AI Going Vertical

01Anthropic's Inflection Points: From Underdog to Vertical AI Leader

Penn traces Anthropic's journey through a series of inflection moments, each compounding the last. The early identity-building moment came with Golden Gate Claude — a research demo spun into a product in 24 hours that, while reaching only ~2,000 users, proved Anthropic could ship experiences its competitors couldn't. More consequentially, Opus 3 training in late 2023 rallied the then-sub-200-person company around a shared goal and built the trust between product and research that persists today. The coding bet — training Opus 3 to excel at long-form code generation — was a small training change with outsized market impact, attracting early Claude enthusiasts and developers (~9:15). Opus 4.5 the following year demonstrated the model-product symbiosis that now defines Anthropic's approach.

See also: consume/concepts/s-curves, consume/concepts/s-curves-following-s-curves

02Labs as a Discontinuous Bets Engine

Anthropic's Labs division operates as an internal startup foundry under Ben Mann, pursuing “discontinuous large bets” outside the core roadmap. Products incubated in Labs include Claude Code, MCP, Skills, and Claude Design. The operating model is distinctive: small pods (sometimes one engineer), a culture that rewards zero-to-one founders, and a willingness to turn off bets that aren't working — with the discipline to revisit them one to two model generations later when the technology may have caught up. The key tension Penn articulates: be “strongly held opinion about the theme or the area and then more weakly held about the exact prototype” (~24:53). This lets Labs explore the 10x, 100x, or 1000x version of an idea without prematurely committing to a specific implementation.

03Evals Are the New PRDs: The PM Role Reimagined

This is arguably the interview's most actionable framework for product people. Penn's team has a saying: “evals are the new PRDs” (~41:42). The shift reflects a fundamental change in how value is delivered when building at the intersection of models and products. In traditional software, a PM writes a PRD to align stakeholders around a spec. In AI-native product management, the primary artifact is an evaluation set — a curated collection of 30-40 concrete examples that capture real user failures. These evals serve three functions: they personify the user need, they provide a measurable, reproducible signal for whether improvements are working, and they translate vague feedback into actionable input for researchers. Penn still writes PRDs for alignment across large groups and for ambiguous greenfield opportunities like computer use, but the application has changed. The core skill is now “shortening the distance to actionability” between user pain and researcher intervention (~43:00).

See also: consume/concepts/taste

04The Research-Product Feedback Loop: Translating User Pain into Training Signal

Penn provides a rare inside look at how product management interfaces with AI research. Researchers at Anthropic are not just iterating on algorithms — the best have “founder-like energy” with long-term visions (computer use was discussed at the company's founding, years before it shipped). On the day-to-day side, Penn's team is embedded in research loops for capabilities with direct user impact: vision, computer use, coding, agentic coding, tool use, test-time compute. The hardest and most important part of the job is translating user feedback into granular, actionable failure modes. “Claude hallucinated” is useless to a researcher. “Claude looked at the right document but extracted the wrong facts” is actionable — it points to search synthesis, not alignment or tool use. This triage — identifying whether a failure is a tool-use gap, a knowledge retrieval gap, or an alignment issue — is the day-to-day language of the research team (~30:00). Penn calls this “test-driven development for PMs” (~46:11).

05The Emerging Capability Curve and the Need for Adaptability

Penn grounds the experience of working inside the exponential in the original scaling law papers. The well-known finding is that loss decreases smoothly as compute and data scale. The less-discussed finding — and the one that shapes her team's operating philosophy — is that capabilities emerge discontinuously. Models jump from not being able to compute 1+1 to reliably computing it. These “emerging capability graphs” mean that even with evals, no one knows exactly when a capability will appear. The implication for product development is radical adaptability: the ability to pull forward a plan originally scheduled months out because a new model can suddenly handle it, and the organizational grace to bring along teams who experience the exponential at different speeds (~17:30).

06Alignment as Product Advantage: The Constitution Makes Claude a Better Thinking Partner

An apparent paradox runs through the interview: Claude's constitutional AI training — the safeguards, the alignment work, the willingness to push back — makes it a better product, not a more constrained one. Penn explains that for Claude to be truly useful, it must know when to disagree. A thinking partner that always says “yes” makes you worse; one that pushes back at the right moments sharpens your ideas. She's used research versions of Claude to help with pricing decisions — “it's a little bit meta” — finding that the model's ability to challenge assumptions leads to better outcomes (~1:04:49). This connects to her personal practice: she forms her own POV before consulting Claude on judgment-intensive tasks, using the model as a sparring partner rather than a replacement for thinking. For routine work like monthly business reviews, she's comfortable delegating writing entirely, with herself as reviewer and verifier (~1:02:00).

See also: consume/concepts/cognitive-surrender, consume/concepts/taste

07Building Careers and Character in the AI Era

Penn's advice for aspiring PMs and researchers is consistent and practical. The hiring bar hasn't changed in three years: first-principles thinking — reasoning from user value and technological reality rather than pattern-matching from prior roles — is the non-negotiable. For managers specifically, Penn is adamant: you cannot lead AI product teams without personally shipping with the technology. Her onboarding for senior hires is identical to junior hires. Beyond skills, she emphasizes that experimentation should be communal, not solitary — pair with excited colleagues, share discoveries, let others' creativity inspire yours (~53:00). For parents, her advice is to develop children's “inner voice” — the ability to form opinions, take a stance, and persist in convictions — before AI becomes their default thinking partner. Her grandfather's motto captures the mindset: “No matter how far you go, there's always another level” (~1:28:14).