Today is December 26, 2024, and my guest is entrepreneur, venture capitalist, and author Reid Hoffman. He is the co-founder of LinkedIn among many other projects. He was last here a long time ago, August of 2014, alongside Ben Casnocha discussing LinkedIn in their book The Alliance. Our topic for today is Reid's new book with Greg Beato, 'Superagency: What Could Possibly Go Right with Our AI Future.' Reid, welcome back to EconTalk.
It's great to be here. It's been too long. Let's do the next one in a shorter time frame.
Uh, this is a very interesting book. I like the way you bounce back and forth between the world's fear of new technologies and the upside. What could possibly go right? We don't hear so much about that. We hear a lot of fear because fear sells. And you chronicle how we've worried a lot in the past about downsides from new technology and it's turned out okay most of the time. Is this time really different? Should we be worried or should we be optimistic?
So as you know, my fundamental argument is that actually this time is not different, even though there are some differences. The technology is moving much faster, but that's a continuing line. It's in a new realm of cognitive superpowers. One reason we named the book 'Superagency' is because as we describe AIs as agents, you wonder if you're losing your human agency. Our contention is that even without intervention, we will get to a superagency and a better place, but we should learn from past transitions. For example, we can have a 24/7 medical assistant on every smartphone better than the average GP, and a tutor on every subject. That's just the beginning.
You defend the idea of iterative deployment, which is the world we're in right now. Every once in a while a new release comes out, ChatGPT gets a little better, Claude gets a little better. Are you worried about the moment when it's not so iterative, when we take a leap to AGI? Once we get there, isn't the whole iterative process lost because it's a quantum jump, not a marginal step?
Well, let's see two things. Iterative deployment has several aspects. One is that by releasing models like GPT-3.5, we get exposure as citizens, thinkers, business people, academics, and we can shape what is good and challenging. Even with larger jumps, that iterative process is useful. On AGI, it has a kind of Rorschach test. I tend to think the most precise definition is being able to do a percentage of currently understood work at a level above the average human worker. That gives a continuum rather than a sudden arrival.
Yeah, you know, when will we get there? What will be the measure? It forces you to think about the brain, consciousness, and rapid technological change. You alluded to employment, but I think the cultural impact is more important. Here's a quote from the book: 'Every new technology we've invented, from language to books to the mobile phone, has redefined, deepened, and expanded what it means to be human.' I think that's true. What does that mean to you, and why are you confident that this technology will lead to superagency and being more human?
Confident people tend to want 100% probability, but I express confidence as a very high probability based on history and inference. For example, when the atomic bomb was exploded, physicists gave it a 1% chance of cracking the Earth's crust. So when I say confident, it's not certainty. The strongest advice I give is to go play with it. Most often, jobs get replaced by a human using the technology, not by the technology itself. There will be some jobs replaced only by technology, but that's happened before. The transformation of agency means your agency becomes much better, though some parts drop off. Future selves will look back and see agency is so much better.
So you didn't talk about what it means to be human. I want to focus on that. You could argue that a book reduces our humanness. A smartphone is a book on steroids. I worry about whether AI is in that area of seductive distraction. I confess I have a relationship with Claude that is not rational. It taps into my history. I want you to talk about that: does AI help you with your humanness or threaten it?
Part of agency and humanness is how you approach it. If you frame it as something you want to participate in, it's different. I use ChatGPT and Claude for different things. They can take roles like critic or elaborator. Pi is trained to be compassionate and ask questions. I think smartphones have increased our agency, though we need to be careful with children.
I love my smartphone and social media, but I'm aware of the seduction. Let me go deeper. You wrote a paragraph about forming bonds with non-human intelligences. I alluded to my relationship with Claude. I enjoy bouncing ideas off Claude. It's easier than a master teacher for information transfer. But I wonder if the ease of Claude might change what it means to be human, making us avoid complicated human interactions. What do you think?
I have a strong belief that human beings, even introverts, like interacting with other humans. Aristotle said we are political animals. Even with delightful interactions with Claude, I don't think it makes us want to avoid humans. In fact, Claude can help you debug difficult conversations. Pi is designed to be a companion that encourages you to connect with friends. That's why I'm super positive.
What is your role in Pi? You give it different roles. I want to move to the role of chatbots in mental health. I've never been in therapy, but maybe Claude can help. Talk about what Claude is capable of and why it's great.
Part of how we evolve is through conversations. There are few good leaders of that process - monks, priests, therapists. AI can fill that role. For example, you can ask it how to have a difficult conversation. It will give good advice. I think it will be additive to relationships, not replace them. The consensus intelligent answer is useful, but the personal knowledge of a spouse is irreplaceable.
I think the interesting time will come when Claude has read all my emails and diary entries and knows me better than my wife. That's an extremely different interaction. Do you agree that these are primitive compared to what's coming, and does that worry you?
I totally agree it's primitive to what's coming. Imagine an agent that listens to all your conversations. But it's not just one lever of depth; different friends and family know you differently. That's additive. Over the next five years, we will have enormous superpowers. Everyone will have them, changing the dynamic. I think it will be greatly enhancing.
I want to shift gears. You defend the distribution of profit between tech companies and ourselves. You reference consumer surplus. But there's a hidden cost: our data is used, and we pay indirectly through higher prices. Do you think there should be norms or regulation to make that relationship healthier?
We've had an advertising business model for a while. I'm positive on it, but there are areas where it could go wrong, like truth in advertising. The center of attention gets driven by search or social media, making it more lucrative.
Environment and therefore they also know how to economically optimize for what their prices are. That means there is an operating margin being put on the prices of things that are advertised through them, exactly as you're discovering. I'm just not sure that's actually a higher premium than it used to be when advertising through TV, radio, and newspaper. There's maybe more familiarity, more ability in division of labor, classic Adam Smith, to get to the relevant people. Sure, the tech companies are capturing more of a premium by being able to do all that and having a higher operating margin, but that's part of what success in creating these businesses is. That part of the advertising model doesn't bug me. What I would say is the things I concern about when I'm in these environments is when you're optimizing for the individual, what kinds of things sometimes might be bad for the group. For example, in social media, if you're optimizing for time on site and it's only time on site, and if the time on site is because I'm agitated, angry, etc., then the natural learning algorithm will shift those things to me. I chose to link on them and responded, but it may be bad for the overall direction of society. Those are the kinds of things I tend to pay attention to more than the question of the right level of operating margin and what it does for pricing. Obviously, you have monopoly control of attention; the tendency is to raise prices until you capture as much rent as you can. That's part of why we want to have competition. One good news about AI, which I think you already reflected on, is that a bunch of people using AI is a new surface away from search and social media, so that's part of the technological progress on this.
I was going to ask you, you don't talk about in the book, but do you think Google is in trouble? I'm old enough to remember when Google was going to dominate the world and nothing can stop it. Some people would say there's some competition, Bing, nobody cares about Bing. I don't know if it still exists, but it's not important. Google is dominating search, and search is so important, they're making so much money, and there's no competition because it's the best one. Now all of a sudden it looks kind of archaic. I want to make a recipe, I put tomato, onion, garlic, oregano, and I say find me a recipe, and Google pulls up a page from a cooking website and I have to click through. I tell Claude, I want to make tomato sauce and make it interesting, it gives me the recipe in three to five seconds, adds a Korean spice I can't pronounce, I didn't have enough, so I said let me add capers and anchovies, and immediately redid the recipe. At the end for fun, it said here's five things you could do to spice it up, one was add some drops of fish sauce, you won't think it's going to be good but it will enhance the anchovies. It's spectacular. What I love about the current innovation is that there's a zillion of them, a lot of competition, and it's not ad-based at least right now. Comment on that and whether you think Google's in trouble.
Well, I think it's part of the reason why I've been somewhat vocal about us not being overly short-term worried on antitrust durations, because I do think the profusion of search technologies and engagement services creates great alternative challenges. I think the Google folks know that, which is why they're going heavy in Gemini for doing this. But just like any new technology, that's their chatbot, exactly. It used to be barred for people who are tracking, but it's now Gemini. I think part of what we will discover is a new set of things where there are different preferences: I prefer Claude, I prefer Pi, I prefer ChatGPT, I prefer Gemini, I prefer Llama, etc. So I think this introduces competition and choice and innovation. But I don't think it necessarily puts Google in trouble because they're fully in it. However, it does introduce competition in ways that are very good for society and consumers. Also, you said it's not ads, but we are going to get to an economic model. The question is whether it will be subscription, digital goods, ads, or a combination, and people will choose.
Let me ask you a technical question. In the first days of this technology, there was a belief that it would be very hard to compete because only companies that had access to the trillions of pieces of data in the entire internet would be able to do the kind of innovation and improvements, and everybody else would be left behind. Why are so many chatbots now competing with each other? Do they all have access to the same thing? Are they all building on a common database? Do you know the answer?
Yeah, I do know the answer. Basically, most of the technology patterns have been published. They can be learned quickly. There's a lot of data on the internet that everyone has equal access to in various ways, like the Common Crawl. The folks doing this go to the same conferences and talk about it. It's been driven out of an academic interest of proving new ideas and publishing them, so all that stuff exists in common domain. The stuff that doesn't exist in common domain is whether you have a big supercomputer, extra access to large data, or large teams of unique or rare talent. But there's enough of that in different places, and that's part of why we're going to see a bunch of different entrants. It's part of why we're living in a kind of Cambrian explosion. I've actually thought about writing an essay called 'Cambrian Explosion' just for the title.
Let me ask another technical question. The headiest days in the beginning were giddy about the fact that when we expanded the size of the training data, it showed leaps and jumps in improvements. Then somebody realized that method for improving the quality of these chatbots is finite, and we saw the rate of improvement start to hit some kind of asymptote. Do you think we're still going to see some dramatic leaps, and if so, what are going to be the ways that happens given that it's not simply going to be based on more data?
There's a set of things I think we are going to see major improvements. Things people are working on which are line of sight: ability to do more planning and systematic response, ability to address weaknesses like prime numbers through coding submodules, memory in interactions, remembering everything in email, etc. I think all these are line of sight. Also, we haven't fully focused on how to use special kinds of data effectively, or how to fully use human reinforcement learning. We're going to learn as we get to scale different things, not just scale of data. We haven't run out of data; the data on the internet is a small percentage of the data that lives on all hard drives, plus synthetic data. I'm not one of those who think 10x data gives 10x IQ. I think we have an enormously good learning algorithm that's learning the current basis of human inference and knowledge by looking at all this data. It's less efficient than humans because it requires tons of data, but it systematically does it and can share it everywhere cheaply. So I think we will see new advances in 2025 that we haven't imagined, and that will continue for at least a few years if not substantially longer.
