The Product Evolution of AI: A Conversation with Anthropic's Diane Penn

In the rapidly evolving landscape of artificial intelligence, the role of product management has undergone a seismic shift. Diane Penn, Head of Product for AI Research and Labs at Anthropic, shares her insights into this transformation, from the early days of AI research to the cutting edge of model development and product incubation. Penn, who joined Anthropic as their first technical product manager over three years ago, has been instrumental in shipping every model from Claude 2 through Fable, and has helped launch key products like Claude Code, MCP, Skills, and Claude Design.

The Genesis of Anthropic and Early Innovations

When Anthropic first emerged, the AI landscape was dominated by established players like OpenAI. Penn recalls the initial skepticism: "I remember dealing, man, these guys have no chance. OpenAI is so far ahead." Yet, Anthropic forged ahead, driven by a strong mission, culture, and a startup-like energy.

A pivotal moment in their early journey was the "Golden Gate Bridge" incident. After publishing research on model interpretability, a feature that made Claude obsess over the Golden Gate Bridge by dialing up a specific model feature, the team decided to release it to the public for 24 hours. This rapid, cross-functional effort, involving engineering, product, design, and research, showcased their ability to bring research to life in an authentic and agile way. "It made us feel like, oh, we can actually bring new user experiences, showcase our research in a way that's different and authentic to us and in a very startupy like pace," Penn explains. This experience was an early inflection point, helping Anthropic identify its unique identity in building products that differed from competitors.

Milestones in Model Development: Opus 3 and Opus 45

The development of Anthropic's frontier models has been marked by significant milestones. The training and testing of Opus 3 was a crucial moment, solidifying the company's commitment to creating a leading-edge model. This period, around December, saw teams working intensely from home, rallying around a common goal. "Everybody that was involved was like really proud," Penn notes, highlighting the foundational trust built during this time.

A year later, Opus 45 emerged as another major inflection point. What made Opus 45 particularly magical was not just the model itself, but the accompanying product experience: Claude Code. "You need frontier products in order to have frontier models and for people to feel the magic of frontier models," Penn states. The synergy between Opus 45's advanced capabilities and the user-friendly interface of Claude Code accelerated adoption and allowed users to experience the full potential of these AI advancements.

Navigating the Exponential Curve of AI Advancement

The pace of AI improvement is often described as an exponential curve, a "hockey stick" phase where progress accelerates dramatically. Penn likens this to the internet's transition from a novelty to an essential tool. She emphasizes the importance of adaptability and first-principles thinking in navigating this rapid evolution.

"The way to drive user value is to figure out the right user feedback," Penn explains, highlighting the emergence of "evals are the new PRDs." This means deeply understanding user pain points, often by meticulously analyzing user interactions and identifying failure trajectories. This detailed feedback loop is crucial for researchers to make actionable improvements.

The concept of "token maxing"—spending generously on token usage to experiment and live in the future—is also discussed. Penn views this not just as a cost, but as an investment in experimentation, leading to better ideas and faster development. She advocates for communal discovery, where teams share their findings and build upon each other's insights.

Anthropic Labs: Incubating Innovation

Anthropic Labs operates on a thesis of identifying and pursuing "discontinuous large bets" that might not fit into the core roadmap. This incubation model focuses on exploring the 10x, 100x, or even 1000x potential of new ideas. Products like Claude Code, Skills, and Claude Design have emerged from this initiative. The key to Labs' success lies in a culture of experimentation, with a strong opinion on themes but flexibility on the exact prototype. This allows for learning even from experiments that don't immediately ship, fostering a culture of seeing around corners.

The Evolving Role of Product Management in AI

The product role in the AI era demands a new set of skills. Penn emphasizes first-principles thinking over pattern matching, urging PMs to deeply understand user value in the context of emerging technology. The shift from traditional PRDs to evals as the primary artifact for defining user needs is a significant change. "Evals are the new PRDs," she states, highlighting the need to "sweat the tokens as much as you sweat the pixels" to understand user pain points.

Building effective eval sets involves meticulously analyzing user feedback, reproducing issues, and standardizing them for researchers. This process is akin to test-driven development for PMs. While PRDs still have value for aligning large groups and exploring ambiguous problems, evals are becoming central to the product development loop.

Furthermore, hands-on leadership is paramount. Managers must be deeply involved in shipping with AI technologies, understanding the details, and "sweating the tokens" alongside their teams. This direct experience is crucial for developing a strong sense of what constitutes a good AI product.

Finding Joy and Avoiding Burnout in the AI Storm

The relentless pace of AI development can be overwhelming. Penn's advice for finding joy and avoiding burnout centers on collaboration and community. She stresses that AI development is not an individual sport, and a strong sense of radical ownership and team collaboration is essential. The ability to "mind-meld" with colleagues, support each other, and take PTO with confidence that the team can manage is vital.

For individuals, finding joy often comes from going deep rather than trying to master everything. Pairing with enthusiastic colleagues or focusing on solving a specific problem can lead to more meaningful engagement. Penn also highlights the importance of curiosity, persistence, and believing in one's own inner voice, traits she aims to instill in her own children.

The Future of AI and Human Value

As AI capabilities grow, the question of where human brains will remain most valuable arises. Penn points to judgment, which is built on nuanced experience and intuition that AI systems currently lack. The ability to make strategic decisions about what AI should build, requiring persistence and proactivity, will continue to be critical.

She also sees continued value in subject matter expertise in fields like biology and life sciences, where AI is just beginning to unlock its potential.

The Constitution of Claude and the Art of AI Writing

The focus on safety and alignment, embodied in Claude's "constitution," has paradoxically made the AI more interesting and capable. Penn explains that an AI that can push back and challenge assumptions, rather than simply agreeing, leads to better outcomes and more robust thinking. This "thinking partner" aspect is crucial for augmenting human intellect.

Interestingly, while AI is adept at processing vast amounts of text, AI writing itself is still an area for improvement. Penn notes that Anthropic is actively investing in making Claude write better, focusing on tone, character, and clarity. The goal is not always to mask AI authorship, but to ensure verifiability and clear attribution of ideas, whether they originate from a human or an AI.

Key Takeaways