[0:00:05] Hello, everyone. Very happy to be here again with Tom. Thanks for being with us, Tom. Today, I have a few questions. Today, I have them written down on my cards to keep them simple. But Tom, maybe if you could just introduce yourself briefly and tell us who you are. Sure. Yeah. So I'm Tom Brown. I'm one of the co-founders at Anthropic, where I work on our core LLM.
[0:00:34] And before that, I was at OpenAI where I led engineering for GPT-3 before our team left to found Anthropic. And created Anthropic. And congratulations for the recent benchmarks on Claude 3, and we'll talk about that in a little bit. But before I got there, maybe...
[0:00:57] You have a particular mission and vision at Anthropic and you have been approaching Gen AI and AI in general in a little bit of a different way. Maybe you can tell us about that mission and vision as well, Tom. I find it fascinating. Totally, yeah. So at Anthropic, our mission is to build trustworthy AI that benefits all of humanity. And so the way that we've been doing that has been building...
[0:01:19] Frontier language models, doing research on how to make them better, and then deploying it through our partnerships like Amazon and with our various products. That's awesome. And on the performance of Claude3, so I was talking about the performance of Claude3 earlier, the benchmarks, I think all of the people in this room have probably seen them on social media. First of all, can you tell us a little bit about that performance and then maybe how you got there? Yeah, totally.
[0:01:49] Yeah, so maybe I could talk through kind of like the three models. Yeah, that would be great. Yeah, cool. Yeah, so there's three models in the Claude 3 family. Kind of our thinking behind that was just that like one model isn't going to be the right model for everything that people are building. And so Opus is our like most capable model. It's like...
[0:02:17] I guess it's at the top of the benchmarks for various different things. I think as of last week, it was top of LMSYS too, which is the first time a model has beaten GPT-4 on the blind taste test kind of thing for chat interfaces. And broadly, it's very good at fine instruction following. If you need it to reason about something before answering it.
[0:02:47] Coding, like all of those types of things that I don't know, kind of are just like you need a smarter model for. So that's Opus. That's Opus. Sonnet is our like medium model. And it's basically like a good balance between like as smart as possible and like as cheap and fast as possible. And then Haiku is our like smallest, fastest model, which is...
[0:03:16] For that one, we tried to make it smart enough to do most tasks that people want, so smarter than GPT-3.5 Turbo, while being as cheap as possible. So it's about half the price as GPT-3.5 Turbo. So it's three times better and twice as cheap. It's hard to measure how much better it is, but notably better than GPT-3.5 Turbo, but half the price.
[0:03:45] So on the engineering side, what do you have to do to make a model that good? There's like a lot of different things where I think kind of like different ingredients for that are like I guess like kind of like price performance is the way I kind of think about it. And the things that contribute to that are like engineering on the like training and inference size.
[0:04:14] That can get you more efficiency. The hardware that you run on too. Your engineering will be specific to the hardware that you run on. And then the algorithmic innovations that you put in your model. Where these modern models are descendants of the original transformer family. Like what we did for GPT-3. But quite different in...
[0:04:42] Both like how they work internally and how they were trained. And so if you think of Claude, how does it differ from other LLM models out there? Yeah. Because I think for many of us, at least for me, sometimes it's hard to understand the differences, right? Yeah. Yeah. So I think the models that we're most similar to are the OpenAI models. Like a lot of us are OpenAI alumni, not that.
[0:05:10] Not that surprising. The main differences, I think, for us have been trying to make the models as trustworthy, reliable as possible. And then I think the price performance is a place where we've put a lot of effort into, and I think that that shows with this new round of models.
[0:05:36] So you talked about Opus, you talked about Sonnet, you talked about Haiku, right? I think I have them categorized, right? Yeah, yeah. I talked this morning, and you were there, I talked in my introduction, I talked about choice, right? And we think, and this comes from the feedback of our customers, that choice is extremely important. Was that choice that was behind the idea of you providing three models? 100%, yeah. I think for a lot of customers, the workflow that... Before this, we had two models.
[0:06:05] So we had Claude 2 and Claude Instant. And what we would see is that people would start out with the smartest model and then get something working. When you're first just playing around, you don't care that much about cost because you're just playing around. You just do it a few times. Yeah, do it. And there's like, I don't know, if the model costs like $5 for a million tokens, you're not going to use a million tokens. And even if you did, it would be $5. So people would like when they're...
[0:06:35] Testing things out, experimenting, use the biggest model. And then once they have this approaching production and if it starts to become high, then people would want to use the smaller model and see if it works. And so we basically just extended that with three models now, where that lets people more finely tune the balance between intelligence, speed, and cost.
[0:07:01] And it's crazy because we're only in the first few months of 2024 in April and we definitely saw and for most of you out there we definitely saw 2023 as the year of proof of concepts and trying things around and since the beginning of 2024 many of those use cases have now been going into production and that's great to see. One of the things obviously we talk a lot about when we talk with Anthropic is the fact that you have a huge sense of responsibility
[0:07:31] In the way you're approaching AI and you're putting guardrails and rules. Can you tell us a little bit about that? Because I think it's a huge difference in the way you approach it. Yeah. So I think one thing that we've been doing... I guess maybe there's three different areas that I could talk about. One is the model itself. And I think for that...
[0:08:00] Previous versions of our models had, and I think everybody has had challenges with, you want the model to not do anything that would embarrass the customer who's using it. And customers want to make sure that their model's not going to do anything embarrassing. But you also don't want the model to annoyingly refuse a harmless thing. And so that's somewhere where we put a lot of effort, was making sure that we reduced the...
[0:08:29] False refusals as much as we could. So that's kind of one area. Another area is guardrails, which a customer can tune to make the model have more fine-grained guarantees about how the model will behave. And that's actually a place where Amazon, I think, has been leading the way. Amazon has the guardrails product right now, which I think is a fantastic way for folks to...
[0:08:58] To like get that extra layer of security and trustworthiness. And then for like the very long term, we have our responsible scaling policy, which doesn't apply really to these models now beyond just like we want to have the best security that we can. But like for models that eventually become like much more powerful, like if a model like none of the models right now could like build a bioweapon.
[0:09:28] Or, like, substantially speed up someone who, like, a bad actor that was trying to build a bioweapon. And so for those, we, like, it would be premature to, like, put in place a lot of safety barriers that would, like, protect against that, but slow down development. And so we don't do that now. But we have a plan for, like, if next year or two years from now, like, the models can make bioweapons.
[0:09:53] How we would handle that. And so that's your responsible AI, your responsible scaling policy. Exactly. And that's the third one. Which is one of your core value. And you're already thinking that the models will be so powerful that you're thinking today about the scaling policies you need to put in place. Exactly. Those are not needed as of today yet, you're saying. Yeah. So the way that we have it is like, there's like...
[0:10:18] We conceive of it as AI safety levels that's kind of similar to the global biosafety level standard. And so if you want to have a bio facility that can just handle non-dangerous bio.
[0:10:38] If you're just doing normal testing of an in vitro fertilization clinic or something like that, you have to handle some bio material. That might be a biosafety level two lab, which means you have to have some level of security on it. But if you have a thing where you're handling smallpox or something like that, then that would be a biosafety level three facility, and you need much more stringent guides for it.
[0:11:07] That's fascinating. So you're taking the inspiration from some of what the industry has been doing already. Exactly, yeah. And so for us, now these models are ASL 2, like AI safety level 2, which is like they're not actually that dangerous. Like even if you removed all of the...
[0:11:22] All of the safety features and guardrails. It would be embarrassing maybe, but no one's going to get hurt. Which is like, you're laughing, but I think that's a good thing, right? Yeah, absolutely. It's also great to hear you thinking about it in the future. Yeah, exactly. And I think everyone needs to think about that too. Because if we're not prepared for that, then the models will be able to make a bioweapon. And then either a bioweapon will get built or everyone is going to have to stop. And nobody wants either of those things to happen.
[0:11:52] And that's one of the very important things, which is how do you keep putting those guardrails and security features on at the same time as allowing innovation to progress and solving some of the other issues. So totally hear you on this. The other one I wanted to ask you about, Tom, because I've heard you talk about it in the past is constitutional AI. And I don't know if some of you have heard the term constitutional AI before.
[0:12:22] Which is not easy to say when you're French by the way. But constitutional AI, what is it? Yeah, so basically constitutional AI is making it easier to take some concrete principles that you want the model to follow and then kind of like bake that into the model's personality. And so that's a thing that we use to give
[0:12:52] Claude, like the personality and behavior that it has. And I think that's just a useful thing for the world. And so we have a published paper that kind of explains how we do it and how other people might do similar work. So if we go to Anthropic website, we can download the paper on constitutional safety? Yeah, I think it's a fun read. It is? It's quite readable. And I'll get some examples of how you do personality. Okay, so I'll go and look that up as soon as we're done with the summit season.
[0:13:21] I think going back a little bit in your development, what are some of the technical challenges related to safety and security that you face when you develop AI and how do you approach it? And maybe give some examples of what you found as you were developing Claude and how you addressed it. Yeah. Let's see. I think that for...
[0:13:53] I was less engaged in the post-training efforts that take the raw language model and then turn it into the Claude personality. I think the constitutional AI stuff seems like it's one big area. We had a pretty large red teaming effort too.
[0:14:22] Where we would have various people go and try to get the model to do bad things. Try to get it to do things like build bioweapons, which generally were prevented because the model's just not smart enough to build bioweapons yet. But we wanted to check, and so we carefully did evaluations on that front.
[0:14:49] Other red teaming of can folks jailbreak the model to get it to do embarrassing stuff, which I think would be more of a risk if you have this as a company, you deploy it to end users, and you don't want the folks to then use your model to write erotica or something like that. And so that's another layer of the safety field that we put on.
[0:15:14] Very interesting. I think, you know, policymakers right now are looking into AI as they should, and I think we all believe here that there is a role in that. You're talking about bioweapons, which is a pretty extreme scenario. I think one of the conversations with policymakers right now is also around misinformation or hallucination, right? And I think one of the big things you did on Claude 3 was working on that aspect. How important is that to you, and what are you doing to address it?
[0:15:43] Yeah, so I think figuring out how to make the model more trustworthy is just a super high priority for us. And hallucinations are one of the biggest things that causes, that means that people can't trust the model. And so figuring out ways to get the model to appropriately
[0:16:11] Caveat, if it doesn't really know something, has been a high priority for us. And then also letting the model use more information sources is another one. And so two different ways that we do that. One is just having a very long context and making sure that the model appropriately remembers everything from its context. So there's like a needle in the haystack eval that folks have and that we did where we took a very long...
[0:16:40] I forget if it was like Moby Dick or something like that. And then in like one sentence in like the whole book, we added a question or like a little, we replaced one of the sentences with like a statement on like what our favorite pizza topping is. And then so this model read all of Moby Dick. Mine is pepperoni. Mine is garlic. So model read like all of Moby Dick and then we asked it like, hey, was there anything weird about this thing that you just read?
[0:17:08] And it was like, yeah, there was this thing about pizza toppings that doesn't seem like it fits well in this book about whaling. So that seemed kind of weird. And this went slightly viral on Twitter, too, where the model then also was like, hmm, maybe that pizza toppings thing was put there to test whether or not I was paying attention while I was reading the book.
[0:17:30] That's incredible. Absolutely incredible. So we talked about policy a little bit, and I know you've been and some of your other co-founders have been extremely involved in a number of meetings that have been organized around the world. I think the first one that was really big was the White House Voluntary AI Commitments, and then a few months back.
[0:17:49] We had the UK Safety Summit. Then just last month, I was with one of your other co-founders, Dario, at the G7 and B7 talking about AI. What is your perspective on regulation and what it means in this space? Yeah. Yeah, so I think regulation is definitely coming. And so I think it seems like that's an opportunity.
[0:18:17] To do great good because I think there are risks both now and in the future that these models will have. But it's also a place where poorly executed regulation could have a hugely damaging effect on the whole world. And so I think it's super important. I thought the executive order in the United States seemed quite reasonable when I read it. And so yeah, I would like to see more stuff like that.
[0:18:46] Okay. And so what level do you think is appropriate versus when it would stifle innovation, as you say? Yeah. Well, so I think the responsible scaling policy that we put together pretty closely matches with what I think, which is just that if there's a real risk, like bioweapons or offensive cyber work,
[0:19:16] Those things should be guarded against because nobody wants... Exactly. Plus you have the rules and you have your core values, but not everybody might have them. So absolutely agree with you. Maybe moving on a little bit from the topic of regulation, even though it's super important. Maybe I wanted your view on the recent advancements of conversational AI and its progress there.
[0:19:41] And whether you think we're approaching a form of artificial general intelligence or general artificial intelligence in whichever way you want to put it, what's your view on that? Because it's been extremely interesting. Yeah. Yeah, so I kind of think that, like, I think definitely capabilities have been super, like, advancing really quickly. So when we built GPT-3, like...
[0:20:10] I think that that was a big improvement over GPT-2, where GPT-2 couldn't count to 10. And Tom, what was the lag of time between 2 and 3? I think it was two years, a year and a half to two years. We finished training GPT-3 a little bit before we released it, so it depends where you count. GPT-2 was like...
[0:20:39] I think like blew people's mind at the time because it could kind of do anything or like it could do something but like it couldn't really do that much like there was like the cool like example that we had where it like told a story about like some unicorns being discovered and it was like a kind of plausible story like the model could kind of write um but if you asked it like hey give me a list of like your top 10 favorite animals it would like be like animal number one is like frog animal number three something
[0:21:07] Animal number four or something. Animal number seven something. Animal number three again. It couldn't count to 10 if it was just trying to enumerate a list. So the model was quite stupid. GPT-3 was a big leap over that. But still, I don't know. One of the impressive things that we found was that if you added it to take two
[0:21:33] 3-digit numbers, add 211 and 244 or something like that, it would get that right 80% of the time, which is much more impressive than not being able to count to 10, but still not that impressive. And then GPT-4 another year and a half later, or Claude Opus, now these models, Claude 2 could pass the bar exam.
[0:21:57] And so we've seen very, very fast improvements there. And I think from that perspective, I think these modern models are, I think, weak artificial general intelligence already. They're pretty general. They have some intelligence, but they're obviously not as smart as a human. But even if they know a lot more facts than a human does. Do you think we're at a weak general artificial intelligence stage?
[0:22:27] And we're moving towards slightly more every month. And so how do you see Claude fitting in our daily lives now? Yeah. Well, I think it depends on different people. I use Claude for all sorts of stuff. I don't know. I'm like an engineer by training, but a lot of my time now is spent on management.
[0:22:56] Dealing with supply chain stuff and stuff that I'm totally unqualified to finance stuff, stuff that I know nothing about, basically. And so I found Claude super useful as a coach for that stuff, where it's like, if I want an opinion from someone who has read everything about that, has all the blog posts, et cetera, that...
[0:23:23] That is something where I suddenly have access to that and I can trust that my conversation is going to remain private. I'm not going to waste this person's time. They're happy to help out. Me personally, that's been a huge thing. I think also I still try to spend as much time as I can doing coding, but the team has moved on on many things and now we'll use languages that I don't know.
[0:23:52] Or that I'm not as good as the other folks on the team. And so for things like that, too, where I don't know the best practices or how to use different libraries.
[0:24:03] Just asking the model because the model knows all of that too. That's been a huge accelerator for me. So you use it as a coach and a personal advisor and a code optimizer and a code teacher. That's awesome. Yeah, exactly. Okay. Actually, you've given me a few things to think about. You know what? Last time I used Claude was at the beginning of this week to help me fix a speech I was working on. Oh, nice. So that worked out well. What do you think are some of the most significant...
[0:24:32] Use cases that you see right now in the use of Claude, and you're obviously focused on the enterprise and on the business side of Gen AI, right? So what are some of the use cases that you've been impressed by and that you think are very promising and where you see a lot of value for businesses? I think there's a ton of stuff. I think it's interesting for me to just hear the wide...
[0:24:57] There's a huge breadth of different things that people use for the model. I guess I could talk about a broad category. I think there's a huge long tail where one broad category that lots of people use it for is customer support tickets. That's a thing where there's lots of them. You need to look over some knowledge base to figure out what the model is actually to get the answer for.
[0:25:24] Use that to accelerate their customer support. And so I think that by itself is a huge thing. I do think that then the long tail of things looks like, and maybe if I was to make that a little more abstract, it's like using retrieval augmented generation RAG on internal knowledge bases of some sort to answer questions. I think that's a huge category.
[0:25:53] I think that the long tail is more like one Pippi description I heard of what is business-to-business SaaS. B2B SaaS mostly is taking a business function that's used across many companies and automating it. That's the majority of what B2B SaaS is.
[0:26:22] And then people will make a whole company to do that. But in order for that to happen, you need this to be enough companies need to care about this or have the same business practice that it justifies a whole new company starting up to solve this problem. And so I think one thing that these models do and that we see people using it for is doing business practices that they have that they will now automate, but that wouldn't justify a whole.
[0:26:52] A whole new startup outside to like go and tackle it. And so like one cool example just from this week that I heard was there's a customer who basically they want to make sure that they basically sell a physical device. And so that physical device is like in the like listed in the
[0:27:18] Like, in the, like, catalogs of, like, Best Buy or, like, the different, like, electronics stores. And so what they want is to make sure that, like, their device shows up at the top of that list. And so, like, what they do is they have the model, like, every day go and, like, read the catalogs from all of the different, like, hardware, like, stores and see if their product is at the top. And if it's not, they get an email and then they go and email.
[0:27:46] Best Buy, and they're like, hey, I thought we were going to be at the top. And so it's a very simple use case. There's no way that this particular thing would justify a whole separate business to do it. But now that's not wasting a person's time to read this every day. It's now an automated business practice. I think this is a great example. Just before getting on stage today, we were discussing with someone and asking us where we were seeing.
[0:28:16] I think most of it is still to be invented in the way we use this. I can't wait to see next year when we come back here some of the new use cases that have been invented and that we don't think about. Just a side question, what's one of your last prompts that you've written? The most recent prompts? The one that you're comfortable sharing. Some of my prompts are pretty personal.
[0:28:47] I tragically don't speak French. As I came here, I wanted to ask the model, how can I say that I don't speak French yet? The model gave me advice on that. I don't speak French yet.
[0:29:13] Encore. Encore. Yeah, yeah, thank you. So that at least was like a useful thing that the model helped me with. So one of the things we just announced this week, or was it end of last week, an extended partnership between AWS and Anthropic. We obviously, at AWS, we love what you do. Our customers love what you do, and they tell us all the time. What does that partnership mean to you at Anthropic, and why did you enter into it, I guess? Yeah, well, I think...
[0:29:42] I think Amazon is the most customer-obsessed company in the world, and so I think that that's a mutual focus that we have on making sure that our models are useful to the people who use them. And I think also about building trust over the long term. And so that seemed like a place where there already was just a lot of cultural resonance.
[0:30:09] And so I think that was one big reason. And then also, I think this shows up in the security of the cloud, too. So AWS, I think, has the best cloud among many different axes. One that I think it clearly, clearly has it is with security. And that's a very fundamental part for us. I mentioned that ASL 2...
[0:30:37] has many fewer requirements than ASL 3. One place that it does have a high requirement is security, where we don't want any of this stuff to get stolen. And so that was the thing that was really attractive to us. And I think that's what Milan was sharing this morning. And I love the analogy, which is where the internet stops is where S3 begins in terms of using the models and the fact that you can, through Bedrock, deploy Anthropic in your virtual private cloud. And then at that point,
[0:31:06] The model is basically your own, and it never gets back to the internet, and your data remains your data, your prompts remain your prompts, your results remain your prompts, and that is so important to our customers right now. And I think one of the reasons why so many customers, and thank you all, but go in production using Bedrock and Anthropic right now, and absolutely the security is critical, and plus all of the great work that you've done on cost, right? And the cost equation is extremely important.
[0:31:36] Good. I think we're coming to the end of our 30 minutes here, Tom. But anything else you wanted since we're in this beautiful room together? Anything you wanted to share? Well, I guess I think we've said before, but right now folks can use Claude on Bedrock in regions, if you're okay with regions outside of France.
[0:32:04] As of end of today, you'll be able to use it in France as well. And so I'm just super excited for you guys to use this and build and to see what you guys put together. Super low latency using Bedrock in CDG for EUS3. I think it's a technical name. I call it CDG. But I think a huge round of applause for everything you do at Anthropic. Thank you because I think...
[0:32:30] What our customers are doing with your technology is just incredible. And also, thank you for the way you are approaching it. I think it's a really refreshing conversation to have in terms of some of the way you look at it and some of the ideas behind constitutional responsibility. So thank you so much for being here. Thank you so much for announcing those things here during the AWS Paris Summit. And hope to see you again in Paris soon. But thanks so much. Big round of applause.
[0:33:00] Thank you, Julien. Thank you all. Thank you all. And I hope you've been having a good day. And it was great to have Tom and Anthropic on stage with us today. Enjoy the rest and see you soon. Thank you.