Self-Awareness, The Key to AI

0:00Sear has all the best active and outdoor brands you need from athletic stuff like a full-court pick-up game Swish to athletic ish stuff like a half-mile stroll get those steps in and for morning hikes up the mountain trail Good pace tonight time ghost stories from the camping chair What a twist whatever level of active seara loves it all head to seara or seara.com for the brands You want it the prices that let you do it all from athletic to athletic ish seara's got it This episode is brought to you by peloton break through the busiest time of year with the brand new peloton cross-training tread plus Powered by peloton IQ with real-time guidance and endless ways to move you can personalize your workouts and train with confidence Helping you reach your goals in less time let yourself run lift sculpt push and go explore the new peloton cross-training tread plus at one peloton.com The last show with David Cooper

1:01Recrastinate your life away with us I'm sure you've heard the buzz about AI making workers smarter faster more productive But a new study suggests the real secret might be something a little more human workers who know what they're good at Who know what their strengths are and when it's time to let AI take the wheel? Those are the ones that benefit the most from these new tools I'm here with someone who worked on that study He is an economics professor at NYU his name is Andrew Kaplan Andrew. Welcome to the program Thank you very much I think there's a lot of doom and gloom a lot of fear mongering about AI taking people's jobs and other people being forced to be Overly productive what I like about your study is it frames things like who are these tools good for? What is it about self-awareness at the helm of me using AI that'll make me a better worker with it? Well, you know what we were looking for is to find out what skills This isn't wide open question what skills help people work with an AI and

2:02We've found that a way underrated skill Is your ability to listen to what the AI tells you? So for example when you're looking at a medical image if the AI says it's 95% likely Yeah, to have a particular condition You better be able to say well. I knew anyway the doctors will often say I knew I'll over 100% that's rubbish You know, I know the answer and typically they're wrong Often they miscalibrate it their way overconfident in their beliefs and they can't listen I do think knowing your own limits and being self-aware is probably a good skill to have Independent of AI, isn't it? I don't know it doesn't pay in many branches of that I've been observing You know, I'm not sure it's been working in politics. I don't think it works in medicine You have to appear very confident and it doesn't work actually in the

3:05false art of predicting the future under AI In which everybody gets that to be nostradamus and tells you what's gonna happen Let's start with this basic idea of calibration When someone says they're well calibrated. What does that mean? Why does that make me excellent using AI if I am well calibrated? Well, it means that when you think something is 70% likely it's roughly objectively 70% likely When you're very confident that something is true 95 to 100 percent It'd be true 95 to 100 percent of the time Those are the kind of things that can be checked and the evidence shows that many many people and In many circumstances on 95% sure and 60% right and That's not a very comfortable thing But imagine then you get an AI that tells you something and it it's only

4:05Rationally like 85% confident. I mean you say it's just a fool. I see so much better. I won't listen to it There's a surprising finding in your study that I love because I'm not a very bright individual It's that lower skill workers who have a good measure of their own ability benefit the most from AI Yes, absolutely and that in that particular thing we studied it was like The AI is helpful and it's a little bit better You know than you if you this isn't your great specialty out of the But you better recognize that and then you can become pretty good yourself. So that's the that's the root So it's a logical following that the AI is making the less able the less skilled among us like better as long as they have a good sense of self awareness Self awareness is critical, but no, I would give a big fat caveat to that. Okay The caveat is and the thing that I'm currently most interested in is

5:05whether in fact It's about being good at asking questions and listening to answers So my own take is that the biggest Generalizing out of that study the bigger skills are can you listen but not only can you listen in the word would allow Large language models. Do you know how to ask questions? Do you understand the limits and the skills of the party you're talking with because with a large language model? That's your job. I mean I I use large language model a lot and I am constantly editing it But I'm nevertheless listening. So that's the skill. Can you listen but can you recognize where you're the expert and whether The alarm is the expert and if I was to say anything, you know a big picture. I would say it helps in More complicated activities. It helps those who plan best in simpler activities It helps those with lower skill who understand themselves. It's a bit of rough continuum there

6:09No, I think that's a fair way to frame it Let's talk about high ability people people who are highly intelligent very good at their craft But overestimate their ability nonetheless. That's a very easy group to hate That's a very easy group to find annoying and it as it turns that group It looks like they're not great at using AI because even when AI is right. They're like no, no, I know better Exactly. I know better thing is a huge deal and I mean I I can't think of The biggest professions in the world have I know better written all over them Uh, you know medicine is a fantastic case Where you'll often have doctors saying think what they've asked doctors to give probabilities In difficult cases and they say why should I bother? I know Yeah, that should never be anyone's answer ever Nobody ever knows anything the only thing that you do know is if they said that they're a bit of an idiot Uh

7:09So what exactly was the experiment like how do we come to all these conclusions that we're talking about? It's actually quite hard to find cases that work Uh, and uh, there are many many constraints on doing a good study We found um judging people of age Whether they're over or under 21 to be a good case study Because even the AI finds that super difficult And there are sometimes cues That um a human might recognize better than an AI So it's not obvious who's the winner in fact in The best of the human subjects were better than the AI with the AI was better than the average subject And we offered them both, you know, why don't you try judging this yourself Then why don't you listen to what an AI has to say then why don't you Listen to an AI what an AI has to say and then judge yourself And produce their answer you'd like which is how likely is this person to be under over 21

8:12Beautiful panic in me which is very very stark. It's not exactly, you know, the biggest task in my life But uh, if you're a bouncer, maybe it is, but that's about it. So so so that was the task But it's a precursor for you know a much bigger Uh, set of things that can now be studied particularly with large language models So much more what that came from This idea of calibration how self-aware I am look I was not born the most self-aware person socially and so what I did was I went to therapy I see a therapist once a week I know that's not the same as like professional calibration But I guess my question is is it trainable? Can it be taught? That's the biggest question that is the biggest question. We're trying a little bit of a game Uh, I can tell you we are trying to train In a particular set of medical images We're trying to train people off saying 99% when they don't have a clue uh, and and we're succeeding a tiny bit

9:15But my much the much bigger game Is to what extent is working well with an AI trainable and it's my conjecture That it is quite a great deal more than we know And that that's where we need to put our energy I believe it's doable But I believe you have to be thinking a great deal harder About questions and answers To what that out and you actually have to start training people to ask good questions and to listen to the answers Well, I think folks would be very interested in your research Especially ones that are scared about AI either replacing them or the requirement to become great at AI to keep their jobs It's an interesting area and Andrew Kaplan is a economics professor at NYU Andrew I've enjoyed this chat. Thanks for coming on the show and sharing your research. Thank you so much This episode is brought to you by athletic brewing company No matter how you do game day on the couch in the crowd or manning the snack table athletic brewing fits right in with a full lineup of non-alcoholic beer styles

10:22You can enjoy bold flavors all game long No hangovers no buzz no subbing out for water in the second half Stuck the fridge for tip off with a variety of non-alcoholic craft styles available at your local grocery store or online at athletic brewing.com near beer fit for all times