#requesting your help for an interview with a law, philosophy and AI expert and writer

3 messages · Page 1 of 1 (latest)

fathom needle
#

Hey guys, requesting your help for another great interview coming for the channel and the podcast 🙂
The interview would be with David Mertz, a data scientist, writer and who has a Ph.D. in philosophy and lots of coding experience! (profile linked below)

The interview would be about the rights that the creators of the works that contribute to the training of generative AIs do or should have. So mainly intellectual property, moral rights, economic rights, and **copyright **concerns with the training data used... More precisely, about what should exist than what does now. Any questions related to this topic would be an amazing help to create the best interview possible! 🙂

David is particularly interested in how technically a weighted attribution might work. So if you guys have any solution ideas we could pitch to him and get his feedback and thoughts, send them here too!

David's Linkedin: https://www.linkedin.com/in/dmertz/

fallen quartz
#

re: ideas on the "how" and feasibility of this kind of attribution and logging: at a very high level, logging the inputs and outputs down to a very detailed level (time/date/user/ip address, etc.) should be possible. Cybersecurity and the toolset/business and technical processes around logging and attribution might be a good place to start looking at a framework. Some caveats: 1.) that kind of logging is computationally "expensive" (and why a lot of businesses settle for come compromise between level of resolution and length/depth of logging and associated metadata versus compute/storage costs and "risk profile" and compliance requirements as the baseline as filters for how they will build and maintain the systems. The good news: that kind of logging of the inputs may already exist by design? 2.) another consideration is users wanting some balance between "anonymity" on the inputs/prompts and the ability to "own" related work product and outputs from the model. If there's absolute 1:1 traceability that users don't have any control over (opt-in?) from a prompt to a result set (like with "search queries" and results) people probably won't be open/honest/unfiltered in the prompts they use and questions they ask, which limits the utility of the system they are using and best training and growth of the model).

fallen quartz