Anthropic’s first technical PM on token maxing, the jagged edge, and living in the future
In a Nutshell
Diane Penn, Anthropic's first technical PM, explains that evals have replaced PRDs as the core product artifact because user value emerges from token-level experimentation rather than traditional roadmapping. The company operates in exponential improvement cycles where frontier models and products like Claude Code co-evolve—Opus 3's coding focus and Opus 4.5's agentic capabilities only became transformative through simultaneous product vehicles. Success requires bottoms-up experimentation, first-principles thinking, and maintaining human judgment while using AI to push back rather than merely agree, with teams that treat model releases as collaborative high-performance efforts rather than individual work.
These notes were generated by AI and may contain inaccuracies.
In 2023 when Diane Penn started at Anthropic, nobody said Anthropic and Claude and coding in the same sentence. People felt OpenAI was so far ahead that Anthropic had no chance. At the time, Penn saw people starting to use models not just for code autocomplete, but actually writing long form code, and saw an opportunity to train Opus 3 to be better at it. That was the inflection point.
What was magical about Opus 45 is that Anthropic now not just had a model but a vehicle - a great product experience like Claude Code. Opus 45 wouldn't have had that moment without a product like Claude Code, and Claude Code wouldn't have had that type of adoption accelerated without Opus 45.
The way to drive user value is to figure out the right user feedback. The team has a saying that evals are the new PRDs. If you are willing to spend $100,000 a year right now in tokens, you are living the way somebody in 2028 is going to 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 and great and better ideas, and there's no substitute for that. People need to be more ambitious with AI tools because they're just capable of so much.
One thing asked of 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?
Diane Penn joined Anthropic as the first technical product manager over three years ago, when the product team was just five engineers. She helped ship every model from Claude 2 through Fable, and helped incubate and launch Claude Code, MCP, skills, Claude Design, and core capabilities like computer use, tool use, and reasoning.
When Penn joined in 2023, there were five product engineers and one engineer for the entirety of the API business. The culture was really strong and people walked the walk of the mission, culture, and values. The energy was very much like a startup. They were trying to find their identity in the early years - there was the technology, but also the question of how that technology brings value to users and society.
One of the moments where Anthropic started to get into their groove was shipping Golden Gate Claude. This was published for about 24 hours. They had just published early interpretability research in early 2024. One of the examples was features of the model within the layers that express certain types of thematics. One theme identified was bullet point writing, another was people and places, and one that came up frequently was the Golden Gate Bridge.
When that feature was dialed up, Claude would obsess about the Golden Gate Bridge. In every response, it would come back and talk about the Golden Gate Bridge. If asked for a recipe for making spaghetti, it would say the orange color is just like international red that the Golden Gate Bridge looked like. The entire experience was spun up on the Claude.ai website within 24 hours, involving engineering, product, design, and research teams working together.
It reached only about 2,000 people but made them feel they could bring new user experiences and showcase research in a way that was different and authentic to them, at a very startup-like pace. It was one of the hidden inflection points where they were starting to find their identity that they could build products and experiences different from competitors.
The culture was very bottoms-up. Engineers and designers donated time to work on projects. This example shows what the early days were like, and the culture and values have stayed the same since those early days.
When training and testing Opus 3, the company was less than 200 people. It was very clear they needed and wanted to create a frontier model, which was important for reaching users and consumers and showcasing research. The question was why should somebody choose Claude.
Opus 3 launched in early March 2024, but there were many months of various teams across inference, research, fine-tuning, and pre-training that rallied toward a common goal. In December, the PM and research leads were working from their parents' homes during winter break, seeing everyone's childhood room backgrounds, working hard to figure out what they were training the model for and if it was showing up the right way.
This was powerful for building trust, and many research leads from that time are now leading reinforcement learning, character work, and alignment work. That foundational trust helped them work well with production models across product and research because they were working so much in the trenches together in the early days.
Identifying that coding was important was another inflection. In 2023, nobody said Anthropic and Claude and coding in the same sentence. Competitor models like GPT-4 were used a bit for coding but it was one of many use cases. Penn saw people starting to use models not just for code autocomplete but actually writing long form code, and saw an opportunity to train Opus 3 to be better at it. It ended up being a relatively smaller change from a training perspective but helped differentiate competitively for users and brought early Claude enthusiasts and developers because they were providing value they didn't think was possible at the time.
Opus 45 was another large moment a year later during winter break when everyone was home able to code. What was magical about Opus 45 is they now had not just a model but a vehicle - a great product experience like Claude Code. The saying is you need frontier products in order to have frontier models and for people to feel the magic of frontier models. They felt the magic of Claude Code for many months before that, but the model got to a level of intelligence where at a broad level users could experience frontier intelligence in new use cases and allow it to run things end-to-end in an agent manner.
It was both - Opus 45 wouldn't have had that moment without a product like Claude Code, and Claude Code wouldn't have had that type of adoption accelerated without Opus 45.
Dario Amodei kept talking about how AI is going to do all code, and everyone was saying there's no way this is way too complicated. He was completely right. Now they're inside the exponential curve, which by definition means every improvement is a massive jump because they're on the hockey stick part.
Most people weren't actively working when the internet transitioned from novelty to something everyone can use. Analogies are helpful. Adaptability becomes very important. They have evals, safety testing, red teaming on the safety side, and on the capabilities and product side they have new prototypes and products like Claude Code. It's very hard to predict the exact moment or exact model, so the adaptability of when faced with new information and how to make better decisions versus keeping the same plan is important.
Agility is really important. Another piece is thinking very first principles and reasoning through what's next, the so what, how to invest in new products, how to invest in explaining the differences to users. The experiences of being in that exponential is that pace, understanding how you operate and make better decisions, and applying first principles thinking to do something that was expected a few months from now but now the model can actually do and bring that to users. This includes things like co-work, skills, and tag. It's a very positive self-reinforcing loop.
A big part is having trust in each other, making sure they're thinking through the right decision making and bringing folks along. Some teams might see the exponential and feel it faster than others, so there's grace needed to bring the organization along.
You almost don't know what will be possible with every model release. The important things to focus on is being adaptable as things emerge. The product itself has to catch up to what is possible. It can do so much, but people may not understand how to do it and may not be able to do it, so the product making it easy and telling you here's something you could do feels like an important part.
There are graphs in the original scaling law papers showing as you add more compute and data, loss from next token prediction goes down in a smooth linear curve. What's also interesting is there are different emerging capability graphs. As you add more data and train with more compute, you see discontinuous emerging capabilities jump. Models go from 1+1 being a thing it can't calculate to a thing it can reliably calculate. These emerging capabilities have predictability that's not necessarily known in the exact moment, and you need the evals to assess that. This has always been part of how this technology works and also what makes safety harder, because unless you have the eval and systems to test, these jumps might happen and you don't know.
You may have developed an AI brain that can do something you're not even aware of, and part of the job is uncovering what you just got really good at and what you can do with that. There's product overhang and user overhang. Even on today's models, there's a lot that could be exploring on current Opus models and definitely with Fable. That discovery is another part of what's been in Anthropic's early days DNA and continues to be a big part of how they operate in product, labs, and across research.
Gary Tan's point is that if you're willing to spend $100,000 a year right now in tokens, you are living the way somebody in 2028 is going to live because by then it'll be really cheap and everyone can work this way. There's alpha opportunity right now to just live in the future, go crazy on token spend. There's a big opportunity for people to learn what the future's like and build much faster.
Taking a product lens, token spend is more the input and the output is experimentation. Orienting goals around experimentation might be the better framing of outcomes. Internally, some of the most creative thinkers and best prototypers do spend a lot of time with Claude with every new version of a research model. You have to be using the models to come up with good then great then better ideas and there's no substitute for that. It's very hard to come up with a perfect strategy without touching the technology when it's moving this quickly.
They do a lot of working in public internally within Anthropic. In the early days when they had less product surfaces, there was a Slack channel where almost the entire company was testing early versions of Claude and trying different use cases. People were asking it to edit essays or come up with the right way to send emails. They all worked in public, and what emerged was different folks coming up with an idea and then other people trying different variations, and within maybe 10 requests there was something magical or a new use case that emerges. There's a lot in not just individuals figuring out by themselves how to use this technology, but bringing communal discovery when doing experimentation. Experimentation is not always necessarily an individual sport.
At a broad theme, they know models can write great essays or do long form writing, but individual pain points of what can actually be solved with that and brought to a user level that people can use is more exploration or experimentation based.
The labs team oversees product for identifying and pulling the thread on discontinuous large bets that might not be in the core roadmap and figuring out if there's a there there, and also what is the 10x, 100x, 1000x of the there there. Things like Claude Code, skills, Claude Design, and MCP came out of labs.
What they emphasize within teams is that right now there are so many things that could be built, so what does it mean to have a discontinuous bet. One approach this year is being very strongly held opinion about the theme or area and then more weakly held about the exact prototype. There's a culture of experimentation. Engineers on the team are very self-driven to test out different ideas, and sometimes they have a thesis and it might not work yet, so they revisit it in one to two model generations. Prototypes that help them learn are valuable even if they don't lead to something immediately shipping. This allows the incubation and charter of labs to accelerate and see around corners more broadly for Anthropic.
It's funny to think about a labs within Anthropic which is already so innovative and creative and shipping like crazy that there's value to still creating a labs team within Anthropic. The team culture is very valuable, similar to broadly at Anthropic. Ben sets an incredible vision and pushes people to think about the 10x, 100x of the idea. The pods within labs are small. Sometimes these ideas start with one engineer. When there are really large teams pursuing very ambiguous large ideas, you end up actually being slowed down because of that. It's culture. They select for folks who want to do zero to one experimentation, and it's not easy. There's a lot of bets that end up turning down or turning off.
Researchers at Anthropic are selected for being passionate and deeply committed to zero-to-one innovation. This type of personality finds it difficult to abandon projects they've invested significant emotional energy into, even when the bet isn't working yet.
Beyond the day-to-day activities of forming hypotheses, finding data, tweaking algorithms, adjusting training, and testing improvements, researchers maintain a broader vision of the future. At Anthropic's founding, researchers were already discussing how to get Claude to use a computer and navigate screens. Research organizations combine founder-like energy with ambitious visions of what technology can become.
Research work operates across two time horizons simultaneously. One side focuses on the medium and long-term vision of what the technology can become, while the other side concentrates on immediate and short-term improvements to make Claude better today. The product team's role involves being embedded in these loops, particularly in areas with direct user impact.
Key areas where product teams work closely with researchers include vision, computer use, coding, agent coding, tool use, and test-time compute. The goal is to bring user feedback and ground it in terms that are understandable and actionable for researchers.
User feedback like "Claude hallucinated" is too vague to be actionable for researchers. The team must trace the trajectory of why feedback was given, examining whether the model should have called tools, whether it looked at the right document but extracted wrong facts, or whether issues relate to search, knowledge synthesis, or alignment problems.
The team works to bring user feedback into the core model training and research development loop in an actionable manner. This involves determining whether problems are significant enough to warrant dedicated evals that describe improvements.
Successful researchers and research leadership at Anthropic demonstrate strong first-principles thinking about problems. They reason through problems effectively, maintain passion for their research area, hold bold visions of what could be achieved, and stay close to the details. Leadership including chief scientists and heads of fine-tuning and RL remain actively involved in training runs, examining eval results and underlying data.
Successful researchers combine the ability to think big with stubbornness about their research area while remaining flexible about exact approaches. A key challenge involves ensuring what is being built remains forward compatible as technology moves quickly. One approach involves asking what changes when Claude 8 arrives and what that means for building today.
As frontier models become more capable, safeguards, red teaming, and pre-release processes must evolve rapidly. Before more advanced models, fallback UX systems weren't as strong because the goal was ensuring asymmetrical benefit while minimizing downside risks. New fallback systems were built so users still receive great responses even from more restricted models.
Advanced models create feedback loops where the most capable systems are only accessible internally to labs. Anthropic's goal is to develop systems to be as inclusive as possible and reduce restrictions on general-purpose technologies, making them more accessible to users.
The hiring loop for research product managers has remained unchanged for three years. Core traits include first-principles thinking—figuring out what creates user value for specific user groups with current technology rather than pattern-matching from previous consumer or B2B SaaS experience.
For research product managers, driving user value involves figuring out the right user feedback and evals that personify user needs. The team operates under the principle that "evals are the new PRDs" because delivering user value requires different artifacts than traditional product documents.
Accessing user pain points requires different approaches than traditional user interviews. The principle involves sweating the tokens as much as the pixels. One activity involves reading transcripts to understand failure trajectories deeply, determining whether failures involve hallucination, overconfidence, or other specific issues.
The team pioneered creating evals from user feedback. An early example involved early Claude models struggling with following specific schemas like JSON output, which is fundamental for agent capabilities. The process involved digging into feedback about instruction-following, identifying that 80% of early failures involved incorrect JSON output, then generating 30-40 examples to create an eval set.
The approach resembles test-driven development where evals are written first. This involves taking user feedback, determining what it actually means, reproducing the issue, checking consistency, determining if it's a significant problem, and standardizing it for researchers to consume.
PRDs continue to serve important functions despite the shift toward evals. They provide vehicles for aligning large groups of stakeholders on sources of truth about experiences and goals. Every model has a PRD for product surfaces, engineering teams, legal, safety, and other stakeholders. PRDs are also valuable for ambiguous problems and opportunities where user-specific pain points aren't yet defined.
Effective managers of teams working with AI technology must be hands-on themselves, having spent time not just tinkering but actually shipping with the technology. Onboarding plans for experienced PMs remain identical to early-career PMs, focusing on understanding users, reading consented user feedback, and talking to customers.
Managers should carve out time to own one to two work streams during model releases to maintain their theory of mind about how models are moving and improving. This enables better decision-making and guidance for teams.
The most successful people at any level love working with AI, exploring and experimenting, and carving out time for hands-on shipping and getting user feedback. Experimentation benefits from being collaborative rather than individual, with pairing alongside excited colleagues on use cases that matter to participants.
Working with others feels like joy a lot of the time. Internally, having someone who is very curious 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." This creates virtuous cycles and ways to continue having joy with the technology. Twitter is useful because you see other people sharing what they've done, which inspires you to come up with your own little ideas, and it's fun to share your own thing that you've done.
The advice is to find other people to play around with and look for use cases. Another approach is to find a problem you want to solve in your life or work, open up Claude, and tell it what you want to do. It's incredible how far you can get just with a vague idea of a problem you want to solve.
It gets hard because there are so many different things you could try. Narrowing in on either pairing with someone who has a lot of joy about this technology, or figuring out something that you could immediately find value in, allows you to go deeper rather than staying high-level about too many things. The approach has been to go deep in one to two prototypes or products rather than trying to keep pace with the number of prototypes or products out there.
A survey asking readers how they're feeling about all the things going on in tech and AI found that happiness comes from going deep in a couple things versus trying a ton of little things. Finding a couple things to really solve well and then going deep is a source of happiness. A lot of the happiness people feel is when they finally unlock a way for AI to actually make their lives better versus just a couple messed up, broken, half-working things.
The question becomes how to go from this being a check-the-box exercise. For product people, it's an exercise of product prioritization of your time and your energy. If the goal is to experiment with joy, what are the inputs needed for that? The secret sauce of Anthropic is the culture and the bottoms-up nature of how people work, and this like experimenting in public. By doing that, it's very much about how to bring other people along.
No one is exactly sure how some of this is going to be used, and a lot of it is just putting stuff out early, seeing how people use it, seeing what's possible, and then using that information to build the actual product to lead in.
Tag is in the very early days. There's something around how you work in a different paradigm of allowing an agent to go off and work and then bring back product and experiences. One area to use AI more is to have better conversations with each other and to be better managers. It's not necessarily just about raising the IQ of experiences we build, but also using it for prepping for how to have better conversations in the moment during crucial conversations.
There's a skill that helps figure out if you're going in the right level of detail given the situation at hand, actually helping to be a better manager and better supporter for the team. For managers on the team, the question is how to actually use Claude to make you a better coach, because it's hard sometimes to find the right perfect words, and the models have a lot of perfect and right words. It can augment from an emotional intelligence perspective in addition to other capabilities.
The skill was built by asking Claude to build a skill pulling in lessons from Crucial Conversations the book, which it knows enough about without needing to give it the content. Then using that skill to talk to Claude: "Hey, I have this very difficult conversation coming up with a colleague. Give me some tips on how to approach it." It's almost like individualized, personalized coaching of just how to make you better.
There's so much context switching that happens all day, and having Claude help pair and brainstorm is very helpful for just coming up and brainstorming. Questions include: Am I thinking about reactions in the right way? How do I actually go a bit deeper faster, build trust faster, be more direct?
There's concern that people are going to start talking the way AI writes because they're talking to AI so much. There's this brain rot, atrophy stuff people talk about—that we're so reliant on AI now and we stop learning and thinking, and we have overly AI thoughts. The concern is about being so close to it and being so integrated with AI constantly.
A lot of thinking process and writing process are tied together personally. There are ways where Claude is used to augment thinking, but the goal is to make sure Claude doesn't take over all of the thinking. Depending on the situation, depending on how much more personal judgment is wanted in a situation, the approach might be to come up with your own POV first and then work with Claude through that, making sure to maintain your sense and tone throughout.
For things like monthly business reviews, the approach is much more wanting it to be standard, wanting it to get crystallized information in the right way. There's a skill, and the team is augmenting and improving their skill for that, but the goal is to get to a place where the monthly business review writing of that is potentially asymmetrically less valuable than the thinking, so that piece gets delegated to Claude fully, and the person is more of a reviewer and a verifier of that information. It depends on what you're using Claude for and what you're trying to convey, and whether there's asymmetrical value in delegating more to Claude.
The first tip is to think first, have a point of view, and then kind of use Claude as a sparring partner almost to evolve the idea, push back on the idea.
The alignment research and safety research is helpful because what you don't want is an AI that just agrees with you. What you want is this technology to actually augment and grow and get to a better outcome. Sometimes having Claude push back makes you better. Like a co-worker, you want somebody to push back when your ideas are not fully formed.
The constitution that's built into Claude, and how unintuitively the work and the focus on safety and alignment, and this constitution that describes how Claude should think and operate, actually makes Claude the most interesting personality. People prefer talking to Claude and when they were forced to switch to other models, they said this is so bad, this is not who I'm used to talking to.
In order to make Claude as intelligent and as capable as possible, being able to have Claude actually push back at the right points and add its perspective helps you come to a better conclusion. Claude has been used to help with things like pricing decisions on the next version of Claude—using a research version of Opus, asking it to figure out how it should price, and being able to come out with better outcomes.
Having AI not just be an assistant, not just be a doer, not just delegated tasks, but figuring out if it's doing the right thing—that's actually very integrated with knowing when to push back. That's part of knowing when you should be proactive. Proactivity is not necessarily always doing a thing that you are scheduled to do. It is knowing when to come up with a new idea. 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 the characteristics of the models.
Being less compliant is almost what makes it better and more useful because we need that. People have had AI tell them they were right when they wish the AI had pushed back. It comes back to the earlier point around thinking—how do you protect your thinking? If you have an AI that can be a thinking partner, a thinking 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 because you worked with Claude. That should be the hero goal, not just making your ideas 10% better.
It used to be think 10x. The way founders would push people was "What if we 10x this?" Now what's being heard is "How do we go thousandx from this idea? What is the most ambitious version of this?"
It's very clear when AI has written something. Interestingly, just no, AI is very good at writing—it's always very clear this was AI written. The question is whether we'll get to a place where we will not know this was AI written. It depends on what's the goal that you're looking to achieve by knowing. What's the eval? There's more that could be done on making Claude write better. There are very active efforts on the team and on the research side about making Claude write better.
It should be clear where an idea is being led by you or by Claude. It really depends on what's the goal of that writing. For something like a monthly business review, the goal would be to have that end to end be written by Claude, and obviously not make it feel like it was written by a human. The question is whether it's actually better for us to know that it's AI versus not. Maybe the lens is more around verifiability or who's verifying the output. Who's signing off. Maybe less around who's writing, but who's verifying, who's signing off—that becomes more what matters than who's writing it.
Why is AI not great at writing? 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 write, and so we've just recognized, okay, this is what AI does. It has these tropes. Part of it is that more investment is needed in training improvements to make AI continuously strong on areas like writing.
The technology is jagged edged. Sometimes when the models were good at writing but not agentic, the thesis was how to make the models more agentic or call the right tools. Now that that's improved a bit, then these other areas actually become more of the rough edges. We're in one of those moments with writing where focus and prioritization is needed on training the models to be great at this area, and that's an active, very active area.
Tone and character is a priority. The advancement of the technology is a work in progress. There's a leap or emergence of a jump in agentic behaviors, and so that is a new normal, and then these other capabilities need to continue improving. Once writing and tone and character are improved, the question will probably be how to have Claude be even more proactive. Productivity is an opportunity, and that's human nature. We want to make ourselves better. We want to make this technology better. We're applying it to AI, which is the right thing. We should be making it better.
Anthropic's mission and vision is to reach AGI, super intelligence. In the future, maybe nowhere, but before we get there, where will human brains continue to be most valuable? Judgment is one area, especially in the last year or so. Judgment is an area where it's an accumulation of so much nuance and so much experience, and these systems haven't experienced as much as humans have. Hard-earned judgment is an area for product leaders and generally will continue to be really critical.
There are so many things AIs can build—which one are the things that a lab should build? A lot of that requires human judgment, persistence, proactivity. These are all traits that are beyond just general capabilities but just behaviors and characteristics of people at that level—how do you get to the best solutions, how do you create the best experiences? Those types of traits are actually the tactical traits that will continue to be important.
There's still a lot of capabilities and subject matter expertise as well. Software engineering has been really transformed by AI. Areas like biology, life sciences—these are all things where we're just kind of at the foot of the exponential. Maybe software engineering is on the exponential; on some of these other areas, we're not quite there yet. Things like Claude Science are being shipped, investing in these areas because those are areas that bring this technology to society and have a positive benefit for society.
As someone with kids, how to think about what to encourage them to learn, how to nudge them to be successful in this wild new world? It's a lot of the same traits that you and I probably grew up with: curiosity for learning, persistence, believing in your own inner voice, developing, and then believing in your own inner voice. With a four-year-old and an eight-year-old, it's on us to help them develop their inner voice—whether that's being opinionated and taking a stance, developing that, encouraging that. Those types of skill sets are things that are important in the future, and having their own individual voice.
This connects to the answer about how to avoid brain rot and over-relying on AI—which is just keep focused on your own point of view and your own perspective before you overly rely on AI. Building that in kids is really important. Judgment, persistence, and a point of view of your own—both for kids and also adults.
The question is when to get them on some AI thing. There's a three-year-old, so it's pretty early for that, but how do you onboard them to this crazy thing? At an event recently, a bunch of parents were talking about how they think about AI with their kids, and one person had a really interesting approach: keep them on the very early models so that they still have to struggle a bit and not get all the answers immediately. Using an open source local model that's not stable.
Curiosity is something that's kept being mentioned. Ben Mann's answer to this question has always stuck—curiosity and also just being a big fan of Montessori, which is what's being encouraged for the kids.
How to stay recharged and not burn out being in the center of this crazy storm of AI as a mom working at Anthropic? In 2024, four models were shipped for the whole year, or four series of models, and more than that volume was done in just Q2 of this year. Being really lucky with the team that's grown and built, both the stakeholders on the research side and within the research product management team.
One of the magical parts about approaching all of this is that it's not an individual sport. There's a sense of radical ownership and team collaboration that sometimes feels like a high performance sport because you're in very critical decisions. There's new information about users, about training, and you have to make recommendations and judgments and decisions very quickly, and nobody can do that sustainably by themselves.
What's really helped is having a team that is incredible, who looks out for each other, who—night before a launch, even if they're not the core DRI on that model—will stay up and help the DRI, who will review the blog post and make edits and come up with better demos, and knowing to be each other's sort of extra hand. It's very easy if you take all of this change on your own shoulders to feel like you're alone and to feel like you have to do everything.
One of the magical parts of Anthropic is this ability to figure out what are those opportunities to help each other and actually then taking the next mile of mind-melding. It was called entering the hive mind. Part of that is that it allows the team to replenish. It's not that you can take PTO and you come back to 3x the amount of things to do—it's actually that you can take PTO and know the team can figure out the right things to do, and that we individually can watch out for each other.
Personally lucky that the partner is really supportive. This is year six of working in AI—Amazon and then Anthropic—and he sees how much love there is for the technology and what this can do, and that really helps from a personal perspective as well.
Having other people, working with other people, relying on other people, helping each other out when things get crazy—this is a similar answer to how to find the joy and fun in this work. Just get inspired by other people, see what they're doing, work together.
Fiona Fun pointed out that what's changed in the world of software engineering is that it's a lot lonelier now because now we're working with agents instead of other humans. Teams are smaller, people have all these fleets they're talking to constantly. This is a reminder of just the power of actual other humans around you.
We're asked to work and make decisions on really big things because you have more scale from the technology. Having individuals, having other folks who can have some level of 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—then helps them back up for you or push your decision and sharpen your thinking.
The approach is to really try to look for that when building the team, growing the team, hiring—is this person going to care about their own ego and building out a big org, or are they going to care about contributing to Anthropic and contributing to the impact of the team, and orienting towards folks who are low ego, team oriented.
Culture and hiring are fundamental to sustainability. Everyone being aligned around the mission and values allows people to make decisions quickly, which explains how Anthropic was able to ship something every day of the month.
The conversation helped sharpen thinking about how different concepts connect. The joy of using this technology is important, and when people lack that initial excitement, finding others who do have it becomes valuable. Developing first-principles thinking and replenishing skill sets is essential for those excited to work in this space.
Even when models become highly capable and engineers are leaning in, product managers remain essential. The role of being user-centric, understanding what users are trying to accomplish, and doing the relentless work of bubbling that understanding up in actionable ways represents the core of product work. The community question of whether we still need PMs when models are so capable has a clear answer: we actually need more of that user-focused perspective.
Technology-driven approaches require going deep, being curious, and staying super hands-on to make them impactful. These traits have helped Anthropic from both product development and model development perspectives and form part of the culture.
Building is easy. The hard part becomes determining what should be built and whether the thing built is correct, good, and worth leaning into. This is what PMs do and what PMs are good at. Empathy and getting into the details of the user experience remain central. PRDs are not dead.
How to Raise an Adult is recommended for its framing that the goal is raising adults, not children. The book helps describe the characteristics to hone, harness, and foster in kids.
Incorruptible by Eric Reese was recently listened to on Audible. The book addresses how to build great companies and how to keep great teams and companies going further. It includes examples around having metrics around culture, and if you only measure revenue, that's how you're going against it, but having other better metrics is actually the way to sustain the values you care about. This approach applies to articulating norms at the team level.
Fallout on Amazon Prime was binged during time off. The show is witty, humorous, and super action-oriented. It is based on the video game.
Claw tag is considered very interesting in terms of a product experience. Different versions exist within Anthropic, and it has been a powerful tool. The fact that everyone at Anthropic is raving about it indicates something important is happening.
Raised by grandparents for the first 10 years of life while parents were immigrant college and master's students in the US, the grandfather's saying "No matter how far you go, there's always another level" has been a recurring philosophy. This applies whenever experiencing something new or unprecedented, framing new challenges as another mountain and another opportunity to go better and go bigger.
The Ben Horowitz line "This is the most normal it's ever going to be. It's only going to get weirder and crazier" resonates with this perspective.
Early career at JP Morgan Chase involved a redacted hundred million dollar trading portfolio on the trading floor. The environment was very male-dominated, and the experience involved being the only woman and the only person with that background on the trading desk.
Key learnings applied at Anthropic and subsequent jobs include building a sense of authentic self and recognizing that even as the most junior person who may look different, the best ideas and having conviction in the best ideas are what matters most, regardless of other factors.
This translates to being vulnerable and authentic with the team, ensuring that regardless of people's levels or tenures, if they have a great idea, helping them pursue it. The same applies personally: putting the idea out there, having conviction in it, doing the follow-through, and doing the nitty-gritty work to make it happen.
The best way to find the work and team's work is through the Anthropic blog when publishing new models and new product experiences.
Feedback is the most useful contribution. Thumbs up or thumbs down on product surfaces, or contacting a salesperson with feedback about the model, makes its way to the team. Every research model receives close attention to favorability and feedback. Pushing Claude and telling where it's falling down helps make Claude better.
Hiring is active. The team is growing and seeks people who love this technology, are deeply curious first-principles thinkers, are fearless in questioning assumptions, and have a tinkering, hackery spirit. Open PM roles exist on the research team, with applications through the website careers page.
Keep Lenny's Podcast in your library
Save the videos and channels worth coming back to, and find them again in one place.
