Growing on Purpose: The Work That Makes You. Jeremy Howard on human flourishing in the time of AI.

We are beginning with a genuine privilege. So Jeremy Howard is one of the really unsung legends. I mean s but I think not as much as he should be in particular because he's Australian. Uh but his work in machine learning and AI go back many years in fact and before that he's the founder of fastmail which [snorts] many of us will use. uh uh he he wrote a paper and developed a technique called ulm fit which is essentially the foundation of modern large language models and the work that he did there sort of helped kick off the revolution that we're all kind of being part of. So we we owe an enormous debt of gratitude to Jeremy and his work there. Uh he is now the founding CEO of Answer.ai AI and the co-founder of fast AI who are looking to make deep learning accessible. So to kick off this morning, would you please welcome and let's express a huge round of gratitude for the work that he's done over many years for the Australian industry and then around machine learning and AI Jeremy Howard. Cheers. Thanks for that. It's uh nice to be back in my hometown. Um it's good to see you all. Um I have a bit of a um unusual talk I wanted to give today. Um the first half of it about psychology rather than AI. um and ends the title which is about uh growing on purpose and the work that makes you um it's such a critical moment in our history right now and the the work that we're all doing is changing and I want to share with you some key findings from the last 50 years of research uh about how your work makes you and so then you can make informed choices about the work that you choose to do. Um, and in particular, I want to draw on uh this this paper. It's not just a paper. It's uh this was a review paper at the end of 30 years of research representing hundreds and hundreds of experiments uh that that led to this huge overarching thing called self-determination theory or SDT. But I I just want to read you the first two paragraphs of this um and and I want you to have a think about it. The fullest representations of humanity show people to be curious, vital, and self-motivated. At their best, they are agentic and inspired, striving to learn, extend themselves, master new skills, and apply their talents responsibly. that most people show considerable effort, agency, and commitment in their lives appears in fact to be more normative than exceptional. In other words, this appears to be how humans are born to be, suggesting some very positive and persistent features of human nature. The very next paragraph continues, "Yet it is also clear," the human spirit can be diminished or crushed and that individuals sometimes reject growth and responsibility. Examples of both children and adults who are apathetic, alienated, and irresponsible are abundant. Such nonoptimal human functioning can be observed not only in our clinics but also among the millions who for hours a day sit passively before their televisions, stare blankly from the back of their classrooms or wait listlessly for the weekend as they go about their jobs. So here we have an interesting bifurcation of the observations about the nature of of human flourishing and the fullest representations of humanity that we observe. Uh there's been um thousands of years of history uh um and uh more recently many decades of psychological research looking at this difference between udemonia and hedonia. So hydonia is where we get the word hydonics or hedonism. There's nothing wrong with it per se. Um it's that frictionless pleasant ease uh pacifisity uh and uh uh kind of easy pleasures. Um uh udemonia on the other hand is what it turns out that these fullest representations of humanity are about. Uh fully actualizing your capacities. that turns out to be what it means to to live well. And as I say, there's there's hundreds of experiments, there's randomized control trials, there's bucketloads of research behind this. This is not just a a crazy idea somebody randomly came up with. So, one of the subpieces of um of SDT, self-determination theory, a key subpiece is around motivation. Now, why is motivation important? Interestingly, when I've read about motivation in books like uh Dan Pink's Drive, which is a great book, and some of this ideas come from his uh the research in that book, tends to talk about like how do you get people to do stuff for you? You know, how do you get people how do you get workers to be productive? But it turns out actually motivation is much much much more important than productivity. It turns out that the research shows people whose motivation is is authentic have more interest, excitement, and confidence. And yes, that does manifest as enhanced performance and persistence and creativity. But it also has enhanced vitality, self-esteem, and general well-being. So motivation is key to this kind of flourishing this udemonia. Um SDT uh has three particular axes and then I've added on one more which is very commonly seen to create these four um which is it comes from autonomy, mastery, relatedness and purpose. Uh relatedness is all about connecting with other human beings uh and feeling supported and part of a group. And purpose is all about what you're doing something for. Is there something worth doing? I'm not going to talk much about those two today. They're very important, but they're rather orthogonal to the points I want to make. So I'm going to focus on autonomy and mastery. Um, interestingly there's another few decades of research from a completely different part of the research community that have looked at the opposite question which is um rather than what helps achieve human flourishing is for those who are very much not the very much not which is those with um clinical depression. How do we pull them out of it? And interestingly the research shows something very similar which is perhaps the most effective action uh even versus um anti-depressants uh codate behavioral therapy um so forth is this thing called behavioral activation which is basically the same thing helping patients to engage with actions which bring a sense of accomplishment. So from both angles, you know, going from kind of yeah, I'm fine, I guess, to I'm thriving, uh, or going from Jesus, life's getting me down to getting by. The same actions, uh, from very different parts of the research community show to be very effective. [snorts] You've probably heard about um, flow and flow fits in here a lot. This is particularly the work of uh, Czechra Mahaley. Um, and I wanted to be careful to define flow here because flow is so key to this this sense um that leads to flourishing. So flow should be a sense that one's skills are adequate to cope with the challenges at hand in a goal- directed rulebound system that provides clear clues as to how well one is performing. So, one of the greatest experiences in my life was getting really good at riding a motorcycle fast around the Philip Island circuit. And that is exactly that very goal-directed rulebound action system. Very clear clues as to how I was performing. Um, and and that sense I'll never forget, you know, of extraordinary flow. Interestingly, however, uh, Chet Mahali also talks about junk flow or dark flow, which is something that can look a lot like flow, but is very bad, which is you can get addicted to a superficial experience that may be flow at the beginning, but after a while becomes something you become addicted to instead [snorts] of something that makes you grow. Um and there's a lot of research and studies around this. And in fact, uh the way um gambling establishments, the way casinos are set up is specifically designed to capture this kind of dark flow to give you what's called an illusion of control um and to to to create this kind of addiction. Um so, uh uh Rachel Thomas, I really encourage you to read this article if you have a chance. um talked about uh breaking the spell of vibe coding in which she noted how certain kinds of interactions, coding interactions with an AI can absolutely harness this kind of dark flow. So you get this kind of positive flow when you have a high level of challenge and a high level of skill. Um so um it's it's it's interesting and a bit scary to note how it's quite possible to end up uh with this kind of uh pulling the slot machine lever version of flow if not careful. So interestingly a lot of people are now saying oh um that's happened to me or that's happened to my friends um um people who have previously been extremely positive about about agents and using uh uh AI and coding and so forth. Um I think uh Ammon was one of the particularly interesting ones. Uh he you might know him as the guy who created Flask uh been a very important software developer. Um, and of course also George Hoz who uh created the comma self-driving AI system and the original iPhone hacker and so forth. Um, I've got some quotes from from Armen here though that I thought was interesting. He said uh when when you know for months he he was in this situation where the dopamine hit from working with these agents is so very real. saying you feel productive, you feel like everything's amazing, and you go deeper and deeper in this belief that it all makes perfect sense, but it's decoupled from any external validation. Um um and so we're kind of starting to see this this these concerns. Um, I saw this like two days ago from um a guy who's been working on this kind of new GPU functional programming system who was saying like how cool it was that he went from naugh to 95% in his me most recent project in five hours and then realized like oh 15 hours later I'm still not there. Um, and I keep finding problems and I actually don't know if there's still problems. And so actually at this point I don't even know where I stand. But getting the first 95% done in five hours sure felt good. But did I actually achieve anything by using AI other than that dopamine anticipation? So I think it's very encouraging that uh thoughtful people in our community are reflecting and sharing their reflections. Um, in fact, one of our own community members um put this on up on our Discord the other day and was asking for feedback from from our community. You're saying talking about the product he works on. It's genuinely interesting. The problem requires deep domain expertise. Um, but it's hard for us to verify because we've got 200,000 lines of of kind of vibe coded um software at this point. and he he's actually realized the pace we've moved at has slowed down um as models get better and token speed spend increases because we generate more and more code with less and less careful engineering uh debugging failures he told us uh you know is super painful uh and interestingly again I guess this kind of dark flow idea most of his colleagues feel they're making great progress um but then when they have quarterly meetings with management they get a reality check when they have to So, what have you shipped? What's the accuracy? How many clients have you signed? And they suddenly realized the results actually weren't good. Um, so, you know, I'm I'm not going to go uh too deep into negativity. We've all seen it. It's in mainstream um newspaper articles nowadays, Wall Street Journal, you know. Um I think just today, you know, Uber is now saying they're putting a strict budget on token use because they're not seeing the ROI. And so rather than uh dive into some kind of like uh AI negativity, I instead actually want to point out something that is you can go in two totally different directions with AI. Uh and I'm looking at two of those key motivation platforms of autonomy and mastery. And it's certainly true that AI can decay those things. So I'm sure anybody who's kind of done uh lot of agentic work and vibe coding has been in that situation where we have what psychologists call an illusion of control. The agents asking you like hey do you want to go with a distributed system here or would you rather use green threads in with a polling loop or whatever whatever and you're like I don't know what any of that means A or B A. Um so this is um this is uh something that actually decays your economy. Uh on the other hand AI can be used to support your growth. It can be teaching you things. You can be trying things. It isn't necessarily creating more outputs more quickly. Um but it's definitely uh something that can happen. So ditto with mastery, right? Um, mastery is not in in uh SDT. It's not about uh creating more outputs, creating more products. It's about creating this uh genuine ability to craft something. Uh it's effortful um and involves uh learning from that effortful work. So with AI you can tackle more complex tasks and you can focus on learning those under demand and underlying foundational principles and master your craft or not right you could focus on uh outsourcing more and more to AI more and more quickly with less and less effortful practice getting less and less learning. So AI is neither good nor bad for you for your psyche. Um, but warning, the people getting you to use AI don't care about your autonomy and mastery. They care about your outputs. And so, they're going to put you in the decay world all the damn time. the people who are selling you the AI models, platforms, harnesses, and your bosses at work who need to be able to show their quarterly token maxing metrics. So, you need to look after yourself in this world. So, I want to show what it looks like to have amazing mastery over a computer and how that's changed over a period of time. So, this is Ivan Southerntherland. I've done this at 2x back in 1963. 1963. And he's shown how he's able to create a direct interface between himself and a computer where he's drawing with a light pen. He's uh you don't see it with one other hand. He's pressing buttons to set constraints as he's drawing. Um he's using it directly against one of these. I think this is one of these fancy vector monitors. And he's showing the the uh interviewer here how he can create an arc for example um using these constraints by drawing directly on the screen and he can adjust it. This is an extraordinary level of deep connection between the human and the computer. Um you might have seen this the mother of all demos. Uh this was uh 1968 that this happened. Um very similar idea. So um in the mother of all demos, Douglas Angelbart show introduced for the first time the mouse piper text, real-time collaborative editing, video conferencing, word processing, screen windowing, and dynamic file linking. Uh this quote was in 1962 towards the start of this project and what the demo was in 1968 and his goal was the same augmenting the human intellect so that the entity to be produced will exhibit more of what could be called intelligence than an unaded human could. We've amplified the intelligence of the human by organizing his intellectual capabilities into higher levels of synergistic structuring. So you see it's very similar between what Southerntherland was doing, what Anglebar was doing and this this was this was their mission was to amplify and augment human intelligence. Um uh one of the most uh underappreciated most extraordinary people in the history of computer science is Kenneth Iverson. I mean not that underappreciated he got the cheering award. Um but uh he designed APL. Uh APL is a new notation or was a new notation for representing um computation and mathematical thinking. Um this is from his Turing Award um uh presentation paper. Um and uh if you don't know APL, it won't look very familiar, but what he's showing here is he's uh proofing some um characteristics of um the inner product. um in his new notation. And one of the really interesting things about this, if you ever get into APL, and I strongly recommend it, is it turns out that this generalizes in a much deeper way than normal mathematical nomature in that the inner product in APL can actually is a is an operator that can combine any two functions. Uh it's not necessarily multiplication and addition. And so suddenly he's proved a whole class of features about a whole class of functions, many of which never been looked at by a mathematician before, just through notation. uh and this can go a really long way. Some of you might have seen this very famous single line of code in APL which is a complete implementation of Conway's gain of life. So here in this video um uh that uh life function is being applied over these two characters and off it goes. Um again it's the same thing right Iver was passionate about creating this connection between between the human and and supporting the human's thinking notation as a tool of thought um perhaps most mind-blowingly um Brett Victor who spent a couple of years as he described being a hermit living on a train and he came out the end of those two years of hermit having built the most extraordinary and inspiring array of real world demos showing about how to understand climate, how to understand electricity, how to understand um how to how to build uh games, how to build graphics, how to understand waveforms. Um and he shared this all with the world. This is his coding environment whereas he changes it graphically. This is this amazing game playing demo where he actually created a time machine for his code. Um, if you haven't seen this, please watch everything Brett Victor has done because it's incredibly inspiring. And all of it, you'll see it's all of it's the same thing. It's creating this connection between the the human and the computer that they're working with so that they can craft. This is this is all effortful craft that he is supporting. Um Chris Latner uh this is his playground system. He's created a whole amazing hierarchy um from from LLVM Plang um Swift Playground MLR Mojo you know at every level trying to improve this ability for humans to to connect to and work with their computers. Um my argument is that actually we're still on this chain. uh we we we can continue working in in along this history from the mother or demos in a really deep and powerful way. Uh uh AI is a marvelous way to connect more deeply with our computers um and achieve this fullest representation of humanity. So this has actually been kind of my mission for the last 30 years and very dramatically for the last 10 and at answer AI it's all of my focus [snorts] this idea that we should be seeking to augment human creativity not to replace it. It's interesting to see hopefully this resonates with you, but also hopefully you see this is almost never what you actually see is being marketed when somebody's trying to sell you a piece of AI. It's like it's it's going to summarize this for you. It's going to write this for you. It's going to do this for you. It's all about being done for you. So, um, I want to quickly demo something we've built, um, to give you a sense of what it looks like to have a tool that that is specifically designed for this augmenting human creativity uh, and understanding indeed. And so, I'm going to do I'm going to just show you a couple examples of actual dialogues I've gone through with the help of AI in the last three days. Um so one was I wanted to learn about recursive language models. So probably a lot of you have um know about recursive language models. They've been kind of taking over the world. So I uh we've got this system called solve it. Um and um so I should show you uh solve its solved.com um and I loaded the paper recursive language models paper into solve it and you can just read it in the normal way um piece at a time um but make sure you understand it. And so here um uh I looked at figure one. It's like, well, I don't know what this is or this is or this is. And so, normally I might just skip over it. But here, I can just say, hey, what's this figure? And it's it tells me like, okay, these are the three evals that the RLM authors are doing here. Um, and and interestingly, it's like why it's going from a constant to a linear to a quadratic complexity, which is not actually captured in the original paper. So, that's really helpful information. Um, now I don't work very well at an abstract level. I need things to be concrete. So, I could ask for an example of each task. And this is much easier than going into the papers and trying to dig them out and so forth. And so, here I'm getting little examples. It's like, okay, I get it. Right? So, I always tell people, don't move on when you're learning a new thing or working on something until you you get it. So, I was like, okay, I get it. Um, so then there's another figure um where they describe it. And I didn't fully understand this figure to be honest at first. Um, and so I just it it tells me what the the key pieces are, which is very handy. [snorts] Um, so I I when I'm reading this, I'm thinking like, okay, I want to push beyond this. I want to understand it, but try and go a bit past it. So I'm thinking like, okay, I wonder if we can replicate uh all of the features of an RLM right now. So I kind of had a hypothesis about how to do that. And I kind of check with the AI like I think we can try this ourselves and it clarified some key points about um where sub agents fit and uh as it turns out solvent has a sub agent. So I said yeah you've got a sub aent why don't you try it? So in this case it spawns an agent um and tells it to use Python to solve the complex square root. And we can also write code here. Right. So I can then try it myself and I can compare. I'm like, "Oh, cool. Okay, so I've kind of confirmed I'm in an environment where I can actually do the same things that the RLM paper did." Um, so another thing I mentioned as I read papers is often it's very easy to skip over citations, right? But here it says, "Well, here's recent recent work." Well, okay, I don't know any of these, so I shouldn't keep moving. So I said, like, stop. Can you please go and read all those papers for me and tell me basically what they are and why they're here? I can decide whether to click on those links, read them. Um, and I want to kind of test my understanding at this point. So anyway, to skip ahead a little bit, um, I'm thinking like, okay, let's, uh, let's try it. So they've got their table of results here for four different tasks. Uh so I'm kind of thinking like I'm not sure that this thing called an RLM even exists. It just feels like a normal tool loop that happens to have particular tools in it. And I seem to have the same tools right now. So I think we can be an RLM, can't we? Um so I said, let's try it. Let's try this code QA thing. I've never heard of this before. So it tells me where to find the data. Um, and so I start so it writes some code for me. Um, which actually didn't work. Um, but then we can it can help me debug it. Um, and then I can test it and I then I could experiment and look at this data carefully, make sure I understand what this eval is and how it works. And then I just tell it like, okay, solve it. Go, you know, go ahead and just solve this right now. And so this is one of the things in one of the emails and it goes ahead and tries to do it and it says I think the answer is B and the check. Oh, it is B. No, tell me how you did that. And then okay, let's try another email. Um and answer is C. Yeah, answer is C. And so this is very interesting. So I did this for a few different um tasks including the hardest um quadratic tasks. Um so I downloaded another of these data sets went through them and um yeah discovered that solve it correctly solved every one of the tasks um and so I've now not only do I understand RLM I've reimplemented it um this is all in the space of a couple of hours uh and in fact discovered that this is a much more powerful platform than even RLM is um another example was I was looking at uh Julia Evans who's a fantastic uh writer always really interesting and she uh I'm very I'm not a fan of Tailwind and I was keen to see how she had done this year this article called moving away from Tailwind um which had a particular structure u this was the structure she had and I decided okay I want to go through her blog post so I loaded the blog post as you can see into solvent and uh started reading it and again so the red is me asking questions as I go through the article. Um, and so she said she used something called Tailwind pre pre-flight as her starting point for her styles. And I was like, well, are there other options? Um, so I went and grabbed it. And actually, one of the cool things about Solit is it's because it's in a browser, I've got actual styles, HTML here. So I'm actually modifying my environment as I go. Um, and so I can see uh see this happening and then I can try it out. So I'm trying out layers. I'd never really I'd never done anything with layers before. So make sure I understand how they work. But again, I'm actually using them. So before I actually used RLM, you know, rebuilt RLM inside dialogue. Here I am rebuilding Julia's styles inside a dialogue. She had a section about components and again same thing. I started creating the components myself as you can see with the help of AI to make sure I understand. So here I've got a badge component for example. Uh and you know this is all real right. I'm creating actual code. I'm seeing it actually um running um button components. Um then very interested in colors. I kind of had some ideas about how to create a new color framework. So I kind of bit of a discussion as you see with the AI but again the AI um didn't write it for me right I kind of came up with this idea of what I think I wanted to do and then I um tried it and you can see I've created my own new color palette that I'm very happy with and I thought oh cool I'll try and map those to kind of semantics now um so like okay which should danger be so like it helps me show me what it looks like different colors on different backgrounds s um inverted ones. Um it helps me create these uh full swatches so I can see how my color palette looks. Um did a similar thing with font sizes. I had an idea of how I wanted to create nice topography. Um so again, kind of write the code, talk through it, and have a look to see and I'm like, "Oh yeah, they all look pretty good. Happy with that." I kind of prefer these tighter line heights than normal. So, I'm really experimenting with different graphical styles to what's fashionable nowadays. Um, so you know, you'll get the idea, right? Um, so you won't be surprised to hear this whole talk was written in solv created none of the narrative, none of the slides, right? I asked it for examples of research. I then read the papers and I was getting pretty tired last night. So I was I was then kind of pasting in um my um slides and saying like here's where I'm up to and it's helps me keep track of like okay this is where you're at on your narrative. As I read through the um SDT paper same thing right I put it in here and I was checking my understanding of it by asking as I went. So, uh, I'm going to wrap it up there, but I I just whether you use this particular tool or some other tool, um, it's not so important, but I just wanted to give you some examples of of like how I work with AI. It doesn't do my work for me. Um, and at the end of every day, I feel honestly energized, excited. I'm learning more things all the time. And, um, it seems to be working. Um, we're building stuff that's no one ever built before. Uh, I've got a um so strictly speaking solvable at the moment. It's been beta tested for the last two years by 3,000 people. But for this conference um we've given you a special um code uh that you can use um before it's officially released. And I've also popped there links to each of the four dialogues. And it's quite cool. If you're on solve it, you can click on any one and it'll open it and solve it. And that's how you should read them. So if you want to read any of these dialogues about um SDT or about creating a um graphics framework, a styling framework or about um recurs recursive language models, don't just read it, right? Open it, ask questions um and engage and then write some code. Um and so yeah, my hope is that um by thinking of AI this way, by focusing on um creating thinking about yourself as hopefully being the fullest representation of humanity, um this is a period of time where you will feel like you and the people around you are truly flourishing. And that's what I hope for all of you. Thank you. [applause]