Recently, Silicon Valley giants have collectively set their sights on a new business: personal AI assistants.

Meta just launched Muse, OpenAI has Dots, and xAI has introduced Grok Bot. From models to computing power, and even user access points, these giants hold nearly all the advantages.

But interestingly, the fastest runners in this race aren't all giants.

Take Instinct, for example. Founded just a year ago, its product only began closed beta testing in February and still operates by invitation only—it doesn’t even have a standalone app.

Yet, this startup has already reached a post-money valuation of $10 billion.

Recently, Noah Shinn, the 23-year-old founder of Instinct, made his podcast debut for a lengthy interview.

Instinct Founder Noah Shinn Interviewed on 'Invest Like the Best'

During the interview, he shared several 'unconventional insights' about personal AI assistants that are worth discussing.

First, insisting that AI products must have an app is self-limiting.

Instinct has neither a standalone app nor a traditional product website. Users can assign tasks to the AI via text, WhatsApp, or email.

In Shinn's view, the ideal form of a personal AI assistant should eliminate the need for users to open a dedicated app.

Second, personal AI assistants have a unique form of data flywheel.

The more personal data users provide, the better the AI understands them, leading to a superior service experience.

Instinct has some striking data: after about three weeks of use, around 40% of users are willing to provide their credit card information. Among users who have shared at least one piece of sensitive information, the retention rate reaches approximately 80%.

Third, AI products can also have network effects.

Many believe AI products struggle to achieve the same network effects as traditional internet platforms. However, Shinn argues that future networks will connect not only people but also the agents behind each individual.

Two AI assistants can directly coordinate schedules, arrange gatherings, or even complete cross-account collaborations on behalf of their owners. The more people around you using the same agent network, the more the AI can accomplish.

Fourth, AI assistants are beginning to take over real-world spending.

According to Shinn, Instinct's annualized transaction volume has surpassed $1 billion, with a user base of just over 100,000.

This averages out to nearly $10,000 in annualized transactions per user. Given that the average annual consumer spending in the U.S. is around $60,000, this represents about one-sixth.

While this doesn’t mean every user truly delegates one-sixth of their spending, the scale is nonetheless impressive.

Fifth, AI assistants could rewrite how the internet makes money.

In Shinn's view, traditional internet platforms rely heavily on advertising revenue, incentivizing them to keep users engaged longer and clicking more ads—behaviors that may not align with user interests.

AI assistants, however, could take a different path: directly helping users search, compare prices, and make purchases, then earning commissions from completed transactions.

Instinct even envisions letting a billion users access its AI assistant for free, covering computing costs through transaction commissions.

Today, Silicon Genesis will use this lengthy interview as a springboard to explore how the business of personal AI assistants actually works.

/ 01 / Treating the Product as the Application is Self-Limiting

In Noah Shinn's view, designing a personal AI assistant as an app that waits for users to open it is inherently limiting. A true assistant shouldn’t have a preset set of functions but should be ready to handle whatever the user thinks of at any moment.

For example, suppose a couple attending the U.S. Open unexpectedly appears on the stadium’s big screen and wants to save the video.

They send their seat numbers to Instinct, which spends about 18 hours contacting event organizers, locating the broadcast recording, consulting customer service, and finally reaching the video editor to obtain the footage.

There was no pre-existing process for this—the AI figured it out on its own.

Numerous similar cases exist.

Some avid online shoppers take photos of every item in their wardrobe—shirts, pants, shoes—and provide their facial and body measurements.

Instinct not only puts together a week’s worth of outfits but also generates images of how the clothes would look on them. If any items are missing, it can shop online for suitable additions and even recommend three new outfits daily.

Canceling subscriptions works similarly. The AI reviews bank statements, identifies unused services, logs into the respective websites with user consent, handles email verification codes, completes the cancellation process, and informs the user of the savings.

The goal of a personal AI assistant is to seamlessly integrate AI into daily life, solving problems for users at any time. The actual user experience matters more than the product’s form.

This mindset is also reflected in specific interaction methods.

Currently, over 90% of interactions for some Instinct users occur via voice. Shinn has even set his phone’s quick-access button as a voice entry point, allowing the AI to send emails without unlocking the phone or opening a chat window.

In the future, he envisions continuous interaction with the AI through earbuds. While walking or cycling, the assistant could report pending tasks, and the user could decide verbally which to keep or cancel.

Of course, text and voice can’t solve everything. For complex tasks like travel planning or wedding arrangements, Instinct can temporarily generate a dedicated webpage for users to review and adjust details.

/ 02 / Personal AI Assistants Can Also Have Network Effects

Many believe AI products struggle to achieve the same network effects as traditional internet platforms. Noah Shinn disagrees.

In his view, personal AI assistants can also form their own networks. However, these networks connect not only people but also the agents behind each individual.

To this end, Instinct has launched an agent collaboration network, enabling AI assistants from different users to communicate and collaborate directly to complete tasks.

For instance, when scheduling a meal with a colleague, both users simply inform their respective AI assistants of their desire to meet. The two agents automatically exchange availability, discuss a location, and finalize the arrangement in the calendar.

Six friends even had their Instinct agents automatically find activities of mutual interest every week, then coordinate timing and arrangements.

This brings about an interesting shift: personal AI assistants are beginning to automate collaboration between people.

Moreover, such connections can generate network effects.

For traditional social platforms, the more friends a user has, the greater the platform’s value and the harder it is for the user to leave.

Personal AI assistants have similar potential. The more friends and colleagues around you using Instinct, the easier it is for your agent to collaborate with theirs, automating more tasks.

Once this collaboration becomes a daily habit, the personal AI assistant can evolve from a standalone productivity tool into a network connecting users’ social relationships and real-world lives.

However, compared to traditional social platforms, agent networks face a more complex issue: permissions. After all, true collaboration between two agents inevitably requires exchanging information.

Here, Shinn highlights an interesting characteristic of personal AI assistants: the more personal data users are willing to share, the better the AI understands them and the better the services it can provide.

Instinct has some revealing data.

Typically, users are reluctant to share sensitive information like credit cards or email addresses with a newly encountered product. However, trust builds over time.

According to Shinn, after about three weeks of using Instinct, around 40% of users are willing to provide their credit card information. Among those who have shared at least one piece of sensitive information, the retention rate reaches approximately 80%.

This suggests that personal AI assistants may be forming a unique data flywheel—a phenomenon entirely unseen in traditional products.

Of course, when AI begins representing users in collaborations with other agents, the question of how much personal data should be shared arises.

To address this, Instinct has designed a mechanism called the 'Trusted Interpersonal Network.'

Different relationships correspond to different levels of information authorization. For example, spouses might share more personal information, while colleagues might only view work schedules and some emails.

This design allows agents to exchange information and collaborate within certain authorization limits but introduces new security challenges.

In Shinn's view, personal AI assistants must protect sensitive data like credit cards and emails while preventing malicious content from inducing agents to perform erroneous actions.

This is another issue that personal AI assistants must resolve to establish network effects.

/ 03 / Targeting the Achilles' Heel of Internet Business Models

According to Shinn, the platform’s annualized transaction volume has surpassed $1 billion.

This figure is quite remarkable.

Keep in mind that Instinct currently has just over 100,000 users. On average, each user completes nearly $10,000 in annualized transactions through the platform.

If the average annual consumer spending in the U.S. is around $60,000, this means users are already delegating about one-sixth of their spending to Instinct.

While this is a rough comparison, it nonetheless illustrates that AI assistants are beginning to truly take over consumer decision-making.

Travel is the most prominent example, accounting for about 50% of Instinct’s spending requests.

For instance, a user might simply say, 'I need to be in New York tonight.' Instinct then finds flights based on the user’s current location, selects airlines, seats, and cabins according to the user’s preferences, books a hotel, arranges airport transfers, and syncs the itinerary to the calendar.

Previously, these transactions would occur across travel platforms, hotel websites, and ride-hailing apps. Now, the AI consolidates them into a single task, providing a simpler entry point.

Shinn also highlights a key difference between AI assistants and traditional internet platforms: their business models may determine whom they ultimately serve.

In Shinn's view, traditional internet platforms have a long-standing issue: platform and user interests are not always aligned.

Take Google, TikTok, and Instagram as examples. These platforms rely heavily on advertising revenue. To maximize ad earnings, they need to increase user engagement, boost ad click-through rates, and even use algorithms to influence consumer decisions, prompting users to buy products they don’t actually need.

AI assistants offer an alternative.

Users simply tell the AI their needs and preferences, and the AI handles searching, price comparisons, and purchases—a service entirely aligned with user interests.

However, for this model to work, a practical question remains: if the AI assistant saves users time and money, how does it earn its own revenue?

Shinn considered a $100 monthly subscription model but ultimately rejected it.

Because what Instinct truly aims to do is let a billion users access personal AI assistants for free and generate revenue through massive transaction volumes.

Specifically, Instinct hopes to earn a commission from transactions facilitated by the AI assistant.

According to Shinn, the volume of payments processed through Instinct is growing at about 10% daily.

His vision is that as more suppliers join the platform, the AI assistant will handle more transactions. Even without charging users subscription fees, commission revenue could eventually cover backend model inference and computing costs—provided the transaction volume is large enough.

This essentially attempts to transform the internet’s advertising-based business into an AI-driven transaction-based business.

/ 04 / The Computing Cost Conundrum

Before this business model truly takes off, however, Instinct must address a pressing issue: computing costs.

Previously, Instinct invitation codes were resold on eBay for around $300 each. Yet, the platform still hasn’t fully opened registration.

The reason is simple: user growth is outpacing computing capacity.

Initially, Instinct was available only to about 200 friends and family, each of whom could invite five new users. As early adopters began sharing their experiences on social media, the product spread rapidly, and the user base grew.

The trouble is, an AI assistant’s computing demands are far more complex than those of a typical chatbot.

While products like ChatGPT primarily compute when users initiate requests, Instinct’s AI may work continuously in the background even when users aren’t actively engaging—checking emails, organizing schedules, tracking orders, or executing previously assigned tasks.

This means that user growth drives computing demand at a rate that could outpace user growth itself.

According to Shinn, he now spends about 40% of his time addressing computing issues. At current growth rates, Instinct’s computing demand nearly doubles weekly.

Yet, computing supply cannot expand synchronously.

Procuring additional computing power often requires months of advance notice. If purchased urgently, costs can be three to four times higher.

This places Instinct in a challenging dilemma: the faster user growth occurs, the higher the computing costs the platform must bear. However, if too much computing power is procured in advance, resources may go to waste if growth slows.

In addition to continuously purchasing GPUs, Instinct has begun reducing costs through task scheduling and model selection.

First, it separates real-time and background tasks.

For example, if a user is waiting for an AI response, even a few seconds of delay can affect the experience—such tasks must be prioritized.

However, background tasks like organizing documents, checking schedules, or summarizing emails can be completed minutes or even hours later without users noticing.

These tasks can be batched and processed through bulk inference.

Since real-time computing resources don’t need to be reserved for each task individually, GPUs are utilized more fully. According to Shinn, processing efficiency for some tasks can improve threefold, fivefold, or even eightfold.

Second, it assigns tasks of varying difficulty to different models.

Complex planning and reasoning can be handled by more advanced frontier models, while simpler tasks like information extraction, classification, and execution can be assigned to lower-cost models.

According to Instinct’s internal testing, some low-cost solutions already approach the performance of frontier models in terms of user experience metrics.

/ 05 / How Can Instinct Challenge Meta and OpenAI?

The competition for personal AI assistants is becoming a showdown between giants and startups.

Meta has Muse, OpenAI has a massive ChatGPT user base, and xAI is developing agents. These giants not only control foundational models and computing power but also rely on existing products to reach vast user audiences.

By contrast, Instinct lacks both its own foundational model and ready-made distribution channels.

More troublingly, it primarily relies on existing communication channels like SMS and WhatsApp to provide services, and these entry points are themselves controlled by other companies.

For example, if Meta restricts access to WhatsApp or Apple adjusts the rules for iMessage, it could impact Instinct. Although currently, more than half of its traffic no longer goes through iMessage, the dependence on external platforms still exists.

These gaps will be hard for Instinct to bridge in the short term.

However, the competition for personal AI assistants is not just about whose model is stronger or whose user entry point is larger.

Shinn has tried Muse and considers it a decent product. He also does not deny that other companies could fully develop AI assistants with similar capabilities in the future.

Therefore, rather than simply pursuing functional superiority, Instinct cares more about how users actually feel when using AI.

A typical example is how AI should respond to users.

Traditional large models often tend to explain problems comprehensively, but Instinct primarily communicates with users via SMS, and overly long responses can easily increase the reading burden.

For instance, when people browse a piece of text, they might read 80% of the first line, only 50% of the second line, and the percentage continues to decrease afterward.

Therefore, Instinct tries to place important information at the front, allowing users to grasp the main points after reading the first 30% of the content. The team even repeatedly tests the length of responses and the use of emojis.

Another, more challenging aspect is whether AI can build trust over long-term use.

For example, when should it proactively contact users, and when is it better to stay quiet? Too many reminders can be annoying, but failing to provide reminders when needed can also affect the experience.

These issues are difficult to judge through model benchmarking.

So, after developing a new feature, Shinn usually tries it out himself first, then hands it over to the team, followed by releasing it to a small group of early users for feedback and adjustments before formally launching it.

Because while it's easy to verify whether AI can complete a task, it's not so easy to test whether users still find it reliable and willing to continue using it after two or three weeks.

Of course, it's still hard to say whether these accumulated improvements in experience can offset the channel advantages of tech giants.

Shinn is not in a hurry to make a judgment on this either.

In his view, if you walk into any coffee shop, there may not be many people who truly let AI handle their daily tasks continuously. The market for personal AI assistants is really just getting started.

As for who will ultimately come out on top, the answer is still far from clear.

By Yuan Yuan