Through the towardsdatascience podcast episode “Making AI safe through debate”
I was recently listening to the Towards Data Science podcast episode called “Making AI safe through debate” where the host Jeremie Harris interviews Ethan Perez. It was a fascinating conversation on AI safety, the future of AI and the possible ways to get there, through Ethan’s expertise in Language models and IDA strategies.
Oh my background? Yeah, I am a recent graduate of the Data Science Bootcamp program from the Flatiron School. The world of seeing data through Machine Learning has been somewhat new to me but very very exciting and interesting. Lately I have been listening to podcasts on different topics of Machine Learning, data science and related. If you are a newbie to programming and data science, I would recommend you start listening too. This article is my summary on the podcast conversation and anything technical should not be taken as truth. I do hope that this article inspires you to do your own search or research to learn more.
The podcast revolves around the idea of how IDA can help enhance AI systems, IDA issues and its capabilities. IDA stands for Iterated Distillation and Amplification and it’s related to an open ended discourse on “AI Safety” (and AI alignment) as AI scales to superintelligent systems. The inspiration for IDA seems to come from the question, “how can we SAFELY create advanced machine learning artificial intelligence systems”.
As I am writing this, I can’t stop myself from thinking about all the SciFi cinema productions I have enjoyed through the years. Have you seen the 2009 animated film “9” or even the very famous movie sequences of “Terminator”. Infact, google this — “Scifi movies on AI”, and the search returns a barrage of films. Whether you work or study in a technology field or not, we humans have been fascinated by AI and the possibilities of AI. And as technology advances today at a faster pace than ever, the discussion on AI safety seems to be significantly important now. The Machine Intelligence Research Institute MIRI gives an overview of what AI safety is and why it is important. You can also listen to a (TDS podcast) on why you should care about AI, and interview with Rob Miles.
Can an AGI or artificial general intelligence be created, that performs well and can be controlled¹; two points that stem from an AI alignment perspective. “The problem of aligning advanced AI systems with human values”, as stated by Jeremie in the podcast blog post², as he introduces the podcast.
The conversation drives from there to the limitations brought on to AI systems through supervised learning techniques, from inputs or data based on the human data provider. The system is only as good as the knowledge provided by the human. The show guest speaker then discusses the possibility of alleviating this limitation from supervised learning techniques by introducing a deep question decomposition process.
Yes, I know it may sound like I haven’t fully grasped the ideas discussed in what I listened to and I admit that I have sooo much more to learn about and read about. But, this podcast episode definitely opened a tiny window into the AI topic arena for me. I am humbled at the vast possibilities of the technology that we sometimes take for granted as well as, oh so curious to read more about it. To end, I just want to thank Jeremie Harris on covering this topic and introducing it in a way that even novices like me can understand.
If you would like to read more on IDA, following are some articles on the topic:
- Iterated Distillation and Amplification by Ajeya Cotra
- or Ajeya Cotra’s medium article on the same topic
- My Understanding of Paul Christiano’s Iterated Amplification AI Safety Research Agenda by Chi Nguyen
: https://ai-alignment.com/two-guarantees-c4c03a6b434f by Paul Christiano