Invited Talk at MilaNLP Lab on Conversational AI (Agents)

conversational, ai, llms, agents


Thanks to my former PhD advisor Kyunghyun Cho, who introduced me to Dirk Hovy, I had the opportunity to visit and give an invited talk to the members of his MilaNLP lab in Milan (I went there immediately after landing from New York). The researchers in his Milan lab are exceptional, focusing on different areas of natural language processing such as alignment, safety, and prompting.

My talk was about the current state of conversational AI. Not long ago, we categorized it into chit-chat bots and task-oriented bots. Now, with terms like “conversational agents” becoming popular, there’s renewed interest and rediscovery of task-oriented chatbots. I took this opportunity to connect these areas and discuss our two projects at AWS: Dialog2API and Self-Talk.

I’m especially excited about our Self-Talk paper (accepted to ACL Findings 2024). I believe more people will focus on self-improvement and self-play (or “self-talk” in conversational AI), inspired by superhuman agents in games like Go and Chess. I believe that experience learned through self-improvement (self-play) alongside ability to plan thanks to the learned world model are the missing recipe to get us to the superhuman conversational agents.

My talk was part of a (coincidental) seminar featuring Anna Rogers on properties emerging in large language models and Faeze Brahman on constrained reasoning, as all three speakers were coincidentally in Milan that week. It was great to discuss these topics, including conversational AI, with the visitors and the MilaNLP group. There are no easy or quick answers regarding the emergence of strong reasoning and conversational abilities in these large language models. However, I believe that there is nothing magical or sudden; with exceptional science and engineering, you can achieve impressive results that are difficult to predict, especially at such a large scale.

Anyway, enough with my thoughts. I hope to get invited to more talks and visit more research labs. Now onto the slides 👇

Slides #

The full slides can be viewed by following this link.

Some of the images from the slides:

MilaNLP slides MilaNLP slides MilaNLP slides MilaNLP slides MilaNLP slides MilaNLP slides MilaNLP slides