CHAPTERS
- 0:02 – 1:06
Panel setup and why investors are coding again
The moderator (Lauren, Sequoia) opens with a light exchange about writing code daily and how internal tools help find founders and ideas. This frames the panel around practical building with AI, not just theory.
- •Panelists are introduced and the session is framed around building with AI
- •Lauren shares Sequoia uses small scripts/products to support investing workflows
- •Sets expectation: focus on how products and building methods are changing
- 1:06 – 2:44
What each company is building: personal CFO, app-building platform, and GTM AI
Each panelist introduces their product and the users they serve. The breadth (consumer finance, SMB app creation, enterprise go-to-market) sets up cross-industry patterns in AI-native design.
- •Sylvia/ProCap: AI personal CFO aggregating accounts and delivering financial guidance
- •Emergent: AI app-building platform producing production-ready full-stack apps
- •Clay: go-to-market AI for finding and converting best customers
- •Emphasis on reaching non-traditional or non-technical users
- 2:44 – 4:13
Capabilities unlocked by modern models: enrichment → action
Clay’s evolution illustrates how LLM APIs expanded from data extraction to taking context-aware action. The conversation highlights moving from “get a datapoint” to “use it to drive real outcomes.”
- •Clay’s original mission: bring programming-like power to more people
- •LLM APIs enable custom web scraping/enrichment tailored to a business
- •Shift from enrichment to actions like personalized outreach and content
- •Models enable workflows that were impractical or impossible pre-AI
- 4:13 – 5:10
Creative AI-native GTM in the wild: the Waste Management dumpster-color story
A concrete example shows AI enabling unconventional go-to-market strategies using multimodal analysis. The panel reacts to how “non-obvious” use cases are becoming feasible and scalable.
- •Using Google Street View/satellite imagery to infer competitor usage via dumpster color
- •Image analysis + automation triggers direct-mail conversion campaigns
- •Demonstrates multimodal models enabling novel targeting signals
- •Highlights AI as a force multiplier for human creativity
- 5:10 – 7:43
Empowering non-technical builders: production-ready apps and workflow discipline
Emergent’s examples show small business owners building sophisticated apps with agent support. Success comes from learning to decompose problems into controlled workflows to manage hallucinations and reliability.
- •Non-technical users can build sophisticated, agent-enabled applications
- •Effective use requires breaking work into small workflows to reduce errors
- •Examples include a Norway builder creating an Airbnb-like real estate platform
- •“Right tools” can let non-traditional builders outperform expectations
- 7:43 – 9:07
Ambient agents and the token-burn lesson: designing proactivity in consumer finance
Sylvia’s team explains why proactive, background agents matter for consumer retention—and the cost pitfalls. They describe shifting from assumed defaults to learning from user behavior to define what automation should run.
- •Consumers don’t want dashboards; they want alerts/insights delivered proactively
- •Early “semantic crons” ran 24/7 and caused massive token spend
- •Key mistake: guessing what users wanted instead of observing behavior
- •Solution: let users do tasks manually first, then infer good defaults
- 9:07 – 11:00
Building agent infra: autonomy + rapid feedback loops over heavy scaffolding
Emergent details architectural choices: avoid overengineering around the model and instead give agents autonomy with fast feedback. Custom infra (sandboxes, logs, runtime visibility) makes agents behave more like cloud-based developers.
- •Don’t over-scaffold the model; let improved models do more work directly
- •Agents need rapid feedback: logs, DB access, runtime inspection
- •Built custom sandbox/container tech and agent infra from scratch
- •Architecture becomes more resilient to new model releases
- 11:00 – 14:31
Pricing and packaging in AI: separating cost coverage from value capture
The panel debates pricing beyond “tokens + margin.” Clay shares a two-metric approach to keep costs covered while still capturing value, while Emergent distinguishes between token-heavy build phases and SaaS-like deployment revenue.
- •AI pricing often regressed to cost-plus due to token costs
- •Outcomes-based pricing is appealing but hard to execute broadly
- •Clay’s approach: bifurcate pricing into low-margin cost coverage + high-margin value metric
- •Emergent: build phase correlates with token usage; deployment resembles SaaS/service pricing
- 14:31 – 14:51
Service is software: positioning, free tiers, and services around AI products
Shane notes Sylvia is currently free, then emphasizes a broader trend: software and service blending. The discussion suggests monetization may increasingly come from services and deployment/operations around AI systems.
- •Sylvia is free (for now), highlighting experimentation in business models
- •Trend: services wrapped around AI products become a core revenue lever
- •Quote/paraphrase: “service is software and software is service”
- •Operational delivery (deployment, support, customization) becomes part of the product
- 14:51 – 16:56
What’s surprised them: smarter models can be cheaper, but trust/brand standards persist
Panelists share recent surprises: model upgrades routinely exceed expectations and can reduce costs through fewer wasted steps. At the same time, enterprises still demand strict brand control and human review in outbound messaging.
- •New model releases often outperform expectations on capability and efficiency
- •Smarter models can reduce cost by avoiding wasteful reasoning loops
- •Enterprise GTM still requires human scrutiny for brand voice and compliance
- •Reliability improvements are felt, but adoption practices lag behind
- 16:56 – 17:51
Memory and file systems: personalization as a major product shift
Shane highlights integrating file-system-like memory to enable persistent, managed user context. This unlocks “hyper-personalization” and developer-facing primitives (e.g., a managed file system in a few lines of code).
- •File-system/memory features enable persistent preferences, skills, and context
- •Silvia manages user file systems behind the scenes for personalization
- •Expectation: memory-driven products become a defining trend into 2026
- •Launch of developer product (Trove Files) to provide managed memory/filesystem primitives
- 17:51 – 20:34
Product affordances + adaptability: building evals and rebuilding constantly
The conversation turns to what builders want next: better memory, better affordances, and continuous adaptability. Shane stresses internal evals as the mechanism to safely iterate and occasionally scrap major architecture decisions.
- •Purpose-built affordances help users act without starting from a blank page
- •Balance: purpose-built UX with flexible primitives as models evolve
- •Teams re-architect frequently; adaptability is a competitive advantage
- •Internal evals enable comparing versions/tools and deciding when to rebuild
- 20:34 – 23:14
Choosing bets in an unpredictable market: stay close to customers and close the loop
As model trajectories get harder to forecast, the panel emphasizes customer proximity and end-to-end problem solving. They highlight trade-offs like APIs/CLI vs product surface, and the need to deliver full outcomes (build + deploy + reliability).
- •Prediction is hard; teams must move fast and adapt continuously
- •Clay weighs surfaces: product UX vs API/CLI access where customers work
- •Emergent stresses customers still struggle with basics; reliability is the real battle
- •“Close the loop”: don’t leave users to stitch together building and deployment themselves
- 23:14 – 26:32
Advice for founders: follow curiosity, don’t fear crowded markets, embrace discomfort
Panelists close with guidance for new builders: pursue curiosity-driven exploration, validate with real usage, and be willing to rip up architectures. Even crowded spaces remain open because many products don’t truly solve user pain.
- •Follow curiosity; greatness often emerges via iterative stepping stones
- •Crowded markets aren’t closed—many incumbents still ‘don’t work’ for users
- •Get hands-on with products/customers to find unmet pain points
- •Seek discomfort: frequent rebuilds and adaptation are normal in AI-native products
