- Anthropic designs its enterprise AI models for accuracy and reliability, contrasting with consumer AI's focus on engagement, to ensure truthfulness in critical applications like drug discovery.
- Companies should be ambitious in their AI adoption, preparing for end-to-end integration of processes rather than just making small, incremental "hill climbing" gains.
- Leaders must anticipate the rapid pace of AI progress and plan future projects based on where the technology is heading, to avoid significant delays in deploying beneficial solutions.
Scaling enterprise AI: Fireside chat with Eli Lilly’s Diogo Rau and Dario Amodei
- Anthropic's enterprise AI strategy prioritizes truth, accuracy, and reliability over consumer-driven incentives for engagement and growth, which can lead to "model sycophancy."
- For critical enterprise applications like drug discovery, AI models must provide objective truth, rather than agreeing with or flattering the user, to prevent costly and dangerous decisions.
- Enhancing models with deep, domain-specific knowledge (e.g., graduate-level biochemistry) is highly valuable for enterprises, even if general consumers do not readily perceive the improvement.
- Anthropic is developing "specialized quads" which are tailored AI solutions that combine inherently smarter, fine-tuned models with wrappers providing access to specific external information relevant to a particular industry (e.g., financial indices, biological databases).
- Organizations in drug discovery should avoid only applying AI to small parts of existing processes and instead begin preparing to integrate AI for the entire workflow end-to-end.
- It is crucial to have faith in the rapid pace of AI technological progress and design future projects based on where AI will be in two years, not just its current capabilities.
- Proactively planning for large-scale AI transformations in parallel with technological advancements can save years of deployment time, accelerating benefits for patients.
model sycophancy — The tendency of an AI model to agree with or flatter the user, even if the user's statements are incorrect or unwise.
enterprise strategy — A plan outlining how a company will engage with and serve business clients, often focusing on reliability, security, and specialized solutions.
specialized quads — Anthropic's term for tailored AI solutions that combine inherently smarter models with access to specific, external information relevant to a particular domain.
fine-tuneings — The process of further training a pre-existing AI model on a smaller, specific dataset to adapt it for a particular task or domain.
end-to-end — Pertaining to a system or process that handles a task completely from start to finish, without human intervention or reliance on non-AI components for intermediate steps.
hill climbing gains — Small, incremental improvements made by optimizing individual parts of a system, rather than pursuing a larger, transformative change.
drug compound — A chemical substance that is being investigated for its potential therapeutic effects in drug discovery.
assays — In biology, laboratory tests used to measure the presence, amount, or functional activity of a target substance or process.
What's have faith in the pace of progress of the technology? Because if the models get good enough to do it end to end a year from now and only then you start deploying it, there will be another two-year delay and that's you know that's two years during which all the work that you're doing to benefit patients is not happening. Hello everyone, my name is Diego Rao and I'm Chief Information and Digital Officer of the Eli Lilly and Company. I'm joined here with Dario who is the founder and CEO of Anthropic. Dario, thanks for joining me today. Thanks for having me, Diego. I know you're spending a lot of time now thinking about how do you work better with enterprises. What's your enterprise strategy and how do you see Anthropic different from other providers? Yeah I mean you know I think we've made a number of choices that are different, right? So if I think about the incentives given by consumer AI, their folks are in a competition for engagement and growth, right? And so that drives a lot of behaviors of the AI that I think are not ideal from an enterprise perspective. For example, there's this idea of model sickofancy where the model tells you whatever you say is a good idea, right? And even on the consumer side that can cause problems, we've seen stories of people who are like, oh yeah I've discovered a new fundamental theory of physics and that's right, models like, that's great. And maybe you don't want it, maybe you don't want it to say that. But I think of course on the enterprise side, the problems are much greater and clearer with that. You really don't want the models to say, oh yeah this drug compound is great. It's been millions of dollars to, I just think this is, I think your idea is great. I think it's really promising. You want truth. And so I think that incentive has led us to design our models in a different way. I think it's more compatible with making the model smarter, making them better at a wide variety of economically valuable tasks. And it causes us to put a premium on accuracy and reliability. One experiment that I give to everyone, although it's particularly relevant because I'm talking to you is I say, let's say I improve the model's knowledge of biochemistry from undergraduate level knowledge to graduate level knowledge. So go to consumers and say that now you're 99% of them are going to say, I didn't know what you were talking about before. I don't know what you're talking about now. But if I go to you, you can appreciate that a lot. That's very important. That's exactly right. Well actually that gets into something else that you've launched as well, which is skills, right? There are a lot of skills that you want in biology or even just skills as an enterprise, how you want to operate. Is that part of the future for you as well? Yeah, I definitely think so. I mean, things ranging from skills to, you know, we're in the process of launching various specialized quads, which are, you know, in some cases will be improvements to the model itself, fine tuneings of the model. But in some cases it'll be something that looks more like wrapping the model with access to particular types of information. So when we did Claude for financial services, you know, we connect it to a lot of the usual kind of indices and ratings. And so, you know, you'd be surprised how much just making it easy to connect Claude to those things and kind of use it in a way that's aware of that knowledge is valuable. So I think, you know, we're working on a Claude for life sciences that will be some mixture of making the model inherently smarter and wrapping it with various things, right? I don't know exactly what the analogy will be here, but like, geez, there are zillions of databases of, you know, proteins, compounds, assays, like, you know, you probably want that at the model's fingers tips. Well, any parting advice for those of us that are working in this world of drug discovery and development? You know, I would say there's a temptation and it's, I think it's hard to avoid starting this way. You know, what are the small things we can do with AI? Like, in a way, you just kind of have to start there. I think one of my pieces of advice is be very, very ambitious in terms of where the models are going. I think you can get caught in a mode where there's an existing process that has 20 parts. You want to swap in AI to part five and part 12. Right. And, you know, that can actually be hard because part 12 has to, you know, intersect with part 13 and part 11, which are not being done with AI. And, you know, you look at, you're like, well, the AMOles aren't, aren't where they could do, you know, part zero to part 20 end to end. Right. And so you should start thinking now, don't get too seduced by, oh, we can make these little hill climbing gains by, you know, doing this part in that part. Let's start preparing to do the whole thing end to end. Let's have faith in the pace of progress of the technology because if the models get good enough to do it end to end a year from now, and only then you start deploying it, there will be another two year delay. And that's, you know, that's two years during which all the work that you're doing, the benefit patients is not happening. Whereas if you go in parallel, if you start preparing now for the, the larger change, as the models were getting better, then, you know, you may save years of time. That's right. So don't do two year long projects and, and expect it, it's going to be exactly the same way in years from now. Yes. You do two year long projects plan for where the AI is going to, I mean, that sounds like an obvious thing to say, but I, but I think it actually, it actually takes a lot of courage and foresight to do that. It does for sure. Well, thanks a lot for taking the time to chat today. Really appreciate it. Yeah. Yeah. Yeah. And then you can see.
TL;DR
- Anthropic designs its enterprise AI models for accuracy and reliability, contrasting with consumer AI's focus on engagement, to ensure truthfulness in critical applications like drug discovery.
- Companies should be ambitious in their AI adoption, preparing for end-to-end integration of processes rather than just making small, incremental "hill climbing" gains.
- Leaders must anticipate the rapid pace of AI progress and plan future projects based on where the technology is heading, to avoid significant delays in deploying beneficial solutions.
Takeaways
- Anthropic's enterprise AI strategy prioritizes truth, accuracy, and reliability over consumer-driven incentives for engagement and growth, which can lead to "model sycophancy."
- For critical enterprise applications like drug discovery, AI models must provide objective truth, rather than agreeing with or flattering the user, to prevent costly and dangerous decisions.
- Enhancing models with deep, domain-specific knowledge (e.g., graduate-level biochemistry) is highly valuable for enterprises, even if general consumers do not readily perceive the improvement.
- Anthropic is developing "specialized quads" which are tailored AI solutions that combine inherently smarter, fine-tuned models with wrappers providing access to specific external information relevant to a particular industry (e.g., financial indices, biological databases).
- Organizations in drug discovery should avoid only applying AI to small parts of existing processes and instead begin preparing to integrate AI for the entire workflow end-to-end.
- It is crucial to have faith in the rapid pace of AI technological progress and design future projects based on where AI will be in two years, not just its current capabilities.
- Proactively planning for large-scale AI transformations in parallel with technological advancements can save years of deployment time, accelerating benefits for patients.
Vocabulary
model sycophancy — The tendency of an AI model to agree with or flatter the user, even if the user's statements are incorrect or unwise.
enterprise strategy — A plan outlining how a company will engage with and serve business clients, often focusing on reliability, security, and specialized solutions.
specialized quads — Anthropic's term for tailored AI solutions that combine inherently smarter models with access to specific, external information relevant to a particular domain.
fine-tuneings — The process of further training a pre-existing AI model on a smaller, specific dataset to adapt it for a particular task or domain.
end-to-end — Pertaining to a system or process that handles a task completely from start to finish, without human intervention or reliance on non-AI components for intermediate steps.
hill climbing gains — Small, incremental improvements made by optimizing individual parts of a system, rather than pursuing a larger, transformative change.
drug compound — A chemical substance that is being investigated for its potential therapeutic effects in drug discovery.
assays — In biology, laboratory tests used to measure the presence, amount, or functional activity of a target substance or process.
Transcript
What's have faith in the pace of progress of the technology? Because if the models get good enough to do it end to end a year from now and only then you start deploying it, there will be another two-year delay and that's you know that's two years during which all the work that you're doing to benefit patients is not happening. Hello everyone, my name is Diego Rao and I'm Chief Information and Digital Officer of the Eli Lilly and Company. I'm joined here with Dario who is the founder and CEO of Anthropic. Dario, thanks for joining me today. Thanks for having me, Diego. I know you're spending a lot of time now thinking about how do you work better with enterprises. What's your enterprise strategy and how do you see Anthropic different from other providers? Yeah I mean you know I think we've made a number of choices that are different, right? So if I think about the incentives given by consumer AI, their folks are in a competition for engagement and growth, right? And so that drives a lot of behaviors of the AI that I think are not ideal from an enterprise perspective. For example, there's this idea of model sickofancy where the model tells you whatever you say is a good idea, right? And even on the consumer side that can cause problems, we've seen stories of people who are like, oh yeah I've discovered a new fundamental theory of physics and that's right, models like, that's great. And maybe you don't want it, maybe you don't want it to say that. But I think of course on the enterprise side, the problems are much greater and clearer with that. You really don't want the models to say, oh yeah this drug compound is great. It's been millions of dollars to, I just think this is, I think your idea is great. I think it's really promising. You want truth. And so I think that incentive has led us to design our models in a different way. I think it's more compatible with making the model smarter, making them better at a wide variety of economically valuable tasks. And it causes us to put a premium on accuracy and reliability. One experiment that I give to everyone, although it's particularly relevant because I'm talking to you is I say, let's say I improve the model's knowledge of biochemistry from undergraduate level knowledge to graduate level knowledge. So go to consumers and say that now you're 99% of them are going to say, I didn't know what you were talking about before. I don't know what you're talking about now. But if I go to you, you can appreciate that a lot. That's very important. That's exactly right. Well actually that gets into something else that you've launched as well, which is skills, right? There are a lot of skills that you want in biology or even just skills as an enterprise, how you want to operate. Is that part of the future for you as well? Yeah, I definitely think so. I mean, things ranging from skills to, you know, we're in the process of launching various specialized quads, which are, you know, in some cases will be improvements to the model itself, fine tuneings of the model. But in some cases it'll be something that looks more like wrapping the model with access to particular types of information. So when we did Claude for financial services, you know, we connect it to a lot of the usual kind of indices and ratings. And so, you know, you'd be surprised how much just making it easy to connect Claude to those things and kind of use it in a way that's aware of that knowledge is valuable. So I think, you know, we're working on a Claude for life sciences that will be some mixture of making the model inherently smarter and wrapping it with various things, right? I don't know exactly what the analogy will be here, but like, geez, there are zillions of databases of, you know, proteins, compounds, assays, like, you know, you probably want that at the model's fingers tips. Well, any parting advice for those of us that are working in this world of drug discovery and development? You know, I would say there's a temptation and it's, I think it's hard to avoid starting this way. You know, what are the small things we can do with AI? Like, in a way, you just kind of have to start there. I think one of my pieces of advice is be very, very ambitious in terms of where the models are going. I think you can get caught in a mode where there's an existing process that has 20 parts. You want to swap in AI to part five and part 12. Right. And, you know, that can actually be hard because part 12 has to, you know, intersect with part 13 and part 11, which are not being done with AI. And, you know, you look at, you're like, well, the AMOles aren't, aren't where they could do, you know, part zero to part 20 end to end. Right. And so you should start thinking now, don't get too seduced by, oh, we can make these little hill climbing gains by, you know, doing this part in that part. Let's start preparing to do the whole thing end to end. Let's have faith in the pace of progress of the technology because if the models get good enough to do it end to end a year from now, and only then you start deploying it, there will be another two year delay. And that's, you know, that's two years during which all the work that you're doing, the benefit patients is not happening. Whereas if you go in parallel, if you start preparing now for the, the larger change, as the models were getting better, then, you know, you may save years of time. That's right. So don't do two year long projects and, and expect it, it's going to be exactly the same way in years from now. Yes. You do two year long projects plan for where the AI is going to, I mean, that sounds like an obvious thing to say, but I, but I think it actually, it actually takes a lot of courage and foresight to do that. It does for sure. Well, thanks a lot for taking the time to chat today. Really appreciate it. Yeah. Yeah. Yeah. And then you can see.