Anthropic’s Krieger, Vora Say Faster AI Coding Makes PMs More Vital

Lenny's Podcast

Because frontier models change every two months and make building software cheap, product managers must abandon upfront UI debates to serve as the human convener and decision-maker across parallel bets.

AI-assisted summary of Why Claude can’t be your PM (yet) | Anthropic CPO Panel — original source. Check the original for full context.

Source published: 2026-09-29

Key ideas

While foundational technology shifts used to occur every 5 to 10 years, AI capabilities now shift roughly every 2 months, making the human PM's role as an operational "convener" and decision-maker more vital than ever as engineering accelerates. (0:06)
Because building software is now fast and cheap, Anthropic PMs skip upfront theoretical debates over button placement to build and test multiple versions directly, while software architecture evolves from chat sidebars toward "agent-native" parity and agent-generated malleable UIs. (6:49)
After finding that Claude Chat and Claude Cowork used separate memory, MCP, and file-storage systems, Anthropic centralized these primitives under a Foundations team so parallel bets share context, while underperforming bets like early Computer Use are "parked" in automated eval harnesses until newer models like Claude 3.7 make them viable. (19:54)
Guided by the "power law of tabs"—where Instagram's main feed drove roughly 80% of usage—Anthropic leaders stress saying no to sidebar clutter and rolling proven experiments into a single core experience to close the gap between frontier model capabilities and everyday usage. (26:17)

Detailed notes

## Background Published by Lenny's Podcast on YouTube on September 29, 2026, this panel was recorded live at the Lenny and Friends Summit in San Francisco. Moderated by Dan Shipper, Anthropic product leaders Ami Vora and Mike Krieger discuss how the PM role is evolving alongside rapid AI development, examining agent-native and malleable software, running parallel experiments, and shipping ideas that users can actually adopt.

## TL;DR Rather than eliminating product managers, rapid AI advancement elevates their importance as human 'conveners' who exercise judgment, coordinate cross-functional teams, and anchor products to enduring human problems. As building becomes cheap, upfront interface debates give way to rapid prototyping, agent-native architectures, and malleable software dynamically maintained by models. To navigate this fluidity, Anthropic structures exploration around autonomous 'bet leads,' unifies shared primitives like memory across surfaces, and parks promising concepts into evaluation harnesses until model capabilities catch up. Ultimately, successful product discipline requires resisting interface clutter, honoring the power law of core entry points, and bridging the widening gap between raw model power and everyday user enablement.

## Atmosphere & style The panel carries an energetic, collegial, and candid tone marked by audience laughter, playful banter, and self-deprecating humor about the chaotic pace of AI development. Dan Shipper balances deadpan wit with sharp industry observations, while Ami Vora and Mike Krieger share candid reflections on unlearning decades of hard-won product habits and navigating the vulnerability of feeling lost together. Watching the discussion feels like sitting in on an intimate, intellectually grounded debrief between seasoned operators who are simultaneously humbled and energized by the dizzying frontier they are building.

## Bulletin - [00:06](https://www.youtube.com/watch?v=sEXdyK6woKU&t=6s) While foundational technology shifts used to occur every 5 to 10 years, AI capabilities now shift roughly every 2 months, making the human PM's role as an operational "convener" and decision-maker more vital than ever as engineering accelerates. - [06:49](https://www.youtube.com/watch?v=sEXdyK6woKU&t=409s) Because building software is now fast and cheap, Anthropic PMs skip upfront theoretical debates over button placement to build and test multiple versions directly, while software architecture evolves from chat sidebars toward "agent-native" parity and agent-generated malleable UIs. - [19:54](https://www.youtube.com/watch?v=sEXdyK6woKU&t=1194s) After finding that Claude Chat and Claude Cowork used separate memory, MCP, and file-storage systems, Anthropic centralized these primitives under a Foundations team so parallel bets share context, while underperforming bets like early Computer Use are "parked" in automated eval harnesses until newer models like Claude 3.7 make them viable. - [26:17](https://www.youtube.com/watch?v=sEXdyK6woKU&t=1577s) Guided by the "power law of tabs"—where Instagram's main feed drove roughly 80% of usage—Anthropic leaders stress saying no to sidebar clutter and rolling proven experiments into a single core experience to close the gap between frontier model capabilities and everyday usage.

## Key takeaways - [00:06](https://www.youtube.com/watch?v=sEXdyK6woKU&t=6s) Foundational technology cycles have compressed from 5–10 years to roughly 2 months, requiring PMs to stay anchored on unchanging human problems while repeatedly discarding outdated technical solutions. - [02:40](https://www.youtube.com/watch?v=sEXdyK6woKU&t=160s) Faster AI-assisted building increases the need for operationally excellent PMs to provide cross-functional "glue"—aligning prosumer and enterprise needs, customer success, and safeguards so builders can stay heads-down. - [04:38](https://www.youtube.com/watch?v=sEXdyK6woKU&t=278s) While Claude helps surface organizational context via Slack search, AI cannot yet replace the human "convener" who brings people and agents together and exercises judgment across rapidly multiplying product paths. - [06:49](https://www.youtube.com/watch?v=sEXdyK6woKU&t=409s) Because building software is no longer slow and expensive, PMs have unlearned decades of upfront theoretical debate over granular UI placement in favor of rapidly building and testing multiple working versions. - [08:38](https://www.youtube.com/watch?v=sEXdyK6woKU&t=518s) Navigating frontier AI creates a "personal innovator's dilemma," requiring leaders to abandon legacy strengths, raise their ceiling on tolerance for change, and "frame the chaos" into a safe container for experimentation rather than enforcing rigid roadmaps. - [11:23](https://www.youtube.com/watch?v=sEXdyK6woKU&t=683s) To balance open exploration with disciplined execution, Anthropic Labs structures initiatives around single Directly Responsible Individuals (DRIs) or "bet leads" empowered to allocate resources, double down, or wind down projects. - [12:09](https://www.youtube.com/watch?v=sEXdyK6woKU&t=729s) AI software is progressing through three distinct eras: from bolted-on chat sidebars, to "agent-native" architectures where agents have full parity with human actions, to "malleable software" where agents dynamically generate and maintain custom UIs on the fly. - [15:34](https://www.youtube.com/watch?v=sEXdyK6woKU&t=934s) Legacy and scaled SaaS companies that cannot overhaul their user interfaces overnight should focus on unified infrastructural plumbing so human UI actions and programmatic agent endpoints flow through the exact same underlying primitives. - [16:42](https://www.youtube.com/watch?v=sEXdyK6woKU&t=1002s) Rather than over-constraining products upfront based on theoretical assumptions, Ami Vora encourages teams to take multiple overlapping "shots on goal" and consolidate winners into a cohesive product once market fit is proven. - [19:54](https://www.youtube.com/watch?v=sEXdyK6woKU&t=1194s) Running parallel product bets successfully requires authentic team conviction and a dedicated Foundations team that unifies shared primitives—memory, Model Context Protocol (MCP), and file storage—across surfaces like Claude Chat, Claude Cowork, and Labs. - [22:19](https://www.youtube.com/watch?v=sEXdyK6woKU&t=1339s) When an ambitious concept like Computer Use fails because current models are too clumsy, Anthropic "parks" the project in an automated evaluation harness and tests it against subsequent releases until a checkpoint like Claude 3.7 crosses the viability threshold. - [22:19](https://www.youtube.com/watch?v=sEXdyK6woKU&t=1339s) Builders must guard against "capability blindness"—writing off a problem space after a single failed test—because new model checkpoints every three to six months routinely overcome prior limitations. - [24:37](https://www.youtube.com/watch?v=sEXdyK6woKU&t=1477s) A central unsolved challenge in AI product development is folding validated parallel experiments back into a single coherent product without offloading cognitive load onto users via fragmented apps or cluttered tabs. - [26:17](https://www.youtube.com/watch?v=sEXdyK6woKU&t=1577s) Because interface usage follows a strong power law—illustrated by Instagram's main feed driving roughly 80% of activity—half of product leadership is saying no to extra tabs and sidebars and integrating proven workflows back into the core entry point. - [27:04](https://www.youtube.com/watch?v=sEXdyK6woKU&t=1624s) Over the coming year, Anthropic's product leaders aim to empower solo builders and bridge the widening gap between frontier model capabilities and everyday users so non-engineers can orchestrate long-horizon, parallel work.

## Detailed walkthrough ### [00:06](https://www.youtube.com/watch?v=sEXdyK6woKU&t=6s) Is the Product Manager Role Going Away or Evolving?

Taking the stage to warm cheers and "good vibes" at the Lenny and Friends Summit, moderator Dan Shipper opens the panel with the question on everyone's mind. Just a year ago, the prevailing narrative was that AI might make product managers obsolete—provoking laughter from the audience—whereas today, the consensus has flipped to the idea that *everyone* needs to become a PM. Because Ami Vora and Mike Krieger work directly with frontier models at Anthropic, Shipper asks them to share where job roles are heading: what does a great PM at Anthropic look like today, and is the role disappearing or transforming?

Ami Vora argues that the fundamental job of product management has always been to act as "a bridge between the real problems that people have in the world and the technology that you can use to solve it." What makes the current moment feel so strange and difficult for humans to process is the sheer speed of technological turnover. Previously, foundational technology changed roughly every five to ten years; now, it feels like it shifts every two months. Despite that dizzying pace, Vora points out that the human side of the equation does not change nearly as fast—people still have the same core problems. Her approach is to remain obsessed with solving those enduring human problems while staying "super fresh-eyed about how the technology works," deliberately throwing away most of what she used to know about how to solve a problem so she can tackle it anew with the latest capabilities.

In many ways, Vora feels the industry is "rewinding the clock" to the early days of product management—"back when the dinosaurs were on the earth"—when strict role definitions did not exist. Back then, a PM was simply a general problem solver who walked around trying to fix things until running into a wall, at which point they would ask someone for help or figure a way around it. Today's environment allows teams to return to that fluid state where traditional roles blur: you push forward to figure out what works, and when you hit a wall, you now have rich new forms of scaffolding to help you through.

- **Old technology cycle:** Major shifts occurred every 5 to 10 years. - **Current AI technology cycle:** Capabilities shift roughly every 2 months. - **Core PM mindset:** Retain deep knowledge of unchanging human problems while repeatedly discarding outdated technical solutions.

[02:40](https://www.youtube.com/watch?v=sEXdyK6woKU&t=160s) Mike Krieger’s Realization: Why Builders Still Need a PM

Mike Krieger shares a personal realization from earlier in the year after he transitioned into an individual contributor (IC) role focused mostly on hands-on building. As his team got close to shipping an exciting project, Cat—one of the PM leads who works with Ami—pulled him aside and told him, "Mike, I know you love to build. This project's going well. Like, it's looking really exciting. Like, we really need a PM on this project." Deep in builder mode, Krieger initially pushed back, wondering if a PM was even necessary: "I don't know. Like, do we need a PM? Like, Claude's got it. Like, we've got a lot going on." Cat stood firm that they truly needed one.

Once a PM actually joined the project, the before-and-after contrast was an eye-opening reminder for Krieger. Watching the PM step in revealed all the "glue and all of the connective tissue" required to ship successfully—and all the critical details that were about to get dropped had they continued without one. The difference was so stark that just one week later, Krieger pinged Cat to tell her, "Cat, you were 100% right—and thank you."

Krieger breaks down everything that connective role encompassed for a company like Anthropic, which serves a massive spectrum of users ranging from individual prosumers all the way up to large enterprises. Even with powerful AI tools, no individual has infinite capacity. Having someone dedicated to operational excellence allows the builders who are heads-down in "full-on Claude mode" to stay focused without constantly popping out of flow to coordinate cross-functional launch readiness. Rather than fading away, the PM hat has become even more vital when recruiting for a team: because engineering moves so much faster now, the person wearing the PM hat must be more operationally excellent than ever before.

- **Audience alignment:** Ensuring the entire customer contingency—from prosumers to large enterprises—is brought along. - **Customer Success enablement:** Preparing customer-facing teams to talk about product changes and answer questions in real time. - **Safeguards and coordination:** Looping in safety and safeguards teams early and keeping all moving parts on track so builders can stay heads-down.

[04:38](https://www.youtube.com/watch?v=sEXdyK6woKU&t=278s) Why AI Agents Cannot Replace the Human "Convener" Yet

Shipper jumps in with a deadpan "serious question," joking askance about why Krieger didn't just have an AI agent—"Fable," or some unannounced "Fable 7" that Anthropic has internal access to—do the PM's job instead. Krieger acknowledges that Anthropic naturally runs heavily on Claude and already benefits from AI-provided connective tissue; for example, Claude routinely surfaces relevant work happening in other parts of the organization via Slack search so teams know what colleagues are doing.

However, Krieger explains that "I don't think Claude is a convener yet." Referencing a talk he gave a couple of weeks prior on the role of humans in a post-super-powerful AI age, he highlights one essential human archetype: **the convener**. Someone still has to bring both the "Claudes and the people together" to actually get the work done. Right now, Claude lacks organizational pull and autonomous scheduling authority. Krieger jokes that it will be quite a moment when a mysterious calendar invite appears and he asks who scheduled the meeting, only to find out Claude decided, "I thought you two should chat." While that hasn't happened yet—perhaps only because they haven't flipped that setting on, prompting a laugh and a "stay tuned" from Ami—convening remains firmly human territory.

```mermaid flowchart TD accTitle: Human PMs and AI in Anthropic's Workflow accDescr: How Claude assists with context discovery while the human PM acts as convener, decision-maker, and operational glue across teams.

subgraph AI["Claude's Current Role"] A["Accelerates heads-down building"] B["Finds cross-org context via Slack search"] end

subgraph HumanPM["The Human PM Role"] C["Acts as 'Convener' across people & Claudes"] D["Exercises judgment across forking paths"] E["Coordinates safeguards, CS & enterprise needs"] end

subgraph Outcome["Product Delivery"] F["Relentless execution tuned to human problems"] end

AI --> HumanPM HumanPM --> Outcome ```

Expanding on why human PMs remain indispensable, Vora notes that because today's tools allow teams to build so many things so quickly, the universe of what is *possible* to build has expanded dramatically. With everything moving fast and information often limited, teams face countless "forking paths." That abundance of options demands even sharper human judgment about *what* to build, paired with relentless execution to make sure it actually happens. A great PM combines a tight feedback loop with end users, a deep grasp of the human problem, and the relentlessness to say, "I'm just going to go for it" in the face of ambiguity.

[06:49](https://www.youtube.com/watch?v=sEXdyK6woKU&t=409s) Unlearning Old Product Habits When Building Is Cheap

Shipper agrees, comparing each new model release to looking out across an ocean wondering what lies beyond the horizon; when you fast-forward past that horizon, you don't find a finished utopia or a disaster, just a vast new landscape to navigate. Recalling their prep conversation for the summit, Shipper brings up an observation Vora made about her own multi-decade career prior to Anthropic: she spent decades learning how to answer specific product questions that she no longer needs to answer at all. He asks which hard-won skills have lost their utility, what she leans on instead, and how that changes her expectations when hiring.

Vora reflects on how much of her career was spent trying to project herself into the mindset of a user upfront—predicting exactly how a feature would fit into their daily life and what to build—Strictly because software used to be so expensive and slow to build and ship. In traditional product reviews, teams would spend significant time debating granular interface questions before writing code:

- *"Ami, where should we put this button?"* - What should the exact interaction pattern look like? - Where will a user's finger naturally reach on their phone screen? - What physical situation is the user in when they pull their phone out of their pocket, and what do they want in that moment?

Having worked hard for years to master that intersection of product and design intuition, Vora admits candidly: "I don't think anyone's ever going to ask me that question again." As Shipper jokes that those questions belong to "Fable" now, Vora explains that upfront theoretical debate is in the past—today, "it's actually faster to just build three versions and try them out and be like, 'Oh yeah, this works for this one.'"

[08:38](https://www.youtube.com/watch?v=sEXdyK6woKU&t=518s) Personal Innovator's Dilemma and Framing the Chaos

Navigating frontier AI requires confronting a deeply uncomfortable psychological hurdle: realizing that the core skills one spent years mastering may no longer be relevant. The speaker draws a direct parallel between Clayton Christensen’s innovator's dilemma for companies and its manifestation at an individual level. Just as successful businesses fail by doubling down on their historic strengths instead of exploring emerging disruptions, professionals fall into the trap of clinging to what they already do well. In an environment where the foundational technology radically shifts every two months, product leaders must be willing to abandon their established identities and dive into areas where they have zero preexisting expertise: "All right, well, let's try this other thing and see if I can figure it out."

When asked what traits allow someone to succeed under those conditions, the answer centers on expanding one's "ceiling on tolerance for change," raw adaptability, sound judgment, and relentlessness. Crucially, leaders must openly discuss these stresses across their teams; otherwise, individuals assume they are the only ones struggling to keep pace, which creates an isolating, exhausting burden.

A vital countermeasure is learning how to "frame the chaos." In moments of deep uncertainty, people instinctively attempt to regain control by charting rigid roadmaps—mapping out precisely what products will be built and how careers will unfold. That instinct locks teams out of discovering breakthrough opportunities. Leadership's job is not to eliminate uncertainty, but to structure it into an emotionally safe, plausible container where teams can experiment freely, tap into the productivity upside of the unknown, and actively participate in building the future.

[11:23](https://www.youtube.com/watch?v=sEXdyK6woKU&t=683s) Bet Leads and Clear DRI Ownership in Rapidly Shifting Environments

Balancing open exploration with operational execution requires a deliberate leadership architecture. While teams need the freedom to explore ambiguous frontiers, the organization cannot devolve into aimless wandering. There must be unmistakable Directly Responsible Individuals (DRIs) who hold decision-making authority over specific domains.

Within the Anthropic Labs environment, these initiatives are organized as discrete "bets," each steered by a dedicated "bet lead": - **Single Point of Decision-Making:** Even when the entire team feels "lost together," the bet lead holds the pen on deciding the next most critical action to derisk, learn, or push forward. - **Resource Allocation Authority:** The lead serves as the ultimate DRI empowered to double down on an initiative, wind it down, or dynamically shift headcount depending on early signal. - **Velocity Driver:** Because frontier technology changes so rapidly, having clear DRI ownership prevents consensus paralysis and keeps exploration disciplined.

[12:09](https://www.youtube.com/watch?v=sEXdyK6woKU&t=729s) The Evolution Toward Agent-Native and Malleable Software

```mermaid flowchart TD accTitle: Evolution Toward Agent-Native Software accDescr: Three progressive eras of AI software integration from sidebars to fully malleable interfaces.

subgraph Era1["Era 1: Disconnected Add-on"] A["Chat Sidebar / Mini Window"] --> B["Customer Support & Q&A"] end

subgraph Era2["Era 2: Agent-Native Architecture"] C["Shared Plumbing / APIs"] --> D["Parity: Agent Can Do Everything a Human Does"] D --> E["Emergent Workflows & Proactive Assistance"] end

subgraph Era3["Era 3: Malleable Interfaces"] F["Agent-Generated Dynamic UIs"] --> G["Personalized, Project-Specific Views"] G --> H["Continuous Maintenance & Iteration by Model"] end

Era1 --> Era2 Era2 --> Era3 ```

The conversation shifts from leadership psychology to the architectural nature of what is being built. Mike Krieger outlines the progression of software design as products transition from human-only interfaces to collaborative environments shared with autonomous agents.

The first era merely bolted AI onto existing workflows as a detached sidebar or mini-window handling Q&A and basic customer support. That model quickly gave way to integrating AI directly into discrete features, and is now advancing toward truly "agent-native" design. Citing the framework popularized by Dan Shipper's publication *Every*—which Krieger notes he turned into an AI prompt skill during his own architectural research—the core premise is simple: "everything a human could do an agent should be able to do as well." While very few products achieve this standard today, unlocking full parity allows emergent behavior to surface, enabling models to connect fragmented systems and proactively suggest workflows.

The frontier beyond parity is "malleable software," where user interfaces are no longer static designs hard-coded by front-end engineers, but fluid surfaces dynamically generated and modified by the agent itself: - **Internal Anthropics Case Study:** For a complex project involving four independent organizational workstreams, Anthropic had Claude monitor real-time cross-team progress and write a bespoke tracking UI for the Technical Program Manager (TPM) and the team. - **Customizability on Demand:** Rather than waiting on internal tools teams or third-party SaaS roadmaps, team members can instruct Claude to alter layout, metrics, and workflows on the fly. - **Living Software:** A significant portion of internal tooling at Anthropic is now built, iterated on, and maintained directly by Claude in continuous response to organizational shifts. - **New Challenges in Governance:** Dynamic software introduces unsolved questions around data provenance and access rights—specifically, reconciling updates made by humans clicking through a UI with background write operations executed by autonomous agents.

[15:34](https://www.youtube.com/watch?v=sEXdyK6woKU&t=934s) Architectural Primitives for Legacy and Scaled SaaS

For traditional SaaS companies that cannot blow up their existing user experiences overnight, Krieger advises focusing on foundational plumbing rather than cosmetic features. When evaluating modern software, it becomes immediately apparent whether a product was engineered so that human-facing actions flow through the exact same underlying primitives as its programmatic endpoints.

While bolting universal agent accessibility onto a twenty-year-old enterprise application is an immense engineering challenge, establishing unified infrastructural plumbing allows companies to evolve at a sustainable pace. Teams can comfortably keep existing sidebar surfaces intact while incrementally experimenting with generative, adaptive UI components underneath.

[16:42](https://www.youtube.com/watch?v=sEXdyK6woKU&t=1002s) Pushing the Frontier vs. Meeting Users Where They Are

Dan Shipper turns the discussion to Ami Vora, noting that while the panel has been exploring the ambitious horizon of agent-native, malleable software built on unified primitives, Anthropic simultaneously serves a broad spectrum of users. Some customers simply want a reliable chat interface to get basic tasks done, while massive enterprise contracts mean Anthropic cannot abruptly overturn core interfaces overnight under the banner of malleability. Dan asks how product leaders balance pushing the technological frontier while preserving a predictable, comprehensible experience that brings along everyday users who are not inclined to adopt experimental workflows on day one.

Ami admits that this tension is an exceptionally difficult product challenge. Anthropic’s core philosophy is simply trying to "meet everyone where they are," especially given how early the industry still is in understanding what interfaces actually succeed. Rather than operating in a stable landscape where teams can design the single "right answer" upfront, the models and user behaviors are shifting every few months, forcing product teams to constantly update their assumptions. Coming from a product background rooted in predictability, simplicity, and reliability—principles designed to foster user muscle memory and intuition—Ami has had to adjust her instincts. In this frontier environment, she encourages teams to "take a lot of shots on goal," even if projects initially look overlapping or duplicative:

> "The fail state is you decide from the beginning like based on a theory of exactly what needs to happen and you like over constrain a product. And I would much rather have like five overlapping products that all work in different ways than like one overly constrained one that we never even gave a chance to actually like meet the market."

- **The shifting landscape:** Model updates and user interaction patterns evolve simultaneously every couple of months, meaning no product architecture is completely settled. - **Exploration vs. consolidation:** Teams must feel comfortable freely exploring unmapped territory, with the explicit mandate to fold successful experiments back into a cohesive, stable product experience once product-market fit is established.

---

[19:54](https://www.youtube.com/watch?v=sEXdyK6woKU&t=1194s) Managing Parallel Bets and Shared Primitives

Dan notes that running multiple parallel experiments is becoming common across AI organizations, but it carries the risk of internal confusion, incoherence, and demotivating teams who feel assigned to a redundant or inferior track. Mike Krieger explains how Anthropic navigates this dilemma internally. First, parallel bets must stem from authentic developer excitement rather than top-down assignments; manufacturing a synthetic alternative track and forcing an unenthusiastic team to build it inevitably results in a subpar product. Second, parallel innovation requires solid architectural foundations. Months earlier, Anthropic conducted an internal inventory and discovered that Claude Chat and Claude Cowork were using completely separate memory architectures, distinct Model Context Protocol (MCP) implementations, and divergent file-storage systems. Any third experimental team building a product was doomed to feel disadvantaged because it lacked shared memory:

```mermaid flowchart TD accTitle: Unified Foundations for Product Experimentation accDescr: Diagram showing how Anthropic's central Foundations team provides shared primitives like memory and MCP to support diverse product bets.

subgraph Core["Foundations Layer"] M["Unified Memory System"] MCP["Model Context Protocol (MCP)"] FS["File Storage Primitives"] end

subgraph Products["Frontier Exploration Tracks"] Chat["Claude Chat Surface"] Cowork["Claude Cowork Surface"] Lab["New Frontier Bets & Labs"] end

M --> Chat M --> Cowork M --> Lab MCP --> Chat MCP --> Cowork MCP --> Lab FS --> Chat FS --> Cowork FS --> Lab ```

By delegating these common primitives to a dedicated Foundations team, experimental teams across Labs and product tracks can pursue overlapping bets while building on a cohesive infrastructure where user context travels seamlessly across surfaces. Ami adds that leaders must simply be candid with their teams about the ambiguity: "We don't know the answer. We don't think anyone knows the answer. We're all going to discover it. And we're all fundamentally on the same team." Naming the friction openly reassures engineers and PMs that parallel bets are an organizational necessity rather than a vote of no confidence.

- **Organic conviction over assigned bets:** Teams only pursue alternative directions if they possess genuine passion and high conviction for that thesis. - **Foundations vs. frontier teams:** Decoupling infrastructure (memory, storage, tool integration) from user-facing exploration ensures experimental features do not end up completely isolated from the rest of the ecosystem. - **Signals of success:** Ami emphasizes that judging experimental wins still relies on traditional product fundamentals: user retention, recurring value, organic word-of-mouth, and whether the tool solves a real problem simply and reliably.

---

[22:19](https://www.youtube.com/watch?v=sEXdyK6woKU&t=1339s) Parking Projects, Computer Use Evals, and Capability Blindness

Dan raises the dilemma of timing: an experiment might fail purely because the underlying model is not quite capable enough, raising the question of when to abandon an idea versus keeping it warm for future model checkpoints. Mike shares how Anthropic Labs handles this by "parking" projects inside an evaluation harness. When Anthropic first experimented with Computer Use internally in 2024, the models were simply incapable of executing the vision. The team initially hypothesized that even if the model struggled with end-to-end automation, it might succeed at educational guidance—such as walking a user through complex tasks in Adobe Photoshop. In practice, the early agentic loop was clumsy and painful: it would try an action, fail, close windows, and after twenty tedious minutes finally complete a step, leaving the user with neither automation nor education:

```mermaid flowchart LR accTitle: The Project Parking and Eval Lifecycle accDescr: Flowchart illustrating how Anthropic parks underperforming product concepts into an eval harness until next-generation models unlock their viability.

A["Prototype Concept<br/>(e.g., Computer Use 2024)"] -->|"Model underperforms"| B["Park Project &<br/>Codify into Eval Harness"] B -->|"Benchmark against new models"| C{"Eval Success Spike?<br/>(e.g., Claude 3.7)"} C -->|"No"| B C -->|"Yes: Success > Failure"| D["Revive into Active Product<br/>Development Track"] ```

Rather than discarding the work, the team converted the project into an automated eval harness and tested it against every subsequent model release. When Claude 3.7 arrived, the eval scores suddenly jumped; reviewing the transcripts confirmed that the model was finally succeeding more often than it failed, signaling that the product track was ready to be reactivated. Mike highlights that proving a model is not yet ready is still a victory if the team externalizes the failure into an evaluation harness that guides researchers.

Dan validates this strategy, warning against what he calls "capability blindness"—a mindset where builders test a capability once, watch it fail, and prematurely write off the entire problem space. A new model release three to six months later can shatter those initial limitations. Dan polls the live summit audience to see who has tested cutting-edge agentic tools, asking for a show of hands for Fable (which garners a strong show of hands) and Astra (which sees fewer adopters). Dan observes that builders frequently tell him a specific workflow is impossible, only to admit they have not tested it against the latest Claude frontier models.

[24:37](https://www.youtube.com/watch?v=sEXdyK6woKU&t=1477s) Integrating Parallel Experiments Without Offloading Cognitive Load

Continuing the thought on adaptability, Dan Shipper emphasizes how important it is to be willing to "throw away everything"—even when it is painful—because otherwise you fall into the trap of assuming that "the last thing that happened is going to be true forever." From there, he raises a thorny structural problem he sees Anthropic grappling with from afar, one that is also happening at OpenAI and that his own team runs into as well: when an organization runs numerous parallel experiments and one of them really starts to work, how do you push that successful experiment back into the core product?

As Dan points out, there are several ways teams try to handle this, and none of them feel like a solved playbook yet. You can add a new tab for every successful experiment, "but then you have a bunch of different tabs and there's all the things that only work in slightly different ways, and that's kind of ugly." Alternatively, you can spin the experiment out as a separate app and never merge them at all. Turning parallel experiments into a single, coherent product remains a critical unsolved challenge across the AI industry.

Ami Vora deadpans, "I don't know what you're talking about," drawing a wave of laughter from the audience before admitting, "You're right. I don't think we've fully figured it out." Returning to an earlier theme from Mike Krieger, she explains that the answer lies in building the right underlying primitives so that products naturally feel the way users expect them to feel. When someone walks up to a product, they want to use it without having to think about it too much. Right now, it is far too easy for AI product teams to transfer a massive amount of cognitive load onto the user by essentially saying, "Here's some tools, why don't you choose?" Instead, Anthropic is iterating on foundational primitives—building a system that genuinely knows the user and presents an accessible UI at all times—by closely observing how people react in practice.

[26:17](https://www.youtube.com/watch?v=sEXdyK6woKU&t=1577s) The Power Law of Tabs and the Art of Saying No

Mike Krieger weighs in with a durable lesson from his years co-founding and scaling Instagram: across every single product, there is a "very strong power law in tab usage." At Instagram, the main feed accounted for roughly 80% of all usage. Even if the team worked hard to make the Explore tab better, its share might only climb from 10% to 15%, because the main feed remained where users landed and represented the heart of the experience.

For AI products today, the fundamental journey is figuring out what that "main tab" experience actually is, making it great, and letting it serve as a natural entry point into a wide variety of downstream experiences. Because of that power law, Krieger notes that "half of the job is saying no to another thing in the sidebar or another tab or another thing." The true art of product development when technology is moving this quickly is giving a parallel experiment enough breathing room to prove itself out—not killing it too early—while ultimately rolling it back into the main experience rather than cluttering the interface.

```mermaid flowchart LR accTitle: From Parallel Experiments to a Coherent Product accDescr: Contrasts fragmented feature bolting with integrating proven experiments into a single core entry point backed by shared primitives.

subgraph Discovery["1. Experimentation"] Exp["Parallel experiments"] Proven["Proven workflow"] Exp --> Proven end

subgraph Fragmented["2a. Fragmented Trap"] Tabs["Extra tabs & sidebar clutter"] Load["High user cognitive load"] Tabs --> Load end

subgraph Coherent["2b. Coherent Product"] Prim["Shared system primitives"] Main["Main tab (80% usage entry point)"] Prim --> Main end

Proven -->|"Bolt on UI"| Tabs Proven -->|"Roll in after validation"| Prim ```

- **The Instagram benchmark:** Roughly 80% of activity happened in the main feed, while secondary tabs like Explore moved between 10% and 15% even after dedicated improvements. - **The product discipline:** Protecting the core entry point by saying no to extra sidebar items and tabs, while resisting the urge to kill promising experiments before they are validated.

[27:04](https://www.youtube.com/watch?v=sEXdyK6woKU&t=1624s) Looking Ahead: Solo Empowerment and Closing the Capability Gap

With time for one final question, Dan asks both guests to imagine being invited back to the Lenny and Friends stage a year from now—joking, "I don't know if he'll invite me back, I've been cursing too much, but he'll invite you two back"—and to share what they expect will be different and how the PM role will have evolved.

Ami hopes that a year from now, people will simply feel like they can do vastly more on their own: building in whatever way they want, getting answers whenever they need them, and starting new businesses. She envisions a wave of small teams and solo empowerment that creates a broad societal uplift. As she puts it, "builders magnify impact for a bunch more people who will never use the tools directly, but who can benefit from people who do build things." Making the act of building dramatically easier is how those benefits reach everyone else.

For Mike, the biggest priority over the coming year is closing the widening gap between what frontier models are capable of and what most people actually use them for. He is emphatic that this gap is not the fault of users: "It's on us. Like we got to build the products to make that possible." That translation challenge was already hard in 2024, became even harder last year, and is harder still today as model capabilities accelerate. Closing that gap is deeply democratizing because it ensures advanced AI isn't reserved only for "your most Claude-pilled software engineer that has this fully working multi-agentic setup." Instead, it brings along the person running parallel research or managing a business—helping them understand the moving parts, check in at the right moments, and execute long-horizon work in a way that feels empowering rather than disempowering.

- **Ami Vora's 1-year vision:** Lowering the barrier to creation so solo builders and small teams can start businesses and build tools that indirectly benefit non-technical users. - **Mike Krieger's 1-year vision:** Bridging the model capability-to-usage gap so everyday researchers and business owners—not just "Claude-pilled" engineers—can orchestrate parallel, long-horizon agentic work.

## Notable quotes > [00:06](https://www.youtube.com/watch?v=sEXdyK6woKU&t=6s) a bridge between the real problems that people have in the world and the technology that you can use to solve it. — Ami Vora

> [02:40](https://www.youtube.com/watch?v=sEXdyK6woKU&t=160s) I don't know. Like, do we need a PM? Like, Claude's got it. Like, we've got a lot going on. — Mike Krieger

> [04:38](https://www.youtube.com/watch?v=sEXdyK6woKU&t=278s) I don't think Claude is a convener yet. — Mike Krieger

> [06:49](https://www.youtube.com/watch?v=sEXdyK6woKU&t=409s) it's actually faster to just build three versions and try them out and be like, 'Oh yeah, this works for this one.' — Ami Vora

> [12:09](https://www.youtube.com/watch?v=sEXdyK6woKU&t=729s) everything a human could do an agent should be able to do as well. — Mike Krieger

> [16:42](https://www.youtube.com/watch?v=sEXdyK6woKU&t=1002s) The fail state is you decide from the beginning like based on a theory of exactly what needs to happen and you like over constrain a product. And I would much rather have like five overlapping products that all work in different ways than like one overly constrained one that we never even gave a chance to actually like meet the market. — Ami Vora

> [19:54](https://www.youtube.com/watch?v=sEXdyK6woKU&t=1194s) We don't know the answer. We don't think anyone knows the answer. We're all going to discover it. And we're all fundamentally on the same team. — Ami Vora

> [26:17](https://www.youtube.com/watch?v=sEXdyK6woKU&t=1577s) half of the job is saying no to another thing in the sidebar or another tab or another thing. — Mike Krieger

> [27:04](https://www.youtube.com/watch?v=sEXdyK6woKU&t=1624s) builders magnify impact for a bunch more people who will never use the tools directly, but who can benefit from people who do build things. — Ami Vora

> [27:04](https://www.youtube.com/watch?v=sEXdyK6woKU&t=1624s) It's on us. Like we got to build the products to make that possible. — Mike Krieger

## Mentioned Lenny and Friends Summit, Anthropic, Ami Vora, Mike Krieger, Dan Shipper, Claude, Slack, Fable, Clayton Christensen, Anthropic Labs, Every, Claude Chat, Claude Cowork, Model Context Protocol (MCP), Computer Use, Adobe Photoshop, Claude 3.7, Astra, OpenAI, Instagram

## Video description What does product management look like when AI changes what teams can build every few months? At the Lenny and Friends Summit, Anthropic’s Ami Vora and Mike Krieger join Dan Shipper to discuss why understanding users, making good decisions, and adapting quickly remain essential to the PM role. They also explore software designed for agents, running experiments in parallel, and bringing the ideas that work into a product people can use.

Recorded live at Lenny and Friends Summit on September 10, 2026, in San Francisco.

[Source](https://www.youtube.com/watch?v=sEXdyK6woKU)

中文简报

Anthropic高管Krieger与Vora：AI编码提速让产品经理角色更加关键由于前沿模型每两个月更迭一次且软件构建成本大幅降低，产品经理必须摒弃前期的UI争论，在多项并行探索中承担起人类“召集人”与决策者的职责。 AI的快速发展非但没有淘汰产品经理，反而提升了其作为人类“召集人”的重要性——负责行使判断力、协调跨职能团队，并将产品牢牢锚定在人类持久的需求痛点上。随着构建成本大幅降低，前期的界面争论逐渐让位于快速原型制作、Agent原生架构以及由模型动态维护的延展性软件。为应对这种不确定性，Anthropic围绕自主的“探索项目负责人”来组织探索，在不同交互界面间统一内存等共享原语，并将有潜力的构想“封存”在评估测试框架中，直至模型能力跟上需求。归根结底，成功的产品把控需要抵制界面冗杂，遵循核心入口的幂律法则，并弥合前沿模型原始能力与日常用户实际赋能之间不断扩大的鸿沟。以往底层技术变革每5到10年发生一次，而如今AI能力大约每2个月就会转变一次；随着工程研发的加速，人类PM作为业务“召集人”和决策者的角色变得比以往更为关键。 (0:06) 由于目前构建软件既快速又廉价，Anthropic的PM跳过了前期关于按钮摆放等理论层面的争论，直接构建并测试多个版本；与此同时，软件架构正从聊天侧边栏演进为与“Agent原生”相提并论的形态，以及由Agent生成的延展性UI。 (6:49) 在发现Claude Chat与Claude Cowork使用各自独立的内存、MCP和文件存储系统后，Anthropic将这些原语集中归至Foundations团队，以便并行探索能够共享上下文；而像早期Computer Use这样表现不及预期的探索项目则被“封存”在自动化评估测试套件中，直至Claude 3.7等更新的模型使其具备可行性。 (19:54) 受“标签页幂律法则”（Instagram主信息流贡献了约80%的使用量）的启发，Anthropic领导层强调要拒绝侧边栏冗杂，并将经过验证的实验功能整合到单一的核心体验中，以缩小前沿模型能力与日常使用之间的差距。 (26:17)

Reading summary

Anthropic product leaders Ami Vora and Mike Krieger argue that while frontier AI dramatically accelerates technical execution, it cannot replace human product managers who act as operational glue, organizational conveners, and decision-makers. As software generation becomes inexpensive, product discipline is shifting from designing static interfaces to maintaining shared architectural primitives and managing experimentation without overloading users. Why AI Cannot Replace the Operational Convener (2:40) Ami Vora notes that while core foundational technology previously shifted every five to ten years, current AI capability cycles turn over roughly every two months, even as core human problems remain unchanged. In response, product management is returning to a fluid problem-solving discipline. Mike Krieger recounts that while working hands-on as an individual contributor at Anthropic, PM Cat demonstrated that technical builders still require dedicated product management to prevent operational gaps. This connective tissue includes aligning diverse user tiers from prosumers to enterprise clients, training customer-facing teams, and coordinating safety safeguards so engineers can stay focused on building. Krieger further argues that AI models like Claude cannot yet act as 'the convener'—the human archetype responsible for bringing people and agents together to execute work with organizational authority and autonomous scheduling. While Claude handles context retrieval, such as discovering cross-team work through Slack search, human PMs remain necessary to exercise judgment across branching paths and maintain alignment under high ambiguity. Unlearning Pre-Code Debates and Framing Ambiguity (6:49) Vora reflects that product teams previously spent decades debating granular interface details—such as button placement and phone-screen reachability—primarily because software was slow and expensive to ship. With AI dramatically lowering the cost of code generation, upfront theoretical debates have become obsolete; teams can immediately generate and test three working variations. Adapting to this reality creates an individual 'innovator's dilemma,' requiring product leaders to abandon hard-won historic skills and embrace domains where they lack prior expertise. To prevent teams from aimless wandering under continuous technological shifts, leadership must frame uncertainty into emotionally safe, structured environments rather than relying on rigid roadmaps. At Anthropic Labs, open-ended initiatives are organized as discrete bets steered by a dedicated 'bet lead' serving as the Directly Responsible Individual (DRI). This lead retains sole authority to prioritize de-risking steps, reallocate headcount, or terminate initiatives, preventing consensus paralysis. Architectural Evolution Toward Agent-Native and Malleable Software (12:09) Krieger outlines three eras of software evolution: isolated sidebar add-ons, agent-native architectures where models achieve parity with human capabilities across shared APIs, and 'malleable software.' In this third paradigm, interfaces are dynamically generated and modified by models rather than statically hard-coded by front-end engineers. For example, Claude monitored progress across four internal Anthropic workstreams and generated a custom tracking UI on demand for a technical program manager. Scaling this malleability requires unified infrastructure. An internal audit at Anthropic revealed that Claude Chat and Claude Cowork initially ran on separate memory architectures, disparate file storage, and divergent Model Context Protocol (MCP) implementations. Centralizing these shared primitives under a dedicated Foundations team ensures context travels seamlessly across surfaces and allows legacy SaaS platforms to incrementally adopt agentic features without fragmenting the user experience. Mitigating Capability Blindness and Interface Clutter (22:19) When early AI prototypes fail, Krieger cautions against 'capability blindness'—the mistake of writing off an entire problem space after a single negative result. When Anthropic tested Computer Use in 2024, early agent loops struggled with automated Photoshop guidance. Instead of discarding the concept, the team converted the project into an automated evaluation harness; when Claude 3.7 was tested against it, benchmark scores jumped, signaling the feature was ready to be revived for active development. Simultaneously, product organizations face difficulties reintegrating successful parallel bets without creating user-facing clutter. Drawing on his experience at Instagram—where the primary feed captured roughly 80% of activity while Explore accounted for only 10% to 15%—Krieger emphasizes the power law of navigation. Rather than adding fragmented tabs or shifting cognitive burdens onto users, product leaders must exercise the discipline to say no to interface clutter and focus on integrating proven workflows into a singular core entry point. Looking ahead, Vora anticipates that lowering software creation barriers will empower solo operators and small teams, while Krieger emphasizes that the central responsibility for product leaders remains closing the expanding gap between frontier model capabilities and mainstream user adoption.

Source quotes

[0:06] Ami Vora: a bridge between the real problems that people have in the world and the technology that you can use to solve it. [2:40] Mike Krieger: I don't know. Like, do we need a PM? Like, Claude's got it. Like, we've got a lot going on. [4:38] Mike Krieger: I don't think Claude is a convener yet. [6:49] Ami Vora: it's actually faster to just build three versions and try them out and be like, 'Oh yeah, this works for this one.' [12:09] Mike Krieger: everything a human could do an agent should be able to do as well. [16:42] Ami Vora: The fail state is you decide from the beginning like based on a theory of exactly what needs to happen and you like over constrain a product. And I would much rather have like five overlapping products that all work in different ways than like one overly constrained one that we never even gave a chance to actually like meet the market. [19:54] Ami Vora: We don't know the answer. We don't think anyone knows the answer. We're all going to discover it. And we're all fundamentally on the same team. [26:17] Mike Krieger: half of the job is saying no to another thing in the sidebar or another tab or another thing. [27:04] Ami Vora: builders magnify impact for a bunch more people who will never use the tools directly, but who can benefit from people who do build things. [27:04] Mike Krieger: It's on us. Like we got to build the products to make that possible.

Original transcript

[0:06] Welcome. How's it going? Good. [0:07] >> Doing great. How are you all? [0:09] >> So nice to see you all. [cheering] [0:11] >> Isn't this a great event? I feel like [0:13] it's just it's so happy. [0:15] >> Yeah. It's good vibes. [0:16] >> Yeah. So, I want to start this [0:18] conversation off with the question that [0:21] I think is on everybody's minds, which [0:23] is a year ago we uh were told that AI [0:28] might make PMs obsolete [laughter] [0:30] and now I feel like the the vibes are [0:34] that everyone needs to become a PM. You [0:36] know, one of the things I love about [0:37] getting to talk to both of you is just [0:40] by being an anthropic and working with [0:42] the models in the way that you do, you [0:44] have a little bit of like a sense of [0:45] where the future is going, especially [0:47] for different job roles. And I'd really [0:48] love for for both of you to talk to us [0:50] about what is a great PM at Enthropic [0:53] look like. Is the PM role going away? Is [0:55] it changing? And if so, how? I think the [0:58] job of product is always to be a bridge [1:00] between the real problems that people [1:03] have in the world and the technology [1:05] that you can use to solve it. And like [1:07] that's always been true. And I think a [1:09] thing that is weird right now is that it [1:12] used to be that the technology would [1:15] change like every 10 years or 5 years [1:18] and now it feels like it changes like [1:19] every two months. And that's really [1:22] hard. That's just like a like as a human [1:24] I feel like it's really hard to learn [1:26] how to deal with that sort of change. [1:28] And so what I always think about is like [1:30] the technology side is changing but the [1:34] human side like the human side doesn't [1:36] change that fast. We have the same [1:38] problems. And so what I always think [1:40] about is like how can I be obsessed with [1:43] like solving the human problems and then [1:46] stay super fresheyed about how the [1:49] technology works? like how can I [1:50] basically throw away most of what I know [1:53] about what used to work to solve this [1:55] problem and then try again like knowing [1:58] everything I do about the problem. [1:59] >> And so I feel like that part hasn't [2:01] changed. Um, what I think has changed is [2:05] it's almost like rewinding the clock. [2:07] Like when I back when the dinosaurs were [2:10] in the earth and I started in product, [2:12] like we didn't know, we didn't have [2:13] definitions. Like we just didn't we [2:15] didn't know what the job was. And so you [2:17] were just like a general problem solver [2:18] and you'd like walk around and try to [2:20] solve a problem until you like ran into [2:21] a wall and then you would ask somebody [2:23] or figure it out. And I feel like that's [2:26] what we can do now is go back to a bunch [2:29] of the roles blurring, but like you just [2:33] keep going and try to figure out what's [2:34] going to work and then when you hit a [2:36] wall, you have all these different sorts [2:39] of scaffolding. [2:40] >> Yeah, that makes sense. Mike, anything [2:41] from you? [2:42] >> I think for me, I had a real like um [2:44] sort of realization moment recently and [2:46] so earlier beginning of the year shifted [2:47] to an IC role. So, I'm mostly building. [2:50] Um, and we were getting close to [2:51] shipping something. And recently, uh, [2:54] and one of the like PM leads works with [2:56] Ammy pulled me aside and was like, [2:57] "Mike, I know you love to build. This [2:59] project's going well. Like, it's looking [3:00] really exciting. Like, we really need [3:02] APM on this project." I was like, "I [3:04] don't know. Like, do we need a PM? Like, [3:05] Claude's got it. Like, we've got a lot [3:07] going on." And she's like, "No, we [3:08] really need a PM." And it like such a [3:10] reminder like once they joined like all [3:11] of the things and all the glue and all [3:13] of the connective tissue and all of the [3:15] things that were about to get dropped if [3:16] I had not done that would that would [3:18] have happened. And it was like just such [3:20] a cool eye opening like before and [3:21] after. Um so much so that the week later [3:23] I pinged her and I was like cat you were [3:25] 100% right. Um [laughter] and thank you. [3:27] And it was just and I think about all [3:29] the things that that encompassed. was, [3:30] you know, actually bringing along, we're [3:32] a company that serves everybody from, [3:34] you know, proumers all the way to really [3:38] large enterprises. Like, have you [3:39] brought all of that contingency along? [3:41] Right? When uh we need to go and figure [3:43] out how to enable our customer success [3:45] team to like talk about the changes as [3:47] questions come up in real time. Has [3:48] somebody thought about that? Has [3:49] somebody looped in safeguards? It's [3:50] obviously very very important. Are [3:52] people getting kept like kept on track? [3:53] All of these things. And nobody has [3:56] infinite capacity even with you know the [3:58] assistance of these tools. So to be able [3:59] to have like call it what you will but [4:02] somebody playing that role of making [4:03] sure that everything is going to go well [4:05] everything is going to get connected. [4:06] we've kept the end user needs in mind, [4:08] Tommy's point, and that the builders who [4:09] are like heads down and are like, you [4:11] know, in full-on cloud mode can keep [4:13] doing that without having to then pop [4:15] out and do the very different activity, [4:16] which is how do you actually make sure [4:18] all these things go well is like I think [4:20] very essential. And I think all that's [4:22] happened now is that because we can move [4:23] faster, that role is like you have to be [4:26] operationally excellent in that role in [4:27] a way that like even beyond what you had [4:28] to be before. But it definitely has not [4:30] gone away. In fact, I have felt this [4:32] like renewed like yes, this is a like if [4:34] you're recruiting for a team and you [4:35] need people to wear these hats, this hat [4:37] is increasingly important. [4:38] >> Can I ask you a serious question though? [4:41] >> So, [4:44] >> only serious questions. Serious. This is [4:47] serious. [4:48] >> Why didn't you just have Fable do it? [4:50] [laughter] [4:50] >> Like, what is And I presumably there's [4:53] some like Fable 7 you've got access to [4:54] that we don't have access to yet. So, [4:56] like what is that person doing that you [4:58] couldn't just have Fable do? I mean it [5:00] is interesting like the we have as you [5:02] can imagine we like run on cloud tag and [5:05] we have a lot of like you know cloud [5:07] provided connective tissue too and it [5:08] does catch things where you say hey like [5:10] there's another thing going on in this [5:11] other part of the organization that you [5:12] should be aware of you know like we [5:13] found it via the slash search. So there [5:15] is that sort of piece as well. I don't [5:17] think claude is a convenor yet. I had [5:19] this um I gave a talk a couple weeks ago [5:21] on like what's the role of the human in [5:22] the like post super powerful AI age and [5:25] there's like a bunch of archetypes I was [5:26] thinking about but one of it was like [5:27] the convenor like somebody still needs [5:29] to bring the clouds and the people [5:31] together to like get the work done and [5:32] so I think that is the role that I cla [5:35] organizational pull or the scheduling [5:37] ability I've yet this is going to be a [5:39] moment Dan when claude schedules a [5:40] meeting with me or for me where I'm like [5:42] who scheduled this meeting it's like oh [5:44] did I thought you two should talk chat [5:45] that has not happened yet [laughter] but [5:46] not because it can't I guess we just [5:48] like turned on that setting, you know. [5:51] >> I [5:53] [laughter] [5:54] >> stay tuned. [5:57] >> I um I think we're also living in a [6:00] world where like the tools are really [6:01] good and you can build a lot of things [6:04] and you can build them really fast. In [6:06] some ways that just like expands the [6:08] universe of what is possible to build [6:11] and you still need to have really good [6:14] judgment about like what to build based [6:16] on sometimes pretty limited information [6:18] cuz it feels like everything's moving [6:19] really fast. So you have to like figure [6:21] out what to build and then I think you [6:23] in some ways you have to be even more [6:25] relentless about making sure it happens [6:27] because there's so many forking paths [6:29] that you could take. Yeah. [6:36] >> I have an idea of what we need to do and [6:38] I have the feedback loop to the people [6:39] who are going to use it and I know what [6:41] the human problems are and I'm just [6:42] going to go for it. [6:43] >> And I think that sort of relentlessness [6:45] and like judgment in the face of [6:47] ambiguity is is really important. [6:49] >> That is definitely something that I have [6:51] found, you know, every time there's a [6:53] shift, every time new models come out, [6:55] it's it's sort of like you're looking [6:56] out on the ocean, you're like, what's [6:58] beyond the horizon? And then you get to [6:59] fast forward beyond the horizon and what [7:01] do you see? Well, there's more stuff. [7:02] There's like a whole new landscape, you [7:04] know? It's not it's not like a utopia [7:06] and it's not like horrible. It's like [7:07] there's just new lots of new things to [7:09] discover and navigate. And I think the [7:11] same is true for just models in general. [7:13] when we were talking Ammy in and in in [7:16] planning for this conversation, one of [7:18] the things that you said and you've had [7:19] a long career in product management um [7:22] but prior to anthropic is that you spent [7:25] decades learning how to answer specific [7:28] kinds of questions as a product leader [7:30] that you now no longer need to answer. [7:34] And so I'm curious about okay, what are [7:36] the skills that you spent such a long [7:38] time building that you actually don't [7:40] feel like are that useful anymore? and [7:42] what are the things that you're leaning [7:43] on now more and then how does that how [7:45] does that filter down into the team and [7:47] your expectations for people that you [7:48] hire? [7:49] >> I think it's really true like I spent a [7:51] really long time trying to understand [7:54] how to put myself in the mindset of a [7:56] user and how something would fit into [7:58] their lives and like what we should [7:59] build because it was so expensive to [8:01] build and ship something. So, you know, [8:04] literally this would happen where people [8:05] would have reviews and what we'd talk [8:07] about is like, am I where should we put [8:08] this button, you know, and we would talk [8:10] about the interaction pattern and where [8:12] like someone's finger would go on their [8:14] phone and what situation they they would [8:16] be in where they pulled their phone out [8:18] of their pocket and what they wanted. [8:20] And I feel like I tried really hard to [8:21] get good at that at that mix of like [8:24] product and design. And I just like I [8:26] don't think anyone's ever going to ask [8:28] me that question again. [8:29] >> That's a fable. That's a fable one. [8:30] [laughter] [8:30] >> That's in the past. Like now it's [8:32] actually faster to just build three [8:34] versions and try them out and be like, [8:36] "Oh yeah, this works for this one." So a [8:38] thing that I prided myself on is like [8:41] not actually that useful anymore. And I [8:43] do think that's really hard just in [8:45] terms of identity. We talk about the [8:47] innovator's dilemma when it comes to [8:49] companies where like you're so good at [8:51] one thing, you keep on trying to do it [8:52] even when there's this other thing that [8:54] maybe you should try doing, but you [8:56] can't because you're already so good at [8:57] one thing. And I feel like that's also [9:00] true on a personal level where I [9:02] sometimes feel like, oh, I know what I'm [9:04] good at. I'm just going to keep on [9:05] trying to do that even when there might [9:08] be a different way to do it that might [9:10] be more important in the future. And so [9:11] that's what I just encourage everyone to [9:13] think about. Like when you think about [9:14] that product bridge, part of the [9:16] technology changing every two months is [9:19] you have to throw away what you knew [9:21] about yourself in the job using that [9:23] technology and just try something [9:26] different. And most of the time I don't [9:28] know if I'm going to be any good at it, [9:30] you know, like I've never worked at a [9:31] model company. Like I just this is not a [9:33] place where I expertise and I I just [9:35] kind of I feel like every couple of [9:36] months I have to constantly be like, [9:38] "All right, well, let's try this other [9:40] thing and see if I can figure it out." [9:42] You know, [9:43] >> what are the characteristics of someone [9:44] that is able to do that well? And what [9:46] have you had to develop in order to do [9:48] that well? [9:48] >> I mean, I think it comes back to the the [9:50] sorts of things we were talking about. I [9:51] think like raising your ceiling on [9:55] tolerance for change and being willing [9:57] to embrace change is super important. [9:59] Like so I think adaptability I mean I'm [10:02] sure this is not new. Like adaptability [10:04] is just such a key thing given how [10:06] rapidly the world is changing. Um, I [10:09] think that sense of like judgment and [10:11] relentlessness as we talked about, those [10:13] are like the things that I keep coming [10:15] back to and really what I try to think [10:17] about for the team is like I think it's [10:20] important to talk about those things [10:22] because otherwise it can feel really [10:24] lonely like you're the one who has to [10:27] constantly adapt but actually everybody [10:28] has to constantly adapt and I think it's [10:30] hard on all of us. But I also talk about [10:34] trying to frame the chaos so it doesn't [10:36] feel as taxing, [10:38] >> you know, like I think I think when [10:40] there's times of uncertainty sometimes [10:43] it's tempting to like try to map out [10:45] here's exactly what's going to happen. [10:47] Like here's exactly the products we're [10:48] going to build. Here's exactly how my [10:50] career is going to go, you know, like [10:51] and that's a way to exert control. And I [10:54] just I I think that locks you out of [10:57] trying all the new stuff. And so what we [10:59] have to do is like frame the chaos so it [11:01] feels safe and it feels plausible to [11:04] engage with it. But that's where like a [11:06] bunch of the magic and productivity is [11:08] going to be. So just like talking about [11:10] it, making it safe, acknowledging that [11:13] there's also emotions around all of this [11:15] and it's not easy, but it is a way to [11:19] just participate and understand and work [11:21] with the future. I think that's [11:22] something we talk about a lot. I think [11:24] there's an organizational leadership [11:25] question that I think Amy you do really [11:26] well which is like you can have the [11:27] chaos but you also need very clear [11:29] leadership and DRI ship within some [11:32] areas. So like you can have a bunch of [11:34] people sort of exploring something very [11:36] frontier but having a clear like this is [11:38] the person who's ultimately going to [11:39] make the call whether this is you know [11:41] the thing to pursue or not. In labs we [11:43] call them leads or like bet we have call [11:44] them bets within labs. So the bet lead [11:46] is a really important role and they are [11:47] the DRRi for saying like I think we [11:49] should double down on this or like [11:50] actually we should wind this down. this [11:51] team needs more people, this team needs [11:53] less. And like that role is even more [11:56] important because we're moving so [11:57] quickly, because things are shifting [11:58] really rapidly to still have that like [12:00] we're all lost together, but somebody's [12:02] got the like, you know, the pen on like [12:04] what what's the next most important [12:06] thing we can do to derisk, to [12:07] understand, to learn, to move things [12:09] forward, too. [12:09] >> We've talked a lot uh already so far [12:12] about uh the psychology and the skills [12:14] of being a product leader. Um, one other [12:17] place that's changing a lot is what you [12:21] actually build and how you think about [12:23] like the features and the functionality [12:25] in particular because we're moving from [12:27] a world where you'd expect only humans [12:29] to use your software to a world where uh [12:32] humans and agents are using software. [12:34] Sometimes it's delegated, sometimes it's [12:35] fully collaborative. And I know Mike, [12:37] you've sort of been at the center of [12:39] figuring out what is what does [12:40] agentnative type software look like? And [12:43] I think you have a lot of thoughts on [12:44] where that is going. So, can you talk to [12:47] us about the uh h how you think about [12:50] what to build, what goes into a product [12:52] versus an agent and uh and what's what's [12:55] the overlap? [12:56] >> I love this question and like I maybe [12:57] there's like a timeline to look at where [13:00] V1 or like er 1 was great, we've got [13:02] this AI thing, maybe it belongs in like [13:05] a sidebar or like a little mini window [13:07] and you can ask questions and it's kind [13:08] of disconnected and if you know maybe it [13:10] helps you with some customer support [13:11] questions. Um and then sort of greater [13:14] integrations where maybe there are whole [13:15] features that are AI powered and then [13:16] this movement towards being more agent [13:18] native and I think every like literally [13:20] wrote the book on it which is and the [13:21] way I for that the whole year where I [13:23] was working a lot in this idea of like [13:25] agent native architectures I would [13:26] basically feed their like write up on [13:28] this and then create a skill out of it [13:29] because I was really good encapsulation [13:31] of it was the idea of like everything a [13:33] human could do an agent should be able [13:35] to do as well and very few products [13:36] actually get this right. I think a lot [13:38] of our products like don't quite get [13:39] this right yet. But when you do that, [13:41] all of a sudden sort of emergent [13:43] behavior gets unlocked where you know [13:45] the agent can sort of piece together [13:46] things that weren't pieced together [13:47] before or can proactively offer um [13:50] different ways of doing things. I think [13:51] we're still on the journey of making [13:53] that good, but I think the next stage [13:54] that I've been thinking about a lot is [13:56] when the interfaces actually become [13:58] something that is malleable by the agent [14:00] too. And we've talked about mal [14:01] malleable software for years. I feel [14:03] like it's actually now like really [14:05] manifesting and becoming real. And um [14:07] maybe back to the sort of role of [14:09] product and connecting to this question, [14:10] one of the things that we have, we have [14:12] like a pretty complex thing that we're [14:13] trying to ship right now. And it's got [14:15] at least four independent orc streams. [14:17] And the way we've operationalized it is [14:19] we've had Claude both sort of keep an [14:21] eye on how everything is going, but also [14:22] create the UI by which first the TPM and [14:25] then the entire team understands what is [14:27] happening in that project. And what's [14:29] really cool about that is that if you [14:30] didn't like that particular display, [14:32] you're not bound to like however it was [14:34] determined to be built either by some [14:36] internal acceleration team or some uh [14:38] sort of third party. You can change it [14:40] um and you can change it along with [14:41] cloud and it's just been a very cool [14:43] thing to see internally. Like one of the [14:45] biggest shifts in how we've worked in [14:46] the last year I think at Enthropic is so [14:48] much of the software we interact with is [14:50] like built maintained and iterated on by [14:52] claude. of course reacting to all the [14:54] things that are happening in the [14:54] organization but very much like changing [14:56] all the time and so that leads to all [14:58] sorts of other interesting questions [14:59] like who can update the data right is it [15:01] just the humans clicking through is it [15:02] also clawed in the background how do you [15:04] have like provenence of data it opens up [15:06] some like interesting questions there as [15:08] well but I feel like that is the piece [15:09] like the both the trend and the thing [15:11] I'm most inspired by right now is how do [15:13] we make that the way almost all of our [15:15] software operates and of course you [15:16] still want building blocks and [15:18] predictability and design systems and [15:20] you don't want you want to frame the [15:21] cats you want to like make it not [15:22] absolutely insane, but for the given [15:24] task, for that given project, for even [15:26] that given user potentially, what is the [15:28] way in which we pull in the right um [15:30] ways and make that really feel both [15:32] personal and like extremely useful [15:33] because of that [15:34] >> and how does someone think about maybe [15:36] you know you're not in anthropic and [15:38] you're you're have you have a SAS app [15:39] that's scaled and you're thinking about [15:41] okay, do I put an agent in it? Do I make [15:43] it open and agent native so that cloud [15:45] can be in it? like how do how do you [15:46] think about those choices and what a [15:49] company should care about versus uh [15:51] leave to leave to a model company like [15:53] Anthropic to to plug in? Yeah, I think a [15:55] lot of it is really building the right [15:57] primitives into the product that almost [15:59] like an infrastructural layer and you [16:00] can kind of tell like I I always poke at [16:02] the whenever I'm using new products um [16:04] I'll ask it to do something that you [16:06] know a human could do but you know might [16:07] be required somebody to think through or [16:10] architect the thing where everything [16:11] goes through some shared plumbing that [16:12] like both agents and like whatever sort [16:14] of rest API you've built can go through [16:16] as well and you can tell when they've [16:18] been thought through that way and it's I [16:19] have a lot of empathy because if you've [16:21] been building an application for 20 [16:23] years just bolting that on is difficult. [16:25] So I think is a journey that a lot of [16:26] companies are going on. But once you get [16:28] it right, I think that question becomes [16:30] you can just evolve it over time, right? [16:31] You can start with, okay, at first I'm [16:33] just going to have maybe more of the [16:34] side paint thing. I don't I'm not ready [16:35] to blow up my whole UI. I'm not ready to [16:37] create fully generative UIs, but maybe [16:39] you start experimenting with like, all [16:40] right, there's also maybe a personal [16:42] landing page that can be more malleable [16:43] and is using the same primitives. You're [16:45] not inventing a whole new sort of [16:46] infrastructure underneath. So I think [16:48] that you can evolve your product along [16:49] the way, but you have like you can tell [16:51] when things feel super bolted on and [16:53] like like that's you know feeling like [16:55] it's not a native part of the product. [16:57] >> So Ammy I'm curious we're hearing we can [17:01] go all the way to you know agent native [17:04] malleable software but uh anthropic you [17:07] have to serve this whole spectrum of [17:08] people. there's this uh challenge of [17:11] pushing the frontier and also you know [17:14] some people just want to use cloud to [17:15] like chat or you've you you have big [17:17] enterprise contracts where you can't [17:19] just like rip out their whole interface [17:20] and then say hey it's malleable like you [17:22] know on on in one day or maybe you can't [17:25] I don't know but I assume you can't so [17:27] how do you balance [17:29] >> [snorts] [17:29] >> uh in terms of thinking about product [17:32] pushing the frontier and creating a a [17:35] consistent experience for users that [17:37] they can understand and bringing along [17:39] the people who are maybe not going to [17:40] try all the new things on day one. [17:42] >> I think it's a really hard problem. Um I [17:44] mean our approach is to just try to meet [17:47] everyone where they are. I think we're [17:49] just so early in like understanding what [17:51] works. It's not like stuff is stable and [17:53] we can just build the right answer. It's [17:56] like we're all trying to explore [17:58] together. And part of the everything [17:59] changing every couple months is like the [18:02] models change, but also like the way [18:03] people use products and the models [18:05] changes and we kind of constantly have [18:07] to update for that. And so a thing that [18:09] I think about is like when we're just [18:11] exploring a a product like like Mike's [18:14] talking about, it's okay to just try [18:16] stuff like we just have to try a bunch [18:18] of things because we haven't actually [18:20] explored the frontier. And then when we [18:22] figure out what works, then we have to [18:25] like make it make sense for all the [18:27] people who want something that feels [18:30] more stable and complete. Um, I think [18:32] this is a really tough trade-off [18:34] because, you know, like raised in like [18:35] the the product like world that I was, I [18:38] was all about like simplicity and [18:41] reliability and total predictability [18:42] because I think that helps people build [18:44] like a muscle memory and an intuition. [18:47] But I also think the reality is like we [18:50] just don't know yet what a lot of things [18:52] should look like. And so I actually [18:53] encourage the team to just like take a [18:56] lot of shots on goal. You know, it's [18:58] okay to have things that feel somewhat [19:00] overlapping or duplicative as long as [19:03] when we figure them out. When we hit [19:05] product market fit, we go back and we [19:07] reinccorporate into something that works [19:09] for everyone. Because I feel like the [19:10] fail state is you decide from the [19:12] beginning like based on a theory of [19:15] exactly what needs to happen and you [19:17] like over constrain a product. And I [19:20] would much rather have like five [19:22] overlapping products that all work in [19:25] different ways than like one overly [19:27] constrained one that we never even gave [19:28] a chance to actually like meet the [19:30] market. And how do you and I definitely [19:32] see this we we do this at every but like [19:34] I see this all around the industry just [19:36] doing lots of parallel experiments and [19:38] seeing which ones one one works and then [19:40] starting to consolidate as you start to [19:42] get uh get signal. How do you do that in [19:44] a way that doesn't feel confusing or [19:47] incoherent or you know if I'm working on [19:50] the third team doing this thing like [19:52] it's can be maybe demotivating like how [19:54] do you do that? Well [19:54] >> I think there's a couple of things. one [19:56] like I think when we were just having [19:58] this conversation this morning [19:59] internally where you know there was like [20:00] two potential like product directions [20:02] and one has like a team with very high [20:04] conviction going after and the other one [20:05] doesn't really and like it feels very [20:07] manufactured to be like we should [20:08] explore this other idea cuz we think [20:10] it's good but nobody's like really [20:11] calling for it then that's like that [20:13] kind of B team feeling that I think I've [20:15] never been like successful tasking a [20:17] team to go like take a product area that [20:19] they're not excited about like the [20:20] product is going to suck at the end of [20:21] that process right um so like you want [20:23] you want the like excitement but if [20:24] there are two three different teams that [20:26] have exciting directions and yeah like [20:27] let that let that explore. I think I I [20:30] think we've made a lot of progress on [20:31] even in the last six months is like [20:33] having the underlying infrastructure to [20:35] support all of that. So, we inventoried [20:37] this. It feels like 6 months ago. It's [20:39] probably only three, but like there was [20:41] a point where like chat and co-work had [20:43] different memory systems and different [20:44] MCP and different um ways of storing [20:46] files. And it was like, okay, like for [20:48] anybody to experiment on top of the [20:50] surface, they're going to have a really [20:52] hard time because if they built a third [20:53] thing that then doesn't share memory [20:55] with any of those, for example, it's [20:56] going to feel disconnected and like [20:57] almost disadvantaged from the beginning. [20:59] So, a lot of the work, you know, we have [21:01] like a foundations team that has like [21:02] been just thinking about like you should [21:03] be able to get your memory wherever you [21:04] are within our products. And like it [21:06] sounds like a simple thing but then it's [21:07] very clearly you pull back the thing [21:08] like oh it's complicated for these eight [21:10] reasons but it's worth solving that and [21:12] solving it in a different team than like [21:13] your labsy innovation or like kind of [21:15] frontier product teams because then like [21:17] that those experiments can feel [21:18] complimentary even if they are [21:20] overlapping too. [21:20] >> I also think like this sounds so simple [21:22] but I also think it's just really [21:24] important to talk about it and just name [21:26] like hey if there's three different bets [21:28] sometimes it can feel demoralizing or [21:30] like one of you is not set up for [21:32] success. That's okay. That's the world [21:34] we're living in. And so what we're [21:35] asking you to do is come along this [21:37] journey with us. We don't know the [21:39] answer. We don't think anyone knows the [21:40] answer. We're all going to discover it. [21:42] And we're all fundamentally on the same [21:44] team. And so we will just like learn [21:46] from whatever works. And just like [21:48] saying stuff like that out loud, I think [21:50] is really important because everybody is [21:52] like trying to figure everything out in [21:55] the dark. [21:55] >> What do you look for to know that [21:56] something's starting to work? [21:58] >> I mean, I think a bunch of that is kind [21:59] of the normal product market fits. Like [22:01] do people use it? Do they like it? Do [22:02] they come back? Do they talk about it? [22:04] Do they get value? Like again going back [22:06] to like the goal of product has remained [22:08] the same. It's like to solve problems. [22:11] Thinking about does this actually solve [22:12] a problem in a way that people can use [22:14] and use easily and feel good about using [22:17] it comes back to the basics for that. [22:19] >> One of the interesting things that feels [22:20] challenging to me about that is [22:22] sometimes a thing looks like it doesn't [22:24] work and the new model comes out and [22:25] then it really starts working and you [22:27] don't want to like throw it away too [22:28] early but you also don't want to be too [22:29] far ahead. So how do you do that? Like [22:31] one of the things we do in labs [22:32] sometimes is just park a project. Like [22:34] when we built we built the first like [22:36] computer use product inside Anthropic in [22:38] 2024. Um and it was so bad because the [22:41] models just weren't there yet. And you [22:42] would be you know our first thesis was [22:44] well it's not going to be good at a [22:45] bunch of automation but it might be good [22:47] about education. Um, so you'd be like, [22:48] "All right, uh, how do I, you know, [22:50] change this, you know, thing in some [22:52] complex software? Like, how do I, you [22:54] know, do this really complex thing in [22:56] Photoshop?" And it would get there, but [22:57] it would get there in like the very like [22:58] early agent cloud way, which was like [23:00] this, did that work? Nope, that didn't [23:01] work. I got to close this. It was like [23:02] 20 minutes later like, I did it. You're [23:04] like, I learned nothing. I learn, you [23:05] know, so like it was neither good for [23:06] automation nor education. But we parked [23:08] it and then we would basically just try [23:10] it with every new model. And we [23:11] actually, the way we would try it is we [23:12] actually just had it like running in an [23:14] eval harness. And the way we actually [23:16] learned that 37 was like a like a leap [23:18] in computer use was like just seeing it [23:20] suddenly took over and then we actually [23:22] looked at the transcript. We're like wow [23:23] it's succeeding more often than it's not [23:25] like this is a real moment. So it is [23:27] really valuable like we see it as a win [23:29] in labs if what you did was all if all [23:31] your project did is show that the models [23:33] aren't ready for it yet but you're able [23:34] to now like externalize that as an eval [23:36] or sort of like create some uh [23:39] connectivity to the research team like [23:41] that's great because it means that 2 3 4 [23:43] 5 6 months from now we'll be able to [23:45] revisit and hopefully we've learned [23:46] something along the way um as well but I [23:48] think you need you need to build early [23:49] so that you can even have that [23:50] intuition. [23:51] >> I definitely agree. I've been calling [23:53] this like it's it's sort there's a sort [23:55] of you can get capability blind if you [23:57] try something and then you don't ever [23:58] try it again or you don't make an eval [24:00] and you're like oh the models are just [24:01] never going to do this and then it just [24:02] turns out a model comes out like in 3 [24:04] months and it just completely chang it [24:06] it improves in ways you never would have [24:08] thought it would. I'm I'm actually [24:09] curious if you're here how many of you [24:11] have used Fable? Raise your hand. Oh, a [24:14] lot of people are using Okay, this is [24:15] great. This is a good What about Astra? [24:18] Fewer. Interesting. Okay, so this is [24:20] great. you guys are you guys are on the [24:22] you guys are on the edge. This is this [24:23] is very helpful because one of the [24:25] things I find a lot is I talk to someone [24:27] and they're like, "Oh, it doesn't do [24:28] this." And I'm like, "Have you used the [24:30] latest model?" And they're like, "No, I [24:31] used like Sonnet 45 or what?" Like [24:33] whatever. And uh that that really [24:35] changes things. [24:36] >> I think that's also part of why it's so [24:37] important to be willing to throw away [24:39] everything, you know. [24:40] >> Yeah. [24:40] >> Even though it's hard because then [24:42] otherwise you just like think that the [24:44] last thing that happened is going to be [24:45] true forever. One one big question that [24:48] I I see you guys working on from afar [24:51] and I see this happening at OpenAI 2 and [24:53] and we run into it a little bit is when [24:55] you're doing all these parallel [24:56] experiments and one starts to really [24:58] work then you have to like push it back [25:00] into your actual product and there are [25:02] several different ways to do it. It's [25:04] like okay you can make a tab but then [25:05] you have a bunch of different tabs and [25:06] there's all all the things like only [25:08] work in like in slightly different ways [25:10] and that's kind of ugly. Or you can have [25:11] it just be a separate app and you never [25:13] merge them. There's all these different [25:14] things that I don't think anyone has [25:16] really come to this is the answer, this [25:18] is how you do it. But it seems like a [25:20] critical problem is how to take those [25:23] parallel experiments and turn it into a [25:25] coherent product. How have you what have [25:28] you learned and how do you think about [25:29] it? [25:29] >> I don't know what you're talking about. [25:30] [laughter] [25:34] Um, you're right. I don't think we've [25:36] fully figured it out. But I think some [25:37] of it comes back to what Mike was [25:38] talking about of making sure that we're [25:40] building primitives so that our products [25:43] feel like the way they should feel, you [25:45] know, when you're using them. Like what [25:46] does a user expect? The user expects to [25:48] walk up to something and like hopefully [25:50] not have to think about it too much. [25:52] Like and right now it's easy to just [25:54] transfer a bunch of cognitive load to [25:56] the user and be like here's some tools [25:58] like why don't you choose? And so one of [26:01] the things we that that Mike's talking [26:03] about is like how do we make sure that [26:05] we build primitives so the right system [26:08] knows you. So a UI that feels like [26:11] really accessible all the time. And I [26:14] think all of that is a bit iterative as [26:15] we see how people react. [26:17] >> I think overall like if I had to bet [26:19] like there's some lessons you'll learn [26:21] and some that are are real. Like the [26:22] thing that we always saw at Instagram [26:23] it's like there's a very strong power [26:25] law in tab usage. It makes sense. every [26:27] single product right it's like main feed [26:29] was like 80% of the use right and you [26:31] can make explore better and maybe you [26:32] would go from 10 to 15% but it was like [26:35] main feed that is where you land that is [26:36] your big thing so I think that is will [26:38] always be the journey that we're all on [26:40] which is like what is the main what is [26:42] the main tab like what is that [26:43] experience how do you still make that [26:44] great and then luckily it can be an [26:46] entry point to a lot of other downstream [26:48] interesting experiences but yeah I think [26:50] half of the job is saying no to another [26:53] thing in the sidebar or another tab or [26:55] another thing and figuring out [26:56] Can this be rolled in once we've proven [26:58] it out but not killing it so early. So [27:00] this is the entire I think art of [27:01] product development I think as things [27:03] are moving this quickly. [27:04] >> All right, we've got time for one last [27:06] question which I'd love to hear from [27:07] both of you which is if if we are on [27:10] this stage here next year and Lenny [27:12] invites us back. I don't know if he'll [27:13] invite me back. I've been cursing too [27:14] much but he'll invite you two back. What [27:16] what do you expect will be different and [27:18] and how will being a PM have changed? I [27:21] hope it just like that that people feel [27:24] like they can just do much more like [27:26] they can just go and build in whatever [27:28] way they want. They can get answers [27:29] whenever they they want. They can start [27:31] businesses and like and we just see a [27:34] lot more like small teams and solo [27:36] empowerment and we we just we start to [27:39] get the uplift of like what that means [27:41] for everyone else. Like I think it I [27:43] think it's always like builders magnify [27:45] impact for a bunch more people who will [27:48] never like use the tools directly but [27:50] who can benefit from people who do build [27:53] things. And so I just I hope that's just [27:55] way easier. [27:56] >> Yeah. I think the biggest one for me is [27:58] you know right now like I often talk [28:01] about like the gap between what the [28:02] models are capable of and like what most [28:03] folks are using them for. Not through [28:05] any fault of their own. It's on us. Like [28:06] we got to build the products to make [28:07] that possible. And like that was a hard [28:09] thing last thing 24. was even harder [28:11] last year. It's getting even harder now. [28:13] And so like the biggest thing I want to [28:15] do is close I think it's very [28:16] democratizing if you can close that gap [28:17] because it means it's not just your most [28:20] like clawpilled software engineer that [28:22] has you know like this fully working [28:24] multi-ogenic setup. It's also the person [28:25] that is doing like a bunch of research [28:27] in parallel. It's also the person that's [28:28] trying to manage their business but they [28:30] feel like they've unlocked it and we've [28:31] brought them along the way in terms of [28:33] how do you like understand the moving [28:35] parts, check in at the right places, do [28:37] long horizon work but not in a way that [28:39] feels disempowering. That's the art. [28:40] Like if we succeed a year from now, [28:42] that'll be closer. [28:43] >> Amazing. Thank you so much and Mike to [28:46] [applause] see you.