The TWIML AI Podcast, originally launched by Sam Charrington in 2016 as This Week in Machine Learning & AI, is a popular podcast highlighting the “top minds and ideas from the world of ML and AI to a broad and influential community of ML/AI researchers, data scientists, engineers, and tech-savvy business and IT leaders.” On Jan. 23, 2023, Charrington sat down with a faculty member of UC Irvine’s Donald Bren School of Information and Computer Sciences (ICS) for a podcast titled, “AI Trends 2023: Natural Language Proc — ChatGPT, GPT-4 and Cutting Edge Research with Sameer Singh.”
The conversation with Singh, an associate professor of computer science, covered everything from chain-of-thought prompting and decomposed reasoning, to causal modeling in NLP and the need for “clean” data, to top commercial developments and predictions for 2023. “The highlights of the conversation were: (1) exploring the developments in NLP research aside from ChatGPT, which included increasing reliance of LLMs [large language models] on ‘clean’ pre-training data and getting trained on supervised data,” says Singh, “and (2) the discussion on ChatGPT, its potential concerns, and how the marketing has not been successful in communicating how it works.”
Targeted at experts, the podcast features interesting discussions of technical papers, novel research and recent developments (summarized nicely in this Twitter thread). However, although the conversation is not necessarily a layman-friendly introduction to NLP, Singh stresses the need for the general public to have a basic understanding of how these technologies work.
“It is crucial to have these conversations,” he says. “We are at a precipice of transformation where a lot of NLP systems, especially LM-based generative AI models, will start becoming commonplace, and it’s important for users to understand, at a high-level, what they’re using, what it can do and, more importantly, what it cannot do.”
So if you’re an NLP expert, be sure to check out the podcast to learn more about the latest developments in ML and AI (as well as previous TWIML podcasts with Singh on language models and on behavioral testing of NLP models). And if you’re not an NLP expert, you still might want to skim through some of the discussion, particularly around the one-hour mark, to learn why Singh compares ChatGPT to a chameleon.
— Shani Murray