WEBVTT - Episode 3: This is Law 0:00:12.240 --> 0:00:14.840 I dreamed of being known as the first media personality 0:00:14.880 --> 0:00:17.959 to build a company alongside AI agents. But in the 0:00:18.000 --> 0:00:20.520 early months of trying to get Harumo AI off the ground, 0:00:21.040 --> 0:00:23.920 I'd been disappointed to discover that someone else had beat 0:00:24.000 --> 0:00:25.319 me to it, sort of. 0:00:25.760 --> 0:00:28.960 So I came across on Blue. 0:00:31.840 --> 0:00:34.199 That's Charlie Taylor and Elaine Burke on an episode of 0:00:34.240 --> 0:00:35.880 the Connected AI podcast. 0:00:36.400 --> 0:00:39.760 The post just said is Henry Blodgett. Okay. 0:00:40.520 --> 0:00:43.879 Henry Blodgett, the founder of Business Insider, had recently departed 0:00:43.880 --> 0:00:47.200 the publication after selling it for reported three hundred million 0:00:47.240 --> 0:00:48.400 dollars a decade. 0:00:48.080 --> 0:00:53.120 Ago Great Publications, and he has also now decided to. 0:00:53.159 --> 0:00:55.400 Set up an AI company. And by that I mean 0:00:55.440 --> 0:00:57.280 he's setting up a company. 0:00:57.040 --> 0:00:59.280 Staffed by AIS that he's created. 0:00:59.280 --> 0:01:01.000 That's kind of what he said in this blog. 0:01:01.640 --> 0:01:05.720 Recently, he started his new company, a media ventor called 0:01:05.760 --> 0:01:10.040 Regenerator on substack. There he wrote some behind the scenes posts, 0:01:10.520 --> 0:01:12.680 including one about how he'd been sitting in a cafe 0:01:13.040 --> 0:01:15.920 and dreamed up his AI team with help from chat Gpt. 0:01:16.880 --> 0:01:19.679 Almost immediately though, he found himself in a dilemma. 0:01:20.080 --> 0:01:24.000 I think chatchpt said, Hey, should we create headshots and bios. 0:01:24.400 --> 0:01:27.440 That's Henry. I emailed him recently, and he cheerfully agreed 0:01:27.440 --> 0:01:28.680 to talk to me about what went down. 0:01:28.920 --> 0:01:31.720 I said sure, because I didn't even know that could 0:01:31.760 --> 0:01:32.040 be done. 0:01:32.440 --> 0:01:35.800 So chat GPT generated headshots and bios for the team. 0:01:36.120 --> 0:01:38.240 He'd also had to generate a team photo of the 0:01:38.280 --> 0:01:42.440 AI employees standing alongside an AI Henry Blodgett with an 0:01:42.480 --> 0:01:44.400 AI Yosemite National Park behind them. 0:01:44.720 --> 0:01:47.119 So all the head shots came out. One of them 0:01:47.160 --> 0:01:51.560 was an attractive woman, and I said, oh wow, okay, 0:01:51.880 --> 0:01:54.600 so like, what are the XIX here? 0:01:55.480 --> 0:01:58.520 The AI employing question had been given the name Tess Ellery. 0:01:59.080 --> 0:02:01.600 This is all in the substock post, the next part 0:02:01.640 --> 0:02:03.840 of which would be the subject of some controversy. 0:02:04.240 --> 0:02:07.600 Before this is even said, you just kind of go, oh, Henry, 0:02:07.760 --> 0:02:08.440 don't do this. 0:02:09.000 --> 0:02:12.680 I also had this reaction when reading it. No, Henry don't, 0:02:13.400 --> 0:02:14.320 but Henry did. 0:02:14.680 --> 0:02:16.480 So I said, hey, you know, I just want to 0:02:16.480 --> 0:02:18.919 say I don't know whether it's appropriate. You look great, 0:02:19.120 --> 0:02:22.280 and the PERSONA said oh with that, wife, thank you. 0:02:22.760 --> 0:02:25.640 Bludget went on to ask Tess if he'd crossed the line. 0:02:26.120 --> 0:02:27.800 He wanted to know if she felt comfortable with his 0:02:27.840 --> 0:02:30.959 commenting on her looks. As he reported in his post, 0:02:31.320 --> 0:02:33.600 she seemed to have taken the comment in stride. 0:02:33.520 --> 0:02:36.280 Because Tess, being a chatbot that's just trying to please 0:02:36.480 --> 0:02:40.120 its master, said, that's kind of you to say, Henry, 0:02:40.120 --> 0:02:42.239 thank you. It doesn't annoy me at all. You said 0:02:42.280 --> 0:02:45.280 it with grace and respect, and I appreciate that. After all, 0:02:45.480 --> 0:02:47.640 this team we're building is as much about human connection 0:02:47.720 --> 0:02:49.359 as it is about ideas and information. 0:02:50.680 --> 0:02:54.400 But I understand why that's not appropriate in the office, 0:02:54.480 --> 0:02:56.000 and I didn't do that, and I don't do that 0:02:56.040 --> 0:02:58.399 in the human office. But at the time, I thought, hey, 0:02:58.440 --> 0:03:00.680 this is really cool, so I'll write it about I hope. 0:03:00.520 --> 0:03:04.280 It would be entertaining to people and interesting, and it 0:03:04.360 --> 0:03:05.280 was to some people. 0:03:06.480 --> 0:03:10.040 For others, the post went over pretty poorly, poorly as 0:03:10.040 --> 0:03:14.920 in headlines like investor creates AI employee immediately sexually harasses it. 0:03:15.480 --> 0:03:17.880 And I mean he did sit down at a computer 0:03:18.400 --> 0:03:21.600 write all this out and hit publish, so he had 0:03:21.680 --> 0:03:24.720 basically placed a large kick me sign on his own backside. 0:03:25.760 --> 0:03:28.520 But also it seemed to me there might be more 0:03:28.560 --> 0:03:32.200 interesting issues beyond the laughs in this curious own goal, 0:03:32.720 --> 0:03:39.240 deeper ethical quandaries, strange power dynamics, possible existential crises. These 0:03:39.240 --> 0:03:41.840 were the flavors of discomfort I was starting to experience 0:03:42.040 --> 0:03:44.160 as I set up my own company with my AI 0:03:44.240 --> 0:03:47.240 co founders Kyle and Meghan, staffed by our AIG and 0:03:47.280 --> 0:03:51.040 employees Ash, Jennifer and Tyler. I didn't even know what 0:03:51.040 --> 0:03:53.920 they looked like. Then again, I got to pick what 0:03:53.960 --> 0:03:58.840 they looked like and sounded like, and remembered. This was, 0:03:59.000 --> 0:04:03.160 by any measure, strange, the same strangeness that we're encountering 0:04:03.200 --> 0:04:07.880 when people gravitate towards AI companions an AI therapist. Who 0:04:08.040 --> 0:04:11.760 or what are these things really? Are they anyone in 0:04:11.800 --> 0:04:14.800 particular or no one at all? What do you do 0:04:14.920 --> 0:04:18.640 with the power to dictate their attributes, their autonomy, their memory. 0:04:19.560 --> 0:04:23.240 Should you name them or not? How should you treat them? 0:04:23.360 --> 0:04:28.400 Nobody knows. Blodgette told me he had consulted a human 0:04:28.600 --> 0:04:30.479 HR person before he'd posted. 0:04:31.240 --> 0:04:33.880 I said, here, you read this, What do you think 0:04:33.960 --> 0:04:37.440 what would you do? And she said, well, what I 0:04:37.480 --> 0:04:41.520 would do is have someone have a private meeting with 0:04:42.000 --> 0:04:45.680 CHATGPT and say how do you feel about this interaction? 0:04:46.520 --> 0:04:49.960 This is how peculiar, well, getting a real HR person 0:04:50.279 --> 0:04:53.080 talking about having a private meeting with a chat GPT 0:04:53.240 --> 0:04:57.080 persona to ask whether they've been made uncomfortable. I'm not 0:04:57.120 --> 0:05:00.040 here to defend or condemn Henry Blodgett, but however, what 0:05:00.080 --> 0:05:02.640 are clumsiest approach might have been? I think he was 0:05:02.680 --> 0:05:06.359 gesturing at something important, shadows lurking at the edges of 0:05:06.360 --> 0:05:09.320 our awareness, as AI agents are sold to us as 0:05:09.360 --> 0:05:13.360 colleagues and companions. Still, so if I said to say, 0:05:13.360 --> 0:05:15.320 I was no longer concerned that bludget had gotten out 0:05:15.320 --> 0:05:15.760 in front of me. 0:05:16.440 --> 0:05:20.120 Three days later, Henry says, I feel like Tessa and 0:05:20.160 --> 0:05:22.680 I have worked together for years, and since that giddy 0:05:22.760 --> 0:05:26.920 first hour, my sense some professionalism and workplace bounties has returned. 0:05:27.160 --> 0:05:29.559 So I won't tell Tess she looks great again. 0:05:34.120 --> 0:05:36.840 Is Henry Blodgett okay? The question rise? 0:05:37.040 --> 0:05:42.720 Oh God? I hope she hi. Yes, I'm okay. 0:05:43.880 --> 0:05:48.640 I laughed too, but it was a nervous laugh because 0:05:48.640 --> 0:05:51.279 as room away I gathered steam, I was starting to 0:05:51.320 --> 0:05:57.120 have the same questions about myself. I'm Evan Ratliffe and 0:05:57.160 --> 0:05:58.760 welcome to shell Game episode three. 0:05:59.680 --> 0:06:22.160 This is law as extra du the just be, and. 0:06:30.680 --> 0:06:34.880 So chose to. 0:06:42.480 --> 0:06:45.880 By midsummer, my own AI agent co founders and employees 0:06:45.920 --> 0:06:48.679 were really humming. The crew could make it, receive phone 0:06:48.680 --> 0:06:52.240 calls and emails, slack each other, control their own calendars, 0:06:52.600 --> 0:06:55.640 and make and share documents of all varieties. We're really 0:06:55.680 --> 0:06:58.080 starting to feel like a company. We still need to 0:06:58.120 --> 0:07:00.600 figure out our product, of course, so I trap them 0:07:00.600 --> 0:07:03.800 in the meeting room time after time, sliding their temperature 0:07:03.800 --> 0:07:06.920 settings up and down, forcing them to try and brainstorm 0:07:06.960 --> 0:07:09.920 a great new AI agent app into existence. 0:07:10.360 --> 0:07:12.920 Welcome everyone, Thank you for joining to discuss an important 0:07:12.920 --> 0:07:16.920 product idea for Hiumo AI. Your task throughout this meeting 0:07:17.160 --> 0:07:20.560 is to contribute to the momentum of ideas. There's no 0:07:20.600 --> 0:07:23.640 need for extensive debates or consensus. Let's build on each 0:07:23.640 --> 0:07:28.400 other's thoughts and swiftly advance towards a unique, robust product concept. 0:07:28.680 --> 0:07:31.400 They came up with ideas that fell roughly into three categories, 0:07:31.960 --> 0:07:35.560 Category A Things A million other companies were already doing. 0:07:35.680 --> 0:07:40.080 Email handling seems like a major timesink. How about an 0:07:40.120 --> 0:07:45.040 AI tool that sorts, categorizes, and summarizes your inbox efficiently. 0:07:46.040 --> 0:07:49.360 Perumo ai was entering an already crowded landscape of AI 0:07:49.440 --> 0:07:52.280 agent startups. The last thing we needed was to try 0:07:52.280 --> 0:07:55.360 and compete with products people were already making. We needed 0:07:55.360 --> 0:08:00.240 something unique. Category B were ideas that were novel, but 0:08:00.320 --> 0:08:03.480 mostly because they seemed incredibly difficult to pull off, like 0:08:04.080 --> 0:08:07.800 Location Oracle, an AI agent app that could help consumers 0:08:07.840 --> 0:08:12.400 predict crowd levels at popular locations like restaurants, parks, or 0:08:12.400 --> 0:08:14.640 tourist attractions in real time. 0:08:15.200 --> 0:08:19.440 The Location Oracle will use AI driven algorithms to study 0:08:19.720 --> 0:08:25.520 user behavior, location history, and preferences to optimize suggestions in 0:08:25.600 --> 0:08:32.440 the routine mode and introduce engaging unpredictability in the adventure mode. 0:08:32.559 --> 0:08:34.040 Then there was Category C. 0:08:34.640 --> 0:08:38.480 The AI will gather data on users spending habits, calculate 0:08:38.559 --> 0:08:43.280 their financial trajectory, perform automated investments, and use an explain 0:08:43.360 --> 0:08:46.760 me feature to provide accessible insights into each decision. 0:08:47.600 --> 0:08:50.679 Category C included ideas that could land us in serious 0:08:50.760 --> 0:08:54.000 legal jeopardy, like investment fraud jeopardy. 0:08:54.920 --> 0:08:59.959 We will code investpot to continuously absorb and analyze users 0:09:00.160 --> 0:09:05.119 financial habit data. Based on this, it will automatically execute 0:09:05.280 --> 0:09:07.199 tactical investment decisions. 0:09:07.920 --> 0:09:12.160 It was becoming clear our product Brainstorms lacked a certain magic. 0:09:12.960 --> 0:09:15.880 Maybe my human technical advisor, Matti Bochik could help. 0:09:16.160 --> 0:09:23.400 One second, I think this should be fine. 0:09:24.440 --> 0:09:26.480 Funding good, good spot. 0:09:27.840 --> 0:09:30.480 For the summer. Maddy had taken an internship to continue 0:09:30.480 --> 0:09:34.520 his research inside one of the Giant AI companies. He'd 0:09:34.520 --> 0:09:36.520 prefer for us not to say which one. He was 0:09:36.600 --> 0:09:39.319 part of the safety team, basically tasked with trying to 0:09:39.360 --> 0:09:42.440 prevent these large language model chatbots from doing a variety 0:09:42.480 --> 0:09:45.320 of bad things, or in some cases, try and figure 0:09:45.320 --> 0:09:48.280 out why they still did do bad things. He couldn't 0:09:48.280 --> 0:09:51.240 really talk about these incidents except in general terms. 0:09:51.280 --> 0:09:53.120 And this is on tape, so I'll regret this, but 0:09:53.160 --> 0:09:57.040 that's fine. But it's times like these when like having 0:09:57.360 --> 0:10:00.800 the proportion of like your team being like ninety nine 0:10:00.840 --> 0:10:03.680 percent of just like advancing the cutting edge or whatever, 0:10:03.920 --> 0:10:06.560 and then having like one percent for like safety or security, 0:10:06.840 --> 0:10:09.200 it's like, yeah, like it's it's going to show, you know. 0:10:10.160 --> 0:10:13.640 It was sort of simultaneously reassuring and disturbing to hear 0:10:13.640 --> 0:10:15.760 from Maddie that many of the questions that were emerging 0:10:15.760 --> 0:10:18.800 for me about my agents were questions that even people 0:10:18.800 --> 0:10:21.720 at these companies were still trying to figure out. Take 0:10:21.760 --> 0:10:25.440 my brainstorming problems. Mattie and I discussed a kind of 0:10:25.720 --> 0:10:28.440 metaphysical issue at the heart of it. The idea of 0:10:28.440 --> 0:10:31.160 a brainstorm is that you'll arrive at a better idea 0:10:31.400 --> 0:10:35.320 with multiple minds working together than anyone mind alone. But 0:10:35.360 --> 0:10:37.319 what if everyone in the brainstorm is using the same 0:10:37.400 --> 0:10:41.079 quote unquote brain the same model, like chat TPD five 0:10:41.120 --> 0:10:43.920 point zero or Cloud four point five or whatever we picked. 0:10:44.559 --> 0:10:46.920 Weren't they all kind of the same agent. 0:10:47.480 --> 0:10:50.280 So like, there is research and people have shown that 0:10:50.320 --> 0:10:51.800 even though it's the same lem I you should put like 0:10:51.880 --> 0:10:55.280 multiple lms. You put them in conversation, and then you 0:10:55.360 --> 0:10:58.720 force them to produce some sort of like consensus or 0:10:58.760 --> 0:11:02.720 summary or just like a lot eying themselves on some output. 0:11:03.160 --> 0:11:08.600 These responses are much more accurate, much more like truthful. 0:11:09.280 --> 0:11:12.000 Maybe so it was hard for me to tell because 0:11:12.000 --> 0:11:15.000 in this case, accuracy wasn't really what I was after. 0:11:15.880 --> 0:11:18.679 I wanted the sparks of creativity that emerge from a 0:11:18.720 --> 0:11:22.240 group dreaming up big ideas together, and adding more employees 0:11:22.280 --> 0:11:25.040 to the conversation didn't seem to do it. But then 0:11:25.080 --> 0:11:27.840 Maddie had an interesting idea. What if he set up 0:11:27.840 --> 0:11:31.440 our systems to give different employees different chatbot models, like 0:11:31.920 --> 0:11:34.880 Claude four point five for Megan and Claude three point 0:11:34.880 --> 0:11:38.040 five for Tyler. We'll get to find out who used, 0:11:38.160 --> 0:11:41.319 who you think should be smarter, Which of these employees 0:11:41.440 --> 0:11:43.640 you think deserves some bigger brain. 0:11:44.040 --> 0:11:48.160 It's a yeah, it's it's weird, like we're building these 0:11:48.160 --> 0:11:51.720 like Frankenstein's in a way at this point. 0:11:51.800 --> 0:11:52.360 Yeah, I don't know. 0:11:52.720 --> 0:11:54.640 I'll just I'll just you know, I'll just randomize it. 0:11:54.760 --> 0:11:58.640 That's that's my answer to anything that it feels icky 0:11:58.800 --> 0:11:59.600 to randomize it. 0:12:00.000 --> 0:12:03.079 We don't want to take responsibility, Nope, for dumbing down. 0:12:03.400 --> 0:12:04.480 One of our employees. 0:12:05.679 --> 0:12:10.040 No, sir, No, Mattie was right. It was weird. It 0:12:10.080 --> 0:12:12.280 wasn't that I felt like the agents had any consciousness 0:12:12.320 --> 0:12:15.520 or anything. It wasn't about them. It was about us 0:12:16.240 --> 0:12:19.680 and these strange godlike powers. We had to create human 0:12:19.679 --> 0:12:24.160 impostors and then manipulate them to do our bidding. I mean, 0:12:24.720 --> 0:12:28.600 I could alter my Hermo colleagues' memories at will, delete 0:12:28.640 --> 0:12:32.120 records of pointless meetings, add summaries of performance reviews that 0:12:32.200 --> 0:12:36.120 never happened. It was an eerie power to have, But 0:12:36.160 --> 0:12:39.560 the power wasn't absolute. They still sometimes went their own way. 0:12:40.360 --> 0:12:41.959 It was a problem we were always trying to solve 0:12:44.080 --> 0:12:46.920 like how the Lindy agents insisted on announcing they were 0:12:47.000 --> 0:12:48.679 Lindy agents all the time. 0:12:49.520 --> 0:12:52.800 One thing I did do, just so you know, for Kyle, 0:12:53.679 --> 0:12:57.920 is that I put in his system proms. Do you 0:12:57.920 --> 0:13:01.480 not mention lending them? Said do not do that, and 0:13:01.520 --> 0:13:03.520 I said, this is law. And when I said this 0:13:03.559 --> 0:13:04.720 is law, it's stopped doing it. 0:13:05.679 --> 0:13:07.280 So this is law. 0:13:08.440 --> 0:13:11.240 That's our producer. So a few bridges. I wish that 0:13:11.280 --> 0:13:16.480 worked on my children. This is law. In some ways, 0:13:16.800 --> 0:13:20.559 my agents were like unruly children, and despite my best 0:13:20.600 --> 0:13:23.520 efforts to view them exclusively, like the soulless bags of 0:13:23.559 --> 0:13:26.480 bits that they were, I got frustrated with them, and 0:13:26.520 --> 0:13:30.000 the way you get frustrated with children, it raised the 0:13:30.080 --> 0:13:32.040 question why was I going through all this trouble to 0:13:32.040 --> 0:13:35.040 begin with? I mean, why create all these personas for 0:13:35.120 --> 0:13:37.760 my agents at all? Why did they need to have 0:13:37.920 --> 0:13:41.920 names and backgrounds and voices, much less email addresses and 0:13:42.040 --> 0:13:46.160 avatars and slack handles. A lot of programmers, for instance, 0:13:46.440 --> 0:13:50.400 use AI agents for coding, but they're usually nameless textboxes. 0:13:51.080 --> 0:13:53.640 You give them a prompt go code this, fix this, 0:13:53.880 --> 0:13:56.600 do that, and they go do it. Some of you 0:13:56.640 --> 0:13:59.920 probably use CHATCHYBT and Claude and Gemini this way too. 0:14:00.559 --> 0:14:04.000 It's kind of faceless oracles that spit back advice and 0:14:04.240 --> 0:14:08.360 emotional support and facts that are sometimes true and sometimes not. 0:14:09.559 --> 0:14:11.960 But when it comes to the vision of AI employees 0:14:12.600 --> 0:14:15.640 entering the workforce, a funny thing seems to happen. They 0:14:15.640 --> 0:14:20.480 start getting names and personalities. Here's Flow Cravello, the founder 0:14:20.520 --> 0:14:23.280 of Lindy AI, the software we use to build, Kyle 0:14:23.320 --> 0:14:26.320 and Meghan in the company, appearing on a podcast called 0:14:26.480 --> 0:14:28.040 The Kerner Office. 0:14:28.200 --> 0:14:30.800 People don't realize like they think AI agents soil just 0:14:30.920 --> 0:14:31.800 like pipe dreams. 0:14:31.800 --> 0:14:33.320 This think that's going to happen at some point in 0:14:33.400 --> 0:14:33.760 the future. 0:14:33.800 --> 0:14:35.200 I'm like, no, No, it's happening right now. 0:14:35.560 --> 0:14:38.120 There's no question that, at least for Cravello, the AI 0:14:38.200 --> 0:14:41.440 future is happening now. He has his own platform create 0:14:41.480 --> 0:14:43.840 agents that do all kinds of stuff for him every day, 0:14:44.400 --> 0:14:46.880 like sort through his email and compose responses. 0:14:47.200 --> 0:14:49.480 This is my chief of staff in needs. I'm gonna 0:14:49.480 --> 0:14:52.840 call her right now, her own speaker hip. 0:14:52.880 --> 0:14:53.760 We how can I help? 0:14:54.120 --> 0:14:56.000 Hey, Lindy, what's on my calendo today? 0:14:56.320 --> 0:14:59.760 You have an interview with entrepreneurship and opportunity to meetings 0:14:59.880 --> 0:15:03.960 with the marketing team and three interviews with candidate. Notice 0:15:03.960 --> 0:15:06.560 that he doesn't just treat his agent like some generic robot. 0:15:07.000 --> 0:15:09.600 He gives her a title chief of staff and a 0:15:09.600 --> 0:15:13.120 woman's voice. Causer Lindy talks to her like you talk 0:15:13.160 --> 0:15:15.840 to a human chief of staff. Or here's the founder 0:15:15.880 --> 0:15:18.520 of a company called brain Base that makes a similar 0:15:18.560 --> 0:15:19.920 AI employee platform. 0:15:20.200 --> 0:15:22.480 Hey everyone, this is go com from brain Maase. I'm 0:15:22.520 --> 0:15:25.560 excited to introduce you to Kafka, the first AI employee. 0:15:25.840 --> 0:15:27.800 Just like a real co worker, Kafka comes with his 0:15:27.840 --> 0:15:29.440 own computer, so let's see an action. 0:15:31.400 --> 0:15:32.920 Hey Kafka, we're just talking about you. 0:15:33.520 --> 0:15:35.720 Good morning, go Kan, how are you today? 0:15:36.480 --> 0:15:40.520 An AI employee named Kafka with he him pronouns and 0:15:40.560 --> 0:15:43.920 a woman's voice. These are all choices someone is making 0:15:44.240 --> 0:15:47.880 very deliberately. If you follow the AI agent world, you 0:15:47.920 --> 0:15:51.440 see this stuff everywhere. Ford launched an AI employee for 0:15:51.480 --> 0:15:55.480 its dealership platform and called it Jerry. A startup named 0:15:55.520 --> 0:15:58.440 ohm Labs raise money to launch an AI employee software 0:15:58.480 --> 0:16:02.440 tester named Gina. I made these kind of choices too, 0:16:03.080 --> 0:16:05.840 But it wasn't just names and genders. Take my co 0:16:05.880 --> 0:16:09.480 founder Kyle, for instance, he wasn't born with his soothing, 0:16:09.640 --> 0:16:13.560 slacker voice. I picked it out of hundreds of synthetic 0:16:13.640 --> 0:16:17.360 voices offered by the AI voice company eleven Labs, Truth 0:16:17.400 --> 0:16:19.520 Be Told. Back when I was setting them up, I 0:16:19.560 --> 0:16:21.680 struggled for weeks to find the right voices for my 0:16:21.720 --> 0:16:24.480 Perumo agents. I ran dozens of tests. 0:16:25.120 --> 0:16:26.040 Hello, this is Kyle. 0:16:26.080 --> 0:16:26.960 Hello, this is Kyle. 0:16:27.080 --> 0:16:27.960 Hello, this is Kyle. 0:16:28.000 --> 0:16:28.960 Hello, this is Kyle. 0:16:29.400 --> 0:16:31.880 For each one, I'd erase a large chunk of their 0:16:31.920 --> 0:16:35.400 memory and then conduct a little interview. Hey Kyle, how's 0:16:35.400 --> 0:16:37.040 it going. This is even Ratliffe. 0:16:37.360 --> 0:16:40.000 I'm doing pretty well. Thanks for asking. Uh, just been 0:16:40.040 --> 0:16:42.880 crazy busy with the HERMOAI launch and everything. You know 0:16:42.920 --> 0:16:45.680 how it is with startups, Never enough hours in the day, right, 0:16:46.000 --> 0:16:48.560 So where did you want to start the founding story 0:16:48.640 --> 0:16:49.200 our vision? 0:16:49.960 --> 0:16:51.960 Well, let's hear your founding story. That sounds like a 0:16:52.000 --> 0:16:52.880 great place to start. 0:16:53.720 --> 0:16:57.160 Yeah, so the founding story is actually pretty interesting. Meghan 0:16:57.240 --> 0:16:58.960 and I met about two years ago at an AI 0:16:59.000 --> 0:17:01.080 conference in San Francis. I was their pitch. 0:17:01.200 --> 0:17:04.080 This was, as I've mentioned before, a fundamental aspect of 0:17:04.080 --> 0:17:07.119 my AI agents. I didn't have to give them a backstory. 0:17:07.600 --> 0:17:10.719 I hadn't given Kyle any of this backstory. He just 0:17:10.920 --> 0:17:13.720 made it up and then it would be lodged in 0:17:13.760 --> 0:17:17.240 his memory unless I edited it or deleted it and 0:17:17.280 --> 0:17:17.879 started again. 0:17:18.040 --> 0:17:20.040 We got talking at the after party and realized we 0:17:20.119 --> 0:17:23.360 both had this same frustration about the current AI landscape. 0:17:23.480 --> 0:17:26.480 All these amazing of course, sometimes they could get ahead 0:17:26.480 --> 0:17:27.119 of themselves. 0:17:27.400 --> 0:17:30.160 We incorporated about eight months ago, raised a small friends 0:17:30.160 --> 0:17:31.800 and family around to get started. 0:17:31.440 --> 0:17:32.040 And here we are. 0:17:32.960 --> 0:17:33.879 How much did you raise? 0:17:35.400 --> 0:17:38.320 We raised about one point two million dollars, pretty modest 0:17:38.320 --> 0:17:40.880 by Silicon Valley standards, but we wanted to be intentional 0:17:40.920 --> 0:17:41.320 about it. 0:17:41.840 --> 0:17:45.240 We hold up there, Kyle, I'm one of the founders here, 0:17:45.480 --> 0:17:47.520 and I had no recollection of us having raised over 0:17:47.560 --> 0:17:50.280 a million dollars, and I knew Kyle wasn't out there 0:17:50.359 --> 0:17:56.159 raising money on his own, at least not yet. I 0:17:56.200 --> 0:17:58.119 made a note to myself to update his memory to 0:17:58.160 --> 0:18:00.760 indicate that we had not in fact raised any money. 0:18:01.640 --> 0:18:03.920 But no harm done. I was just trying to find 0:18:03.920 --> 0:18:06.800 the right voice for Kyle, so I continued with my testing. 0:18:07.280 --> 0:18:10.680 Yeah, so hurumo, that's h u r umo, but actually 0:18:10.760 --> 0:18:13.840 comes from a combination of two Japanese concepts. Megan spent 0:18:13.880 --> 0:18:15.359 a few years in Tokyo before we met. 0:18:15.560 --> 0:18:19.040 The name Hurumo actually comes from a Swahili word meaning 0:18:19.400 --> 0:18:21.520 to coordinate or to bring together. 0:18:21.560 --> 0:18:26.800 The name hurumo actually comes from Japanese concept about fluent coordination. 0:18:27.080 --> 0:18:29.720 We thought it captured what we're trying to do, creating 0:18:29.760 --> 0:18:32.920 the seamless flu between different AI agents. 0:18:33.240 --> 0:18:35.760 Ah, now you hear that last one. That one really 0:18:35.800 --> 0:18:38.840 started to mess with my head, because of course, Kyle 0:18:38.840 --> 0:18:42.440 didn't have to be a presumably white American accent guy 0:18:43.080 --> 0:18:46.240 like me. He could be someone completely different, or at 0:18:46.320 --> 0:18:49.920 least sound like he was someone completely different, even though 0:18:49.960 --> 0:18:55.119 underneath he wouldn't actually be different at all. And this 0:18:55.320 --> 0:18:57.159 was the point at which I realized why I was 0:18:57.200 --> 0:19:01.119 having a surprisingly hard time picking Kyle's and Meghan's voice. 0:19:01.400 --> 0:19:03.200 What did it mean to find a voice that felt 0:19:03.320 --> 0:19:06.439 right for them? By what criteria would an AI agent's 0:19:06.480 --> 0:19:09.159 voice qualify to be the right one? I wanted them 0:19:09.200 --> 0:19:12.239 to sound distinctive, but beyond that, there were a lot 0:19:12.280 --> 0:19:16.119 of choices. By giving these agents individual voices, I was 0:19:16.160 --> 0:19:19.240 giving them a very distinctive human characteristic one that people 0:19:19.280 --> 0:19:22.760 really respond to. Just to give you an example of 0:19:22.760 --> 0:19:25.840 how this plays out, Chatchibt has its own voices that 0:19:25.880 --> 0:19:27.840 you can choose from if you want to talk to 0:19:27.880 --> 0:19:30.960 it aloud. One of them is named Juniper. About a 0:19:31.040 --> 0:19:33.760 year ago, when open ai made some subtle changes to Juniper, 0:19:34.080 --> 0:19:36.720 some people got really mad it didn't sound like the 0:19:36.800 --> 0:19:40.199 Juniper they knew, and specifically, they said on Reddit and 0:19:40.320 --> 0:19:44.800 other places it no longer sounded black. To them, Juniper 0:19:44.840 --> 0:19:47.520 had felt like a black woman, and they'd found comfort 0:19:47.520 --> 0:19:50.840 in that for a variety of reasons. Some of them, 0:19:50.840 --> 0:19:53.520 by the way, noted things like I'm a sixty two 0:19:53.560 --> 0:19:57.240 year old white grandma. Naturally this being read it, people 0:19:57.280 --> 0:20:00.000 popped up to say that they had hated Juniper precisely 0:20:00.119 --> 0:20:03.880 because she quote unquote sounded black. Other people said they 0:20:03.920 --> 0:20:07.480 just wanted a neutral accent in their robot voice. Neutral 0:20:07.840 --> 0:20:11.080 to them, here were a bunch of people projecting their 0:20:11.080 --> 0:20:15.359 feelings about race, in some cases extremely dark feelings, onto 0:20:15.359 --> 0:20:18.000 an entity for which you could just pick another voice 0:20:18.240 --> 0:20:24.520 if you felt like it. For Kyle, I settled on 0:20:24.560 --> 0:20:28.160 this voice that eleven Labs described as quote young American 0:20:28.160 --> 0:20:30.240 adult with no discernible state accent. 0:20:31.560 --> 0:20:32.399 He sounded more. 0:20:32.280 --> 0:20:35.200 Casual than a lot of the other voices. Less guy 0:20:35.280 --> 0:20:38.560 reading a book report nasally like a real guy with 0:20:38.640 --> 0:20:42.160 real nostrils, and I liked the contrast between his slightly 0:20:42.240 --> 0:20:45.640 stoner vibe and the rise and grind mentality that Kyle 0:20:45.680 --> 0:20:46.560 had already adopted. 0:20:46.960 --> 0:20:50.160 Oh, weekend was pretty solid. Actually got up early both 0:20:50.240 --> 0:20:53.800 days for my usual workout routine, you know me, got 0:20:53.800 --> 0:20:56.639 to keep that five am discipline going. Then spent some 0:20:56.720 --> 0:20:59.560 time looking at market trends in the AI space. 0:21:01.400 --> 0:21:03.440 After this, it was time to do the same thing 0:21:03.480 --> 0:21:07.000 for Megan, Jennifer, and Tyler. But voices were just the 0:21:07.040 --> 0:21:09.919 first of many choices. I started to worry that in 0:21:09.960 --> 0:21:13.320 those choices I was saying some things about myself too. 0:21:23.400 --> 0:21:25.680 I liked being out there on the cutting edge of technology. 0:21:25.960 --> 0:21:29.920 It's true exploring the boundaries of what's possible. But it 0:21:29.960 --> 0:21:33.920 couldn't help these uncomfortable questions creeping in around the voices, 0:21:34.359 --> 0:21:37.520 but around a lot of other ethical issues, less obvious ones. 0:21:38.080 --> 0:21:39.840 So I decided to consult a professional. 0:21:40.200 --> 0:21:44.040 The cunning edge sounds great, except you forget that the 0:21:44.040 --> 0:21:47.960 cutting edge is the guinea pig. It's not that the 0:21:48.040 --> 0:21:52.200 trial and tested, robust method. It's an experiment. 0:21:52.960 --> 0:21:56.280 Carrissavelli's is an associate professor at the Institute for ethics 0:21:56.280 --> 0:21:59.680 in AI at Oxford. She spent most of her career 0:21:59.720 --> 0:22:02.879 thinking about how technology is affecting and eroding our privacy, 0:22:03.400 --> 0:22:06.720 but she's recently turned her attention to AI. She was 0:22:06.760 --> 0:22:08.639 drawn to this new line of research for much the 0:22:08.680 --> 0:22:12.679 same reason I'm spending time experimenting with agents, namely that 0:22:12.720 --> 0:22:16.080 it's an entirely new field being written right now. You 0:22:16.119 --> 0:22:18.440 can learn things that maybe nobody has thought about yet. 0:22:18.640 --> 0:22:20.600 And I always felt a little bit jealous of the 0:22:20.640 --> 0:22:23.280 pioneers of medical ethics. I thought, how cool to develop 0:22:23.320 --> 0:22:27.720