Collective Learning With Generative AI
Me, Myself, and AI · 17 Dec 2024
Source: https://sloanreview.mit.edu/audio/bonus-episode-collective-learning-with-generative-ai
Transcript: https://podcasts.opennash.com/shows/me-myself-and-ai/collective-learning-with-generative-ai-f970b7/ (via OpenNash Podcasts, machine-generated)

AI is quite literally everywhere. Sam and I come on here every other week to talk with smart people about how they use it at work and in their personal lives. In the last year, we've heard a lot of predictions and watch outs, and we'd like to spend some time reflecting on these. What's changed, what stayed the same, what concerns are still top of mind? We were fortunate to speak with Microsoft's Eric Boyd right as Chat GPT 3.5 hit the scene, and everyone suddenly became aware of the potential of LLMs.

It's one of the things I push my team on a bit is I want us to think about scenarios and products where the portion of people using AI-powered features is 100%, right? You can think of something like, yeah, in this team's call, we have transcription, and so we could turn transcriptions on, but not everybody uses that. You could go your whole day never using transcription. What are the products that you absolutely can't avoid AI because it's just intrinsic to the product?

Search is, of course, like that. You can't avoid using AI in search. When you are talking to your phone to compose a text message with speech, you're using AI 100% of the time you do that. And so increasingly, as we see these scenarios, there's so many things that are just not possible, right? If I'm gonna now start with my three bullet points and have that expanded into a paragraph for me, like, well, you can't do that without AI.

So every time you're using that functionality, you're using AI. It's pretty ubiquitous, but that's something that there're really a lot more scenarios coming on that. I mean, there's this whole field of AI-powered applications that's really about to start to blossom, where the application just doesn't exist without the AI that powers it. It's really kind of crazy just how quickly this field is moving. The speech quality that we deliver through our speech API literally improves every month.

We measure it, we've got data to back that up. Our vision models have just exploded in quality recently, and we've seen lots of crazy things. And then let's not even get started on language, right? The large language models are just incredibly powerful these days, and so just a massive explosion going on there. Someone was making a joke the other day that they said, we're not too far from the days where a CEO is gonna have four bullet points and ask an AI to create a two-page memo for his staff, and the staff is gonna use an AI to reduce this two-page memo to four bullet points and then go and read them.

I thought that was very funny, but that type of process, though, right? I've used AI already to say, you know, I've got some tough email I need to write. I wanna make sure I'm getting the tone right. Here's how I wrote it. Can you make this more polite? Or can you give me a suggestion on how I should change? You know, just being able to get valuable input and feedback on that, it's kind of amazing. That's really empowering and so very central then to the work that you're gonna try and do as a result of that.

We have a lot of scenarios. We've got AI embedded into dynamics, which is now Microsoft dynamics. It's a CRM and advertising tool. And one of the things we're using it for is to help people create advertising copy. They can create advertising copy that gets generated for them. A similar use case we worked with Carmax and Carmax has every single car on the planet and they wanna have a unique page describing every single car on the planet.

And so whenever I talk about Carmax, I say, well, my first car was a 1986 Ford Tempo. The 1986 Ford Tempo is a piece of garbage. It was a piece of garbage in 1986 and it's absolutely one now. But that was my first car. So CarMax wants to have a page describing the 1986 Ford Tempo, and they have user reviews about it, but they want a page that describes it.

That'll do really well for search engine optimization. And so they use GPT-3 to go and summarize reviews that they've got on each make, each model, each year, and then generate a page for it. And it would have taken them years, literally years. They did the math on how you could sort of go and do that. But now they have this really high quality, valuable content that's directing people to their site as a result of that.

And what's exciting about that is the scale part that you've alluded to, that this is individualized and personalized, but it's scale at the same time. And I think there's where we see a lot of the promise.

Exactly right. I think that's very exciting that you can do this and you can run it for a couple of hours and get lots and lots and lots of work done. And again, sort of with that theme of this is AI that's helping people do things better. As an editor, you can review these and see how they've come and really be able to go so much faster and so much more productive than you would in a different manner.

We see examples all over the place of how this AI is really helping people do things that they couldn't before. We work with a lot of companies and they come just in all shapes and sizes, right? And and you really have to take them with the problems they have today and give them solutions that they can use today. I think there's already just such a democratization of technology that's going on. When you think about how powerful your cell phone is versus who was able to sort of get access to that type of computing power three decades ago.

When you think about the access to the information on the internet and online versus three decades ago, we see that transformation happening just everywhere where it's you're putting power in the hands of so many more people. And so AI is both going to be a part of that and an accelerant of that, as I think about it. When you think about the ability to create an application that right now requires knowing computer science and how to write code and knowing a programming language and having access to the tools to go and do that.

And then you look at something like GitHub's copilot, which is just scratching the surface of how powerful it can be to describe a concept and have an AI literally translate that into code for you. We're gonna have so many more people who, because of AI, they're now able to create applications. They're able to get work done that they previously couldn't imagine getting done that they might have needed to go hire someone to build something for them.

I think we're gonna see a lot of that democratization continue to happen. Even with ChatGPT, I think we're starting to see some of the democratization of let's expose to the world. Hey, this is the type of thing that's possible. It may be the ultimate homework cheater, and so we'll have to deal with all the essays now need to be filtered against, though ChatGPT didn't write this. But just getting people exposed to the ideas of what you can do and then thinking about how that's going to turn into the next companies, the next ways that are going to go and empower more and more people.

That, to me, is where that democratization is really going to take off.

Pretty fascinating to think that Eric was kind of talking to us right at the moment that this seemed to change the world. Artificial intelligence became the kind of thing that a few people talked about to the kind of thing that it seemed like everybody was talking about. Shervin and I, as part of our research with Boston Consulting Group and MIT Sloan Management Review, have just released a report that finds that the use of artificial intelligence has jumped 20% since the last time we talked.

It was 50% before, to 70% now, and it's hard not to think that generative AI is a big part of that. But this pervasive use comes with its own set of concerns, perhaps. Last season, in May of this year, we spoke with Mario Rodriguez from GitHub. He explained Copilot, another generative AI tool that's broadly being used now, and how it got started. We're seeing that these tools can benefit not just individual uses, but they're getting embedded in products, and those products are showing up everywhere.

Let's hear what Mario had to say.

Product management, maybe this is a little bit controversial out there, but it's a very new discipline, in my opinion. It's not something super big, like we know how to do agriculture as an example. Developers have been doing stuff for a very long time too, and in one way or another, but product management as a discipline is, I think, is probably new. I think of product's role overall is to take something like an outcome that you want to achieve and then figure out how to get the product to be able to accomplish that outcome.

And it's magical to do that. People think it's a set of features and you you then get into this thing called feature factories, if you have heard that term, and I think that's that's not right. Like, the product discipline itself is more art than science. Many times, you're not there just to run stand ups, or you're not there just to talk to customers either. There's this beauty of the product discipline that is all about taking what does a customer want to accomplish, or what does the business want to accomplish.

But for them are super valid, right? And then figuring out how to get there. At least from my end, I love to do it incrementally. It's very hard to be right all the time. And for me, what I always tell the product team is how can we actually figure out what is the next incremental thing that we could do to validate a set of assumptions to try to get us there? And that incrementality doesn't have to show up to the customer right away.

We do this thing at GitHub called staffship, so we ship it first to ourselves. But that is a point where we could validate the idea, and that is a point where we could then figure out that this work or this didn't, and what's the next ref of the idea to accomplish that outcome. I usually said a product team is really there to take these ideas, marry with outcomes, and they fill the middle of that equation. GitHub Copilot is our AI assistant, and we're kind of a little bit proud about this.

We were the first copilot out there, and if you were internally within GitHub at the time that we started this, it was a big memo at one moment done by a set of researchers at GitHub. We have this division called GitHub Next, and they oversee Horizon two, Horizon three set of investments for GitHub. And we had just started to play with some of the models, some of these new LLM models coming from OpenAI. And the paper was about that is, can we create this AI coding assistant that can really help you and supercharge you overall?

So GitHub copilot is the brand name for that, and we've been, you know, very successful with again, going back into that iteration, iterating towards amazing value to developers. And, you know, we have stats like 55% improvement in productivity, and this one I'm really proud of. We try to ask a lot of our customers to really measure developer happiness overall. We know that if developers are unhappy in an organization, not a lot of things are happening, right?

You wanna keep them happy. Talent out there is scarce, so you definitely wanna keep your developers happy. We've been very successful in both the adoption side and the value side, but even more importantly, at least for me at times, is that developer happiness and increasing that across the world as well. Mario talks about using GitHub tools internally with staff ships to test, which also motivates and inspires employees when they see their products working. We're seeing AI tools democratized and being used by so many people at work and for personal use.

Another theme we've seen is around sharing the knowledge we're gaining and the mistakes we're making. Last episode, I referenced a YouTube channel I found where employees literally show the mistakes they made in front of customers and how they fixed them to get happier customers in the end. How is the tech community doing this, especially in the wake of this insanely fast-moving world of gen AI? Our recent discussion with Rebecca Finlay of the Partnership on AI talks about openness and sharing.

We know that in order to be innovative, in order to be opening up new markets, in order to be thinking about new beneficial outcomes, you need to be thinking about how you are doing this safely and responsibly as well. The whole opportunity for generative AI to become much better, because today it still has real challenges, whether it's hallucinations or other ways in which it is deployed, just hasn't really gotten to where it needs to be. But if we're starting to think about once it's there, how it can be deployed to really deal with some of the biggest global challenges of our time.

That's why I'm at the Partnership on AIs, because I believe that AI does have that transformative potential to really support important breakthroughs, whether it's in healthcare or really the big questions in front of us around our environment and about sustainability. We're already seeing this in the predictive AI world, right, where we're starting to see it just becoming integrated into the scientific process across all sorts of disciplines. I do think getting back to this question of, you know, the trade off between responsibility and innovation, that one of the things that I hear from companies right now is they feel alone as they're trying to disentangle the risks of deploying these technologies and the benefits to their productivity and the innovation and

how they serve their customers as well. And so one of the reasons why I think the work that we do at PAI is important is I want to say there is a community of organizations that are wrestling with exactly these same questions that are trying in real time to figure out what does it mean to deploy this responsibly in your workforce. What does it mean to think about safety of these systems and how they're operating, whether that's auditing, oversight or disclosure or otherwise.

How do you experiment and what is best practice? And so I think more and more if we can let companies and organizations know that there's a communities who are actively working on these questions. Where you can get some insights and really in real time develop what will become best practice, that that's a good thing for them to know.

The partnership on AI's focus on community and knowledge sharing aspects seem really big. There's a wave of trend here that's going on that shifted from AI projects being super secret to maintain competitive advantage, and then maybe a shift towards openness, but that's not entirely true because some organizations are open and then closing again. We're looking at open AI when we say that. Can we find a balance between competitiveness and thinking about how we all can learn together?

There are many, many benefits to openness, for example, open algorithms. Many, many models are available online, free for everyone to use. MapReduce, the Hadoop distributed file system, these algorithms came out of companies who shared them. And data is important too. We've seen that open data accessed enabled computer vision progress with ImageNet and the surge of deep learning that happened in 2012. Shervin, I had an episode with Nightingale, who are working to improve open data access in the healthcare context.

How can we foster all this? We talked with Joelle Pineau at Meta about how they're using the algorithms and developing algorithms, but then also sharing them with other people to benefit from.

There's a strong culture towards open protocols at Meta that predates the AI team. The basic stack, the basic software stack is also based on many open protocols. And so that culture is there to this day. That culture continues. It goes all the way to the top of the leadership, and that commitment to open sourcing the models is strongly supported by Mark Zuckerberg and his leadership team. So I don't see that this is gonna stop very soon.

What is going to be important is that we continue to release models in a way that is safe. And that's a broader conversation than just one company. The governments have several points of views of how should we think about mitigating risks for this model. There's also a lot of discussions about how to deal in particular with very frontier models, the largest, most capable models. And so we're going to have to have these conversations as a society beyond just the labs themselves.

I think one of the challenges is determining what we want out of these models, right? We've seen some pretty egregious examples recently of groups, which from I assume is well-meaning intent to rebalance data sets, especially with representation of, for example, different racial groups in images. You know, of course, if someone asks for like an image of an engineer, you don't want only men to show up. You would, uh, hope to have a few women show up.

And there's ways to rebalance the data. There's ways to sort of re compensate at the algorithmic level, but sometimes you end up with very unusual results. And so it's also a question of what are the distribution of results that we expect and that we we tolerate as a society. And and in some cases, that's not very well defined, especially when the representation is biased within the real world as well. In many ways, deciding as a society what we want these models to optimize for and how we want to use them is a very complicated question.

That's also the reason why at Meta we often open source the research models. We don't necessarily open source the models that are running into production. That would open us up, I think, to undue attacks, and it's something we have to be careful about. But we often open our research models, and so that means very early on, if there are major opportunity to improve them, we learn much faster. And so that gives us a way to essentially make sure that by the time a model makes it into product, it's actually much better than the very first version, and we will release multiple versions as the as the research evolves, as we've seen, for example, with the Llama language models.

You know, we release Llama one, Llama two, Llama three, and so on. And every generation gets significantly better. Some of that is, of course, the work of our own fabulous research teams, but some of that is also the contributions from the broader community. And these contributions come in different forms. You know, there's people who have better ways of mitigating, for example, safety risks. There are people who bring new data set that are allowing us to evaluate new capabilities.

And there's actually some very nice optimization tricks that allow us to train the models faster. And so all of that sort of converges to help make the models better over time.

To close out here, it's hard not to spotlight a really cool example. Last December, we spoke with Vandi Verma, roboticist at NASA. She shares a really specific example of how NASA uses AI to simulate conditions on Mars. But I think a key point here is the rover has some parameters on what to do and what to navigate, but also a lot of degrees of freedom to explore and to find out what's going on. Vandi talks about, you know, we've never been to Mars, so we actually don't know what we're looking for.

So they allow the rover quite a lot of degrees of freedom to explore and to stop and take pictures of whatever it finds interesting, no one having ever defined for it what interesting is.

I am the deputy manager for the mobility and robotics system at NASA's Jet Propulsion Laboratory, and I'm also the chief engineer for the Mars, uh, 2020 mission, which consists of the Perseverance Rover and the Ingenuity Helicopter. JPL is a NASA center that specializes in building robots for space exploration, and NASA's mission is to explore, discover, and expand knowledge for the benefit of humanity, and what we do is the robotics aspect of that. AI has sort of transformed what we call that over a period of time, and there are things that we do on the ground and there are things that we do on board our robots, and so I'm gonna touch on some of those.

So, in general, we are more on the side of autonomous capability, closer to what you might think of as what self driving cars use, and not a lot of it is potentially classically machine learning per se, although we use that to inform a lot of our work. In fact, with Perseverance, 88% of the driving that we've done is autonomous driving, and so the rover has cameras, it's taking the images, it's detecting the terrain and figuring out what's hazardous and navigating around obstacles.

And it's actually quite interesting because it's driving on terrain that no human has ever seen, so we can't even give it that kind of information, right? So that is definitely a form of autonomous navigation. We also, at the end of drives, are trying to make a lot of progress because we're in this really harsh environment and we have a mission to collect and cache a certain number of samples with Perseverance. Because for the first time, we are actually going to bring them back to Earth.

But we want them to be from as distinct places as possible, so we want to do a lot of driving. So if you stop all the time, you're not gonna make as much progress. But who knows if there's something really exciting along the way that we're just gonna miss? In our world, we call it the dinosaur bone. So we have AI capabilities on the rover where it'll take a wide angle image, look at a large swath of terrain, and then try to figure out what is the most interesting feature in there.

We have a whole slew of instruments, but one of the instruments is the Super Cam instrument, which does a lot. It has a laser, and from a distance, you can shoot a laser at a rock and it creates a plasma, and we study that with a telescopic lens. That is such a narrow field of view, you know, a milliradians. And so if you were to try and do that to the whole view you see, you'd spend days there.

So essentially, we'd use the AI to figure out what's the most interesting thing that we should zap, and then you can, you know, send the data back and tell the scientists on Earth. That's been very valuable as well. I think one of the most interesting thing about defining what's interesting is it puts it on the humans. We actually have a really hard time telling machines what we want them to do, right? In order for us to tell what's interesting, we have a lot of different parameters that scientists can use to specify.

I am looking for light toned rocks of a particular size, of a particular albedo and shape that are interesting in this area, and we can change that. So we have these different templates, depending on the terrain we are in, that scientists on the ground help us determine. We send that to the robot to say we're looking for this kind of thing. We have done some research as well, where we tell it you now track all of the things we have seen, and it's called novelty detection, which we don't actually yet have deployed, but find what we haven't already looked at.

That's another one, you know? But there are two things in here. When we are doing exploration, we are both looking for things that are new, but we also try to characterize things we have seen with multiple different instruments, cause we are trying to collect a statistically significant amount of data for the hypothesis we have. We're trying to figure out could life have existed on Mars, and, uh, especially ancient life, and so that puzzle. There are hypotheses, and you're trying to answer specific questions, and that's what the scientists then will tell the robot what they're interested in.

We've actually used supercomputers to translate that into parameters that we can then uplink to the robot. This is, I think, an area where AI can help a lot, right? Because we are still in that phase in robotics in a lot of areas where we have a lot of knobs, we can do a lot, but the art is in tuning this multi variant space. In fact, you know, just on Perseverance, we call them parameters in software. This isn't even taking into account hardware design and other things.

We have over 64,000 explicit parameters. These are saved in non volatile memory. This is not even taking into account the arguments to commands you can send. So there's just so many ways in which you can express what you have to say, and that's where we can use a lot of capability to know what the right combination is for what we intended to do.

I really like the NASA episode. It was one of my favorites, but that's like asking which of your children is your favorite. If my children are listening, both of my children are my favorite. NASA is still exploring, but we're all exploring. Exactly zero people in the world have ten years experience with generative AI, no matter what the job postings ask for. We're all learning as we go. We're gonna increasingly use generative AI and AI tools in places that are not obvious and are not clear at first.

That's okay. We all are learning together about these things. There's a lot of uncertainty about the things we're learning. Shervin, I've just got a report that came out about this exact relationship between organizational learning and uncertainty. We'll link to that in the show notes, as well as to each of these full episodes. We've had inspiring conversations with leaders in supply chain, marketing, software development. But if there are other topics that you'd like us to do a deep dive on, please comment in your reviews or send us an email.

Again, we're all collectively learning how to use these tools. They never existed before, and we'll have more guests coming along that we can learn from. Thanks for listening. We're back soon with a conversation with Ronald D. Nelson, Chief Digital and Technology Officer at the Heineken Group. Please join us.

Thanks for listening to me, myself and AI. Our show is able to continue in large part due to listener support. Your streams and downloads make a big difference. If you have a moment, please consider leaving us an Apple Podcast review or a rating on Spotify and share our show with others you think might find it interesting and helpful.