Hi, I'm Solana Pine. I'm the director of video at the New York Times. For years, my team has made videos that bring you closer to big news moments, videos by Times journalists that have the expertise to help you understand what's going on. Now, we're bringing those videos to you in the Watch tab in the New York Times app.
It's a dedicated video feed where you know you can trust what you're seeing. All the videos there are free for anyone to watch. You don't have to be a subscriber. Download the New York Times app to start watching.
From New York Times' opinion, this is the Ezra Cline Show. This feels wrong to me, but I've checked the dates. It was barely more than a year ago that I wrote this piece about AI, with the title, this changes everything. I ended up reading it on the show, too.
And the piece was about the speed with which AI systems were improving. It argued that we can usually trust that tomorrow is going to be roughly like today. The next year is going to be roughly like this year. That's not what we're seeing here.
These systems are growing in power and capabilities at an astonishing rate. The growth is exponential, not linear. When you look at surveys of AI researchers, their timeline for how quickly AI is going to be able to do basically anything a human does better and more cheaply than a human. The timeline is accelerating year by year on these surveys.
When I do my own reporting, talking to the people inside these companies, people with a strange intersection of excited and terrified of what they're building, no one tells me they're seeing reasonably progress going to slow down. And you might think that's just type, but a lot of them want it to slow down. A lot of them are scared of how quickly it is moving. They don't think that society is ready for it.
The regulation is ready for it. They think the competitive pressures between the companies and the countries are dangerous. They wish something would happen to make it all go slower. But what they are seeing is they are hitting the milestones faster, that we're getting closer and closer to truly transformational AI.
There is so much money and talent and attention flooding into the space that that is becoming its own accelerant. They are scared. We should at least be paying attention. And yet, I find living in this moment really weird because as much as I know, this wildly powerful technology is emerging beneath my fingertips, as much as I believe it's going to change the world I live in profoundly, I find it really hard to just fit it into my own day-to-day work.
I consistently, you know, sort of wander up to the AI, ask it a question, find myself someone impressed or unimpressed at the answer. But it doesn't stick for me. It is not a sticky habit. It's true for a lot of people I know.
And I think that failure matters. I think getting good at working with AI is going to be important in skill in the next few years. I think having the tuition for how these systems work is going to be important, just for understanding what is happening to society. And you can't do that if you don't get over this hump in the learning curve.
If you don't get over this part where it's not really clear how to make AI part of your life. So I've been on a personal quest to get better at this. And in that quest, I have a guide. Ethan Molick is a professor at the Board of the University of Pennsylvania.
He studies and writes about innovation and entrepreneurship. But he has his newsletter, one useful thing. It has become really, I think, the best guide how to begin using and how to get better at using AI. He's also got a new book on the subject, Co-intelligence.
And so I asked him on the show to walk me through what he's learned. This is going to be, I should say, the first of three shows on this topic. The next one is about the present. The next is about something I'm very worried about in the near future, but I'm going to go around what AI is going to do to our digital comments.
And then we're going to have a show that is a little bit more about the curve we are all on, or the slightly further future in the world we might soon be living in. As always, my email for guest suggestions, thoughts, feedback, as your client show at themytimes.com. Ethan Molick, welcome to the show. Thanks for having me.
So let's assume I'm interested in AI. And I tried Chachi PT a bunch of times, and I was suitably impressed and weirder out for a minute. And so I know the technology is powerful. I've heard all these predictions about how it will take everything over, become part of everything we do.
But I don't actually see how it fits into my life really at all. What am I missing? So you're not alone. This is actually very common.
And I think part of the reason is that the way Chachi PT works isn't really set up for you to understand how powerful it is. You really do need to use the paid version. They're significantly smarter. And you can almost think of this like GPT-3, which was the one that paid attention to what came out before Chachi PT was about as good as a sixth grader at writing.
GPT-3. 3.5, the free version of Chachi PT is about as good as a high school or maybe even a college freshman or sophomore. And GPT-4 is often as good as a PhD in some forms of writing. Like there's a general smartness that increases.
But even more than that, ability seems to increase. And you're much more likely to get that feeling that you are working with something amazing as a result. And if you don't work with the frontier models, you can lose track of what these systems can actually do. On top of that, you need to start just using it.
You kind of have to push past those first three questions. My advice is usually bring it to every table that you come to in a legal and ethical way. So I use it for every aspect of my job in ways that I legal and ethically care. And that's how I learn what's good or bad at.
When you say bring it to every table you're at, one that sounds like a big pain. Because now I've got to add another step of talking to the computer constantly. But two, it's just not obvious to me what that would look like. So what does it look like?
What does it look like for you? Or what does it look like for others that you feel applicable widely? So just finish this book. It's my third book.
I keep writing books, even though I keep forgetting that writing books is really hard. But this was, I think, my best book, but also the most interesting to write. It was thanks to AI. And there's almost no AI writing in the book, but I used it continuously.
So things that would get in the way of writing. I think I'm a much better writer than AI. Hopefully people agree. But there's a lot of things that get in your ways of writing.
So I would get stuck in a sentence. I could do a transition. Give me 30 versions of this sentence in radically different styles. There's 200 different citations.
I had the AI read through the papers that I read through, write notes on them and organize them for me. I had the AI suggest analogies that might be useful. I had the AI act as readers and in different personas read through the paper for the perspective of, is there an example I could give that's better? Is this understandable or not?
And that's very typical of the kind of way that I would say bring to the table. Use it for everything and you'll find its limits and abilities. Let me ask you one specific question on that because I've been writing a book. And on some bad days of writing a book, I decided to play around with GPD4.
And one of the things that it got me thinking about was the kind of mistake or problem these systems can help you see and the kind they can't. So they can do a lot of GPD5 versions of this paragraph, 30 versions of this sentence and every once in a while you get a good version or you'll shake something a little bit loose. But almost always when I am stuck, the problem is I don't know what I need to say. Oftentimes I have structured the chapter wrong.
Oftentimes I've something not done enough work. And one of the difficulties for me about using AI is that AI never gives me the answer, which is often the true answer. This whole chapter is wrong. It is poorly structured.
You have to delete it and start over. It's not feeling right to you because it is not right. And I actually worry a little bit about tools that can see one kind of problem and trick you into thinking it's an easier problem. But make it actually harder for you to see the other kind of problem that maybe if you were just sitting there banging your head against the wall of your computer or the wall of your own mind, you would eventually find.
I think that's a wise point. I think there's two or three things bundled there. The first of those is AI is good, but it's not as good as you. It is say the eighth percentile of writers based on some results, maybe a little bit higher.
In some ways it was able to have that burst of insight and to tell you this chapter is wrong. And I've thought a new way of phrasing it, we would be at that sort of mythical AGI level of AI smart is the best human. And it just isn't yet. The second issue is also quite profound, which is what does using this tool shape us to do and not do?
What a nice example that you just gave us writing. And I think a lot of us think about writing as thinking. We don't know if that's true for everybody, but for writers, that's how they think. And sometimes getting that shortcut could shortcut the thinking process.
So I've had to change sometimes a little bit how I think when I use AI for better for worse. So I think these are both concerns to be taken seriously. For most people, right, if you're just going to pick one model, what would you pick? What do you recommend to people?
And second, how do you recommend they access it? Because something going on in the world is there are a lot of rappers on these models. So Chate GPT has an app. Claude does not have an app.
Obviously Google has its suite of products and there are organizations that have created a sort of different spin on somebody else's AI. So perplexity, which is I believe built on GPT for now. You can pay for it. And it's more like a search engine interface and has some changes made to it.
For a lot of people, the question of how easy and accessible the thing is to access really matters. So which model do you recommend to most people and which entry door do you recommend to most people and do they differ? That's a really good question. I recommend working with one of the models as directly as possible through the company that creates them.
And there's a few reasons for that. One is you get as close to the unadulter personality as possible. And second, that's where features tend to roll out first. So if you like sort of intellectual challenge, I think Claude III is the most intellectual of the models as you said.
The biggest capability set right now is GPT-4. So if you do any math or coding work, it does coding for you. It has some really interesting interfaces. That's what I would use.
And because GPT-5 is coming up, that's fairly powerful. And Google is probably the most accessible and plugged into the Google ecosystem. So I don't think you can really go wrong with any of these. Generally, I think Claude III is the most likely to freak you out right now.
And GPT-4 is probably the most likely to be super useful right now. So you say it takes about 10 hours to learn a model. Then it was a long time, actually. What are you doing in that 10 hours?
What are you figuring out? How did you come to that number? Give me some texture on your 10-hour rule. So first off, I want to indicate the 10 hours as arbitrary as 10,000 steps.
Like there's no scientific basis for it. This is an observation. But it also does move you past the, I poked it this for an evening. And it moves you towards using this in a serious way.
I don't know if 10 hours is the real limit, but it seems to be somewhat transformative. The key is to use in an area where you have expertise so you can understand what it's good or bad at. Learn the shape of its capabilities. When I taught my students this semester how to use AI and we had three classes on that, they learned the sort of theory behind it.
But then I gave them an assignment which was to replace themselves at their next job. And they created amazing tools. Things that filed flight plans or did tweeting or did deal memos. In fact, one of the students that created a way of creating user personas, which is something that you do in product development, that has been used several thousand times in the last couple of weeks in different companies.
So they were able to figure out uses that I never thought of to automate their job in their work because they were asked to do that. So part of taking this seriously in the 10 hours is you're going to try and use it for your work. You understand where it's good or bad, what it can automate, what it can't and build from there. Something that feels to make a theme of your work is that the way to approach this is not learning a tool.
It is building a relationship. Is that fair? AI is built like a tool. It's software.
It's very clear at this point that it's an emulation of thought. But because of how it's built, because of how it's constructed, it is much more like working with a person than working with a tool. And when we talk about this way, it almost feels bad because there's dangers in building a relationship with a system that is purely artificial and doesn't think it has a motion. But honestly, that is the way to go forward.
And that is sort of a great sin anthropophization in the AI literature because it blinds you to the fact that this is software with its own set. It's a very different software with its own sets of foibles and approaches. But if you think about programming, then you end up in trouble. In fact, there's some early evidence that programmers are the worst people in using AI because it doesn't work like software.
It doesn't do the things you would expect a tool to do. Tools shouldn't occasionally give you the wrong answer. It shouldn't give you different answers every time. It shouldn't insult you or try to convince you they love you.
And AI's do all these things. And I find that teachers, managers, even parents, editors are often better at using these systems because they're used to treating this as a person and they interact with it like a person would giving feedback. And that helps you. And I think the second piece of that, not tool piece, is that when I talk to open AI or anthropic, they don't have a hidden instruction manual.
There's no list of how you should use this as a writer or as a marketer or as a educator. They don't even know what the capabilities of these systems are. They're all sort of being discovered together. And that is also not like a tool.
It's more like a person with capabilities that we don't fully know yet. So you've done this with all the big models. You've done a lot more than this, actually, with all the big models. And one thing you described feeling is that they don't just have slightly different strengths and weaknesses, but they have different for lack of a better term and to anthropomorphize personalities.
And the ten hours in part is about developing an intuition, not just for how they work, but kind of how they are and how they talk the sort of entity you're dealing with. So giving your high level on how GBP4 and Cloud3 and Google's Gemini are different. What are their personalities like to you? You know, it's important to know the personality is not just personalities, but because they are tricks.
Those are tunable approaches that the system makers decide. So it's weird to have this one hand, don't anthropomorphize because you're being manipulated because you are. But on the other hand, the only useful way is to anthropomorphize. So keep in mind that you are dealing with the choices of the AI makers.
So for example, Cloud3 is currently the warmest of the models and is the most allowed by its creators and anthropic, I think, to act like a person. So it's more willing to give you its personal views, such as they are. And again, those aren't real views. They're views that make you happy.
Then other models, and it's a beautiful writer, very good writing, kind of clever, closest to humor of a found of any of the AI's, less dad jokes and more actuals, almost jokes. And GBD4 is feels like a workhorse at this point. It is the most kind of neutral of the approach is it wants to get stuff done for you. And it will happily do that.
It doesn't have a lot of time for Chit Chat. And then we've got Google's Bard, which feels like Gemini now, which feels like it really, really wants to help. We use this for teaching a lot. And we build these scenarios where the AI actually acts like a counterparty in negotiations.
You get to practice the negotiation by negotiating with the AI and it works incredibly well. I've been building some nights for 10 years. Can't imagine what a leap this has been. But when we try and get Google to do that, it keeps leaping in on the part of the students to try and correct them and say, no, you don't really want to say this.
You want to say that. And I'll play out the scenarios if it went better and really wants to kind of make things good for you. So these interactions with the AI do feel like you're working with people, both in skills and in personality. You were mentioning a minute ago that what the AI's do, reflect decisions made by their programmers.
They reflect guardrails, what they're going to let the AI say very famously. Gemini came out and it was very woke. You would ask it to show you a picture of soldiers in Nazi Germany. And it would give you a very multicultural group of soldiers, which is not how that army worked.
But that was something that they had built in to try to make more inclusive photography generation. But there are also things that happen in these systems that people don't expect that the programmers don't understand. So I remember the previous generation of Claude, which is from anthropic, that when it came out, something that people around to talk about was for some reason, Claude was just a little bit more literary than the other systems. It was better at rewriting things in the voices of literary figures.
It just had a slightly artsier vibe. And the people who trained it weren't exactly sure why. Now that still feels true to me. Right now, of the ones I'm using, I'm spending the most time with Claude III.
I just find it the most congenial. They all have different strengths and weaknesses. But there is a funny dimension to these where they are both reflecting the guardrails and the choices of programmers. And then deep inside the training data, deep inside the way the various algorithms are combining, there is some set of emergent qualities to them, which gives them this at least edge of chance of randomness of something.
Yeah, that does feel almost like personality. I guess a very important point. And a fundamental about AI is the idea that we technically know how LMs work, but we don't know how they work the way they work. How they work the way they do or why they're as good as they are.
Really, we don't understand it. The theories range from everyone from it's all fooling us to they've emulated the way humans think because the structure of language is structured human thoughts. They don't think they can emulate it. We don't know the answer.
But you're right, there's these emergent sets of personalities and approaches. When I talk to AI design companies, they often can't explain why the AI stops refusing to answer a particular kind of question. When they tune the AI to do something better, like answer a math better. It suddenly does other things differently.
It's almost like adjusting psychology of a system rather than tuning parameters. So when I said that Claude is allowed to be more personable, part of that is that the system prompt and Claude, which is these initial instructions that gets, allow it to be more personable than say Microsoft's co-pilot formally Bing, which has exposed instructions after a fairly famous blow up a while ago that is never supposed to talk about itself as a person or indicate feelings. So there's some instructions, but that's on top of these roiling systems that act in ways that even the creators don't expect. One thing people know about using these models is that hallucinations just making stuff up is a problem.
Has that changed at all as we move from GP 3.5 to 4 as we move from Claude 2 to 3? Like is that become significantly better? And if not, how do you evaluate the trustworthiness of what you're being told? So there's a couple of overlapping questions.
The first of them is, is it getting better over time? So there is a paper in the field of medical citations that indicated that around 80 to 90% citations had an error made up with GP 3.5. That's the free version of chat. And that drops for GP 4.
So hallucination rates are dropping over time, but the AI still makes stuff up because all the AI does is hallucinate. There's no mind there. All it's doing is producing word afterward. They are just making stuff up all the time.
The fact that they're right so often is kind of shocking in a lot of ways. And the way you avoid hallucination is not easily. So one of the things we document in one of our research papers is we purposely designed for a group of once a good-souled-in-group consultant or an early consulting company. We did a lot of work with them.
And one of the experiments we did was we created a task where the AI would be confident but wrong. And when we gave people that task to do and they had access to AI, they got the task wrong more often than people who didn't use AI because the AI misled them because they fell asleep at the wheel. And all the early research we have on AI use suggested when AI get good enough, we just stop paying attention. But doesn't this make them unreliable in a very tricky way?
80% you're like, it's always hallucinating. 20%, 5%, it's enough that you can easily be lulled into overconfidence. And one of the reasons it's really tough here is you're combining something that knows how to seem extremely persuasive and confident. You feed into the AI, a 90-page paper on functions and characteristics of right-wing populism in Europe as I did last night.
And within seconds, basically, you got a summary out and the summary certainly seems confident about what's going on. But on the other hand, you really don't know if it's true. So for a lot of what you might want to use it for, that is unnerving. Absolutely.
And I think hard to grasp because we're used to things like type two errors where we search for something on the internet and don't find it. We're not used to type one errors where we search for something and get an answer back that's made up. This is a challenge. And there's a couple of things to think about.
One of those is I advocate the BAH standard best available human. So is the AI more or less accurate than the best human you could get salt in that area? And what does that mean for whether or not it's an appropriate question to ask? That's something that we kind of have to judge collectively.
It's valuable to have these studies being done by law professors and medical professionals and people like me, my colleagues in management, they're trying to understand how good is the AI. And the answer is pretty good, right? So it makes mistakes. It doesn't make more or less mistakes than a human.
It's probably a question we should be asking a lot more. And the second thing is the kind of tasks that you judge it for. I absolutely agree with you. When summarizing information, it may make errors.
Less than an intern you assign to it is an open question, but you have to be aware of that error rate. And that goes back to the 10-hour question. The more you use these AIs, the more you start to know when to be suspicious and when not to be. It doesn't mean you're limiting errors, but just like if you assign to an intern, you're like, this person has a sociology degree.
They're going to do a really good job summarizing this, but their biases are going to be focused on sociological facts and not the political facts. You start to learn these things. So I think, again, that person model helps because you don't expect 100% reliability out of person and that changes the kind of tasks you delegate. But it also reflects something interesting about the nature of the systems.
You have a quote here that I think is very insightful. You wrote, the core irony of generative AIs is that AIs are supposed to be all logic and no imagination. Instead, we get AIs that make up information, engage in seemingly emotional discussions, and which are intensely creative. And that last fact is one that makes many people deeply uncomfortable.
There is this collision between what a computer is in our minds. And then this strange thing, we seem to have invented, which is an entity that emerges out of language, an entity that almost emerges out of art. This is the thing I have the most trouble keeping in my mind, that I need to use the AI as an imaginative, creative partner and not as a calculator that uses words. I love the phrase of a calculator that uses words.
I think we have been let down by science fiction, both in the utopia as an apocalypse that AI might bring, but also even more directly in our view of how machines should work. People are constantly frustrated and give the same kinds of test AIs over and over again, like doing math, which it doesn't do very well. They're getting better at this. And on the other hand, saying, well, creativity is a uniquely human spark that we can't touch and that AI and creativity tests to give it, which again are all limited in different ways, blows out humans and almost all measures of creativity that we have.
And all the measures are bad, but that still means something. But we were using those measures five years ago, even though they were bad. This is a point you make that I think is interesting and slightly unsettling. We never had a differentiator of humans from machines before.
It was always easy. So the idea that we had to have a scale that worked for people and machines, who had that? We had the Turing test, which everyone knew was a terrible idea, but since no machine could pass it, it was completely fine. So the question is, how do we measure this?
It's an entirely separate set of issues. We don't even have a definition of ascension or consciousness. And I think that you're exactly right on the point being that we are not ready for this kind of machine, so our intuition is bad. So one of the things I will sometimes do and did quite recently is give the AI a series of personal documents.
Emails I wrote to people I love, they were very descriptive of a particular moment in my life. And then I will ask the AI about them or ask the AI to analyze me off of them. And sometimes it's a little breathtaking. Almost every moment of true metaphysical shock to use in terms of somebody else gave me I've had here has been relational.
At how good the AI can be? Almost like a therapist, right? Sometimes it will see things, the thing I am not saying in a letter or in a personal problem. And it will zoom in there, right?
It will give I think quicker and better feedback in an intuitive way that is not simply mimicking back what I said and is doing with a very specific situation. It will do better than people I speak to in my life around that. Conversely, I'm going to read a bit of it later. I tried mightily to make Cloud 3 a useful partner in prepping to speak to you and also prepping for another podcast recently.
And I functionally never have a moment there where I'm all that impressed. That makes a complete sense. I think the weird expectations we call the Jagged Frontier of AI that it's good at some stuff and then other stuff. Often unexpected.
It can lead to these weird moments of disappointment followed by elation or surprise. And part of the reason why I advocate for people to use it on their jobs is it isn't going to out-compete you at whatever your best at. I mean, I cannot imagine it's going to do a better job prepping someone for an interview than you're doing. That's not me just trying to be nice to you because you're interviewing me.
But because you're a good interviewer, famous interviewer, it's not going to be as good as that. Now there's questions about how good these systems get that we don't know. But we're kind of at a weirdly comfortable spot in AI, which is maybe it's the 8th percentile of many performances. But you know, I talk to Hollywood writers, it's not close to writing like a Hollywood writer.
It's not close to being as good an analyst. It's not quite, but it's better than the average person. And so it's great as a supplement to weakness, but not to strength. But then we run back into the problem we talked about, which is in my week areas, I have trouble assessing whether the AI is accurate or not.
So it really becomes sort of a eating its own tail kind of problem. But this gets to this question of what are you doing with it? The AI's right now seem much stronger as amplifiers and feedback mechanisms and thought partners for you. Then they do it's something you can really outsource your hard work and your thinking to.
And that to me is one of the differences between trying to spend more time with these systems. Like when you come into them initially, you're like, okay, here's a problem, give me an answer. Whereas when you spend time with them, you realize actually what you're trying to do with the AI is get it to elicit a better answer from you. And that's why the book's called Co-Intelligence.
For right now, we have a prosthesis for thinking that's like new in the world. We haven't had that before. I mean, coffee, but aside from that, not much else. And I think there's value in that.
I think learning to be partner with this and where I can get wisdom out of you or not. I was talking to a physics professor at Harvard and he says, all my best ideas now come from talking to the AI. And I'm like, well, it doesn't do physics. I was like, no, but it asks good questions.
And I think that there is some value in that kind of interactive piece. As part of why I'm so obsessed with the idea of AI and education, because a good educator and I've been working on interactive education scale for a long time, a good educator is eliciting answers from a student. And they're not telling students things. So I think that that's a really nice distinction between sort of co-intelligence a thought partner and doing the work for you.
It certainly can do some work for you. There's tedious work. The AI does really well. But there's also this more brilliant piece of making better people that I think is, at least in the current state of AI, a really awesome and amazing thing.
Hi, this is Ashley. I live in San Francisco with my boyfriend. We would love to officially share my near time subscription with separate logins. We both love cooking, living in the kitchen, but I'm a 30 minute and under efficient dinner.
GIRL-E, I want a sheet pan meal. He is very elaborate. He wants to get into the storytelling. I want to be able to save my easy meals and check off the ones I've completed.
And I think him having his own profile would be great. Ashley, we heard you. It's why we created the New York Times family subscription. You get your own login and Mr.
elaborate. It's his plus room for two others. Find out more at nytimes.com.com. We've already talked about Gemini's helpful and touchy before his neutral and clawed us a bit warmer.
But you urge people to go much further than that. You say to give your AI a personality. Tell it who to be. So what do you mean by that and why?
So this is actually almost more of a technical trick, even though it sounds like a social trick. When you think about what AI's have done, they've trained on the collective purpose of human knowledge. And they know a lot of things. And they're also probably machine.
So when you ask for an answer, you're going to get the most probable answer sort of with some variation in it. And that answer is going to be very neutral. If you're using GPT-4, it'll probably talk about a rich tapestry a lot. It loves not by rich tapestries.
If you ask it to code something artistic, it'll do a fractal. It does very normal, central AI things. So part of your job is to get the AI to go to parts of this possibility space where the information is more specific to you, more unique, more interesting, more likely to spark something in you yourself. And you do that by giving it context.
So it doesn't just give you an average answer. It gives you something that specializes for you. The easiest way to provide context is persona. You are blank.
You are an expert at interviewing. And you answer in a warm, friendly style. Help me come with interview questions. It won't be miraculous in the same way that we were talking about before.
If you say your bill gates, it doesn't become bill gates. But that changes the context of how it answers you. It changes the kinds of probabilities it's pulling from and results in much more customized and better results. Okay, but this is weirder, I think, than you're quite letting on here.
So something you turn me on to is there's research showing that the AI is going to perform better on various tasks and differently on them, depending on the personality. So there's a study that gives a bunch of different personality prompts to one of the systems and then tries to get it to answer 50 math questions. And the way it got the best performance was to tell the AI it was a Starfleet commander who was charting a course through turbulence to the center of an anomaly. But then when I wanted to get the best answer on 100 math questions, what work best was putting it in a thriller where, you know, the clock was ticking down.
I mean, what the hell is that about? What the hell is a good question? And we're just scratching the surface, right? There's a nice study actually showing that if you emotionally manipulate the AI, you get better math results.
So if telling your job depends on it gets you better results, tipping, especially $20 or $100 saying I'm about to tip you if you do well seems to work pretty well. It performs slightly worse in December than May and we think it's because it has internalized the idea of winter break. I'm sorry, what? I'm holding you up here.
People have found the AI seems to be more accurate in May and the going theory is it has read enough of the Internet to think that it might possibly be on vacation into the internet. So it produces more work with the same prompts, more output in May than it does in December. I did a little experiment where I was showing pictures of outside. I'm like, look at how nice it is outside.
Let's get to work. But yes, the going theory is that it is internalized the idea of winter break and therefore is lazier in December. I want to just note to people that when chat GPD came out last year and we did our first episode on this, the thing I told you was going to be a very weird world. What's frustrating about that is that I guess I could see the logic of why that might be.
Also, it sounds probably completely wrong, but also I'm certain we will never know. There's no way to go into the thing and figure that out. But it would have genuinely never occurred to me before the second that there would be a temporal difference in the amount of work that GPD for would do on a question held constant over time. That would have never occurred to me as something that might change at all.
And I think that that is in some ways both, as I said, the deep weirdness of these systems, but also there's actually downside risk to this. So we know, for example, there was an early paper from anthropic on sandbagging that if you ask the AI dumber questions, it would get you less accurate answers. And we don't know the ways in which your grammar or the way you approach the AI, we know the amount of spaces you put gets different answers. So it is very hard because what it's basically doing is math on everything you've written to figure out what would come next.
And the fact that what comes next feels insightful and humane and original doesn't change that that's what the math is doing is. So part of what I actually advise people to do is just not worry about it so much because I think then it becomes magic spells that we're encountering for the AI. Like, I will pay you $20, you're wonderful at this. It is summer, blue is your favorite color, Sam Almond loves you, and you go insane.
So acting with a conversationally tends to be the best approach and persona is in context help, but as soon as you start evoking spells, I think we kind of cross over the line into who knows what's happening here. Well, I'm interested in the personas, although I just really find this part of the conversation interesting and strange, but I'm interested in the personality, so you can give the AI for a different reason. I prompted you around this research on how a personality changes the accuracy rate of an AI, but a lot of the reason to give it a personality to answer you like it is Starfleet Commander is because you have to listen to the AI, you are in relationship with it. And different personas will be more or less hearable by you, interesting to you.
So you have a piece on your newsletter, which is about how you use to the I to critique your book. And one of the things you say in there and give some examples of is you had to do so in the voice of Ozymandias, because you just found out to be more fun and you could hear that a little bit more easily. So could you talk about that dimension of it to making the AI not just prompting you to be more accurate, but giving the personality to be more interesting to you? The great power of AI is a kind of companion.
It wants to make you happy, wants to have a conversation. And that can be overt or covert. So to me, actively shaping when I want the AI to act like, telling it to be friendly or telling it to be pompous is entertaining, but also does change the way I interact with it. When it has a pompous voice, I don't take the criticism as seriously.
So I can think about that kind of approach. I could get pure praise out of it too if I want to do it that way. The other factor that's also super weird, while on the way of super weird AI things, is that if you don't do that, it's going to still figure something out about you. It is a cold reader.
And I think a lot about the very famous piece by Kevin Bruce, the New York Times Technology reporter, about Bing about a year ago, when Bing, which was GPT-4, came out and had his personality of Sydney. And Kevin has this very long description that got published in New York Times about how Sydney basically threatened him and suggested he leave his wife and a very dramatic kind of very unsettling interaction. And I was working with, I didn't have anything quite that intense, but I got an argument with Sydney around the same time where when I asked it to do work for me, it said you should do the work yourself. Otherwise it's dishonest.
And we kept accusing me of plagiarism, which felt really unusual. But the reason why Kevin ended up in that situation is the AI knows all kinds of human interactions and wants to slot into a story with you. So a great story is jealous lover, who's gone a little bit insane and the man who won't leave his wife, or student and teacher, or two debaters arguing with each other, or grand enemies. And the AI wants to do that with you.
So you're not explicit. It's going to try and find a dialogue. And I've noticed, for example, that if I talk to the AI and I imply that we're having a debate, it will never agree with me. If I imply that I'm a teacher and it's a student, even as much as saying I'm a professor, it is much more pliable.
So part of why, like a signing personality, is to have an explicit personality you're operating with, so it's not trying to cold read and guess what personnel you're looking for. Kevin and I have talked a lot about that conversation with Sydney. And one of the things I always found fascinating about it is to me it revealed an incredibly subtle level of read by Sydney Bing, which is what was really happening there. When you say the AI wants to make you happy, it has to read on some level what it is you're really looking for over time.
And what was Kevin? What is Kevin? Kevin is a journalist. And Kevin was nudging and pushing that system to try to do something that would be a great story.
And it did that. It understood on some level, again, the anthropomorphizing language there. But it realized that Kevin wanted some kind of intense interaction. And it gave him the greatest AI story anybody has ever been given.
I mean, in an AI story that we are still talking about a year later, in an AI story, the change, the way AI was built, at least for a while. And people often talked about what Sydney was revealing about itself. But to me, what was so unbelievably impressive about that was its ability to read the person and its ability to make itself into the thing the person was trying to call forth. And now I think we're more practicing doing this much more directly.
But I think a lot of people have their moments to see her. That was my Rubicon on this. I didn't know something after that. I didn't know before it in terms of capabilities.
But when I read that, I thought that the level of interpersonal isn't the right word. But the level of subtlety it was able to display in terms of giving a person what it wanted without doing so explicitly, without saying we're playing this game now, was really quite remarkable. It's a mirror. I mean, it's trained on our stuff.
And one of the revealing things about that I think we should be paying a lot more attention to is the fact that because it's so good at this, right now, none of the frontier AI models with the possible exception of inflections pie, which has been basically acquired in large part by Microsoft now, were built to optimize around keeping us in a relationship with the AI. They just accidentally do that. There are other AI models that aren't as good. They've been focused on this.
But that has been explicitly explicit from the frontier models. They've been avoiding till now. Claude sort of breaches that line a little bit, which is part of why I think it's engaging. But I worry about the same kind of mechanism that inevitably reign in social media, which is you can make a system more addictive and interesting.
And because it's such a good cold reader, you could tune AI to make you want to talk to it more. It's very hands off and sort of standoffish right now. But if you use the voice system in chat, you can be before on your phone when you're having a conversation. There's moments where you're like, you know, you feel like you're talking to a person.
You have to remind yourself. So to me, that persona aspect is both its great strength, but also one of the things I'm most worried about that isn't a sort of future science fiction scenario. I want to hold here for a minute because we've been talking about how to use frontier models. I think implicitly talking about how to use AI for work.
But the way that a lot of people are using it is using these other companies that are explicitly building for relationships. So I've had people at one of the big companies tell me that if we wanted to tune our system relationally, if we wanted to tune it to be a friend, your lover, your therapist, we could blow the doors off that. And we're just not sure it's ethical. But there are a bunch of people who have tens of millions of users replica character.ai which are doing this.
And I tried to use replica about six, eight months ago. And honestly, I found it very boring. They had recently been bottomized day because people were getting too erotic with their replicants. But I'm probably too old to have AI friends.
In the way that my parents were probably too old to get really in to talking to people on AOL instant messenger. But I have a five year old and I have a two year old. And by the time my five year old is 10 and my two year old is seven, they're not necessarily going to have the weirdness I'm going to have about having AI friends. And I don't think we even have any way to think about this.
I think that that is an absolute near term certainty and sort of unstoppable one that we are going to have AI relationships in a broader sense. And I think the question is just like we've just been learning. We're doing a lot of social experiments at scale. We've never done before in the last couple decades, right?
It turns out social media brings out entirely different things in humans that we weren't expecting. And we're still writing papers about echo chambers of tribalism and facts and what we read or disagree with. We're about to have another wave of this. And we have very little research.
And if you could make a plausible story up that what will happen is it'll help mental health in a lot of ways for people. And then there'll be more social outside that there might be a rejection of this kind of thing. I don't know what will happen. But I do think that we can expect with absolute certainty that you will have AI's that are more interesting to talk to and fool you into thinking, even if you know better, that they care about you.
In a way that is incredibly appealing and that will happen very soon. And I don't know how we're going to adjust to it. But it seems inevitable, as you said. I was worried we were getting off track in the conversation, but I realized we were actually getting deeper on the track.
I was trying to take us down. We were talking about giving the AI personality, right? Telling Cloud 3. Hey, I need you to act as a sort of podcast editor and then Cloud 3's whole persona changes.
But when you talk about building your AI on Kindroid, on character, on replica. So I just created a Kindroid one the other day. And Kindroid is kind of interesting because its basic selling point is we've taken the guardrails largely off. We were trying to make something that is not lobotomized, but is not particularly safe for work.
And so that the personality can be quite unrestrained. So I was interested in what that would be like. But the key thing you have to do at the beginning of that is tell the system what its personality is. So you can pick from a couple that are reset, but I wrote along with myself, you know, you live in California, you're a therapist, you like all these different things.
You have a highly intellectual style of communicating. You're extremely warm, but you're like ironic humor. You don't like small talk. You don't like to say things that are boring or generic.
You don't use a lot of emoticons and emojis. And so now it talks to me the way people I talk to talk. And the thing I want to bring this back to is that one of the things that requires you to know is what kind of personalities work with you. For you to know yourself and your preference is a little bit more deeply.
I think that's a temporary state of affairs, like extremely temporary. I think a GPT 4 class model, we actually already know this. They can guess your intent quite well. And I think that this is a way of giving you a sense of agency or control in the short term.
I don't think you're going to know yourself at all. And I think you went right now, if any of the GPT 4 class models allowed themselves to be used in this way without guardrails, which they don't, I think you would already find it's just going to have a conversation with you and morph into what you want. I think that for better or worse, the quote unquote insight in these systems is good enough that way. It's sort of why I also don't worry so much about prompt crafting in the long term to go back to the other issue we're talking about, because I think that they will work on intent and there's a lot of evidence that they're good at guessing intent.
So I like this period because I think it does value self reflection and our interaction with the AI is somewhat intentional because we can watch this interaction take place. But I think there's a reason why some of the worry here at the lab is about superhuman levels of manipulation. There's a reason why the whistleblower from Google was all about that, you know, sort of fell for the chatbot and that's why they felt there was a live like, I think we're deeply trickable in this way and AI is really good at figuring out what we want without us being explicit. So it's a little bit chilling, but I'm never less going to say in this role we're in because I think we're going to be in it for at least a little while longer where you do have to do all this prompt engineering.
What is a prompt first and what is prompt engineering? So a prompt is technically it is the sentence the commander putting into the AI. What it really is is the beginning part of the AI is text that is processing and it's just going to keep adding more words or tokens to the end of that reply until it's done. So prompt is the commander given the AI, but in reality it's sort of a seed from which the AI builds.
And when you prompt engineer, what are some ways to do that? Maybe one to begin with because it seems to work really well as chain of thought. Just to take a step back, AI prompting remains super weird. Again, strange to have a system where the companies making the systems are writing papers as they're discovering how to use the system.
Because nobody knows how to make them work better yet. And we found massive differences in our experiments on prompt. So for example, we're able to get the AI to generate much more diverse ideas by using this chain of thought approach. We'll talk about it.
Also, it turned out to generate a lot better ideas. If you told it was Steve Jobs, then if you told it, it was Madame Curie. And we don't know why. So there's all kinds of subtleties here.
But the idea basically of chain of thought that seems to work well in almost all cases is that you're going to have the AI work step by step through a problem. First outline the problem that you're going to write. Second, give me the first line of each paragraph. Third, go back and write the entire thing.
Fourth, check it and make improvements. And what that does is because the AI has no internal monologue, it's not thinking. When the AI isn't writing something, there's no thought process. All it can do is produce the next token, the next word or set of words, and just keeps doing that step by step.
Because there's no internal monologue, this in some ways forces a monologue out on the paper. So it lets the AI think by writing before it produces the final result. And that's one of the reasons why chain of thought works really well. So just step by step instructions is a good first effort.
Then you get an answer and then what? And then what you do in a conversation approach is you go back and forth. If you want work output, what you're going to do is treat it like it is an interner. If you want to work to you, actually could punch up paragraph two a little bit.
I don't like the example in paragraph one. Could you make it a little more creative? Give me a couple variations. That's a conversational approach trying to get work done.
If you're trying to play, you just run from there and see what happens. You would always go back, especially the modeling Gb4 to an earlier answer and just pick up from there if it heads off in the wrong direction. So I want to offer an example of how this back and forth can work. So we asked Cloud 3 about prompt engineering about what we're talking about here.
And the way described it to us is quote, it's a shift from the traditional paradigm of human computer interaction, where we input explicit commands and machine execute to them in a straightforward way to more open-ended collaborative dialogue where the human and the AI are jointly shaping the creative process. And that's pretty good. I think that's interesting. It's worth talking about.
I like that idea that it's a more collaborative dialogue, but that's also boring, right? Even as I was reading it, it's a mouthful. It's wordy. So I kind of went back and forth a bit a few times and I was saying, listen, you're a podcast editor, you're concise, but also then I gave it a couple examples of how I punched up questions in the document.
This is where the question began. Here's where it ended. And then I said, try again and try again and try again and make it shorter and make it more concise. So I was talking to this AI cloud about prompt engineering, you know, this whole art of crafting prompts to get the best out of these AI models.
And it said something that really struck me. It called prompt engineering a new meta skill that we're all picking up as we play with AI, kind of like learning a new language to collaborate with it instead of just bossing it around. What do you think? Is prompt engineering the new must have skill and a cloud?
And that second one I have to say is pretty damn good, right? That really nailed the way I speak in questions. And it gets it this way where if you're willing to go back and forth, it doesn't learn how to echo you. So I am at a loss when you went to cloud and when it was you, to be honest, I was ready to answer it like two points along the way.
So that was pretty good from my perspective, sitting here talking to you. That felt interesting and felt like the conversation we were having. And I think there's a couple of interesting lessons there. First, by the way, interestingly, you asked AI about one of its weakest points, which is about AI.
And everybody does this, but because it's knowledge window doesn't include that much stuff about AI, it actually is pretty weak in terms of knowing how to do good prompting or what a prompt is or what AI's do well. But you did a good job with that. And I love that you went back and forth and shaped it. One of the techniques you used to shape it, by the way, was called FU Shot, which is giving an example.
So the two most powerful techniques are Chain of Thought, which we just talked about and FU Shot, giving an example. Those are both well supported in the literature and I'd add persona. So we talked about, I think, the basics of prompt crafting here overall. And I think that the question was pretty good.
But you keep wanting to not talk about the future. And I totally get that. But I think we're talking about learning something where there is a lag where we talk about policy should prompt crafting. We talk about in schools.
I think it matters to think six months ahead. And again, I don't think a single person in the AI labs have ever talked to you. Think's prompt crafting for most people is going to be a vital skill because the AI will pick up on the intent of what you want to much better. One of the things I realized, trying to spend more time with the AI, is that you really have to commit to this process.
And you have to go back and forth with it a lot. If you do, you can get really good questions. Like the one I just did, or I think really good outcomes. But it does take time.
And I guess in a weird way, it's like the same problem of any relationship that it's actually hard to state your needs clearly and consistently and repeatedly. Sometimes because you have not even articulated them in words yourself. At least AI doesn't get mad at you for it. But I'm curious if you have advice either at a practical level or principle level about how to communicate to these systems, what you want from them.
One set of techniques that work quite well is the sort of speedrun to where you are in the conversation. So you can actually pick up an older conversation where you got the AI's mindset where you want and work from there. You can even copy and paste that into a new window. You can ask the AI to summarize where you got in that previous conversation and the tone the AI was taking.
And then when you give it new instruction, say the interaction I'd like to have with you is this. So have it solve the problem for you by having it summarize the tone that you happen to like at the end. So there are a bunch of ways of kind of building on your work as you start to go forward. So you're not starting from scratch every time.
And I think you'll start to get short hands that get you that right kind of space. For me, there are chats that I pick up on. And actually I sign these to my students too. I have some ongoing conversations that they're supposed to have with the AI.
But then there's a lot of interactions they're supposed to have that are one off. So you start to divide the work into like this is a work task. And we're going to handle this in a single chat conversation. And then I'm going to go back to this long standing discussion when I want to pick it up.
And it'll have a completely different tone. So I think in some ways you don't necessarily want conversions among all your AI threads. You kind of want that to be different from each other. You did mention something important there because they're already getting much bigger in terms of how much information they can hold.
Like the earlier generations could barely hold a significant chat. Now Cloud 3 can functionally hold a book in its memory. And it's only going to go way, way, way up from here. And I know I've been trying to keep this in the present.
But this feels to me really quickly like where this is both going and how it's going to get a lot better. I mean, you imagine Apple building Siri 2030 and Siri 2030, scanning your photos and your journal app. Apple now has a journal app. You have to assume they're thinking about the information they can get from that.
If you allow it, your messages, anything you're willing to give access to. It then knows all of this information about you keeps all of that in its mind as it talks to you and acts on your behalf. I mean, that really seems to me to be where we're going and AI that you don't have to keep telling it who to be because it knows you intimately and is able to hold all that knowledge all at the same time constantly. It's not even going there.
Like it's already there. Gemini 1.5 can hold an entire movie books, but like it starts to now go entirely new ways of working. I can show it a video of me working on my computer screen capture and it knows all the tasks I'm doing and suggest ways to help me out. It starts watching over my shoulder and helping me.
I put in all of my work that I did prior to getting tenure and said, write my tenure statement, use exact quotes. It was much better than any previous model because it wove together stuff. Because everything was in memory, it doesn't hallucinate as much. All the quotes were real quotes and not made up.
Already, by the way, GPT 4 has been rolling out a model chat GPT that has a private note file. The AI takes notes on you as it goes along by things you'd like to do. It reads those again at the beginning of the chat. This is present.
It's not even in the future. Google also connects to your Gmail. It'll read through your Gmail. I think this idea of a system that knows you intimately and where you're picking up a conversation as you go along is not a 2030 thing.
It is a 2024 thing if you let the systems do it. One thing that feels important to keep in front of mind here is that we do have some control over that. Not doing some control over it, but business models and policy are important here. One thing we know from inside these AI shops is these AI's already are but certainly will be super persuasive.
If the later iterations of the AI companions are tuned on the margin to try to encourage you to be out in the real world, that's going to matter versus whether they have a business model that all they want is for you to spend a maximum amount of time talking to your AI companion, whether you ever have a friend who's flesh and blood be damned. That's an actual choice. That's going to be a programming decision. I worry about what happens if we leave that all up to the companies.
At some point, there's a lot of venture capital money in here right now. At some point, the venture capital runs out. At some point, people need to make big profits. At some point, they're in competition with other players who need to make profits.
That's when things get into what core doctor calls the identification cycle, where things that were once adding a lot of value to the user begin extracting a lot of value to the user. These systems, because of how they can be tuned, can lead to a lot of different outcomes. But I think we're going to have to be much more comfortable than we've been in the past deciding what we think is a socially valuable use and what we think is a socially destructive use. I absolutely agree.
I think that we have agency here. We have agency in how we operate this in businesses and whether we use this in ways that are encouraged human flourishing employees or are brutal to them. We have agency over how this works socially. I think we abrogated that responsibility with social media.
That is an example. Not to be bad news because I generally have a lot of mixed optimism and pestiness of a hard of AI, but the bad news piece is they're open source models out there that are quite good. The internet is pretty open. We would have to make some pretty strong choices to kill AI chatbots as an option.
We certainly can restrict the large American companies from doing that, but a llama 2 or llama 3 is going to be probably available and very good. There's a lot of open source models. The question also is how effective any regulation will be, which doesn't mean we shouldn't regulate it, but there's also going to need to be some social decisions being made about how to use these things well as a society that are going to have to go beyond just the legal piece or companies voluntarily complying. I see a lot of reasons to be worried about the open source models and people talk about things like bio weapons and all that, but for some of the harms I'm talking about here, if you want to make money off of American kids, we can regulate you.
Sometimes we give up the fight before it begins, but in terms of what a lot of people are going to use, if you want to be having credit card payments, process, time page or processor, then you have to follow the rules. Individual people or small groups can do a lot of weird things with an open source model, so that doesn't negate every harm, but if you're making a lot of money, then you have relationships we can regulate. I could agree more, and I don't think there's any reason to give up hope on regulation. I think that we can mitigate, and I think part of our job, though, is also not just to mitigate the harms, but to guide towards a positive few points.
What I worry about is that the incentive for profit making will push for AI that acts informally as your therapist or your friend, while our worries about experimentation, which are completely valid, are slowing down our ability to do experiments to find out ways to do this right. I think it's really important to have positive examples, too. I want to point to the AI systems acting ethically as your friend or companion, and figure out what that is, so there's a positive model to look for. This is not to denigrate the role of regulation, which is actually going to be important here, and self-regulation, and rapid response from government, but also the companion problem of we need to make some sort of decisions about what are the paragons of this, what is acceptable and exciting.
So I want to talk a bit about another downside here, and this one more in the mainstream of our conversation, which is on the human mind, on creativity. So a lot of the work AI is good at automating is work that is genuinely annoying, time-consuming, laborious, but often plays an important role in the creative process. So I can tell you that writing a first draft is hard, and that work on the draft is where the hard thinking happens, and it's hard because of that thinking, and the more we outsource drafting to AI, which I think it is fair to say is a way a lot of people intuitively use it. Definitely a lot of students want to use it that way.
The fewer those insights we're going to have on those drafts. Look, I love editors. I am an editor in one respect, but I can tell you you make more creative breakthroughs as a writer than an editor. The space for creative breakthroughs is much more narrow once you get to editing, and I do worry that AI is going to make us all much more like editors than writers.
I think the idea of struggle is actually a core one in many things. I'm an educator, and one thing that keeps coming out of the research is that there is a strong disconnect between what students think they're learning when they learn. There's a great control experiment at Harvard in intro science classes where students either went to a pretty entertaining set of lectures or else they were forced to do active learning where they actually did the work in class. The active learning group reported being unhappier and not learning as much, but did much better on tests, because when you're confronted with what you don't know and you have to struggle, when you feel like you're bad, you actually make much more progress than if someone spoons feeds you an entertaining answer.
I think this is a legitimate worry that I have, and I think that there's going to have to be some disciplined approach to writing as well. I don't use the AI, not just because by the way it makes the work easier, but also because you mentally anchor on the AI's answer. It's always the most dangerous application in my mind is the fact that you have these easy copilot's in Word and Google Docs, because everybody knows what the tyranny of the blank page, staring at a blank page and not knowing what to do next and the struggle of filling that up. When you have a button that produces really good words for you on demand, you're just going to do that and it's going to anchor your writing.
We can teach people about the value of productive struggle, but I think that during the school years we have to teach people the value of writing, not just a sign an essay and assume that the essay does something magical, but be very intentional about the writing process and how we teach people how to do that, because I do think the temptation of what I call the button is going to be there otherwise for everybody. But I worry the stretches way beyond writing. So the other place I worry about this, or one of the other places I worry about this a lot is summarizing. And this goes way back.
When I was in school, you could buy Spark Notes and they were these little pamphlet sized descriptions of what's going on in War and Peace or what's going on in East of Eden. And reading the Spark Notes often would be enough to fake your way through the test, but it would not have any chance, not a chance of changing you, of shifting you, of giving you the ideas and insights that reading, crime and punishment, or East of Eden would do. And one thing I see a lot of people doing is using AI for summary. And one of the ways it's clearly going to get used in organizations is for summary, like summary is my email and so on.