't nail the landing? If we do fail, can we still provide viewers with something valuable? Asking these questions reduce the risk of scrapping video idea halfway through and also prep us for when things go wrong. In terms of what to pivot to if we do fail (which we so often do lmao), it really depends on the point of failure. It usually means we have a call and work together to figure out how we can still make the video valuable without starting from scratch. Despite what it sounds like, it's pretty easy because (usually) we just need to identify the point of failure and extract lessons from it. The motivation being that if that's where we failed, it might be where others might as well. So, if it isn't a lesson - at least it's a warning. - Sabrina
where did you get the beige short sleeve collar shirt? 'tis a nice shirt (clearly asking all the important questions here, i know)
I'm pretty sure it's from Uniqlo lmao. - Sabrina
That's fair enough. What you were able to create is still pretty impressive! And imo, I don't think the cat problem was really a problem at all. :P So then the obvious follow up question, and this might be too big of a question so no worries if you don't have an answer, if you were able to create a purely objectively ethical AI, should we? It would probably make decisions that we wouldn't agree with, but would that make those decisions wrong? For instance, maybe there are 10 people aged 80 on one track, and 1 person aged 10 on the other track. The AI might conclude that there are more years of human experience left on the second track and kill the 10 people.
Honestly, this question might be a bit of a non-starter. A "purely objectively ethical AI" can't exist without everyone coming to agreement on what objectively ethical means - and the continuing existence of philosophy as a field of study suggests we're never getting that agreement. - Sabrina
Did you have any discussions about the difference between an ethical AI and a Human like AI? Essentially humans do a lot of unethical things, even small things in day to day life. Would your goal be for the AI to make decisions like a human or purely the most ethical decision?
While researching, I definitely considered the differences in approach - do I create an "objectively" virtuous machine that follows a doctrine of my choosing VS do I create a machine that mimics that acts of relatively ethical people. I ended up choosing the latter because (1) on a theory level, I'm not well-versed enough to pick "the best moral philosophy" let alone create a value system that encompasses that domain well enough for a machine to be able to use it to make decisions and (2) if I was able to do that, implementing that knowledge is currently outside the scope of my ability. I would've needed to figure out how to do reinforcement learning on top of everything else. It's definitely a technique I want to pick up for future videos though - might make an AI to learn chess so I don't have to :) - Sabrina