Adept CEO David Luan and Stanford's Percy Liang | Words into Action

Hi everyone, welcome to Gray Matter, the podcast from Greylock where we share stories from company builders and business leaders. I'm Heather Mack, head of editorial at Greylock. Today we're rebroadcasting our episode featuring Greylock General Partner S.'s conversation with David Luan and Percy Liang. David is the co-founder and CEO of AI startup Adept, and Percy is a computer science and statistics professor at Stanford. While text and image-generating AI tools like ChatGPT and DALL-E are all the rage right now, Adept is developing tools that take things a step further by actually executing actions based on text commands. The company just raised 350 million in Series B funding to further its development of a tool that can be thought of as an AI teammate trained to use every software tool and API for every knowledge worker. Greylock contributed to the latest funding round, and the firm has been partnering with Adept since co-leading the company's Series A in 2022. In this interview, David, Percy, and S. discuss how advancements in large language models are paving the way for the next wave of AI. This interview took place during Greylock's Intelligent Future event in August 2022, the summit featured experts and entrepreneurs from some of today's leading artificial intelligence organizations. You can read a transcript of this interview on our website, and you can also watch the video on our YouTube channel. Both are linked in the show notes. And if you aren't already a subscriber to Gray Matter, you can sign up wherever you get your podcasts.

Okay, David, Percy, I'm excited about this. There's no doubt that large-scale models are topical for all of us here, and I'm really excited to have the two of you to discuss them with. For those of you in the audience who aren't familiar with these two gentlemen, Percy is the associate professor of computer science and statistics at Stanford, where among other things he's the director for the Center for Research on Foundation Models, and David is one of the co-founders and CEO of Adept, an ML research and product lab building general intelligence by enabling humans and computers to work together. And before Adept, David was at Google leading a lot of large models effort, and before that at OpenAI. And we're fortunate to get to partner with David and the team at Adept here at Greylock. Percy, David, thank you guys for being here and for doing this. So I want to start high level and just start with the state of the play. There's a lot of talk about large models and it's easy to forget that a lot of the recent breakthroughs and models that we're all familiar with, like DALL-E and GPT-3, are actually fairly recent. And so we're still in the early innings of these models running in production and delivering real concrete customer and user value. Maybe just give us the state of the play, David, starting with you. Like, where are we with large-scale models and what's the state of deployment of these models today?

Yeah, I kind of want to highlight, I agree with everything that David said. I want to emphasize one distinction he made, which is already with all the applications out there, these foundation models can just lift all boats and just make all the numbers kind of go up. I think another thing which is even more exciting is that there's a huge sea of applications that we're not even maybe dreaming of because we're kind of stuck in this paradigm where what is ML? Well, you could get some data, you train on it, but with prompting and all these other zero-shot capabilities, I think you're going to see a lot more new types of applications. So I think we should be looking not just for how to make faster horses or faster cars, but new types of applications.

First, maybe to follow up on that, I totally agree, and I think it connects to David's point around something like Copilot. And the thing that's amazing to me about something like Copilot is both how new of an experience it is and how quickly it's taken off and gone into end-user adoption. What are some of the other areas that you're looking forward to and are excited about in terms of net new applications that become possible because of these large models?

Yeah, so I mean, maybe one general category you can think about is creation. So this includes code, text, proteins, videos, PowerPoint slides, anything that you can imagine humans kind of doing right now which could be a creative or sort of a more task-oriented activity. You could imagine these systems being very helpful, with you in the loop, taking you much farther and giving you many more ideas. So I think the space is quite broad, and underscoring the multimodal aspect of this which David touched on is really important. We shouldn't just think about language models and code models and image models. Think about things that you could do when you mix these together, creating different illustrated books or films or things like that. One thing that you have to deal with is long context dependence. Relatively, right now you're generating single images or text up to maybe 2,000 or 8,000 tokens depending on your model, but imagine generating full films. That's going to require pushing the technology further, but we have the data, and if we can harness that and scale up, then I think there's a lot of possibilities out there.

David, what would you add? I mean, at Adept, you guys spend a lot of time thinking about how to use these models to unlock new ways of collaborating with computers and software. I'm curious what some of the use cases you think about are.

So maybe if we take something like actuation as a key net new capability or we take longer contexts as an important net new capability, there's the form of the question is where do we still need to see key research blocks and where are the key areas of focus to actually make these products a reality?

I think there are maybe two sides of things. One is pushing up capabilities and one is making sure things are pushed up in a way that's robust and reliable and safe. The first one is in terms of scaling. If you think about video and the ability to scale to hundreds of thousands of sequence lengths, you're going to have to do something different. The Transformer architecture has gotten us surprisingly far, but you need to do something different there. And I think David mentioned this briefly, but I think these models are still in some ways chatterbots. They give you the illusion that there's something good going on, and in certain applications this is okay if there's another external validity check on things with the human in the loop. But I think there's a deep fundamental research question on how to make these models actually reliable. There's many strategies that people have tried, using reinforcement learning or more explanation-based or retrieval-augmented methods, but I feel like there's still something deeper missing. I think this is one thing I hope the academic community and researchers will work on to ensure that these foundation models have good and stable foundations as opposed to shaky ones.

So let's get to safety in a moment. Continuing on data, because I think that is a really important topic here. David, at Adept, you all are thinking about how to build products that humans collaborate with, and I think one of the nice consequences of that becomes this data flywheel. Can you maybe add a little bit about how you're thinking about that and how you're approaching designing products that end users will work with?

Just add to that. I think the interaction piece is really interesting here because these models are in some ways the most interactive ML models we have. You have a playground, you type in a prompt, and you immediately get to play with a model, as opposed to the previous cycle where someone gathers some data, trains a model, and then you experience it from the user side. So the line between developer and user is actually getting blurred, which I think is a good thing because if you can connect these two up, then you get better experiences.

Is there anything interesting from the Stanford both the HCI perspective and the foundation models perspective on the research side that you all are working on around interaction?

Yeah, so one thing that we've been doing at Stanford is, as a part of a larger benchmarking effort, trying to understand the ways in which these models—what it means for humans to interact with these models. The classic way that people think about these models is you train them and then there's 100 benchmarks and you evaluate, taking the automation approach. But as we know, a lot of the potential here is in copilot or autocomplete experiences where there is a human in the loop. Humans, and I think Adept is also a good example of this. What does that mean? Should we be building our models differently if we know that humans are going to be in the picture, as opposed to doing full automation? That's interesting because maybe in some cases you want a model not just to be accurate, but you want it to be more interpretable or more reliable or understandable. And for creative applications, you may want a model to actually have a broader distribution of outputs. We're seeing some of this where what is good for actual interaction is not necessarily what's good just for the standard benchmarks. So that's really interesting.

How is that going to get resolved? In classical machine learning applications, there's some point of view on benchmarks and standards. There are different products out there that can actually measure these things around bias and auditing. As we massively blow up the scope around creativity, all of that kind of shifts. So how do you think this is going to resolve?

Yeah, so first order, scale definitely is helping, so we're safe on that. If you scale up the models, it lifts all boats. And then given a particular scale, you have a question of where you're investing your resources. I think what we want to do is develop effective surrogate metrics which you can actually evaluate that correlate well with human interaction. We don't really have a good handle on this quite yet. But having humans in the loop for an inner loop is also potentially problematic and hard and not reproducible. So you want something that's easy to evaluate but at the same time actually tracks what you care about.

I want to shift to building products and companies around large-scale models. David, maybe I'll start with you. There are people in the audience who are in the early stages of building these companies, and one fundamental question is, do you go build on top of an OpenAI API? Do you go build on something in the open source? Do you go build your own large model? How should a founder navigate making that decision?

Maybe just add one thing: the nice thing about having these APIs is it's extremely easy to get started and try something. You can sit down in an afternoon, punch in some data, and get a sense of the possibilities. In some cases, it's a lower bound on how well you can do, because you spend an afternoon, and if you invest more, if you fine-tune and build custom things, it can only get better. So I think that has opened up the challenge to even formulate what is the right problem to go after. Typically, you don't know, and you have to collect data and then train a model, and that loop becomes very expensive. But you could just sit down in an afternoon, try a few things, and maybe few-shot your way to something that's actually reasonable. That gets you into a different part of the space, and you can iterate much faster.

That makes a lot of sense. In terms of prototyping quickly and trying to take out product-market fit risk, one question becomes, and I'm curious for your take on this: if you start that way, how do you over time build durability into your product? Because I could make the argument that maybe you're just a thin layer on top of someone else's API. You can quickly de-risk product-market fit, but is there real durability in your layer of the stack?

Yeah, I think the transition out of API is a very discrete one in some sense. People also do like human wizard-of-Oz experiments. You put a human there and have the human do it, and then you see work out all the interface issues and whether this makes sense at all, and then you try to put something else, take the human out. Now you could put an API there and get a sense of what things are like. And then in some cases, maybe few-shot learning is actually for some things not that strong. If you have data, maybe a fine-tuned T5 model or something much smaller can actually be effective. I don't think the last thing on your mind should be to go pre-train a 500-billion-parameter model when you don't know what application you're building.

Maybe continuing on the theme of building on top of these models, despite the magical qualities of these things, there's still limitations. One of the limitations is falsehoods, and there are others that developers need to navigate as they think about building these applications. David, maybe starting with you, what do you think some of the key limitations are, and how do you guide people around navigating those?

Yeah, I mean, these models are exciting, and the flip side is that they have a ton of weaknesses. Falsehoods, generating things that are not true, biases, stereotypes—basically all the good and bad and ugly of the internet gets put into them. I think it's actually much more nuanced than just 'let's remove all the incorrect facts and debias these models.' Their efforts on filtering can filter out offensive language, but then you might end up marginalizing certain populations. And what is truth? We like to think there's the truth, but there's actually a lot of text which is just opinions, and there's different viewpoints, and a lot of it's not even falsifiable. So you can't talk about truth without falsifiability. In a lot of cases, for example creative cases, you do want things which are a bit more edgy, like how you create fiction. If everything has to be true, there's no easy way to throw out all the bad stuff. So I think one good framework to think about is that there's no one way to make these models obviously much better. I think what you want is control and documentation. You want to be able to understand, given a model, what is it capable of, what should it be used for, and what it should not be used for. This is tricky because these models have a huge capability surface. You get a prompt, you can put any string and get any other string back. So what are the guarantees? I think as a community, we need to develop a better language for thinking about what are the possible inputs and outputs, what are the contracts like the things you have in traditional APIs and good old-fashioned software engineering. We need to import some of that so that downstream application developers can look at it and say, 'Oh yeah, okay, I'll use this model, not that one, for my particular use case.'

There's another side to this which, for lack of a better word, I'll use the word 'risk.' If I'm a product builder and I'm building these models and building different products, there's different levels of risk I might be willing to take in terms of what I'm willing to ship to end users and the level of guardrails I need in place before I'm willing to ship. How do you guys think about that and frameworks around that?

Yeah, to build on top of that, I think there's another category of risk which is adversaries in the world. Whenever you have a product that's getting enough traction, there's probably people who want to mess with you. One example is data poisoning. This side hasn't really been born out yet; there's some papers on it. But if you think about it, these models are trained on the entire web crawl, so anyone can go put up a web page or put something on GitHub, and that can enter the training data. And these things are actually pretty hard to detect. So from a security point of view, this is a huge gaping hole in your system. A determined attacker could probably figure out a way to screw over your system. And the other thing to think about is misuse. All powerful models like these are dual-use technologies. There's a lot of immense good that you can do with them, but they can also be used for fraud, disinformation, spam, and all the things that we already know exist, but now amplified. That's a scary thought.

Yeah, definitely a lot of asymmetric capabilities here. Just hooking up one of these giant code models to some RL agent to get into systems, there's so many things that become way easier for malicious actors to do as a result. Attack is so much easier than defense.

This is a question for Percy. From an academic standpoint, what is an ideal wish list that you might have in ways that corporations and big companies that are building a lot of the ecosystem could help?

Yeah, I mean, I think one big thing is openness and transparency, which is something I think is really sorely missing today. If you look at the deep learning revolution, what's been great about it is that it's benefited from having an open ecosystem with toolkits like TensorFlow, PyTorch, datasets that are online, tutorials, and people can just download things, tinker, and play with it. It's much more accessible. And now we have models which are behind APIs with fees, and things have become such that you can't really tinker as much with it. So I think that. Also, a lot of organizations are addressing the same issues around safety and misuse, but there's not really an agreement on what the best practices are. So I think what would be useful is to develop community norms around what is safe or what are the best practices for mitigating some of these risks. In order to do that, there has to be some level of openness so that when a model is deployed, you get a sense of what these models are capable of, and benchmark them and document them in a way that the community knows how to respond, as opposed to 'Here is a thing you can play with. It might shoot you in the foot or not. Good luck.'

We have time for one more question. So maybe I'll ask a final question which goes back to creativity. I think one of the important things to inspire in all of us is what's going to be possible, the magic of these models. I think DALL-E was an important moment for people to see the type of creative generation that's possible. For each of you, what are some of the things that you think are going to be possible in the way we interact with these models in a few years that you're most excited about? Maybe starting with you, David.

Yeah, I agree with that and have definitely thoughts, but maybe I'll end with a different perspective, another example. If you think about it, we get excited about generating an image, but it's just an image. It's also something that you imagine an artist be able to do. And if you think about humans aren't changing that much; computers are. A year ago or two years ago, we weren't able to do that. Now we can. So if you extrapolate, now image is only just so small. Think about videos, or 3D scenes, or immersive experiences. Maybe with personas, you could imagine generating worlds in a sense. That's scary but also sort of exciting. There could be many possibilities there. But I think the bigness of things that you can create—if you think about these models as excellent creators of large objects now, and you think about what are big objects—well, there are environments in some sense. What if we could do that? What would that look like? What kind of applications could that unlock that would be interesting to think about?

Yeah, there's a commonality there which again connects back to multiple modalities and continuing to push scope. It's a really exciting glimpse of the future. Percy, David, thank you guys so much for doing this.

Thank you.

That concludes this episode of Gray Matter. Like what you hear? We encourage you to rate and review Gray Matter on your favorite podcast platform. We sincerely appreciate your feedback. You can also find all content on our website at greylock.com. I'm Heather Mack. Thanks for listening.