Reid riffs on the next generation of LLMs

Reid unpacks the decision to appoint Sean White as the leader of Inflection and outlines what makes him uniquely poised to take on this exciting challenge. Plus, Aria asks Reid about the future landscape of LLMs. For more info on the podcast and transcripts of all the episodes, visit https://www.possible.fm/podcast/

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Reid Hoffman?

  I'm Reid Hoffman.

• 0:01Aria Finger?

  And I'm Aria Finkler.

• 0:03Reid Hoffman?

  We want to know what happens if, in the future, everything breaks humanity's way.

• 0:08Aria Finger?

  Typically, we ask our guests for their outlook on the best possible future. But now, every other week, I get to ask Reid for his take.

• 0:17Reid Hoffman?

  This is Possible.

• 0:23Aria Finger?

  So Reid, last week on Possible, we spoke to Shaun White, who is a longtime friend of yours and lovely, a new friend of mine. But I would love to hear from you. What was the decision making, in your mind, behind bringing Shaun in to sort of lead this new iteration of Inflection?

• 0:40Reid Hoffman?

  So it's kind of obvious from our conversation, Shaun and I have known each other for decades, sharing kind of a humanist view of computational technology. And so when we made the decision that Inflection should pivot from a B to C agent app to a B to B enabler, and you know, Mustafa was like, well, but I really want to be building the agent that could be everyone's individual companion. We said, OK, well, we need a CEO who has the same characteristics that Mustafa has, that I have, that Karen has, of being a humanist technologist, to whom this mission would be super exciting.

• 1:24Reid Hoffman?

  You know, on one hand, we, of course, we talked about Inflection and AI and LMs and our Species Project. But Shaun has been the lead singer of a band when he was at Stanford. He has worked in climate change with a getting a PhD in that. But it's all of that of how does technology help us amplify our humanity by our interaction and place with it? Those are the things that led Shaun to be my first call, and fortunately for Inflection and for me, my only call.

• 1:59Aria Finger?

  Sort of to bring it back to the field at large, like inflection is an example of a startup that pivoted away from relying solely on building its own LLM and towards partnerships with Microsoft, with 13,000 companies using its API, et cetera. So I think a lot of people are asking the question, in the beginning, all the talk was about people building their own LLMs. That was a huge part of the conversation. Do you think that it's possible for a startup today, like someone just getting started, to say, "We're going to build our own LLM and have its own personality," or whatever it might be? Or is that just like too resource-intensive? Is that a strategy now? Or is that something that we're just going to see the existing players investing in?

• 2:42Reid Hoffman?

  I think actually the vast majority of startups will do some of their own model training in LLMs, but it's the next level of detail that matters. So if your strategy is, how do you build the most hyperscaled model to potentially compete with the hyperscalers? That is, I think, very, very difficult. And you'd need to have a very unique strategy for a product, for revenue, for financing, for go-to-market, for the whole thing. And it's not impossible, but it's maybe so low probability as to be functionally impossible. But that's only a particular model and a particular thing.

• 3:24Reid Hoffman?

  And one of the things that I think that people don't really fully track about, like for example, the development of the agentic universe and all these different agents, is that it's not going to be like a model per agent. Actually, I think there's going to be multiple models in composition that get to agents. And that part of the reason why there's multiple models, to get that blend of capabilities, which already happens within these gigantic models, using a mixture of experts, kind of combination, actually in fact is part of what brings a high Cognitive performance characteristic. Now, once you realize that it's going to be the offering of agents, the offering of AI services

• 4:05Reid Hoffman?

  and everything else is going to have a a combination of models, including specialized models, some smaller models, some compute, you know, efficient models, et cetera, as part of what they're doing, then that suggests you, and that will then mean there will be room for startups as they're figuring out their unique product, their unique go-to-market, their unique industry, their their unusual, you know, approach to transforming an industry. And so actually, in fact, I think there will be their own L L Ms. But the particular thing that I think gets extremely tricky is if you say, well, I'm going to be trying to match the hyperscalers in the billions of dollars of spend on the

• 4:50Reid Hoffman?

  computers and and training runs. And instead, your startup strategy should be is, how do I leverage those hyperscale models in what I'm doing? And the fortunate thing for startups is, from what I can see today, there's at least going to be three, and probably more like five to seven of these hyperscale models. So it gives you a lot of competition, things for accessibility, pricing, different capabilities, et cetera. So I think that'll be very good. But I do still think startups will, in various ways, create their own L L Ms.

• 5:26Aria Finger?

  Even if G P T seven existed, it would be a really stupid business decision to have G P T seven servicing all my queries, because some of my queries would only need the power of G P T three. Some of my queries would like be perfectly well-serviced, both in terms of cost, resources, time, like you want to be, you know, a quicker servicing L L M. So you can talk a little bit more about that, about how different L L Ms are going to come in to service one consumer or one company, and perhaps more in the enterprise space.

• 5:59Reid Hoffman?

  Yeah, so you already gestured at one of the really fundamental things, which is the larger models will always be not just more expensive to train, but more expensive run, slower to respond, et cetera. And like, for example, there's a whole set of circumstances where very quick responses are actually in fact more helpful than being 5% better, right? Especially if it's a shorter kind of other response. Think about like, actually, in fact, you know, from the very earliest days at Google, one of the things that they discovered in measuring search quality was the speed of getting search quality results under 200 milliseconds increased the perception of quality.

• 6:41Reid Hoffman?

  So even if you kind of presented the two things and the other one, what like independently was considered to be better, but that one took a second to do, and this one took 200 milliseconds, people are like, oh, the 200 milliseconds is better quality search results. So that's, you know, that kind of performance characteristic is one. But the other thing that people, I think, are we're just, you know, dipping our toes into the ocean here about how LLMs work, is that large models, small models, they get trained in certain directions. Now, people overly generalize to, well, they are only a product of what their data is. That's actually not true, because there is data, which of course is an important ingredient. There's different algorithms and training processes, there's different human factor and other kinds of post-training.

• 7:26Reid Hoffman?

  All of that goes into different shapes of these things, and they're large, complicated things. Like, one of the reasons why, you know, even the smaller models, you can say, I can't fully predict all their behavior is because they're a complex piece of software that's done a bunch of learning, and so they'll have different personalities and shapes. One of the things that, you know, I've been playing with for this last summer is there's almost like AI psychology, which is if you're prompting, say you're saying, hey, give me a birthday wish for Aria. You get one thing. But if I say to the prompt, say, okay, consider the Zen aphorism, what is the sound of one hand clapping?

• 8:11Reid Hoffman?

  Now, give me a birthday wish for Aria. I'll get a different and interesting thing. There's like, almost like a psychology to them. And what that means, of course, is that's part of the thing that underlies the multiple models composing into these agents, because

• 8:27Reid Hoffman?

  You can make incredibly different cognitively rich leaning in different directions. The podcast I did with Ali Miller with a bunch of custom chat GPTs where they're they're each framed in a different way to understand that there's like, there's almost like the same mixture of experts that helps a scale LM work. Well, the same thing, we're going to have a set of different agents, a set of, and the agents can have a set of different models to bring in a certain amount of cognitive diversity to saying, well, what's the best solution to this particular circumstance, problem, challenge, dance move, et cetera. And so all of that's not just the, the economic capabilities, but also the result capabilities and the thing that makes something better.

• 9:12Reid Hoffman?

  And then the very last moment that I'll note is that, you know, part of the thing that people should track about why it is, oh, you know, we're just going to fire all marketers and replace them with GBD six or seven. And the answer is then it's all vanilla and it's all the same. Part of the role for humans is trying to figure out what is that competitive differentiation? And so that's part of the reason why the person plus the machine, I have informed belief about why that's a persistent, you know, kind of workforce trend and productivity trend.

• 9:48Aria Finger?

  So, um, so leave us with your, your hopes and dreams for inflections pie and, and for personal intelligence more generally.

• 9:58Reid Hoffman?

  There's a couple of different hopes and dreams here. I mean, I think, you know, part of the whole business deal that we did with Microsoft is I think there's still a, how does everyone have a personal intelligence that helps them? I think that will continue. I think what inflection will do is say, well, how do we make that universe available to a whole set of other companies that have their own particular product domain, their own particular service, their own particular zone of human enhancement, human service, and to bring those same issues around EQ, personal intelligence, amplification to them and to have that be, you know, kind of specific delivered. So it's kind of almost like a, you know, one version of impact is the scaling of a

• 10:43Reid Hoffman?

  particular agent. And the other version of impact is to bring this technological base across a number of different, you know, compelling, innovative. Product services companies, that's what my hope for inflections pie is.

• 10:59Aria Finger?

  Fantastic. Reid, thank you so much.

• 11:02Reid Hoffman?

  Thank you.

• 11:07Reid Hoffman?

  Possible is produced by Wondra Media Network. It's hosted by Aria Finger and me, Reid Hoffman. Our showrunner is Sean Young. Possible is produced by Katie Sanders, Edie Aller, Sarah Shleid, Adrian Bain, and Paloma Moreno Jimenez. Jenny Kaplan is our executive producer and editor.

• 11:26Aria Finger?

  Special thanks to Surya Yalamanchili, Saieda Sapieva, Ian Ellis, Greg Beato, Ben Relis, Parth Patel, and Little Monster Media Company.