How To Break Into Designing "Inclusive" Microsoft LLMs + Navigating Bias & Privacy Concerns - w/ Advitya

In this episode, featured not-expert is  @AdvityaGemawat . Advitya works on Responsible AI at Microsoft as a Machine Learning Engineer 2. I am extremely stoked to present this incredibly nuanced discussion on the ethics and responsibilities that come with building mass-market generative AI tools. We discuss the technical foundations of Advitya’s career that brought him to this cutting edge of technology, and also go over how GenAI tools by big tech navigate critical concerns such as privacy, ethics, inherent biases that creep in all the time; and the aspect I was most curious about - WHO is the actual decision maker when it comes to flagging something as problematic and/or offensive.

-----------------------------------------------------------------------------------------------------------------------------------------------------------------------Support the Ready Set Do podcast:-------------------------------------------------------------------------------------------------------------------------------------------------------------------------

Resources:

Follow Advitya on the handles below:LinkedIn https://www.linkedin.com/in/agemawat/YouTube https://www.youtube.com/@AdvityaGemawatMicrosoft Research https://www.microsoft.com/research/people/agemawatInstagram https://www.instagram.com/advitya_17/Facebook https://www.facebook.com/Advitya17/

Timestamps:

(00:00:00) Intro + Background(00:03:51) Advitya’s career snapshot(00:05:22) Landing Microsoft’s ML rotation program role spanning 4 orgs(00:08:25) Experience of studying Data Science at UC San Diego(00:11:11) Day in the life starting out at Microsoft(00:14:52) Hands-on experience working on products at Microsoft(00:20:16) Responsible AI on image classification(00:23:44) Transitioning to working on LLMs(00:25:34) Pillars of Responsible AI(00:30:53) The issue of privacy in LLM responses and LLMs trained on copyrighted content(00:35:46) Regulating (slamming the brakes) on AI and solving for privacy(00:40:44) Who decides if a LLM response is problematic?(00:47:17) Industry landscape of LLM Evaluation tooling(00:52:19) Balancing inclusiveness AND accuracy for evaluating LLMs(00:57:30) Adversarial prompts and defenses against them(00:59:08) How far are we from AGI?(01:05:05) Gen AI sometimes sucks at Math(01:06:59) What a time to be alive(01:08:00) Outro + GratitudeDisclaimer: Opinions expressed by the speaker are solely his own and do not reflect his employer.