Anthropic’s first technical PM on token maxing, the jagged edge, and living in the future | Dianne Penn

0:00 In 2023, when I started, nobody said anthropic and clawed and coding in the same sentence. I want to go back to the beginning of Anthropic. I remember dealing, man, these guys have no chance. OpenAI is so far ahead. By the time I saw people were starting to use these models not just for code auto-complete. But actually Writing long form code. And that an opportunity for us to train Opus III to be better at. That was the inflection. I always think about Opus Four Five a year later during winter break when everyone was home, able to code. What was magical about Opus 45 is we also now not just had a model, but a vehicle, a great product experience, like Cloud Code. Code, Opus Four Five wouldn't have had that moment without a product like Cloud Code. And Cloud Code wouldn't have had that type of adoption accelerated without Opus Four Five. I want to talk about how the product role is changing. For my team, the way to drive user value is to figure out the right user feedback, the evals. We actually have a saying on the team of Evals are the new PRDs. Something Gary Tan's been talking about. If you're willing to spend$100,000 a year right now on tokens, you are living the way somebody in 2028 is gonna live. You have to sweat the tokens as much as you sweat the pixels. You have to be using the models to come up with good, then great, then better ideas.

1:19 substitutes for that. People need to be more ambitious with AI tools these days because they're just capable of so much. One thing I asked the team is let's say Claude 8 comes around. What changes in what users do? What does that mean for how you're building today? Today my guest is Diane Penn. Head of product for the AI research and labs teams at Anthropic. She joined Anthropic as the first technical product manager. Over three years ago, which is a lifetime in AI time.

1:48 When the product tumus five engineers She's helped ship every model at Ananthropic from Claude II through Fable. She's also helped incubate and launch claw code, MCP skills, claw design. And also core capabilities like computer use, tool use, and reasoning. It is always such a treat and so mind expanding.

2:06 To get to talk to someone who's at the very center of AI and product management. It's hard to imagine someone who has seen more of where things are going. than the head of product for Anthropics research and labs teams. Before we get into it, don't forget to check out Lenny'sProductPass.com. For a year free.

2:23 Of the hottest and most beautifully crafted AI products in the world. Available exclusively to Lenny's newsletter subscribers. With that I bring you Diane Penn. Diane, thank you so much. For being here and welcome to the podcast.

2:40 Thank you, Lenny. It's so nice to see you again. I wanna go back to the beginning. Uh Anthropic.

2:48 uh the early days. I remember when Anthropic first launched, this was I don't know, yeah, the first model when it launched years ago, three years ago, something like that. It works. Three years. I remember just like feeling that Man, these guys have no chance.

3:01 Open AI is so far ahead. Everyone they're just like how what are they thinking? How is this possible? Open A as one. It's too late. Uh things are very different now. The latest number I saw was Anthropic was making like, I don't know, fifty billion dollars in ARR.

3:16 That's like what companies used to go public at. Like very successful companies want public at fifty billion. Invaluation. Anthropic reportedly is making that every single year. You joined. As one of the earliest PMs.

3:30 There were something like five engineers when you joined. The model hasn't And even launched when you joined. What was it like in those early days? Event dropic.

3:40 What's something that might surprise people about what it was like at the beginning. I think a big part of What's made anthropic. today actually has been very much the core of even the early days. So I joined in twenty twenty three.

3:55 Like you said we had Five product engineers. There was one engineer for the entirety of our API business. If you if you believe. Um and I think A big portion of it was the culture was really strong.

4:09 And I think this is something I emphasize for folks who are interested in the company. Um, really do walk the walk of um the mission and the culture and the values. Um And the energy was very much like a start up. And I think you're right, we were very much trying to find

4:26 our identity. In the early years. Like I think there's one's piece around the technology, but how does that technology Bring value to users.

4:36 bring value to society and what can it possibly be. And I think the early years were us exploring that in different ways. Like we did start with like Claude.ai, like another chat bot chat assistant. and evolving into things like tool use. Um I think one of the moments where

4:55 really we started to get into our groove was shipping things like Golden Gate Claude. I don't know if you like remember that. No. Um so this this was actually up for about twenty four hours or so. Uh we had just published one of our um early interpretability research. In early twenty twenty four. And one of the examples was essentially you could have What's called like features.

5:20 of the model within the layers which s uh express certain types of uh Thematics. So one of the one of the themes that the researchers was able to identify was Uh let's say Bullet point writing.

5:34 Another one was people and places. And one that really came up frequently that uh resonated was The Golden Gate Bridge. And so when you actually I essentially dialed up that feature.

5:46 Claude would obsess about the Golden Gate Bridge. So meaning in every one of its responses, it would come back and talk about the Golden Gate Bridge. So if you said like give me a recipe for making spaghetti. Uh it was saying. Hier is a recipi. And the orange color is just like international red that the golden bridge.

6:06 Golden Gate Bridge look like that. Um and so it was like really quirky and We We we very much wanted to in that situation just bring that user a bring bring it to the masses and bring it to people who are starting to use Claude. And

6:21 Uh so the entire uh experience actually we spun up on our website. within twenty four hours. And that took like engineering. Product design.

6:33 Uh are like research teams all working together. And We were really, really proud of it. I think it maybe reached only two thousand people, to be honest. Uh, but it it made us feel like, oh, we can actually bring new user experiences showcase our research In a way that's different.

6:50 And authentic to us. And in a very startupy like pace. that to me was like one of those like maybe hidden inflection points of we were starting to find our identity that we could build products, build experiences that were different from what our competitors had seen. What was already out there. I think that

7:11 Obviously labs, clogged. Et cetera, like we then started to identify ourselves as would we actually think the world Uh How to think about AI, how to bring that closer to the public. And but it was a very bottoms up culture.

7:26 And so that entire experience was very bottoms up. I see engineers, I see uh designers donating time to work on. Um and so I I like to always use that as example of like what the day early days were like. But the culture and and and the values have very much I think stayed the same since since those early days. This episode is brought to you by our season's presenting sponsor, Work OS.

7:51 What do OpenAI, Anthropic, Cursor, Vercelle, Replit, Sierra, Clay, and hundreds of other winning companies all have in common? They are all powered by WorkOS. If you're building a product for the enterprise, you felt the pain of integrating single sign-on, skim, Rback, audit logs, and other features required by large companies. WorkOS turns those deal blockers into drop-in APIs with a modern developer platform built specifically for B2B SaaS. Literally every startup that I'm an investor in that starts to expand up market ends up working with WorkOS. And that's because they are the best. Whether you are a seed stage startup trying to land your first enterprise customer,

8:30 Or unicorn expanding globally. Work OS is the fastest path to becoming enterprise ready and unblocking growth. It's essentially stripe for enterprise features. Visit WorkOS.com to get started or just hit up their Slack where they have actual engineers waiting to answer your questions. WorkOS allows you to build faster with delightful APIs, comprehensive docs, and a smooth developer experience. Go to workOS dot com to make your app Enterprise ready today.

8:56 What are some of the other um big inflection moments as you think about just anthropic going from just this like lab that's trying to compete with us. Juggernaut of open AI at that point. to what it is today. What are some moments that stick out of like wow, that really changed things. Definitely when we were training and

9:14 Uh Testing. Uh Opus three. I think that was the moment when the company I think we were less than two hundred people still at that point.

9:23 And It was very clear that we needed and wanted to create a frontier model. And a uh that was very important in terms of like our ability to reach like users. Consumers.

9:37 And uh to showcase our research. And We were looking for ways for also why sho somebody choose. Claude. And that was like a core question, and that was a core question we were getting asked in the early days.

9:53 And I think with Opus three. Yeah, it launched, I think, early March twenty twenty four, but there was many, many months of various teens across Inference across research, fine tuning, pre-training that rallied at different points. And towards a common goal. And

10:11 Uh I think everybody that was involved was like really proud. I remember I'd be in the PM us. uh the research leads myself we were all in our um this was from December. So we were all at home in our various uh um parents homes and seeing everybody's background of like their childhood room.

10:31 And everybody's working really hard. Uh, to figure out the like what are we training the model for? Is it showing up the right way? So I think that was really powerful in terms of just Building a lot of trust. And a lot of our research leads have actually uh from that time.

10:48 are now like leading Reinforcement learning, leading our character work and lineman work. So the that foundational trust, I think. Осоп. Now with any of our production models across product and research, because we were working just so much in the trenches together. In the early days.

11:08 And then I think there were things like Identifying that coding was important. Right. In twenty twenty three, when I started Um

11:16 Nobody said Anthropic and Claude and coding in the same sentence. I think competitor models like GPT four at the time was used a bit for coding, but it was One of many use cases.

11:29 And one thing that For example, I saw was People are starting to use code, uh, these models not just for code. Not just like code auto complete. but actually writing long form code.

11:42 And is that an opportunity for us to train you know, Opus three to be better at. And it ended up being a relatively smaller change from a training perspective. But it ended up helping us differentiate.

11:55 in the early days, uh competitively for users. And actually bring a lot of the Very early. Claude enthusiasts and developers. because we were uh providing a value that they didn't really think was possible at the time.

12:10 It's so interesting you talk about Opus three, like I that's so long ago and just like it's hard to think that was a big Inflection. And so this is really interesting to hear that that was internally a big milestone. It almost feels like this confidence you all built that wow we could really ship a frontier model. Which is now today so not great if you compare it to what we've got today. What I always think about is Opus four five, which was

12:31 And interestingly, like a year later, also during winter break when everyone was home. Able to code. Uh. Was that another big milestone? Yeah, um Opus four five was definitely another Large moment. I think

12:44 What was magical about magical about Opus four five is We also now Not just how to model. But a vehicle. Which is like a great product experience, like Claude Code.

12:56 Um one thing we say a lot on the team is You need frontier products. In order to have frontier models and for people to feel the magic. Frontier models.

13:07 And I think You know, we felt the magic of claw code. For very for for uh for many months before that. Uh but the fact that

13:17 The model. Есенті га лево в інтеліжен. Where at a very broad level. Users can experience

13:27 both frontier intelligence In new use cases. Allow it to run things end to end in an e genetic manner. I think that was the inflection. It was actually both. I I think Opus four five.

13:39 wouldn't have had that moment without a product like Claude Code. And claw code, I think. Wouldn't have had that type of adoption accelerated. without Opus four five.

13:50 So kind of speaking on on this On its thread. Uh, Dario, interestingly, if you look back at all his predictions, he's just like okay, coding's gonna be solved. It was a hundred percent in like a year. Something like that. He kept talking about how

14:03 We're gonna do c like AI's gonna do all our coding. I remember everyone Uh Being like there's no way. This is way too complicated. How is how is AI ever gonna get really good at this very complex thing that Humans do not this is gonna be humans for a long time. He was completely right.

14:17 Something else that he talks a lot about is this exponential that we're now we're now that we're on. That's the way he describes it now. We're like we're on the exponential curve. I remember not long ago where new models were being released. And everybody was like, Okay, we're done. There's no more upside. It's plateauing. It's over. There's no more room to grow. Uh And now it's like the opposite. Now we're inside like if you think about the curve of the exponential, we're like

14:38 Inside of the exponential now. Which by definition means every improvement is m a massive jump because we're like on that hockey stick part. What's it like just being on the inside of this crazy Historic Moment.

14:52 When AI is improving So fast. So much is being unlocked. Uh What is it like and how should people prepare for the coming

15:02 acceleration of more and more improvement from AI. One thing I like to say on the team is Most of us weren't like actively working yet. when the internet transition from this novelty to something that everyone can use.

15:17 And it feels Like that's Just taking humans. I I think analogies are helpful and so Like the analogy of that is

15:26 I think a couple of things. Um, number one is Adaptability. becomes very important. Um

15:34 I think We We have evals, we have, you know, on the safety side, safety testing, red teaming on the capabilities and product side. You prototypes. Products.

15:47 Like cloud code, tag and others. But it's very hard to predict the exact moment or the exact model. And so the adaptability of when you're faced with new information. How do you then make better decisions? versus keeping the same plan.

16:03 And so like that agility is really important. I think another piece is With that. How do you actually be Thinking very first principles.

16:13 And reason through What's next? What's a so what? How do we invest in new products? How do we invest in explaining? The differences to users.

16:23 So a lot of the Латарінцы I think of being in that exponential is that pace. Understanding how you

16:31 operate and make better decisions. And then applying that first principle's thinking to then Do something that maybe We pull up. A

16:42 plan that uh we would ex were expecting a few months from now, but now the model can actually Do uh and work on and actually bring that to user. So this is things like Co work. Skills.

16:55 Tag. Yeah, was d the it's a very positive self enforcing loop. And I I I think a big part of it also is just having

17:04 but like trust in each other, like making sure we have like We're we're thinking through the right decision making, we're bringing folks along. Some teams might see the exponential feel it faster than others. So how do we kind of have the grace to bring the organization, the growing organization and company along on that? So what I'm hearing here is

17:24 You almost don't know what will be possible with every model release. And so the important things to focus on is being adaptable. As Things emerge. Uh to your point, the product itself has to

17:36 Stay up to d has to catch up to what is possible. To your point again. Just like He can do so much, but people may not understand how to do it and may not be able to do it.

17:46 So the product making it easy and even just like telling you here's something you could do feels like an important part. Is that roughly what you're describing? I I think so. I think um There's some really interesting graphs in the original scaling law papers. And I think Folks are very familiar with the scaling loss in in the lens of

18:04 I'm as you add in more compute and data. What's called loss. Aka the loss from next token prediction. uh goes down. And so it's a very smooth linear curve with like the models get more intelligent as you scale them up. What's actually also interesting.

18:19 Ah in that paper is There are these like very A friend? emerging capability graphs. And so for example

18:28 I as you add in more data. And you train the models with more compute. You essentially see these Actually discontinuous. emergent capabilities jump.

18:38 So the models go from one plus one being a thing that it can't c calculate to a thing that it could reliably calculate. And so these emerging capabilities. This like some nature of like Predictability is is is

18:53 not necessarily everyone knows the exact moment. Like you need the e bells to be able to Аса газ акції ось би апар. of uh how this technology works. And also what makes like things like safety harder. Because unless you had the evals, unless you had the systems.

19:10 To test Um, these jumps might actually happen and y you don't know. Mm. That's so interesting that you may have developed this like AI brain that Uh can do something you're not even aware of.

19:22 And so part of the job is just uncovering, wow, we just got really good at this thing. What can we do with that? I think there's like product overhang and user overhang, like to to maybe put it in our um PM language. Even on today's models. And I think there's like a lot that

19:39 Uh we could be exploring on like our current opuses and definitely with like Fable, for example. And that that discovery is actually Another part of what's been in the early days of anthropics DNA.

19:54 And I think is also continuing to be a big part of How we operate in product. in labs and and across research. This makes me think about something Gary Tan's been talking about. Uh president of Y C. I don't know what his title is. Uh he's he had this interesting point that

20:11 If you're willing to spend a hundred thousand dollars a year right now on tokens. You are living the way somebody in twenty twenty eight. Is gonna live. Because by then it'll be really cheap. Everyone can work this way.

20:22 But if you there's this alpha opportunity right now to just Live in the future. Go crazy until it could spend. Uh and so there's a big opportunity for people to learn what the future is like and also just build Much faster.

20:33 Thoughts on this idea of And the value of Token maxing, let's call it. Yeah, I think I I take more of like a almost product lens. It's almost like token spin is more the input.

20:44 And really the output is what you described of Experimentation. And I think if we were orienting like goals around experimentation. I feel like that

20:54 that might be the better framing of the outcomes and therefore there might be different ways of achieving that outcome. I will say internally Some of the most creative thinkers the best like prototypers. do spend a lot of time with Claude.

21:09 with every new version of a research model that we have. And so there is something around You have to be Light. Using the models.

21:19 to then come up with Good. Than great. Them better ideas. And there's no substitutes for that.

21:26 Um it's very hard to come up with a perfect strategy without touching the technology when it's moving this quickly. At the same time. I think there's other things that we could be doing. Like so one thing that

21:40 We um do a lot is actually working in public. and internally within Enthropic. And so in the early days when we had less product surfaces. There was a Slack channel where

21:53 Everyone almost the entire company was testing early versions of Claude. And trying different use cases like People were not calling them use cases, but you might be asking it. to edit an essay or uh to come up with

22:08 the right way to send this email. Like they're all different use cases. But We all worked in public. And then what you would see magically. is

22:18 different users or different different folks on the team coming up with an idea. And then other people trying different variations of that idea. And then within maybe Ten or so requests. there was something magical or potentially in a use case that emerges.

22:36 And I think there's a lot in not just individuals. figuring out by themselves how to use this technology. I think We could be doing more. to actually bring like that communal discovery.

22:48 uh when we do experimentation like experimentation is not always necessarily A individual sport. It's so interesting. Yeah, this idea that we're just going to We're not sure what this is capable of or what we could do with it and

23:00 It takes all this poking around and People trying things, hearing what other people are trying to figure out what's possible. Such an interesting uh I don't know, technology. We're just like, Okay, here's what oh, I figured out I could do this thing. What are you gonna do with that? I think in a broad theme we know, right? We know that the models can write great essays or can write long form writing. But individual pain points of

23:21 What can you actually solve with that? And bring it to like a user level that people can use. Um I think is something that is more exploration or experimentation. Uh based. So following this thread, you uh

23:35 You ever see product for the labs team? Which uh is extremely cool. We've had Ben Man on the podcast, Mike Grieger, who both work on labs now. Talk about labs. What is labs? What's come out of labs?

23:48 Many people have heard of these things. And w how do they work that enables them to Create such innovative ideas. outside of even the core anthropic product team.

23:58 The thesis of Labs. In many ways is identifying and pulling the thread on the thread of discontinuous large bets.

24:07 that might not be in the core roadmap. And figuring out is there a there? And also What is a ten X, a hundred? A thousand X of the there.

24:21 And so for example, uh things like cloud code, um, I think I've heard of it. Uh things like clock code. uh things like uh skills and most recently cloud design.

24:36 MCP The thing that we really try to emphasize within the teams is Especially right now, there are so many things that could be built. What does it mean then to have a discontinuous spet? And

24:50 I think one approach that we're taking this year is It can be very strongly held opinion. About the theme or the area.

24:59 and then more weekly held about the exact prototype. And so like there is a culture of experimentation. Um, there's a lot of The bottoms up like Engineers on the team are very self enabled.

25:12 Um self driven to test out different ideas. And sometimes Uh we have a thesis and It might not work yet. And so we then might revisit it in one to two model generations.

25:25 And so this idea of like these prototypes that actually end up just helping us learn like that's also valuable. even if it doesn't lead to something immediately shipping. And so I think that allows the incubation and like the charter of labs to really accelerate and see around corners. More broadly for anthropic.

25:44 It's so funny to think about a labs within an anthropic, which was already so innovative and and creative and just, you know, shipping like crazy, that there's value to still creating a labs team within. Anthropic. What enables labs to work as well as it has, because you listed all these products. And it's in that's like what else has Anthropic ship? Feel like all the biggest wins almost. I'm sure there are many that I'm not thinking about right now. What's what's kind of core to creating a successful labs org within

26:11 Within a larger company. I think that team culture Like similar to broadly at anthropic, I think the team culture. It's Very valuable.

26:21 I think Ben sets in uh incredible vision and Pushes people to think about the ten x, a hundred ex. of the idea. And

26:32 You know, are the teams The pods within labs is small. Sometimes these ideas Start with one engineer. Right.

26:40 And I think uh sometimes when there's almost really large teams pursuing very ambiguous large ideas, you end up

26:50 I be Slow down. Because of that. Um so I think it's Culture, I think

26:57 Yeah, we actually Also select for folks who actually want to Do that zero to one experimentation. And it's not easy. There's a lot of bets that we end up turning down or Turning off.

27:11 Um and maybe you know, we revisit them in the future. Uh but that's hard. That's hard when you pour your heart and soul, you're acting as a founder for a bet and it's not working yet. Um so

27:23 I think it's like that type uh selecting for that type of personality, folks who are really passionate and deep about the zero to one. So you lead product for the research team. You work with the researchers at Anthropic. A lot of people kind of get an idea sense of what is research. What are research what researchers do. I think a lot of people don't totally understand. These very valuable people uh at all the AI labs.

27:43 Uh the way I think about it and I wanna under help people understand, help me understand just what are the researchers doing all day. What I imagine is they have a hypothesis for how to improve the model. They find data. They tweak some algorithms, they

27:56 Yeah. But just how it's trained. And they test it. See how did Keep iterating and keep trying to find ways to improve the model.

28:03 Is that roughly right? Slash. Help us understand what researchers are doing all day. That's really I I think that's a lot of Uh maybe the the like the more day to day.

28:14 I think one piece around uh researchers and like research organizations like at Anthropic Is there's also a vision of the future. Like More broadly. So

28:27 For example, things like I I I think even at the founding of the company researchers were talking about how do we get Claude to you? use a computer? How do we get AI to like navigate a screen?

28:39 Right. So there's a lot of actually Very founder like energy is how I describe it within researchers or really bold and ambitious researchers. Um, and we have a ton of those at at Anthropic. So there's one layer of Vision?

28:57 technology can go. And then I think on this other side of the loop. There's also now that this technology or Claude is in people's hands, how do we make it better? Today. So it's a medium and long term.

29:09 and a lot of energy thinking about that lens of the future. And also in the immediate and short term, what are the improvement areas we can make? And so like I think you're describing a really good sense of How do we make Iterative improvements on different versions of Claude.

29:26 The way that like my team works with researchers is kind of being very integrated and embedded in in those loops, particularly areas where there's a lot of impact. On Users. So this is things like

29:40 Vision Computer use coding, agente coding, tool use. Test Hum compute. Things where there's a direct user impact.

29:50 And then figuring out what are the ways to Uh bring the user feedback. And ground it. In a level that is understandable

30:02 for user uh for researchers. And also actionable. For researchers. And I think that's the second piece is actually a big part of the job in

30:12 Sometimes a hard part of the job. So For example We might get feedback on cloud.ai. Claude hallucinated.

30:21 It's very vague. If you bring that to a researcher and you say Please fix. Claude from being hallucinated. It's not very actionable. And so part of the time of the team is understanding

30:33 Okay, what's the trajectory of why that user gave that feedback? And it's like consented. And so we we we look at, okay. What should Claude have called tools? In that moment. Or from its current knowledge or It called the right looked at the right document, but it

30:50 Look at the wrong facts. In the first case That would have been a failure on tool use. On the second case. It would have been a failure on let's say search or uh knowledge and search and search synthesis.

31:04 Or it could be something around alignment. And so Bring that level of detail. to researchers. coming up with like is this a big enough problem? Figure out things like evals.

31:16 to then describe we've improved it. Like those are the levels of actionability. And it's uh day to day language of their researchers. And so we try to stay very close to how to bring that in an actionable manner.

31:30 uh between users to to the core model training and the research development. I was talking to someone the other day about how feels like research, AI research is uh the place to be now if you want to be very successful in life. What does it take to become a really successful researcher? From which you can tell.

31:49 Uh, you know, not everyone can get in not everyone's brain is gonna work this way, but just Say people are like, Hey, I wanna explore this career path. From what you've seen, what does it take to To make it there. Researchers generally are research and product managers working with research or both.

32:03 Let's do both. But uh the researchers like you know, PM's working researcher is also gonna be very successful. But it feels like everyone's trying to you know. Poach all the top researchers across every company. So just I I know you're not an AI researcher, but just what from what you've seen, just like what does it take to make it in that in that career path.

32:21 Yeah. I think a lot of the most successful researchers and research leadership at Anthropic. Are folks who Are really strong first principles thinkers.

32:32 About problems, like they reason through problems. really well. Um, who are just passionate about their research area. And have a Bold.

32:43 Description. Of what that could look like. And then who are actually close to the details. And so Uh

32:52 No are like leadership our chief scientists, our heads of like fine tuning and like RL. Folks are actually really close to their training runs. And actually look at things like how the training run is going. evals looking at the underlying data.

33:09 So like actually staying really close and Be excited to be in the details. I think have been like a sign of like really strong researchers and developing taste. And I think s like another piece is just

33:23 Like their ability to think big. Over time. And be like very ambitious, right? Like the Dario like we can transform software engineering. And and and the

33:33 And I think that's a good thing. uh going in that direction you learn so much. You get You had to shoot for the stars. in in in many ways across uh your ideas, I think in order to be a

33:46 Um a successful researcher. I I love just the s me of just be more ambitious comes up so often now. Which is so hard. Like it's it's easy to say that it's hard to actually just like how big can you think? And how that's so much of what AI now unlocks, just be more ambitious. Uh yeah.

34:02 Yeah. I think It's Thinking through it once or twice. And to end.

34:08 And then Being I think Stubborn about the Uh area.

34:14 And maybe more Uh loose around the exact like approach. Um It it is a question we challenge ourselves with.

34:25 Uh but The technology is moving so quickly. And so how do you make sure what you're building is actually Uh.

34:34 Forward compatible. And so it's also actually part of like I think the core product development loop to think bigger. Right. Um One thing I ask the team frequently or how I think about when we're building a product.

34:49 Is Let's say Claude eight comes around. What do you what changes in what users do. And then what should

34:57 what does that mean for how you're building today? Is it gonna be forward compatible to that experience? Right. So like just grounding it's I think um Being ambitious is very broad. And so trying to like ground it in

35:10 In some ways of describing Scribing that. And also, yeah, everything heading in a direction that all is cohesive and makes sense versus just Ambitious in a completely different direction. Speaking of ambition and Claudate. Um

35:22 Slash Mythos recently. Feels like hit this very new kind of Tipping point with models. Where used to be You have an awesome model.

35:32 Release it. Hey everyone welcome. Opus four five is out. Everyone can use it. Nythos went in a very different direction. We got Blocked. There was a lot of scrutiny, a lot of concern about what was capable of. Uh all the companies had to go make sure it wasn't gonna pack into all their systems.

35:47 And it feels like now every model. because they continue to get better will now have a lot more scrutiny and there will be more restrictions on who can use them, which feels like a big deal. How do you think about that? How does that change the way you operate?

36:02 I'm going to maybe leave the policy and the export control side to to folks that um on that and work on that. Um I think the product question and how we interact with these internally is

36:16 I think as you mentioned, as frontier models become more capable. The safeguards and the ways of red teaming and testing and The pri release процес. uh also needs to evolve and adapt quickly to to address that.

36:31 And so one example is You know, before Fable models, we ha didn't have a strong of let's say fallback UXs and systems. Because our Our our goal was to make sure that like

36:46 There is a symmetrical benefit. For this technology. and to minimize like the downside or like a severe risk of A fit. And so

36:56 We ended up building like fallback systems so that Users will still get a great response from Opus four point a immediately. And so I think there's a piece around

37:07 Uh as we evolve and like improve safety systems, how do we s continue to develop and deliver great user experiences? I think There's more that we can do on both sides, and so. You'll see us innovating, improving on what we called now the model safeguards package.

37:25 uh more and more in the coming coming weeks and months. What's really interesting and just like unexpected here. is creates this really interesting advantage for anthropic where you have access to the latest stuff. And this is gonna happen at every lab. Everyone's gonna And it's

37:40 it creates this unfair advantage within the labs to have access to the best stuff that other people can't yet Outside of your control. You prefer everyone use it. So it's a really interesting this new feedback loop that's gonna start where models that are so advanced are only accessible to certain companies. And that's gonna be a whole new unexpected it's like a second order effect of Of all these restrictions.

38:00 Uh to develop these systems and the models to be us. Inclusive as possible. Um I think Our goal is to not have that. happen uh for the general purpose, general use like technologies.

38:13 And to make him more accessible, I think you know, it this is like one of our top priorities right now to kind of reduce w what we're seeing there. Yeah. That makes sense. I would imagine you'd want as many customers if people using this thing as possible. This episode is brought to you by Mercury, radically different banking loved by over three hundred thousand entrepreneurs, and now with command. I've been a customer of Mercury's for over six years. I have never once thought about leaving. Mercury is basically what happens when banking is built by product people, not by bankers. They make it so easy.

38:45 Dare I say fun? To send invoices, move money around, set up virtual cards for folks on my team. Does your bank have an API, a terminal native CLI, or an AI ready MCP server? I don't think so. And just recently they launched Command, a conversational interface built directly into Mercury, which acts as your financial operator.

39:07 I've been using command to transfer money around to figure out what categories I've been spending the most money in, analyze my cash flows. And just today I used it to find out how much I've made from a specific sponsor over the past year. I just ask, how much have I made from X over the past year? Ten seconds later I have an answer. It is so freaking cool. Visit Mercury.com to learn more and apply online in minutes. Mercury is a FinTech company, not an FDIC insured bank. Banking services provided through Choice Financial Group and Column NA members FDIC.

39:38 I wanna talk a little bit about how the product role is changing and who Who is Doing well in this new world. Uh, now that AI is such a core part of

39:47 Uh of our Life. When you're Hiring. PMs.

39:52 product people when you're looking at people that do well in today's world. What are some things that you notice? What are you looking for more most? What are you looking for more? What's kinda like trending up in what you find is important and what's kind of trending down? We actually on my team have not changed our hiring loop. Uh for three years now.

40:11 Um So what we actually look for And the traits and how we evaluate Uh generalists like PNs, generalists like research product managers have actually been the same.

40:24 Um So I think some of those traits Number one Is first principles thinking. And this is really

40:34 Uh rather than pattern matching what you used to do. In let's say consumer product or B2B SaaS. Um but actually figuring out In this moment for this user group with this technology, what What is the user value?

40:50 Is there an example that's not true? A lot of people hear first principles thinking. They're like, Yes, I got it. I'm good at this. What is it? What's an example of someone having really demonstrated really good first principles thinking? I think one example is I think you think of a product manager as I own Product strategy.

41:06 And Delivering user value. A But I demonstrate day to day. By writing a PRD.

41:13 or try writing a product vision doc. And for for my team as like research product managers The way to drive user value. Is to figure out

41:25 the right user feedback. The evals. Right. That then can be A personification of that.

41:34 user need. So like We do write some product documents and PRDs, but we actually have a saying on the team of Evals are the new PRDs.

41:45 Right. Cause in order to deliver that user value. Uh it's not that exact artifact. that people used to write in the last like One to two decades.

41:54 It's a new way of working. And so the first thing principle thinking would be Let me figure out what is the thing. I should do. To achieve my goals.

42:03 Rather than here is a set of activities. that I've done and therefore I will continue to do. So the idea here is Used to be have kind of an idea, create a P R D

42:14 Talk to people about it, align on the plan. Design it, build it, ship it. See how it goes. Iterate. What I'm hearing here is it's like, Okay, here's some feedback. About something that's wrong or an opportunity.

42:25 Step one is the eval is now how you define what Yeah. Work is versus a PRD. Maybe maybe step one would be Uh

42:35 Understanding the user pain point. And so the way to even access a user pain point is different, right? In the past, we might do a user interview. And I think if you go like deep enough, you you might have the user walk you through their user flow, the pixels. Here You have to sweat the

42:53 Tokens as much as you sweat the pixels. And so one activity we have on the team is reading the transcripts. and understanding Uh, what was the trajectories that failed? very deeply to then say, was this?

43:09 Like a hallucination. What's this? clapping overconfident. So like the theme of the failure actually has a lot of nuance. And then that allows you to build.

43:21 A description A like sustained description of that. Pain point. Uh so That could be essentially in a new eval.

43:32 And is it eval on distribution? Right. Is it capturing both the positive situations where this is failing. And also Areas when it should actually not fail.

43:43 And then bring that back. to let's say research. So then we can make the improvements and actually measure the quality of Okay, when we have opus five point five. Is this area improving or not? Is Claude now able to

43:57 Uh identify the right places in the document. Uh And pull the right synthesis out. So it's just the actionability.

44:07 Like and and shorten the distance to actionability. Um for for our stakeholders and partner teams like researchers. um to take action on. Is there an example of something like this where you Found an issue or opportunity and then read the about.

44:22 And what is what is the eval looking like in in most cases? Uh what when people want to picture an eval, what is that? What is what do they picture? We actually uh pioneered this concept within anthropic So uh one of the early examples is the early cloud models. We're not very good at following specific

44:41 schema so like things like outputs in JSON. And Uh now. That is fundamental to Claude being able to be a good agent.

44:51 Right. If you can output a certain format, you don't know how to like access APIs, you can't call tools. Excetera. And so The initial uh And to end was

45:02 I was hearing feedback around Yeah, Claude Two Days. Claude was not very good at following instructions. So then digging in with users.

45:12 What do you mean? By Claude is not good at following instructions. Give me uh what situations this was happening, like what's the exact Like paragraph, what did you ask? What was Claude's response? going to like that level of detail.

45:25 And what It's all was Something like eighty percent of what people meant in the early days for this failure was Claude would not Write the right JSON.

45:35 And so then okay, let's generate maybe to start. Just Thirty to forty examples. of when Claude was not doing this thing correctly. And then that actually is your e ball set.

45:48 And you could have essentially Uh a prompt. And a response. And If that is not working

45:57 uh in the right golden answer that you might have, then that means that the the eval essentially uh it it's beneficial'cause it's identifying a pain point consistently. And so then we added that to our Um

46:11 repositories for evals and When we have uh versions of Claude, we actually run that eval and just check. I think at this point it's always 100%. Or like ninety nine point nine, and so it's no longer a pain point. Uh, but in the early days it was taking the user feedback, figuring out actually what they mean. Can we reproduce it? Is it consistent?

46:33 Is it a big issue? And then figuring out how to uh standardise it in a way that can be consumable. for researchers. It's basically test driven development for PMs is is the world we're having now.

46:46 Uh we write the test first. So is this just the core part of the product management job now at Anthropic Writing Ebels? I think so. I I also think It's um something uh I've talked to other pianza other companies about. And I think it's also more and more of the skill set more broadly. Cause a lot of the products that we're building.

47:07 Is at the intersection. Of models With harnesses. with a set of contacts.

47:14 For setup users. And so having things like evals actually is an is a way not just for uh folks working on models, but generally within product. uh to to get to better user experiences. Cause you can't improve what you can't measure.

47:30 And a lot of this is very still tactile based is still very judgment based. And so you have to stay close to the details. And also very non deterministic. Which is a big part of this, just like it's not gonna give you the same answer every time, so you get it.

47:44 Describe it kinda more broadly. It's not gonna be. Yeah, an exact match. So this is a really interesting change in the way the product happens. And will happen. is evals, writing evals.

47:54 versus PRDs is is a big part of this. Do you guys still do PRDs? Is there still like a one pager describing a problem or is it replaying Okay. You're shaking your head, yes. We we we we are. We do. I think Um, when there's a very defined problem, I think things like evals might be almost a shorthand. I think there's other cases where purity's are really valuable. Um PRDs are Great vehicles for

48:17 Getting a very large group. of people aligned on a Set of sources of truth about experience and set up goals. So when we do have a model, we actually for every model we do have a PRD.

48:30 Лесаріфор ресерчер, мор. Are Growing product surfaces. For our engineering teams. For our uh

48:41 stakeholders like uh legal and safety and others. As just a source of truth of putting together what we're aiming to achieve so that a big group of people can row in the same direction.

48:54 The other place where I do think PRDs are valuable are on the more ambiguous. problems and opportunities. Right. So we if we haven't shipped a thing. Like computer use.

49:05 We don't necessarily have a set of like user specific pain points always. And I think there's some value in the product vision portions of a PRD. To explore What could Even if a technology is not yet.

49:21 Ready to work for everyone. How do you get it to work well for some group? So you can explore the value. You can actually bring something that is

49:32 uh coherent to a user group. So we do have PRDs. Um I think the application's a little different now. Okay, this is great. There's I just had uh uh Andrew for he's the head of the codex app at OpenAI and he's you guys are aligned. Uh PRD is not dead, still very useful. For specific projects and ideas. Uh great. Okay, we've closed. Purity is still kicking.

49:55 Okay. So we've been talking a bit about just what kind of skills are kind of emerging for product People. Um Is there anything else that you find

50:04 Is shifted in what Patterns. Uh are common across people that are doing well in this new AI world in terms of product managers and folks on the product teams. There anything else that you're like, okay, the sum of you got it shift.

50:16 Or something you look for more. People. I think maybe specifically Uh. For

50:24 Folks might be mid career or folks who have been more in a managerial like product like leadership seat Um one thing that I think I feel pretty strongly about is

50:37 In order to be good. Managers of teams MPM's working with his technology. You have to be really hands on. Yourself.

50:46 And has spent not just time tinkering, but actually shipping with this technology and And and again being in the Details. And

50:57 Sweating the tokens along with your PMs and your engineers and your teams. And so Even for folks that I hire who have more tenured PM experience. The onboarding plans

51:12 Are exactly the same as somebody who is like more uh early career. And it's around Understanding users, reading Like

51:21 Consent and user feedback. Talking to customers. I think there's something around Uh being able to like understand what to do with this, what what good looks like.

51:34 And having developed that in a very Hands on matter that's important. Um It's not necessarily easy for someone to

51:44 Uh Agree or be able to see what a What good or great? AI product or AI feature could look like if they haven't kind of experience building.

51:56 themselves. Um so I think I think there is a I I I do feel pretty strongly that like you know, if you're a manager, you have to be hands on, you have to spend a portion of your time actually shipping.

52:09 Yeah you have to kind of walk in the shoes of your teams. Uh and and that's I I always try to carve out a portion of time uh to to actually like own one to two work streams when we have models. In order to keep like keep my theory of mind, keep my sense of how the models are moving, how quickly

52:29 It's improving. Uh Uh so I can help the team make make decisions and and make better decisions. What I'm hearing here is if you're not No matter where you are in the Ladder of hierarchy at a company.

52:40 If you're not building yourself, if you're not actually talking to Clive, talking to Codex, building stuff. You're not gonna make it. And you should have fun. working with his technology. I think that's the other piece. I think the folks that will be most successful, regardless of their level.

52:55 are people who love working with AI. And and are exploring and experimenting. And Carving out the time not just for the experimentation, but actually hands on shipping end to end, getting the user feedback.

53:10 I think has to be fundamental for everyone. I a hundred percent know what you mean there. Just like Me sitting on my newsletter and this podcast just talking about stuff and like, Yeah, yeah, that sounds great. Like every time I actually build something and I tinker with all kinds of little projects. You're just like, okay, I see what's happening here. And you just get so much more It's like hard to exactly describe what you' what you what you experience.

53:30 actually working with the models and building stuff, but it's like a whole different world. Like, okay, I see. Here's where the here's what they're talking about computer use. Here's what they're talking about with This limitation of this UX situation. Yeah. Yeah. So it's just like and I you made this really interesting point that you have to have fun with it. Which

53:46 Isn't that easy for a lot of people because they're pushed to use AI or they just don't know exactly what to do with it. For people that are just like, I don't know, it's just so annoying, I just have to do this. I don't know what So like I hate this friggin' thing. Why do I have to work with this? Things are changing so much. I'm tired. Uh advice for helping people find that find that joy.

54:05 in this work. I think maybe I'll reemphasize something I said earlier around just that Experimentation is not an individual sport. Like some of the moments where

54:15 I think I've touched it practically every version of research models across twenty vers versions of production claw at this point. And I think part of the joy comes from seeing other people discover use cases too.

54:31 And so Maybe one Idea here. would be pairing with somebody who is excited. And seeing what

54:40 And working together. Versus um Uh identifying or trying to figure out the perfect use case yourself because

54:51 That might feel like work. Working with others feels like joy. A lot of the time. And is there more that we could do to bring that bring other people along.

55:00 That's something like a lot of times internally, we have somebody who is like very curious and them sharing an idea of a new prototype actually brings a ton more people who are like, Oh, I didn't know this could Work now with Claude. So there's just some virtuous cycles here. Um and

55:18 waste of yeah, bring continuing to have joy with with this technology. That's such a good point. I think that's also why Twitter, so useful for a lot of this is you see other people Sharing what they've done. And it inspires you to come up with your own little ideas. And also it's just like fun to share your own thing that you've done.

55:36 So that's a really good point. Just like find other people to kind of play around with and Look for use cases. The thing I've also heard a lot is just find like a problem you want to solve in your life or work. And just Open up Cloud, Cloud Code, tell it here's what I want to do and

55:48 It's incredible how far you can get just with like A j vague idea of a problem you want to solve. Yeah. I think it gets hard in that there's so many different things that you could try. Yeah. And so you just like narrowing in. On either pairing with someone.

56:02 Working with somebody who who is who have a lot of joy about this technology. Or figuring out something that you could immediately find value. Like Either of things those things allow you to go deeper. rather than like more high level about too many things.

56:18 I'm I I find it hard to keep pace with the number of prototypes. or products that are out there. And so my lens has been, how do I go deep? In one to two of them.

56:29 Myself. That's uh that's so interesting you say that, because that's exactly it. We just had the survey. Uh that I I ran with uh my colleague Noam. Uh asking My readers just how they're feeling about all the things going on in the tech right now and AI.

56:43 And uh One of the most interesting takeaways we had was uh to find that happiness is exactly what you said is go deep in a couple of things. versus China just ton of little things. Find a couple of things to really salt well and then go deep in that.

56:58 Is a source because a lot of the happiness people feel is when they finally unlocked a way for AI to actually make their lives better. versus just like a couple of messed up, broken half working things. Yeah. It's it's um how do you go from this being a check the box? Right.

57:13 And so Like us as product people, it's then A exercise of product prioritization of your time. And your energy and and if the goal is to experiment with joy. Then how do you

57:26 What are the inputs that you need for that? Um But yeah, I I I think A lot of the Um

57:35 I think the secret sauce of anthropic is the culture and the bottoms of nature of how People work. And this like experimenting in public. Um and by doing that It's very much about

57:49 How to bring other people. Along Um that ends up being I think really valuable. Yeah, I've heard this so many times from all the labs. Just like

57:59 No no one's exactly sure how some of this is gonna be used and a lot of it is just putting stuff out early. Seeing how people use it, seeing what it's what's possible, and then using that information to build. That product to lead in. Yeah. I'm curious how to

58:12 Kind of on the thread of finding ways AI. For AI to help you in your work in life. Are there any interesting ways you've Been using Claude. lately in your work as a as a PM.

58:22 I think there's a lot of things with Uh you know, fable and things like tag. So there they're I I think tag is Um in in the very

58:33 like early days, I think there's something around how you work in a different paradigm of allowing this an agent to go off and work and then bring back uh product experiences to you. I think one area that It's not more recent, but one that

58:48 Um I bring up a lot with the team and I think We could do more on using AI is just like How to

58:57 Use it to also be more Uh To have better conversations with each other, to be better managers. I I don't think it's necessarily Uh

59:08 just about raising the IQ of Like experiences we build, but also I used it a lot and actually like prepping for How to have better conversations. um in the moment during like crucial conversations. So I love that book.

59:23 And so I actually have a skill that helps me. Figure out Am I having am I going in the right level of detail given the the situation at hand and actually helping me be a better manager and better supporter for the team. Um so for for like managers on the team. That's actually a thing that I've been sharing more with with a t uh with our managers of okay, how how do you actually use use Claude to

59:48 To to make you a better coach. 'Cause it's hard sometimes to find the right perfect words. And the model have a lot of perfect and bright words and U I think there is something about how how it can actually augment us from like an E Cube perspective in addition.

1:00:04 Oh man, there's so much interesting stuff there. So just to understand what you're doing there. So you you built a skill, you're just like Claude, build a skill. Pulling in uh lessons from Crucial Conversations, the book. Which it knows enough about you don't have to even give it the content. And then you use that skill to talk to Claude. Hey, have this very difficult conversation coming up with a colleague.

1:00:23 Give me some tips on how to approach it. Yeah. And it's it's a great I It's almost like

1:00:30 Uh Coaching? Like individualized, personalized coaching of just how to make you And and there's so much context switching that we do all day.

1:00:39 And having Like Claude. Help me pair. And help me. And maybe there are times where I end up not using suggestions from Claude.

1:00:48 Uh but it actually is Uh ends up being very helpful for for just coming up and brainstorming. Am I thinking about reactions in the right way.

1:00:58 How do I actually Uh go a bit deeper, faster, build trust faster. Uh, be more direct. Yeah. Man, I have so many questions here. This is so interesting. Uh one is just like there's concern people are gonna start talking the way AI writes. Because they're talking AI so much and it's gonna be like

1:01:15 Diana, it's not this, but it's that. Uh I know that you're not doing that, but that's all you know, a concern people have. Let me just ask about that, I guess. Do you fear this there's this, you know, brain rot uh atrophy stuff people talk about or just so reliant on AI now? And we stop. Learning and thinking and

1:01:32 Yeah, overall N AI thoughts. on that being so close to it and being so integrated with. With AI constantly. A lot of actually thinking process and writing process are tied together. For me personally.

1:01:43 And so I think There are ways where I use claw to augment my thinking. But what I wanna make sure, and maybe this is what you're describing is Claude doesn't take over all of my thinking for me.

1:01:57 And so I think depending on the situation, depending on how much more personal judgment I want to have in a situation. I my uh Um Come up with my own. P O V first.

1:02:10 And then work with Claude through that. Um And making sure that like I maintain my sense and tone. Throughout. I think there are then other things like updates, right? We have like monthly business reviews.

1:02:25 And then in those cases It's much more I want she want it to be standard. And I want it to be much more like it gets a crisp crystallized information in the right way. And maybe and I have a skill and like we're augmenting and improving our skill for that. But I want to get a to a place where like the monthly business review, the writing of that. Is

1:02:45 Potentially. Asymmetrically less Valuable than the thinking. And so how do I get that piece? delegate it to Claude fully.

1:02:55 And I'm more of a reviewer and a verifier of that information. So I think it depends on like what you're using Claude for and what you're trying to convey and like. Is there Is there uh asymmetrical value in And Delegating more to Claude.

1:03:11 What I'm also hearing the first tip is really great, which was Think first. Have a point of view. And then kind of use Cloud as a uh sparring partner almost to evolve the idea of

1:03:20 Pushback on the idea. Yeah. Yeah. And I think this is where things like actually our alignment research and safety research is helpful. Because it

1:03:30 What you don't want is like A AI that just agrees with you. Right. What you want is this technology to actually augment and grow and like get to a better outcome. And so sometimes it's

1:03:42 having Claude push back Makes me better. And so that's great. Like a co worker. I want somebody to push back when I my ideas are not fulfilled.

1:03:53 I wanna hear more about that. I've heard that when Bad Man was on the podcast he talked about The constitution that is built into Claude. And how Unintuitively. The work.

1:04:03 And the focus on safety and alignment, as you said, and this Constitution that describes How Clotch are thinking. Operate.

1:04:11 That actually you would think that would limit the abilities of Claude and make it less fun and interesting. It's exactly the opposite. Claude is the most interesting personality. I hear that constantly. It's just like I much prefer talking to a c open claw famously. Uh was built on Claude and then People were s forced to switch we won't get into it. We're forced to switch to JPT and they're like, This is so bad. This is not Yeah.

1:04:35 Uh So that is I think a really interesting point. I just want to make sure we spend a little time on. W why is it why is that the case? Just this focus on alignment safety. having this clear constitution. Why does that make Claude better and And more interesting to talk to also.

1:04:49 In order to make Claude as like im intelligent and as capable as possible. being able to have Claude actually push back in the right. Points? And

1:05:00 Then add. It's like a yes. Or no and Actually helps You

1:05:06 come to a better conclusion. So I've used cloud to help with things like Are we making the right pricing decision on the next version of Claude? I little by meta, but using a research version of Opus, asking it to figure out how it should price. And

1:05:22 Being able to come out with better outcomes is a goal at the end of the day. And so having AI not just be an assistant. Not just be

1:05:32 A Do we're not And being delegated. Task. By figuring out, is it doing the right thing?

1:05:39 That's actually very integrated. with knowing when to push back. Mm-hmm. Right. That's part of knowing when you should be proactive.

1:05:47 Proactivity is not a necessarily always. doing a thing that you are scheduled to do. It is knowing when to come up with a new idea. And so in order for Claude to be More useful. The general Approach has to be that it knows when to push back. It's a core part of

1:06:05 the characteristics together. Uh of the models. That is so interesting. It's so interesting that that is what A big part of it. Like it be it being Less compliant is almost what makes it better and more useful.

1:06:17 Because we need that. Like I've had so many people where they're like hey, like AI told me I was right. And like no, I wish Yeah, we should listen to other people. Yeah, and it comes back to our earlier point around thinking, right? How do you protect your thinking? Um if you have

1:06:33 A AI that can be a thinking partner. I think in partner doesn't just agree with you, it should add to you. And you should come away at the end of the day. having better ideas.

1:06:45 Because you worked with Claude. That should be the hero goal. not just making your ideas ten percent better. Yeah, I love this and like it used to be think ten X. I used to be the the way, you know, founders push people. Like what if we technically this? And I love

1:06:59 What I keep hearing is like it's like how do we go a thousand X from this idea. What is the Most ambitious version of this. I wanna come back to something that I ha I was thinking about as we were talking about Uh talking to Claude constantly. Um

1:07:11 It's very clear when AI has written something still. It's funny that it's a large language model. You would think of all things. It would be very good at writing. And interestingly, just no AI is very good at writing. It's always very Clear this was AI written. Do you think

1:07:27 We'll get to a place where We will not know. This was AI. I think it depends on What's the

1:07:35 uh goal that you're looking to achieve with by knowing or yeah. Uh what's the eval? Um I actually do think there's more that We could be doing on making Cloud Write better. There's actually very active

1:07:49 Efforts um on my team and on the research side about making Claude Wright better, just generally. I think it should be clear. Where An idea Is I don't is

1:08:02 Or bike. You Lenny or me, Diane? I think it really depends on

1:08:10 Uh, what's the goal of that writing? Like for something like a monthly business review. I would actually love to have that end to end be written. By cloth. Uh not make it feel like it was written by a human. It's such an interesting point you're making, like

1:08:26 Is it actually better for us to know? That it's AI versus not. Yeah. But but it's it's Um

1:08:33 But it's also for maybe The lens is more around like Verifiability or who's verifying. Mm. the output. Right. Right.

1:08:41 Like who's signing off? Uh, maybe less around who's writing, but who's verifying who's signing off. That becomes like more what matters. then who's writing it. Why why do you think AI is not? Great at writing.

1:08:56 Like my guess is It has studied all of the best writing in all of humanity. It's figured out here's the best way to write. And now that we And it's just there's only so many ways to to write. And so we've just recognized okay, this is what AI does.

1:09:11 It has these tropes. Is that the core of it? Is there something else that's keeping it from being a great writer? Ironically, being a large language model of all things you'd think could be really great at language. I think part of it is also Uh

1:09:24 We need to invest more in training improvements to make AI continuously strong on areas like writing. Um I think it's also You know, like the technology is jagged edged. Like like we mentioned. So sometimes when the models were good at writing but not a genetic. Our our thesis is.

1:09:44 How do we make the models more genetic or call the right tools? Now that that's improved a bit, then it's well, now it's these other areas actually become more Of the rough edges. And so I think we're in one of those moments worth writing where Uh.

1:09:59 We need to actually just focus and prioritize on training the models to be Like Great at this area and Like that is an active a very active area for us.

1:10:09 Oh that you mentioned. Okay, I'm glad. I'm glad and also Uh it's gonna be interesting once AI is so good. We're like, I don't know who wrote that. But um to your point, sometimes we actually wanna know that it's AI. That's really interesting. I never thought of it that way. The other interesting part of this is that there's that comedian who was joking that we're like on a plane and the Wi Fi's down.

1:10:27 And we're just like, What the hell? The Wi Fi's not working on this plane the sucks, how dare you When you're like Yeah, In a tube in the sky flying like a bird. And uh how dare you complain that the Wi Fi doesn't work.

1:10:39 Like your point is There's so much advancement. And so much power. Uh we can't fix it all. We can't make it all work. The best. Yeah.

1:10:47 possible and so Uh basically AI writing has been not the priority and feels like there's more investment happen there. Yeah, I think like tone and character is a priority. I think it's this advancement of The technology is a work in progress. And so

1:11:03 We made we we see a leap or emergence of like a jump. In agenic behaviors. And so That is a new normal. And then these other capabilities need to continue like improving. Yeah. And I think

1:11:17 once we improve let's say writing and like tone and character. Uh we probably will say like How do we have Claude be even more proactive? Like proactivity is an opportunity. And that's human nature. Like we want to make ourselves better, we want to make this technology better. Um so yeah, it I I think we're applying it to to AI, which is the right thing. We should be making it better.

1:11:41 I'm gonna ask you a couple questions I like to ask folks working at the very center of the future of That is coming. Um one is Where do you think human brains will continue to be Most valuable.

1:11:53 Over the years. I know Anthropic's mission and and vision is we'll reach a GI a uh superintelligence. So in the future Maybe nowhere. But before we get there.

1:12:04 Where do you think human brains will continue to be most valuable? As we approach that. The timeline. We started to talk about making Claude and models better at judgment.

1:12:15 Um Especially in the last Um you're so I think judgment is one And it's an area where

1:12:23 It's a cumulation of so much nuance. and so much experience. And these systems haven't experienced As much as humans have. And so I think that

1:12:34 Hard earned. Like judgment. is a A a area for for product leaders and just generally

1:12:42 Um continue to be really critical. There are so many things. AIs can build. Which one are the things? That

1:12:51 you know, an org like lab should build. Right. A lot of that requires like human judgment. Persistence. So proactivity, these are all traits are are beyond just general capabilities. But just behaviors and characteristics of like people.

1:13:05 at that level of like How do you get to the best solutions? How do you create the mo the best experiences? So I think those types of traits are actually The tactile. Uh traits. I think will be

1:13:18 uh continue to be important. Um I think there is also Uh Still a lot of like

1:13:27 capabilities and subject matter expertise as well, I think. Yeah. Software engineering has been really transformed by AI. I think there's areas like

1:13:37 Biology, life sciences, these are all things that Um We're just kind of at like The foot of the exponential on.

1:13:46 Like maybe software engineering, we're on the exponential on some of these area, other areas, we're not quite there yet. And so um, I think you're seeing us. ship things like lots of science investing in these areas because those are areas that Um, I think it's just bring the this technology to

1:14:05 society and having a positive benefit for society. So I think there's a lot more to go there. Another question I want to ask is Um and someone with kids How do you think about what you You're encouraging them.

1:14:19 To Learn What do you think you're gonna Nudge them. to be successful in this

1:14:25 Wild new world that we're entering. I actually think it's a lot of the same traits. Like you and I probably grew up with which is Curiosity for learning.

1:14:35 Persistence believing in your own inner voice, developing and then believing in your own inner voice. Like I have a four year old, I have a eight year old. It's on us to help. Uh it's on me to help them.

1:14:49 Develop their inner voice. And whether that's being opinionated and Taking a stance. To me.

1:14:58 Right, and developing that, encouraging that. Uh, I think that those types of skill sets are things that Um is important in the future. And I like.

1:15:07 having their own individual voice. That is so interesting. It's so in related to the answer you had when asked about how to avoid a brain rod essentially and over relying on AI, which is just a keep focused on your own point of view and your own perspective before you over relying AI and just this idea you're describing of

1:15:25 Building that in kids is is really important. Uh that is so interesting. And I love how this All this kind of connects judgment persistence, you know. A s a point of view of your own. Yeah.

1:15:35 Both for kids and also adults. Yeah, anything we think of L um for your Well, like the question I'm thinking about is just when to get them on, like some AI thing, you know, when I I have a three year old, so it's pretty irritable for that. But you know, how do you get how do you onboard them to this? crazy thing. I had I was at an event recently and

1:15:52 bunch of parents were talking about how they think about AI in their kids and One person had a really interesting approach, which is Uh Keep them on the very early models. So that they still have to struggle a bit and not get all the answers immediately?

1:16:05 Mm. Like an open source local model. Not table. Yeah. Yeah. Yeah.

1:16:12 And curiosity is something uh I always I keep mentioning Ben Mann, but his answer actually to this question has always stuck with me, which is Um Curiosity and also just like he's a big fan of Montessori, which is what I'm We're encouraging for our kids, so there's something there.

1:16:27 Maybe a last question, just along kinda along these lines. Something Fiona Fong actually suggested ask you. Uh who's recently on the podcast. How do you stay just recharged and not burn out, being in the center of this crazy storm of AI.

1:16:39 A as a mom. Uh working in, you know. Uh we're seeing the research work at anthropic. Uh I just like we're living through the most unprecedented time working at

1:16:50 Just like Being you know, being on the outside Anthropic is crazy. I don't even know what it's like to be on the inside. Um what have you learned about Avoiding burnout. Staying recharged, staying sane.

1:16:59 During the middle all this. In twenty twenty four, we shipped. Four models. for the in the whole year, or for series of models. And I think we did more than that volume in just Q two of this year.

1:17:12 I think I've been really lucky. With uh the team that we grown and built. both the stakeholders on the research side and

1:17:23 within our research. product management team. Um I think that Magical parts.

1:17:31 about approaching all of those is that it's not an individual sport. Um there's like a sense of radical Ownership and team. Collaboration.

1:17:43 That I think Sometimes it does feel like a por high performance sport. 'Cause you're in very critical decisions, there's new information about users. About training.

1:17:55 And you have to make recommendations and judgments and decisions very quickly. And nobody can do that sustainably by themselves. Um and so I think what's really helped is having a

1:18:09 Team that is incredible. who Looks out for each other. Who You know, the night before a launch, even if they're not the core DRI on that model will stay up and help the DRI who

1:18:22 Uh to review the blog post. And make edits and come up with better demos and Knowing To be each other's sort of extra hand. I think it's very easy if you take all of this change on your own shoulders.

1:18:35 To feel like you're alone. And to feel like you have to do everything. Um but I think one of the like magical parts of anthropic is this ability for us to

1:18:47 uh figure out what are those opportunities to help each other and actually then taking the next smile of like Mind melding. We call it like Entering the hive mind, there was a article about this. And I think like part of that is just that allows like the team to replenish. It's not that you

1:19:03 I I was just on PTO in June. It's not just that you can take PTO and you come back to like three X the amount of things to do. is actually that you could p take PTO and know the team can Figure out the right things to do. and that we individually can like watch out for each other.

1:19:21 Um so I think that's a big part. I'm really lucky just personally. Um, also my partner is really supportive. Um, this is year six of me working in AI, so. Amazon and then Anthropic. And so he sees how much I just love the technology and what this can do. And that really helps, I think, also um from like a personal perspective as well.

1:19:44 I love I love how many of these answers connect. So What I'm hearing here is just the having other peop working with other people. Relying on other people, helping each other out when things get crazy.

1:19:55 Uh It which is a similar answer you had for just how to How to find the joy in in And fun in this work. Just

1:20:02 Gets and be inspired by other people, see what they're doing. Work together. Yeah. And it's interesting, when Fiona was on the podcast recently, she I was asking her just like what's changed in the world of software engineering and she pointed out. It's a lot lonelier now.

1:20:15 Because now we're working with agents instead of other humans. Teams are smaller. People are having all these fleets they're talking to constantly. And so this is just a reminder of just the power of Just actual other humans around you. We're asked to

1:20:27 work and make decisions on really big things because you have more scale from the technology, right? And I think Having individuals having other folks more who can have some level of like mind meld with what you work on, how you approach maybe not exactly every detail, but what are the first principles, what are the assumptions you make.

1:20:51 then helps them Uh You know, back up for you. Or uh push your decision and sharpen your thinking. Um so I think

1:21:00 You know, we really try to like I really try to look for that when like Building the team, growing the team, hiring. Like, is this person going to care about their own ego? and building out a big org, or are they gonna care about contributing to anthropic and contributing to the like impact of the team.

1:21:17 and orienting f towards folks were like low ego. Team oriented. Um, I think that's Yeah. It's a big part of I think the sustainability.

1:21:30 Yeah, just always a lot of it's always just comes down back to culture. And hiring and And I know I've heard a lot just the reason Anthropic is able to move so fast. I remember that Moment when like something shipped every day of the month. There's like a calendar of launches. And people were talking about how is this possible?

1:21:46 And what I heard a lot is just because everyone is so aligned around The mission. And the values it allow people to make decisions really quickly. Before we get to our very exciting lightning round. Is there anything else, Dan, that you wanted to share? Anything else you wanted to touch on? Anything you want to

1:22:01 Maybe double down on of things we've talked about. This was actually really fun'cause I feel like your questions actually sharpened some of my thinking around how the thoughts kinda connect. I'm your real human clot over here. One thing that I really uh Come or um have

1:22:19 People take away is I think one in the ways of working. But also just two that like This is a This is a lot of like growth and change and

1:22:32 having the joy in using this technology and like If you're feeling like in this moment you don't have As much of that feeling of initial joy, how do you find people who do?

1:22:44 uh if this is an area that that you're excited and like want to work on. And I think developing skill sets, replenishing skill sets in many ways of things like thinking from a first principles manner about what you solve. I think fundamentally

1:23:00 You didn't ask me this, but there's this question in the community of Do we still need PMs? When the models are so capable when engineers are leaning in. Um

1:23:10 I think that Role of People who are user centric.

1:23:15 who go into the details of understanding what users are trying to accomplish. in an actionable manner. And doing the relittinous

1:23:26 work to do that. Like That to me is a core of a product person. And I actually think we need more of that. I think we are becoming very

1:23:37 technology layered. Driven. And actually to make that impactful, it's You have to go deep, you have to be curious, you have to be super hands on. And those are things that I think

1:23:48 are also trades that have I think helped anthropic. from a product development and model development perspective. And as part of the culture and Hopefully that's valuable for others as well. Amazing. What an inspiring way to

1:24:01 And Oh man. Yeah, and this is I've been saying this too for a long time, just now that building is easy. the hard part becomes, as you said, what should we build and is the thing we have built correct and good and worth leaning into? And to me that's what PMs do and what PMs are good at. Yeah.

1:24:18 Yeah. Yeah. And it's getting into the details of the user. Mm-hmm. Yeah. Empathy.

1:24:24 Okay, great. PMs are gonna make it. Okay. PRD is not dead. All kinds of all kinds of uh important lessons here. Uh, Dan, with that we've reached our very exciting lightning round. I've got five questions for you. Are you right? First question, what are two or three books that you find yourself recommending most to other people? One personal one

1:24:45 I really like How to raise an adult? So I I'm a mom.

1:24:54 I think a lot about What is a things that I want to instill in in in my kids. And that book is really helpful for describing, we're not trying to raise children, we're trying to raise adults. So just the framing of what does that mean. And what does it mean? What are the characteristics that we want to hone and like harness and foster in our kids?

1:25:13 Um, the other book that I uh was listening to on Audible recently is Incorporable. By Eric Reese. So the the corruptible incorruptible yes, yes. Yeah. This recent podcast cast.

1:25:26 Um Yeah, and I I I just I think The question of how to build great companies is important.

1:25:34 I personally just been most fascinated with how to keep great teams and great companies going further. And it was very interesting to just kind of see his framing and reframing of the question. Um I loved some of the examples around having metrics around culture. You if you can if you only measure revenue.

1:25:54 And then that's kind of how your goaling against, but if you have other better metrics, that's actually the way uh to to to sustain the the values you care about I've been kinda trying to think about how to actually bring that to the team level of like how do we better articulate right our norms, a lot of the things we talked about on the team. So I think that's also a really good read.

1:26:15 There you go. Uh that'll be your next watch, everyone, as you're listening to this the Eric Greece episode. Uh such a good episode. Yeah. Uh and his book just came out, Incruptible. Thanks. And I think it was like a n uh New York Times bestseller. Like it's actually doing incredibly well, which I was really happy to see.

1:26:30 Yeah, exactly. Next question, favorite recent movie or TV show. You really enjoy it. Most people at Anthropic do not have time to do what to watch things, but I'm curious if you have an answer. I would say um drink Uh some time off last month I did Get to g like binge watch Fallout.

1:26:48 On Amazon Prime. So that was actually uh I kind of like Um it's kind of Uh have you heard of it? Yeah, yeah, it's based on the video game.

1:26:58 Yes, it's based on the video game. Uh I think it's a uh it was really um It's witty, it's humorous, it's also like super action oriented. So highly recommend. Okay, next question. Do you have a favorite product you've recently discovered that you really love? I really do think like Claude Tag is very interesting in terms of a product experience. Um, we actually have like different versions of this.

1:27:21 uh within Anthropic and I I I think it's actually been really uh really, really uh powerful tool. Yeah, it feels like I think some people are like what's the big deal? The fact that everyone at Anthropic is like raving about it. Tells me something important is going on here. And I'm trying to actually get it working within my Slack community that I have for paid newsletter subscribers. How cool would that be?

1:27:43 Yeah. We're trying to figure out how it works when it's not a company when it's just a bunch of People that don't know each other and how that might work. But we're trying it out. Okay. Uh two more questions. Your favorite life motto that you find yourself Often coming back to in work or in light. So I was actually raised by my grandparents.

1:28:00 uh for the first ten ten years of my life and my parents were immigrant uh college and master students in the US. And Um My grandfather always says No matter how far you go, there's always another level.

1:28:17 Which uh um is I think um A really good way, though like a pretty uh intense way of describing uh his his life or philosophy. But I go back to that whenever there's something new or unprecedented that we Experience.

1:28:35 And I think, you know, first half of this year there was definitely a lot of That like There was a lot of new things that we were learning, I was learning. Um, so just feeling like There's always like another mountain, another

1:28:48 Uh. Not good enough, Dan. We need to go better. We need to go bigger. Uh. Makes me think about actually another Ben Mann line from his podcast episode. That this is the most normal it's ever gonna be. It's only gonna get weirder and crazier.

1:29:04 Okay, final question. Uh with Pokenair and your LinkedIn. You were a high yield bond trader, JP Morgan Chase. early in your career.

1:29:14 Uh You had like Uh you have this like redacted uh a hundred million dollar trading portfolio of some kind. Uh what did you learn from that?

1:29:25 Stuck with you. And or is there a crazy story from that period? It was four years of your life. I think I learned actually a lot that I uh apply here. Uh at Anthropic and

1:29:37 Other uh jobs thereafter. Um so when I was at JP Morgan Um The trading floor uh you you could kind of envision like sort of wallful washry, that's very different.

1:29:49 Uh most traders I think are in front of a terminal. They're uh much more doing analyses uh on their computers. Um, but it's still very I would say like male dominated. And so Uh

1:30:02 I was the only woman. I was the only a um person with like my background uh on the trading desk. And I learned that Um It was a very good environment to kind of building one my sense of authentic self.

1:30:21 And two. Uh That Even if I was the most junior person. Even if I may look

1:30:29 Different. Uh They're the best ideas. And having conviction in the best ideas. Uh

1:30:39 In regardless of all of those other factors. Like It's the most important thing. And so I think Just bring that sense of Uh

1:30:47 How I show up. More at work. I'm I'm pretty vulnerable and authentic. With my team. Uh I try to really make sure that regardless of people's levels or tenures, if they have a great idea.

1:30:59 How to help them pursue that. And To do all still the same. Um so to like put the idea out there. to actually

1:31:08 Um have conviction in it to do the follow through to do the like nitty gritty work to make it happen. Um, so those were all things that I learned from trading. Um And yeah, I think applies to any any job in many ways. And it's beautiful. Where can people find you online if they wanna follow you?

1:31:26 And how can listeners be useful to you? I don't have a large presence on like Uh social uh I think the best way to uh find My work.

1:31:37 Uh my team's work is really uh the anthropic blog. And when we're publishing new models, new product experiences. I think in terms of uh useful uh for me. I think the best thing number one is our feedback. Like

1:31:55 we actually if if you thumbs up or thumbs down on any of our product surfaces, if you contact your salesperson with feedback about the model, it will make its way to me. Uh, we actually with every like research model, I actually get pretty close into understanding favorability and feedback. Um so giving us that feedback, pushing Claude, telling us where it's falling down. Um, those help us make Claude Batter.

1:32:21 Uh the other the other thing is like if you have folks in your network who seem like this type of profile person that I just talked about. I'm hiring the team is growing. We really will love just people who love this technology. W are deeply curious first principles thinkers.

1:32:39 who are fearless in questioning assumptions. Um and who have like a tinkering hackery spirit. Wow, what a dream job. So basically open open PM roles at Anthropic on the research team. Yes.

1:32:53 And they apply, I assume on the website, the careers page. Holy moly. All right, here we go. Enjoy the flood of resumes you're about to receive. Thank you, Lenny. Uh, Dan, thank you so much for being here. Thank you so much for having me. Thank you for um

1:33:09 Really helpful, thought provoking questions. I'm helping leave and connect the dots on how how we work. how how this whole technology is coming together and being product people in it. I really appreciate that.

1:33:22 But thank you, Dan, for real. Okay. Well bye everyone. Thank you so much for listening. If you found this valuable, you can subscribe to the show on Apple Podcasts, Spotify, or your favorite podcast app. Also, please consider giving us a rating or leaving a review, as that really helps other listeners find the podcast. You can find all past episodes or learn more about the show.

1:33:44 at Lenny's podcast dot com. See you in the next episode.

What you see above is a preview of the first minutes. One unlock costs 10 credits and covers this episode forever: full segment and word-level timestamps on this page, plus .txt, .srt, .vtt and word-level JSON downloads, as many times as you like.