Greg Bennett’s the final Conversations2 interviewee focused on large language models!
When I heard that Salesforce have adopted LLMs at scale (and also heard that they had started to design and build this prior to ChatGPT’s release) I wanted to know more!
I’ve known Greg for years and have always found him to be warm, knowledgeable and an expert at making complex subjects understandable. He didn’t disappoint here - we spoke about thinking in systems and how that relates to conversational AI, how LLMs affect conversation design (and how conversation designers are key to the best usage of LLMs), how his team of designers work with the wider team, and Salesforce’s recent push into Generative Conversational AI.
So listen up to Greg’s thoughts on the best ways to use LLMs in conversational AI! Salesforce have gone all-in on generative conversational AI and Greg’s team have made that possible.
00:00\tStart
00:29\tGreg's introduction
03:55\tWhat's your favourite bot?
05:11\tWhat's the most useful thing you use at work?
06:21\tWhich conversation designer should we all know about?
08:15\tConversations are systematic, and design should be too
10:17\tWhat does 'generative conversational AI' mean to you?
11:38\tHow have you been incoporating generative conversational AI into your work at Salesforce?
16:05\tWe need expertise when writing prompts for LLMs
17:45\tHow does your team collaborate with the wider Salesforce team?
20:50\tWhat benefits did you envision and how close are you now?
24:23\tHow has this affected the day to day work of conversation designers?
27:46\tNow we need to be able to educate people about how to do conversational AI well
29:52\tWe're thinking conceptually about language
32:43\tIs this a new phase of conversation design?
36:32\tHow do you minimise the risks?
39:49\tDo you continue to iterate together with engineering?
41:37\tThe model has a risk of learning from bad behaviour
42:46\tHow do you stay up to date with generative AI?
44:05\tResearching attempts to break LLMs on social media gives great REAL data
45:49\tWhen will you next speak live?
46:46\tOutro