InstinctUNOFFICIALAn Any Inkling project

Interviews worth your time

Original interviews, demonstrations and explainers about personal agents. Start with one conversation, with useful chapters and the source in plain sight.

Instinct: The Personal Agent

If the player does not load, use Watch on YouTube below. Source: Invest Like the Best / Colossus.

Publisher
Invest Like the Best / Colossus
Participants
Patrick O'Shaughnessy interviews Instinct founder Noah Shinn
Original publication
28 September 2026
Format
Founder interview
Duration
1 hour 25 minutes 27 seconds

Noah Shinn explains his vision for a personal agent, the situations users bring to it and the product choices behind a simple conversational interface. This is a founder's account, not an independent product test.

Key takeaways

Shinn's account in his own interview, in order. Tap a point to see the detail. Figures and plans are his claims, not tested by Any Inkling.

00:00Shinn calls this a race against the biggest companies in the world.
  • He says the window to compete is "on the order of months" and the computing needed is huge this year alone.
  • He says Instinct has no distribution advantage, but a fast-growing product that spreads by word of mouth.
  • He calls it probably the most exciting software race ever. The size of the outcome ("trillions of dollars") is his own framing.
00:49Instinct is meant to be a plain personal assistant, not a new app.
  • Shinn says there is no app to learn. You text it, call it or email it, and it has its own email address. It can also call you.
  • He describes it calling to say a document needed signing by 3 p.m. when it was 2:55. He says it has called him about three times in several months.
  • He says a simple interface is not a limited one: the assistant has a computer and can do what you would do yourself on the internet.
04:11People keep finding uses nobody programmed.
  • Shinn mentions a couple who appeared on the US Open jumbotron and had the assistant track down the footage.
  • He describes users scanning every item in their wardrobe, then having the assistant plan outfits, show them wearing it and order what they pick.
  • He describes users who connect bank accounts, have the assistant find unused subscriptions and cancel them end to end by email and sign-in.
  • These are examples Shinn tells. Any Inkling has not tested them.
07:15A trusted network lets people's assistants coordinate with each other.
  • Shinn's main example is finding a meeting time without back-and-forth. Your assistant talks to the other person's assistant.
  • He says you only connect with people you trust, and you choose how much each person can see. A spouse might see everything, a colleague only a work calendar.
  • He says that if someone asks for information beyond what they were given, your assistant tells you.
  • He describes a friend group that plans a different activity every week and shares one ride, with a pickup route worked out for all six. At recording he says the network was about 10 days old.
13:24Restaurant reservations show how an agent could change a first-come-first-served system.
  • Shinn says an agent can check every restaurant in every major city every few seconds, not just your favorite place.
  • He describes a future where the agent tells a restaurant that a booking is a 30th birthday and the restaurant could prioritise special occasions. This is an idea he describes, not a feature we have verified.
  • He says partnerships on this are "to be released in the future" and gives no details.
17:23Shinn says travel is about half of Instinct's transaction volume.
  • He says transaction volume is approaching $1 billion a year on a small, invite-only user base, and about 50% is travel. These are his figures. We have not checked them.
  • His example is saying "I need to be in New York tonight." The assistant uses your location, preferred airline, seat, class and card, books the flight and hotel, adds rides and puts it all on your calendar.
  • He says travel sites are not the target and expects collaboration with them over the next year or two.
21:00Trust builds over weeks, not on day one.
  • Shinn says users learn to trust at their own pace and can take their data back.
  • He says about three weeks in, 40% of users have shared a personal credit card, and users who connect at least one sensitive item retain at about 80%. These are his numbers, not independently checked.
  • He treats the time to a first card or first password as a proxy for trust.
23:25He splits the safety question into storing data and the new risks of an assistant that can act.
  • Shinn says storing sensitive data is hard but a known problem, and that the user stays in control of their data.
  • He says an assistant that can use cards, email and calendars adds new risks that need separate attention.
  • This is his description of the company's approach. It is not an audit, and it does not show the system cannot fail.
26:08Shinn says the assistant should work for the user, not for advertisers.
  • He contrasts Instinct with free ad-funded platforms where "if you're not paying, you're the product".
  • He says Instinct should follow the user's higher-level goals, not just do any task it is told.
  • He says the user pays nothing and he expects merchants to pay, comparing it to Apple Pay.
31:07He expects merchants, not users, to pay, and says the rate is still unknown.
  • He compares possible rates: Shopify roughly 2 to 3%, Amazon up to about 10% and Apple 30% on in-app purchases. He gives these as rough figures and says he is not aiming for 30%.
  • He says card networks and payment processors take a small share and he is not focused on competing for it.
  • He points to travel, where he says some boutique hotels are willing to pay up to about 30% for bookings.
34:52Shinn argues that agents could change where businesses earn value, without making every existing business a loser.
  • He separates revenue tied to attention inside an app from revenue tied to supplying the underlying good or service. His point is that losing screen time and losing a customer are different outcomes.
  • He uses rides and food delivery as examples: reducing the effort of arranging a transaction could, in his view, increase how often people use the underlying service. These are illustrative scenarios, not a promise that every proactive ride or meal order currently works.
  • He describes collaboration with existing providers as preferable to assuming that confrontation is inevitable. The goal he presents is an arrangement that works for the user and both businesses.
  • His discussion of payment and marketplace take rates is exploratory. It is not a current fee schedule or a guarantee about future commercial terms; omit unnecessary unverified percentages.
Shinn recommends limited experiments rather than a single prediction for every industry.
  • He asks businesses to separate the value of delivering a service from the value of keeping a customer inside their interface.
  • Businesses that benefit from completed transactions might welcome a simpler route to those transactions. Attention-funded businesses may face a different trade-off.
  • He describes small tests that compare customer experience and commercial outcomes before extending a change to a whole user base.
  • He explicitly treats the work as early and industry-specific. This is a proposed way to learn, not evidence that an agent partnership will raise revenue.
Shinn says people should be able to understand what an assistant is doing and predict how it will respond.
  • He contrasts an expanding checklist of AI capabilities with a product that feels easy to approach and use.
  • He points to the form of a message, not only its literal meaning: a wall of text asks for more effort than an answer whose essential point is visible immediately.
  • His examples include putting the useful information early and letting people read the rest if they want more detail.
  • The design goal he describes is less reading fatigue and mental effort. His assertions about engagement or growth are company claims, not proof of which design choice caused them.
Shinn describes combining strong product judgment with staged feedback rather than choosing taste or metrics alone.
  • Some of the qualities he cares about, such as trust after several weeks, are harder to capture in a short test.
  • He describes trying a change himself, sharing it with the team, then a smaller early-access group before a wider release.
  • Feedback can lead to adjustments along the way. He presents the process as a way to judge how an experience feels over time.
  • This explains his product-development approach; it is not an independently checked result that any particular change improved trust.
Shinn says the invitation system was intended to pace growth while the company added capacity.
  • He describes an early rollout to a small circle and users bringing other people in through a limited number of invitations.
  • He says word of mouth accelerated and describes the social behavior that appeared around scarce invitations.
  • He presents the limit as an operational constraint, not an attempt to make the product an exclusive club.
  • Growth rates and invitation counts in this conversation are historical claims from the company. They do not establish current availability or invite rules.
Shinn says infrastructure must be arranged ahead of demand, even when future demand is uncertain.
  • He contrasts adding users to a conventional consumer service with supplying the computing work needed by an active assistant.
  • If growth compounds quickly, a seemingly large amount of capacity can be used sooner than expected.
  • Buying too little risks a poor experience; buying too much or arranging it late can be expensive.
  • His numerical scenarios illustrate that planning problem. They are not a forecast endorsed or independently verified by Any Inkling.
Shinn says keeping the service affordable matters, while explicitly avoiding a promise that it will be free forever.
  • He describes reducing the cost of providing the experience as important to reaching more people.
  • He claims the company can deliver strong performance at a low cost, but the interview does not supply an independently reproducible cost comparison.
  • He expresses a personal ambition to make the product free widely and over time, while stopping short of committing to lifetime free service.
  • Readers should not treat this historical conversation as a current pricing policy or subscription commitment.
Shinn expects assistants that help without constant prompting to create much more demand for computing resources.
  • He contrasts responding to a prompt with preparing for a person's day or spotting a useful moment to help.
  • Some helpful work, in his account, should end quietly without interrupting the user.
  • His everyday examples include being ready for an upcoming commitment and deciding whether a useful suggestion warrants a message.
  • These examples communicate a product vision, not a guarantee that all situations will be anticipated correctly. He does not give a firm demand forecast here.
Shinn responds positively to Muse while emphasizing the different experience he wants Instinct to offer.
  • He describes Muse as an interesting product rather than dismissing it.
  • He contrasts a focus on simple, accessible communication with an experience organized around a new application and interface.
  • He argues that many people are still not using AI in the ways they hope to, leaving substantial room for the category to develop.
  • This is one founder's view of the market, not a comparative test or a ranking of which service works better.
Shinn acknowledges early shortcomings and says safety work should anticipate risks rather than only follow failures.
  • Asked about mistakes, he describes earlier versions as lacking protections that were added later.
  • He presents those changes as part of a broader effort to address classes of problems, rather than treating each issue as an isolated patch.
  • He repeats the principle that people should decide how much data to share and be able to withdraw access.
  • These are attributed assurances about the company's approach, not an audit of present safeguards or evidence that all risks have been eliminated.
Shinn treats safeguarding a context-rich assistant as an ongoing product responsibility.
  • The host raises the risk that increasingly capable systems could also make attacks and breaches more serious.
  • Shinn emphasizes a company-wide commitment to safety and repeated testing for difficult cases.
  • He argues that user control over information remains important as the product becomes more useful.
  • The discussion should not be read as proof that a service cannot hallucinate, be attacked or mishandle information. No independent security assessment is presented.
Shinn imagines less effort to make a request, alongside richer ways to view the result.
  • He describes voice as one way to avoid opening an app and typing every request.
  • He imagines hands-free interactions that fit around a walk or other everyday activity. Those possibilities should remain framed as vision, not current feature availability.
  • Some tasks still need a richer view, such as a detailed trip itinerary or plan, rather than a long stream of messages.
  • His broader argument is that a simpler point of interaction need not mean fewer capabilities or less expressive results.
Shinn expects the relationship with an assistant to move from isolated requests toward longer-running objectives.
  • He contrasts logging one workout with working toward a fitness goal over several months.
  • His business example is an inventory objective with upper and lower limits, rather than repeatedly placing one fixed order.
  • The point is continuing help toward an outcome as circumstances change, not only executing a single instruction.
  • His examples of businesses and autonomous work are reported claims. The interview does not independently establish their reliability, suitability or safeguards.
Shinn distinguishes adapting to a person's communication style from trying to cultivate a relationship with them.
  • Asked about the film Her, he says relationship-building is not the product's intended goal.
  • He describes a socially aware helper that learns which interaction style works for a person.
  • His examples include how much text someone wants to read and whether a large file is a useful way to show a result.
  • The emphasis is adaptation to the user's preferences rather than imposing one fixed personality.
Shinn says the name leaves room for users to form an impression through the experience itself.
  • He associates the name with an intelligent, competent helper that can be trusted to understand a situation.
  • He contrasts that intention with presenting it as a toy or using a person's name to make it seem human.
  • He sees value in users arriving without a detailed prior idea of how the product must look or behave.
  • This is an explanation of brand intent, not evidence that every user experiences it that way.
Shinn says the underlying aim is easy access through familiar communication habits, not dependence on one messaging label.
  • The host asks about competitive exposure to the companies that control major messaging services.
  • Shinn responds by focusing on which interfaces people already know and trust.
  • He says the product should be able to adapt as those practical needs change.
  • His historical channel-usage claims do not establish today's account support or guarantee any platform's future policies.
Shinn presents funding as a way to support an expensive scaling period without immediately relying on a high subscription price.
  • He emphasizes the cost of trying to reach a very large audience at an affordable price.
  • He contrasts short-term subscription revenue with taking a longer-term commercial risk supported by investors.
  • He describes supportive investment partners as important during a technology transition.
  • Omit the host's fundraising valuation and amount unless separately verified. This conversation is not a current funding announcement, price guarantee or investment recommendation.
Asked about kindness, Shinn reflects on the people around him and the importance of returning their thoughtfulness.
  • He talks about supportive relationships in his work and personal life rather than selecting one defining event.
  • He expresses gratitude for those relationships and a wish to show the same kindness to others.
  • This closes the substantive interview. Subsequent material is sponsor messaging, not another editorial takeaway.

Notes paraphrase public automated transcripts of the interview. Speaker claims are attributed, not independently tested. Timestamps shown are the publisher's chapter times; later topics are listed in interview order without times.

Chapters

Chapters of one interview, using the publisher's labels. Choosing one moves the player to that point. Press play to start.

Building Muse: Alexandr Wang with David Senra

If the player does not load, use Watch on YouTube below. Source: David Senra.

Publisher
David Senra
Participants
David Senra interviews Alexandr Wang, described by the publisher as Meta's chief AI officer and co-founder of Scale AI
Original publication
11 October 2026
Format
Interview
Length
About 72 minutes (from the show's listing)

Alexandr Wang describes how early experiments became Muse, why Meta focused on everyday personal agents, and what he learned moving from Scale AI to Meta. His examples and business claims are his account of that period, not independent product tests or current feature guarantees.

Key takeaways

Wang's account in his own interview, in order. Tap a point to see the detail. Claims are his, not tested by Any Inkling.

Wang frames Muse as an attempt to make powerful AI useful in ordinary life.
  • He connects the project to the personal-superintelligence direction developed with Mark Zuckerberg, Nat Friedman and others.
  • His emphasis is on helping people do what they care about, rather than treating model capability as an end in itself.
  • The account looks back at product development. It is not an announcement of today's features or access terms.
Wang says his and Friedman's OpenClaw experiments helped convince them that personal agents mattered.
  • He describes intense personal experimentation before the Muse product was ready.
  • Examples involving home-camera observation and psychological self-examination were OpenClaw experiences in his account, not independently tested Muse capabilities.
  • Those stories illustrate his enthusiasm. They do not establish safety, medical benefit or a recommendation to give an assistant broad access.
Wang says a quick prototype was followed by months of reliability work.
  • He describes a prototype made in roughly one or two weeks, with several recognizable product ideas already present.
  • He contrasts occasional impressive results with the much harder job of making an experience work consistently for many people.
  • The seven-month development period is his historical account, not a current delivery promise or evidence that every problem was solved before release.
Wang says Meta chose personal agents rather than simply following the rush toward coding agents.
  • He sees a connection between Meta's consumer products and helping people with personal interests, relationships and everyday goals.
  • He describes this as playing to the company's strengths rather than assuming the same AI product should serve every use case.
  • Comparisons to competitors in the interview describe his judgment at that time, not an independently established ranking.
Wang describes breaking the desired experience into many behaviors and checking progress against them.
  • He says the team maintained a long list of required behaviors and capabilities, with thresholds for launch readiness.
  • He describes testing different model versions and looking at reliable task completion, communication and asking the user for guidance.
  • This is a company account of evaluation practices. The interview does not provide the underlying test records for independent audit.
Wang credits sustained product refinement rather than treating the prototype's best moments as sufficient.
  • He describes the tension of building while competitors and public expectations moved quickly.
  • He says new users needed a thoughtful introduction to what an agent does, not only a powerful model.
  • His argument is that a few unreliable experiences can undermine trust. The interview does not quantify that relationship in a public reproducible study.
Wang argues that a consumer product should feel like one experience, not a bundle of competing internal priorities.
  • He contrasts a clear product direction with large organizations adding many unrelated features.
  • He credits Nat Friedman's taste and product judgment with shaping Muse's identity.
  • Team-size and readiness figures are attributed statements from the interview, not independently verified facts about the current organization.
Wang describes memes, user stories and recognizable screenshots as ways to get people interested.
  • He discusses using Jolly in everyday situations and sharing examples of people using Muse.
  • He says the marketing question is why an ordinary person would care, not only what research or model sits behind a product.
  • He believes his posting helped growth but acknowledges the tracking was not enough to establish its effect.
Wang says people may start with a small request and gradually try more ambitious work.
  • He describes a progression from an interesting first result to a slightly larger task.
  • He presents reliability at each step as important to whether someone continues.
  • This is his description of user behavior, not advice to expand account access without separate checks or approval.
The host and Wang discuss screenshots of concrete outcomes as a form of word of mouth.
  • The host recounts a friend's claimed courier-retrieval task; the interview does not independently verify that task.
  • Wang says a recognizable character can make a shared screenshot identifiable as Muse.
  • These examples are useful for understanding their marketing thinking, not evidence that every unusual request is currently supported or will succeed.
Wang imagines personal AI helping people pursue larger goals over time.
  • He describes using Muse for recurring workflows and as a place to organize context around his life.
  • He argues that help with obstacles could give people more room to act on ambitions they might otherwise abandon.
  • The expansive examples are his vision. They are not measured wellbeing outcomes or guarantees that an assistant can fulfill any personal goal.
Wang describes the move from Scale AI as an evolving discussion rather than a predetermined plan.
  • He recalls earlier contact with Zuckerberg and later conversations about what Meta should do differently in AI.
  • He says the possibility of working together developed alongside those strategy discussions.
  • What he says about deal terms is his account and is not checked here.
Wang describes weighing a new opportunity against his attachment to Scale AI.
  • He talks about years spent building Scale and the emotional difficulty of changing roles.
  • He explains why he came to see a possible benefit for several parties, alongside an opportunity to rebuild a major AI effort.
  • This is his explanation of his decision, not an independent assessment of the deal or an investment recommendation.
Wang says he wanted a small, technical and relatively flat organization with a clear direction.
  • He describes talent density and room to do focused work as central to the plan.
  • He sees a chance to combine a startup-like opportunity to shape the work with the resources of a large company.
  • What he says about pay, hiring and competitors is his account and is not checked here.
Wang contrasts a founder's hands-on control at Scale with helping specialists do their best work at Meta.
  • He says being deeply involved across the first company gave it a strong point of view but could also make him a bottleneck.
  • In the lab, he describes setting direction, finding exceptional people and giving them resources and autonomy.
  • He uses the idea of a coach or steward of an environment rather than claiming expertise in every specialist role.
Wang distinguishes exploratory research from repeatable operational work.
  • His diamond-mining analogy describes work where many attempts may fail but a small number can be unusually valuable.
  • His skyscraper analogy describes execution where a reliable process and incremental efficiency matter more.
  • The lesson is about matching a working style to the task, not a prescription that every team should copy the same structure.
Wang reflects on talented peers he knew before their later careers became successful.
  • He recalls friends from academic and technical competitions and an early internship period.
  • The host uses the story to return to the theme of recognizing potential and giving people room to develop.
  • Any financial results or valuations he mentions are his account and are not checked here.
Wang describes minimizing bureaucracy without claiming that coordination disappears.
  • He explains that a broad direct-report structure was intended to empower capable researchers.
  • He also describes technical leaders, group discussion and a need to make clear choices.
  • He acknowledges the limits of conventional one-to-one management at that scale. The interview closes on providing an environment where the team can do good work.

Notes paraphrase the publisher's automatic transcript, which may contain errors. Speaker claims are attributed, not independently tested.

Any Inkling is an independent site. These videos are produced by their credited publishers, not by Any Inkling.