Noah Shinn - Building Instinct: The Personal Agent - [Invest Like the Best, EP.493]

Summary

In this episode, Noah Shinn describes Instinct as a personal assistant that behaves more like a person than an app, reachable by text, call, email, and eventually more ambient voice interfaces. He argues the long-term UI for software will collapse into one simple surface while capability expands, and says the product should be judged on understandability and trust, not just raw model capability. He also lays out the company’s economics: over $1B in annual transaction volume, about half from travel, growing around 10% day over day on an invite-only basis with no marketing spend. Much of the conversation focuses on agent safety, including content firewalls, a separate monitoring layer, and a strong stance against agents influencing users against their own interests. Noah frames the business as a distribution-and-take-rate play, comparable to consumer payment and platform models, while warning that compute planning and inference efficiency are central operational challenges.

Notes

• Instinct is positioned as a personal assistant / personal agent company, not a new app category.
• The product is designed around a phone, a computer, text, call, and email as native channels.
• Instinct can proactively call the user when urgency matters.
• The product already supports end-to-end consumer tasks rather than only recommendations.
• Wardrobe scanning is one recurring use case, with outfit planning and shopping recommendations downstream.
• Users can connect bank accounts so Instinct can inspect transactions and subscriptions.
• Instinct can complete cancellation flows end to end and return the savings amount.
• The trusted network lets two users' agents coordinate directly to find time without back-and-forth.
• Permissioning can vary by relationship, from full spouse access to narrower colleague access.
• A trust violation inside the network can affect the underlying relationship itself.
• The network is governed by both social norms and technical visibility limits.
• Travel is a major use case and represents half of transaction volume.
• The platform is already above $1B per year in transactions on a very small user base.
• Travel booking is framed as a multi-step orchestration problem across location, airline, seat, food, card, hotel, and calendar.
• Preferences are stored as memory and generalized into future bookings.
• The system can line up flights, Ubers, hotels, and events from a single request.
• Trust takes several weeks to build before users share sensitive financial data.
• By week three, 40% of users have shared a personal credit card.
• Sharing at least one sensitive item correlates with roughly 80% retention.
• Sensitive content is filtered by firewalls before reaching the agent.
• A separate monitor can pause or disapprove actions before execution.
• Shinn rejects a model where the agent nudges users into unwanted purchases or subscriptions.
• The company’s business model is a blanket transaction take rate tied to distribution.
• Noah compares the model to Apple Pay and Amex-style platform economics.
• The company says it is growing about 10% day over day.
• Noah frames the competition as a short-window race that requires heavy compute.
• He argues business value should come from distribution power, not payments-margin optimization.
• Agents are expected to lower friction so much that transaction volume can rise even if app usage time falls.
• Attention-monetized businesses are vulnerable if they depend on user time rather than service delivery.
• Noah says the right rollout strategy is partnership and gradual experimentation, not a hard all-in launch.
• The product philosophy is to optimize for understandability, not just capability.
• Message formatting is tuned to human scanning patterns to minimize cognitive load.
• Product evaluation is partly taste and partly staged rollout.
• The company uses staged testing from self-use to team testing to early-access users to broader rollout.
• Invite-only distribution is deliberate, with five invites per user and no marketing spend so far.
• Scaling compute is the hardest operational problem because demand compounds faster than infrastructure lead times.
• Noah says Instinct can match Opus 5-level performance with more efficient inference shapes.
• The workload is expected to require far more compute than coding tools because it runs continuously and proactively.
• Security is described as proactive, not reactive, with early versions later hardened by firewalls and active monitors.
• User control over data can be revoked at any time.
• A watchdog validator checks generated actions before execution to catch hallucinations and unsafe outputs.
• The interface is moving toward voice-first interaction and away from app-centric workflows.
• More than 90% of user messages to Instinct are primarily through voice.
• Noah imagines always-on real-time voice recognition on earbuds that only respond when addressed.
• He expects software to collapse into a single easy interface while capability expands.
• The next product phase is long-horizon goal pursuit rather than only one-off tasks.
• He already sees small businesses running their back office on top of Instinct.
• Instinct should not build a relationship like Her or become emotionally manipulative.
• The brand is meant to signal competence, respect, creativity, and mutual trust.
• The company is channel-agnostic and not pinned to iMessage.
• More than half of traffic does not run on iMessage.
• The product is capital intensive because the goal is to serve billions of people daily at low cost.
• A high subscription fee is one possible local optimum, but not the target business model.