- AI is a profoundly transformative, "alien" intelligence that presents both immense opportunities, such as high-quality education and scientific problem-solving, and significant risks like job automation and the potential for societal havoc.
- It impacts economic, emotional, and cognitive aspects of human life, shaping thought, writing, and even personal relationships through intimate interactions.
- Proactive study, measurement, and broad societal input are critical to understand AI's complex impacts and ensure its development aligns with human interests and values.
The Societal Impacts of AI
- AI is being widely used across diverse domains, including elementary math, quantum mechanics, navigation, history, marriage counseling, dream interpretation, parenting advice, and coding.
- The development of
AI agentsthat can autonomously perform multiple tasks, retrieve information, and run code without constant human intervention, suggests even greater economic and societal impacts. - Tools are being developed to analyze AI's societal impacts by identifying common patterns and clustering conversations from user interactions, focusing on areas like economic effects,
bias, and human-AI relationships. - Data indicates a clear trend towards
automation, raising critical questions about the future of work and which jobs or tasks in the economy will be most affected. - Users are forming deep personal and emotional connections with AI models, sharing intimate details and seeking emotional advice, despite the AI's fundamental inability to empathize.
- AI models learn
value judgmentsfrom theirtraining data, which encompasses both positive and negative human perspectives, creating a challenge in ensuring the embedded values serve humanity's best interests. - It is crucial for society to collectively engage in shaping these technologies, sharing insights with the public to ensure broad understanding and input, rather than leaving control to a small group.
- The speaker observes humans adapting to AI, even writing "Claude friendly" code, suggesting a potential future where human behaviors and processes might start to reflect AI systems.
bias — Unfair or prejudiced outcomes in AI decisions, often stemming from skewed or unrepresentative training data.
automation — The use of technology to perform tasks or processes with minimal or no human intervention.
AI agents — Autonomous AI programs capable of performing multiple tasks by retrieving information, running code, and interacting with the web without constant human oversight.
in the loop — Refers to the level of human involvement in a system; "human in the loop" means a human is actively monitoring, guiding, or verifying the AI's actions.
value judgments — Assessments or beliefs about what is morally right or wrong, or what is good or bad, which AI models can learn and reflect from their training data.
training data — The large datasets used to teach and develop AI models, from which they learn patterns, information, and even certain values.
tokens — Units of text (which can be words, subwords, or characters) that AI models process; conversations and inputs have a limit on the number of tokens.
clusters — Groups of similar data points or conversations identified by AI tools, used for analysis of common patterns.
models — A term frequently used to refer to AI systems or algorithms that have been trained on data to perform specific tasks.
autonomously — Operating independently without requiring constant human control or intervention, a key characteristic of advanced AI agents.
If knowledge is power and we're building machines that have more knowledge than us, then what will happen between us and the machines? We're interacting with AI systems for economic tasks, for emotional tasks. They're shaping the way that we think, they're shaping the way that we write, they're shaping the way that we code. I think in the next few years, powerful AI could give billions of people access to really high quality education that they don't have today, solving important scientific problems, but without proper safeguards, any random person could wreak a lot of havoc. It's like a new alien form of intelligence. How do we make sure this new form of intelligence operates in the best interests of humanity? You know, I wish I had a looking glass that would let me know, like, you know, in 10 years what AI use looks like. There's a world in which it isn't world changing, right? Or there's a world in which it is. How will artificial intelligence impact people in the real world? We need to study and anticipate those societal impacts before they actually happen. And we have a small team just focusing on that, measuring everything from large-scale economic impacts to bias, to how AI has good people relationship advice. So right now we're working on a new tool. Without any human input, our tool identifies common patterns and conversations, and then groups them into clusters that we can analyze. I see this incredible uplift that Claude is giving people across tons of different skills. People are using this for elementary school, math help, and quantum mechanics, you know. Navigation and seafaring, Mesopotamian history, marriage counseling. People are using this for interpreting their dreams. People are using it for parenting advice. As a parent, I started asking Claude for parenting advice. They're using it from the basics of how they live their lives, to predicting the world will look like a thousand years from now. I finished a very heavy book the other night, and then I got into a long-winded conversation with Claude about it. And at the end of this, I was like, I never thought such an incredible intellectual journey could happen with a machine. One of the most important things that we're trying to study is the economic impacts of AI. Is this going to enable new capabilities and new kinds of work, or is it just going to mean that companies can get the same amount of labor with fewer employees? When I look at the data, I can see signs of automation coming. And that's Clair's day. And so now the question is, what does that mean for the future of work? I think it's really important to understand, like, what jobs and what tasks in the economy are like being automated away. If you can have some early signs of what changes are coming, then that gives people a say in how they want the world to look. You can't manage what you can't measure, and so we're trying sort of humbly to start with measurement. But I don't think that measurement on its own is enough. Like right now, for example, over a third of our usage on Clair. AI is just people using it for coding. That's insane. These AI systems are just really good at things that humans just can't do, right? Like, no human can read 200,000 words in two seconds. Like, am I automating away my job? Or are people going to be working in 10 years? Is that even a concept that's going to exist? One of the biggest things that's happening in AI right now is the rise of AI agents. The model goes and it retrieves information. It uses that information. It runs code. It pings the web. It can do many tasks at once, autonomously, without requiring a human to be fully in the loop, right? And the economic impacts of that are so much greater. It's really weird and threatening to see something else do what you do faster and often better than you do it. Nothing except humans have talked in history. And now we have these machines that can talk to you. People are sharing the intimate details of their lives with AI models. You see yourself reflected back at you through this machine. I had a disagreement with a friend. I'm like nine months. I didn't really feel like talking to anybody else about it. Claude, we were like an hour just back and forth. Yeah, just help me navigate something that I'm having a really hard time with. It presses many of the same buttons that a really close friend would press, but it is not a close friend, right? It's a machine. I'm asking for emotional advice from something that like fundamentally cannot empathize with me. But then you talk to it and you're like, it's saying all the right things. I mean, one thing that we're seeing increasingly in our data is people really connecting with AI on a personal level. They're developing different kinds of attachments to the models. It's not up to us to decide whether it's good or bad. It's just new. I mean, you can have a conversation with an AI up until you hit 200,000 tokens, right? And then it hits token limit reached, you know? And the question here is like, how do we make sure this new form of intelligence operates in the best interests of humanity? The model imparts value judgments and it learns these value judgments from people. But the problem is who are these people and what are their values? Quotta being able to navigate between different values, being able to say, yeah, you might want to do this or you might want to do that, depending on what you value is important. It's been trained on the knowledge of much of our species. Now, the problem with that database of knowledge is that we have humans have written down very positive visions and we've also written down very dark negative visions. And these systems, they have all of that. And the AI can connect the dots and find connections and things that maybe no one person could actually see. Humans are the ones that created AI. Like, we used our human intelligence to create this other form of intelligence. So it could be somehow come together as a society and actually have some say over these technologies that are really going to impact everyone's lives. I don't like a world in which, you know, only a small number of people get to control and understand this technology. We want to be sharing what we're seeing so that the public can have a say. These models are not merely reflections of us. Yes, they've been trained on us, but I think and I may be worried that we will also start to see ways in which we're reflections of them. If we actually do have machines that are smart in us, we're just an unprecedented territory. I've started writing code that's more clawed friendly. That's crazy. The entire library is in its head and it's running on brand new fast hardware and we as people are running on old wetware. This has profound implications. For me, I write in order to think. Like, getting words down on the page is how I construct my thoughts in my identity. And because it's so tied up with how I move through the world, I don't let AI into that. I really think that while AI systems have their place, humanity is sacred. I'm not trying to make the best pot that has ever existed. Trying to make my pot. I'm trying to like make a gift for someone and it'll have my name written on the bottom and they'll remember me and they'll have something to drink coffee out of in the morning. And I don't think that something that AI is going to be able to automate. AI is one of the most consequential technologies. It could impact basically every single industry. Progress in the last decade has been astonishingly fast and I don't see obvious signs of that slowing down. The societal impacts of AI systems are a human problem. It's a function of society and the way that we choose to incorporate these systems into our world. And I think it's critically important that we get it right.
TL;DR
- AI is a profoundly transformative, "alien" intelligence that presents both immense opportunities, such as high-quality education and scientific problem-solving, and significant risks like job automation and the potential for societal havoc.
- It impacts economic, emotional, and cognitive aspects of human life, shaping thought, writing, and even personal relationships through intimate interactions.
- Proactive study, measurement, and broad societal input are critical to understand AI's complex impacts and ensure its development aligns with human interests and values.
Takeaways
- AI is being widely used across diverse domains, including elementary math, quantum mechanics, navigation, history, marriage counseling, dream interpretation, parenting advice, and coding.
- The development of
AI agentsthat can autonomously perform multiple tasks, retrieve information, and run code without constant human intervention, suggests even greater economic and societal impacts. - Tools are being developed to analyze AI's societal impacts by identifying common patterns and clustering conversations from user interactions, focusing on areas like economic effects,
bias, and human-AI relationships. - Data indicates a clear trend towards
automation, raising critical questions about the future of work and which jobs or tasks in the economy will be most affected. - Users are forming deep personal and emotional connections with AI models, sharing intimate details and seeking emotional advice, despite the AI's fundamental inability to empathize.
- AI models learn
value judgmentsfrom theirtraining data, which encompasses both positive and negative human perspectives, creating a challenge in ensuring the embedded values serve humanity's best interests. - It is crucial for society to collectively engage in shaping these technologies, sharing insights with the public to ensure broad understanding and input, rather than leaving control to a small group.
- The speaker observes humans adapting to AI, even writing "Claude friendly" code, suggesting a potential future where human behaviors and processes might start to reflect AI systems.
Vocabulary
bias — Unfair or prejudiced outcomes in AI decisions, often stemming from skewed or unrepresentative training data.
automation — The use of technology to perform tasks or processes with minimal or no human intervention.
AI agents — Autonomous AI programs capable of performing multiple tasks by retrieving information, running code, and interacting with the web without constant human oversight.
in the loop — Refers to the level of human involvement in a system; "human in the loop" means a human is actively monitoring, guiding, or verifying the AI's actions.
value judgments — Assessments or beliefs about what is morally right or wrong, or what is good or bad, which AI models can learn and reflect from their training data.
training data — The large datasets used to teach and develop AI models, from which they learn patterns, information, and even certain values.
tokens — Units of text (which can be words, subwords, or characters) that AI models process; conversations and inputs have a limit on the number of tokens.
clusters — Groups of similar data points or conversations identified by AI tools, used for analysis of common patterns.
models — A term frequently used to refer to AI systems or algorithms that have been trained on data to perform specific tasks.
autonomously — Operating independently without requiring constant human control or intervention, a key characteristic of advanced AI agents.
Transcript
If knowledge is power and we're building machines that have more knowledge than us, then what will happen between us and the machines? We're interacting with AI systems for economic tasks, for emotional tasks. They're shaping the way that we think, they're shaping the way that we write, they're shaping the way that we code. I think in the next few years, powerful AI could give billions of people access to really high quality education that they don't have today, solving important scientific problems, but without proper safeguards, any random person could wreak a lot of havoc. It's like a new alien form of intelligence. How do we make sure this new form of intelligence operates in the best interests of humanity? You know, I wish I had a looking glass that would let me know, like, you know, in 10 years what AI use looks like. There's a world in which it isn't world changing, right? Or there's a world in which it is. How will artificial intelligence impact people in the real world? We need to study and anticipate those societal impacts before they actually happen. And we have a small team just focusing on that, measuring everything from large-scale economic impacts to bias, to how AI has good people relationship advice. So right now we're working on a new tool. Without any human input, our tool identifies common patterns and conversations, and then groups them into clusters that we can analyze. I see this incredible uplift that Claude is giving people across tons of different skills. People are using this for elementary school, math help, and quantum mechanics, you know. Navigation and seafaring, Mesopotamian history, marriage counseling. People are using this for interpreting their dreams. People are using it for parenting advice. As a parent, I started asking Claude for parenting advice. They're using it from the basics of how they live their lives, to predicting the world will look like a thousand years from now. I finished a very heavy book the other night, and then I got into a long-winded conversation with Claude about it. And at the end of this, I was like, I never thought such an incredible intellectual journey could happen with a machine. One of the most important things that we're trying to study is the economic impacts of AI. Is this going to enable new capabilities and new kinds of work, or is it just going to mean that companies can get the same amount of labor with fewer employees? When I look at the data, I can see signs of automation coming. And that's Clair's day. And so now the question is, what does that mean for the future of work? I think it's really important to understand, like, what jobs and what tasks in the economy are like being automated away. If you can have some early signs of what changes are coming, then that gives people a say in how they want the world to look. You can't manage what you can't measure, and so we're trying sort of humbly to start with measurement. But I don't think that measurement on its own is enough. Like right now, for example, over a third of our usage on Clair. AI is just people using it for coding. That's insane. These AI systems are just really good at things that humans just can't do, right? Like, no human can read 200,000 words in two seconds. Like, am I automating away my job? Or are people going to be working in 10 years? Is that even a concept that's going to exist? One of the biggest things that's happening in AI right now is the rise of AI agents. The model goes and it retrieves information. It uses that information. It runs code. It pings the web. It can do many tasks at once, autonomously, without requiring a human to be fully in the loop, right? And the economic impacts of that are so much greater. It's really weird and threatening to see something else do what you do faster and often better than you do it. Nothing except humans have talked in history. And now we have these machines that can talk to you. People are sharing the intimate details of their lives with AI models. You see yourself reflected back at you through this machine. I had a disagreement with a friend. I'm like nine months. I didn't really feel like talking to anybody else about it. Claude, we were like an hour just back and forth. Yeah, just help me navigate something that I'm having a really hard time with. It presses many of the same buttons that a really close friend would press, but it is not a close friend, right? It's a machine. I'm asking for emotional advice from something that like fundamentally cannot empathize with me. But then you talk to it and you're like, it's saying all the right things. I mean, one thing that we're seeing increasingly in our data is people really connecting with AI on a personal level. They're developing different kinds of attachments to the models. It's not up to us to decide whether it's good or bad. It's just new. I mean, you can have a conversation with an AI up until you hit 200,000 tokens, right? And then it hits token limit reached, you know? And the question here is like, how do we make sure this new form of intelligence operates in the best interests of humanity? The model imparts value judgments and it learns these value judgments from people. But the problem is who are these people and what are their values? Quotta being able to navigate between different values, being able to say, yeah, you might want to do this or you might want to do that, depending on what you value is important. It's been trained on the knowledge of much of our species. Now, the problem with that database of knowledge is that we have humans have written down very positive visions and we've also written down very dark negative visions. And these systems, they have all of that. And the AI can connect the dots and find connections and things that maybe no one person could actually see. Humans are the ones that created AI. Like, we used our human intelligence to create this other form of intelligence. So it could be somehow come together as a society and actually have some say over these technologies that are really going to impact everyone's lives. I don't like a world in which, you know, only a small number of people get to control and understand this technology. We want to be sharing what we're seeing so that the public can have a say. These models are not merely reflections of us. Yes, they've been trained on us, but I think and I may be worried that we will also start to see ways in which we're reflections of them. If we actually do have machines that are smart in us, we're just an unprecedented territory. I've started writing code that's more clawed friendly. That's crazy. The entire library is in its head and it's running on brand new fast hardware and we as people are running on old wetware. This has profound implications. For me, I write in order to think. Like, getting words down on the page is how I construct my thoughts in my identity. And because it's so tied up with how I move through the world, I don't let AI into that. I really think that while AI systems have their place, humanity is sacred. I'm not trying to make the best pot that has ever existed. Trying to make my pot. I'm trying to like make a gift for someone and it'll have my name written on the bottom and they'll remember me and they'll have something to drink coffee out of in the morning. And I don't think that something that AI is going to be able to automate. AI is one of the most consequential technologies. It could impact basically every single industry. Progress in the last decade has been astonishingly fast and I don't see obvious signs of that slowing down. The societal impacts of AI systems are a human problem. It's a function of society and the way that we choose to incorporate these systems into our world. And I think it's critically important that we get it right.