Summary

Sandwich Lab treats “we need to be expensive first” as demand validation: Lexi launched without a free trial at an entry price of $199/month (roughly $200 in software fees, across different plans), and users still paid—郭振宇 says that at least shows the market has been found, though F remains uncertain. Three months after launch, paying users came from 94 countries and monthly revenue grew more than 150% month on month; the company completed two financing rounds totaling more than $10M over the past four months and remains gross-margin positive after customer acquisition and compute costs. The first round was led by Jinqiu Fund and Mobvista, and the second by 5Y Capital and Gobi Alibaba Entrepreneurs Fund.
Lexi is targeting not only traditional advertisers, but at least 50M SMBs globally that need customer acquisition yet lack the ability to run Meta ads. About 70% of current customers had never advertised before; traditional local agencies charge roughly $6,000-$20,000 per month, versus about $200 for Lexi. 郭振宇 sees these users as a purely incremental advertiser base for Meta.
The product’s differentiation is not generating a single ad image, but closing a continuous feedback loop across fundamental analysis, creative production and the “technical trading” that follows deployment. Lexi dynamically adjusts budgets, merges or deletes ad sets and regenerates creative based on ROI, CPC, CPM and estimated “probability-adjusted odds”; because creative usually fatigues within 2-3 days, one-off delivery cannot produce durable results. He says Lexi has at least reached the average human agency level, with the vast majority of sales targets above the corresponding acceptable threshold and the best cases reaching 8-10.
郭振宇 chose Meta over expanding to Google and TikTok based on the information structure of the three platforms, not channel coverage. Google captures active search intent, making automation more like centralized allocation within the platform; TikTok already understands users, leaving human creative as the bottleneck; Meta is built on the social graph and has relatively limited insight into latent purchase intent, leaving advertisers and Lexi to play the black-box game of “where to find whom.”
Lexi deliberately minimizes process visibility and prevents users from editing copy, shifting trust from “watching the Agent work” to sustained operating results. Devin- and Manus-style process displays fit hands-off tasks with a clear delivery endpoint, while advertising is an “infinite game”; 郭振宇 defines the product as a “revenue generating machine”—“you don’t need to care what I’m doing”—because users ultimately care more about long-term operating results than any single ad.
The Email Marketing Agent will test whether Sandwich Lab can expand from a single product into a multi-category revenue machine. The new product will not merely generate emails; it will continuously test “what to send next” around landing-page conversion. 郭振宇 estimates that shared users, GTM, customer service, algorithms, backend and frontend components reduce the cost of a new product to roughly 10 person-months. Future products may extend into financial insights, supply chain, management and hiring.
Meta building fully automated advertising itself would not necessarily eliminate Lexi, but the multi-product expansion rests on shared infrastructure and an estimated R&D cost of roughly 10 person-months. 郭振宇 believes Lexi brings incremental advertisers to the platform, Meta must confront the same black box, and the platform’s use of differentiated strategies for different advertisers is “not necessarily reasonable or necessarily legal”; the deeper technical confidence comes from a “non-zero-sum game”—mistakes can be iterated on, and growth need not come at a competitor’s expense. These judgments still require ongoing product validation.

Deep dive

1. Sandwich Lab Turned an Automation Background into an SMB Growth Product

郭振宇 is 38, with an undergraduate degree from Zhejiang University and a PhD from UBC. He has spent most of the past 15 years in IoT, robotics and autonomous driving—not advertising—and calls this venture an exercise in “methodology entrepreneurship.”
During his PhD, the IoT hardware company he worked with became a supplier to Tesla SolarCity and was later sold. He subsequently created a sidewalk-delivery robot project inside Postmates; the project later spun out as Serve Robotics, a Nasdaq-listed company that he estimates was valued at roughly $800M-$1.2B.
After returning to China, he led commercialization and scaling for the “Xiaomanlv” autonomous last-mile delivery project at Alibaba’s DAMO Academy. Sandwich Lab now has more than 20 employees, and its first product, Lexi, automates Meta customer acquisition end to end.
The company completed two financing rounds over the past four months, totaling more than $10M. The first was led by Jinqiu Fund and Mobvista; the second was backed by 5Y Capital and Gobi Alibaba Entrepreneurs Fund.

2. $199, No Trial and 94 Countries Formed the First Demand Test

Lexi charges roughly $200 per month for software, with different plans. Its early entry price was $199, and it has offered no free trial since launch. Advertisers pay their Meta budgets separately.
Three months after launch, paying customers came from 94 countries, while monthly revenue grew more than 150% month on month. 郭振宇 is not disclosing absolute revenue, but says the business remains gross-margin positive after customer acquisition and compute costs.
Customers include a team-building drum studio in the UAE, an ADHD clinic in the US, an accessibility-modification service provider in Australia and a dreadlock-wig merchant in Cameroon. Some may have only 3-5 employees, but Lexi gives them executable advertising capability for the first time.
Lexi is also “advertising Lexi.” The team has no dedicated growth function and mainly uses its own product to acquire customers and iterate on the algorithm. 郭振宇 says this has kept CAC relatively low, an important reason unit economics are positive.

3. The Real Market Gap Is “Strong Demand, No Prior Advertising”

In markets such as the US and Canada, local agencies often handle account setup, recommendations and guidance without guaranteeing results, charging monthly retainers of roughly $6,000-$20,000. In smaller or less-developed markets, the service may not exist at all.
About 30% of Lexi customers have advertised before and are looking for a better tool. The other 70% have strong demand but have never successfully run Meta campaigns. Even when handed creative, they cannot operate the platform; the system has to complete the last mile.
郭振宇 cites a market estimate of roughly 130M SMBs with a presence on Meta globally. Narrowed by geography and revenue capacity, at least 50M potential advertisers remain. What matters more to him is that these businesses form an incremental advertiser base for the platform.

4. Advertising Is Not a One-Off Forecast but Continuous “Technical Trading”

Lexi starts with the “fundamentals”: understanding the business and product, the competitive landscape and its marketing position, then inferring from the demand side which countries, markets and demographics genuinely need the product and can afford it. 郭振宇 describes advertising as “a process of filtering.”
A downstream Agent then uses advertising experience to generate selling points, audiences, tags, copy, images and video. Most AI advertising products stop at this creative handoff, which only helps people who already know how to use computational advertising platforms.
Lexi continues with the “technical trade”: it monitors ROI, CPC and CPM in real time, adjusts budgets and parameters based on estimated “probability-adjusted odds,” merges what should be merged, deletes what should be deleted and creates what should be created, then feeds the results back into the fundamentals.
There is no “optimal ad” that can be forecast once and for all; creative often fatigues within 2-3 days. In slower cycles, the system iterates several times a week; in faster markets and campaigns, several times a day. Continuous feedback is the core of the system’s work.

5. Serving SMBs Is Both a Productivity Allocation Choice and a Scale Bet

After years working on automation, 郭振宇 is less interested in distributing wealth directly than in “allocating the productivity gains from advanced technology to the people who deserve them most.” The additional 50M SMBs are therefore the venture’s underlying motivation.
Large enterprises can hire excellent talent, so for now fully automated AI is mainly a productivity and cost-efficiency tool rather than a must-have. When the host asked whether this sacrificed commercial outcomes, his answer was that lowering the barrier to advanced technology could produce larger long-term results.
His value judgment is explicit: “You can’t be a self-interested person and still do something worth hundreds of billions of dollars.” In B2B, genuinely new demand is rare; huge companies usually solve operating needs that have existed for centuries in new ways.
Sandwich Lab is not permanently ruling out key accounts. The team is discussing a pilot with a globally recognized consumer-goods group, but it wants to test more than ordinary media buying: reaching potential users through advertising to validate demand, which may be more valuable than operating an existing user base.

6. China’s Local Business Ecosystem Reveals the Structural Gap AI Should Fill

The team has deliberately biased its targeting toward local businesses rather than starting with e-commerce. 郭振宇 acknowledges the personal preference, but the product was built to “lower the bar” and give people who previously lacked operating capability greater accessibility.
The Meituan and Alibaba ecosystems showed him how Chinese small stores developed sophisticated operations through three layers of accumulation: talent and capabilities flowing out from platforms, SaaS and SOPs cascading down, and millions of ecosystem operators handling cold starts, customer acquisition, coupons and bundled offers for stores.
The barber he regularly visits in Hangzhou says his main job each day is managing his private traffic pool—“more important than cutting hair.” To 郭振宇, an ordinary small shop understanding growth hacking is “a commercial miracle in human history.”
Outside China—including the US, UK, Australia, Canada, Africa and South America—there are not millions of operators who combine strategic ability with affordable pricing. This is not manual labor that can be solved by copying SOPs. 郭振宇 believes this generational supply gap is exactly where fully automated AI should step in.

7. Lexi Deliberately Prevents Users from Treating Creative as the Final Product

During the host’s trial, Lexi asked only about the product and budget before launching the campaign; the creative interface was deeply buried. This was the opposite of Agents such as Devin and Manus, which display every command and build trust through the process.
郭振宇 confirmed that this was intentional. Users can inspect creative, copy and conditions deep in the interface, but “users have no opportunity to change any copy or tags.” Seeing the system does not mean being able to intervene.
He agrees that process visibility helped establish trust in the category of “fully automated Agents” worldwide, but does not believe every application needs to copy it. Sandwich Lab chose “Shadowing Mode” for its AI Agents from the outset.
For hands-off tasks such as coding, the endpoint is handing software to the user, after which the Agent has limited ability to keep improving the work. Advertising has no one-off delivery endpoint: users operate continuously, and what they care about is the operating result, not whether a single output looks attractive.

8. The “Infinite Game” Makes Outcome Trust More Important Than Process Trust

郭振宇 defines every product as a “revenue generating machine”: “You don’t need to care what I’m doing. I’m doing a few things, through shadowing, to help increase your revenue.” It sounds too good to be true, but that is precisely the product promise.
A customer spending $100,000 continuously may receive tens of thousands of creatives—too many to review and impossible to approve one by one. The approval process would become a burden and undermine the system’s ability to iterate daily.
He acknowledges that showing creative during onboarding may increase initial trust and could be tested when resources allow. But the sensitivity of monthly revenue to any single line of copy is not as high as people imagine, especially when the customer is not a large company with a powerful brand.
The more aggressive view is that every product direction should serve must-have demand with “no alternative.” When users have to complete an operating task and the market offers no viable substitute, formal trust matters less than whether the product can deliver.

9. Must-Have Demand Came from Travel Samples; Configuration Friction Was the Post-Launch Surprise

When traveling, 郭振宇 asks coffee shops, restaurants and boutique retailers in Tokyo, the US East and West Coasts, Canada, Australia, Asia and Southeast Asia about their businesses. These merchants often have polished Instagram accounts, but in his sample “99.99% do not advertise”; they rely only on organic traffic.
He personally tried setting up and running Meta ads and still could not get the hang of it in a day. SEO or Google tools could probably be learned within a day. Searching for local providers turned up only traditional digital marketing agencies, often more than a decade old and focused on design and website building.
He repeatedly stresses that this is only an “anecdotal small-sample impression,” even “very unrigorous.” But social presence was already widespread while paid amplification went unused, leading to the initial hypothesis of “strong demand that nobody is serving.”
What exceeded expectations was how many customers could not even open and configure a Meta account. 郭振宇 revised his view: the most important SaaS solution is not the subscription fee, but eliminating CD-ROM-style configuration. There are still many barriers that are “not about intelligence, but configuration,” and Lexi will fill them in as product features.

10. Google, TikTok and Meta Are All Called Advertising, but the Underlying Problems Differ

郭振宇’s breakdown of Google is that search exposes active intent, giving the platform extremely precise information about what users want to buy. Automation is therefore mainly centralized allocation inside the platform, with little need for a heavily structured intermediary trading strategy.
TikTok is a recommendation feed whose foundation is “it recommends accurately.” The bottleneck in ad automation is therefore not finding people, but identifying what creative attracts them and testing it quickly. Human life and human creativity remain sources of strong creative; in the near term, he sees this as neither AI’s strength nor his own area of interest.
Meta is first a social network. It knows who knows whom, who follows whom and who is followed, but receives far fewer signals from knowledge search and transactions. Combined with privacy restrictions, this forces it to leave part of the question of “where potential users are” to advertisers, leaving Lexi to play the black-box game.
The host warned that listeners might remain skeptical. 郭振宇 acknowledged that he is not deeply familiar with these platforms, which is precisely why he tried to deconstruct their underlying principles. He presents these conclusions as his own judgments rather than established facts; until those judgments change, the team will not naturally expand into Google or TikTok.

11. ROAS Has Crossed the Usability Threshold, but Statistical Maturity Remains Limited

The team tracks three types of goals each day: traffic for page visits, leads for forms and phone inquiries, and sales for completed transactions. Sales are measured by ROAS, or return on ad spend.
郭振宇’s rule of thumb is that ROAS above 2 is generally acceptable to human advertisers, while SMBs with few alternatives may accept anything above 1.5. He says the vast majority of sales targets exceed their applicable thresholds; the best cases reach 8 and 10, meaning each $1 spent produces the corresponding return.
He says most users are very new and that the company wants users with less than a week of history to outnumber older users at any point, showing that growth is fast enough in both pace and structure. It is therefore too early to draw highly definitive conclusions. Based on results after campaigns have settled, he says Lexi has at least reached the average human agency level.
Traffic has continued to perform well. Leads remain incomplete because many local businesses acquire customers by phone and phone tracking is not yet fully built out. The team plans to provide tools that identify which phone inquiries came from advertising.

12. Users Not Seeing Optimization Does Not Mean the System Should Manufacture Activity on Day One

The host summarized a class of complaints from Discord: Lexi seemed to open the account and launch ads without visible A/B testing or ongoing optimization. 郭振宇 believes many of these comments came from early users who shut down their ads on the first or second day.
Meta needs at least 3-7 days to observe group behavior, while traditional agencies may ask for two weeks. Lexi generally increases the frequency of interventions after the third day; changing things too early can itself damage learning and performance.
When the host asked whether the team could deliberately create activity during the first two days to make users feel the Agent was working, 郭振宇 rejected the idea outright: “We shouldn’t do that, because it wouldn’t be reasonable.” The team chose to improve explanations and use guarantee-based rebates to encourage customers to keep campaigns running.

13. High Pricing Validates the Market First; Marketing Tactics Come Later

郭振宇’s pricing principle is: “I think we need to be expensive first.” If users still pay $199 when the product is not yet that easy to use and offers no free trial, that at least shows the market has been found—though F remains uncertain.
He does not treat MAU or ARR growth as the only PMF signal, and sees price as the fastest pressure test of demand. Around August or September, the team may open trials, discounts and other marketing tactics to expand registrations and usage, rather than using free access from the start to conceal willingness to pay.
A customer spending a few hundred dollars and one spending $100,000 cumulatively currently pay the same subscription fee. The next phase may introduce a traditional agency-style package: customers hand their budget to Lexi, which then pays part of it to Meta, allowing the company to explore a more diversified revenue model.

14. The Email Marketing Agent Targets the Next Action, Not Another Shovel

Faced with an email marketing industry that has existed for roughly 50 years, 郭振宇 chose neither Google nor TikTok. The most direct reason is that Lexi needs to operate its existing user base, and he could not find a product that would actually do that work for the company.
Existing tools can tell him the final landing-page conversion rate. Even expensive A/B testing tools merely “sell you a shovel”: users still have to decide what to send next, what to test and how to improve clicks and conversion.
The new Agent’s goal is for the user to tell the system only which landing page they want recipients to reach. The system then continuously searches for the timing and content that will improve clicks and conversion. The logic is the same as advertising, allowing the company to reuse its objective-feedback algorithms, server-side infrastructure and frontend.
郭振宇 previously said the company would release 2 new products in August. The Email Marketing Agent is one of the clearly identified directions.

15. The Company’s Boundary Is Not Advertising but Every Operating Step That Directly Generates Revenue

郭振宇 believes there are “not many new demands” in operating a business. Business schools have taught marketing, finance, management, supply chain and human resources for hundreds of years. AI’s opportunity is to automate work that software has never truly completed, using new methods.
In tax, finance and legal, the company would not focus on compliance. It would identify revenue opportunities and rational growth strategies from a financial perspective, then turn those strategies into customer acquisition, user operations or supply-chain actions. Finance finds the opportunity; execution remains in the operating chain.
The supply-chain product has not yet taken concrete form. Human resources may lead to recruiting. The directions look scattered, and none is necessarily enormous on its own, but each must satisfy the same constraint: “Every step is a necessary part of directly increasing revenue, and without us the user might not be able to do it.”
The email product could even be free, with “reaching 1M MAU in 3 months” as the validation target rather than revenue. 郭振宇 treats this as a new causal hypothesis, not a forecast already achieved.

16. Meta Building Fully Automated Ads Does Not Mean It Can Flip a Switch and Eliminate Lexi

The host cited Zuckerberg’s vision: advertisers tell Meta only what they sell and how much they want to spend, while the platform handles audience targeting, creative generation and campaign strategy—almost exactly the full chain Lexi is building.
郭振宇’s first response again concerns information structure. Meta does not possess demand information as completely as Google, so even if it builds the system itself, it may need to operate outside the black box using experimentation and game-theoretic logic similar to Lexi’s, rather than simply opening “a small branch” inside its internal system.
The second layer is incentives: Meta’s main revenue still comes from key accounts, while Lexi serves the roughly 50M SMBs that previously could not advertise. For Meta, they are mainly incremental rather than a threat to existing revenue. The third layer is platform boundaries: a rule-maker applying different strategies to different advertisers is “not necessarily reasonable or legal.”
These are 郭振宇’s judgments, not established facts. He also stresses that Meta still owns customers’ attention. In many emerging markets, merchants have no website and rely on a Facebook or Instagram page to establish a presence, then communicate and even receive payment through WhatsApp or Messenger.

17. The Multi-Product Strategy Rests on a Marginal-Cost Assumption of “10 Person-Months”

Investors repeatedly ask why a company with more than 20 employees is pursuing multiple lines at once. 郭振宇 says these needs were listed on day one, but the team did not begin until it confirmed the reusable structure; it did not launch a new line every time it spotted a direction.
The same global SMB base allows GTM channels, operations, customer service and maintenance to be reused. The products all iterate around objective feedback, while the algorithms, backend and frontend components overlap heavily. He estimates that a new product takes roughly “10 person-months.”
Low development cost is not enough; the truly expensive part is operating seriously through PMF. Lexi has already validated customer acquisition, payment and user learning. A multi-product strategy also forces each functional team to standardize and automate in order to serve different product lines.

18. 郭振宇 Abandoned the Magic Button in Favor of Letting Products Grow Separately

The initial vision was radically unified: users enter a URL, press one button and “your revenue starts growing tomorrow,” without even knowing what functions existed inside the system. He once wanted all Shadowing Mode capabilities hidden inside a single product.
His recent change of view was informed by companies at the stage of Microsoft and Baidu, which had entry points and relatively stable traffic rather than rapid growth, and therefore tended to aggregate functions. Fast-growing ByteDance-style teams, by contrast, often let each app stand alone so that the team can tell what needs to change when performance moves.
AI applications remain a rapidly developing new market. 郭振宇 therefore decided to split each Agent into an independent product, with each line running its own data-driven acquisition and iteration while sharing the underlying infrastructure.

19. The Reward of Entrepreneurship Is Validating Causality, Not Merely Building a Company

The team only completed and launched its A/B testing system recently—on the day before the interview. 郭振宇 admits that it previously relied mainly on manual A/B testing, and insists that tests should examine the hypotheses behind orthogonal features rather than merely compare button details.
He summarizes the method with a piece of “inspirational wisdom”: “Correlation lets you predict the future; causality lets you change the future.” Every day of entrepreneurship involves hypotheses about customers, sales, technology and algorithms, followed by real-world results that determine whether the causal chain is right or wrong.
“Making it” is therefore both the goal and an experimental way to understand the world. If Lexi succeeds, it suggests the earlier assumptions about Meta and SMBs may be valid. If a free Email Agent reaches 1M MAU in 3 months, it will validate another rule that could be applied to something larger.

20. Past Narrow Escapes Made Him Trust Must-Have Demand More Than Surface-Level Team Quality

郭振宇 had never truly been a CEO. His student venture was led by a classmate from business school, while the Postmates robotics project was an internal venture still operating under a parent-company CEO and CTO. This is the first time he has carried full responsibility for a company.
His first venture went a long time without outside funding, surviving for six months on $500,000 from Kickstarter. It had no clear PMF, yet it “somehow signed contracts and somehow got sold.” Because he was still a student, the experience felt more fun than frightening.
Before several projects exited, the teams still looked by Alibaba and ByteDance’s top functional standards like they “had nothing and could do nothing.” Yet each eventually exited. The lasting feedback was: “As long as you find must-have demand, you can always exit.” The current team is far stronger than those earlier ones.

21. High Objectives and Low Expectations for Any Single Move Are His Discipline for Distinguishing Trading from Gambling

郭振宇 rarely uses “success” to define himself and dislikes competing where everyone is taking the same exam. He would rather work on problems that look competition-free from his perspective but are sufficiently interesting and complex.
He calls entrepreneurship the activity with the highest “density and concentration of dopamine”: “Nothing is more satisfying than entrepreneurship—not even gaming.” Promotions, 16-month bonuses and 20,000-stock refresh grants at large companies can become false feedback that distracts from actual operating results.
His objective is extremely high, but his expectation for any particular action or near-term result is very low. High expectations crowd out contingency plans, Plan B and operating slack. “The difference between gambling and trading is whether you are calculating with discipline and leaving yourself room.”
He takes an almost contrarian approach to being misunderstood. If a path can truly lead to a $100B company, not everyone should understand it early; otherwise competition would already have filled the market. Investors used examples such as Amazon initially selling books, Kuaishou initially making GIFs and Facebook initially starting on campuses to teach him to treat “not yet visible” as potential space rather than proof of success.

22. The Non-Zero-Sum Game Is Both an Algorithmic Condition and the Company’s Value Boundary

One candidate used Pareto optimization to explain computational advertising: even with all the data, a platform cannot achieve the system optimum through centralized planning. Only when each bidder competes from the advertiser’s self-interested objective can the mechanism potentially improve returns for both the platform and advertisers.
Pareto optimality means that once the boundary is reached, any further gain for one party necessarily harms another. Before that boundary, there remains non-zero-sum space for mutual improvement. This definition made 郭振宇 reconsider what the company should build.
Meta advertising, email operations and internal supply-chain optimization all allow the team to make a mistake today and keep experimenting tomorrow; no single action causes total failure. Football competition, a secondary-market counterparty trade or helping a neighboring store wage a price war involve stronger zero-sum dynamics and more irreversible risk.
Sandwich Lab therefore tries to choose tasks that can be explored repeatedly without making someone else worse off. 郭振宇 hopes the ultimate result is growth for both SMBs and platforms, even “growth in global GDP,” rather than moving revenue from one merchant to another.

23. Global Small Businesses Bring Macroeconomic Pessimism Back to Concrete Lives

After serving customers in Palestine, Pakistan, Israel and many emerging markets, 郭振宇’s impression is that “the world is much better than we imagine.” Even amid war and economic cycles, people continue to run their businesses seriously and hope life will improve.
He says World Bank data he found with AI shows that inequality among countries and regions has generally narrowed over the past 30 years. The US and China may show different internal trends, but the gap between people in the Democratic Republic of the Congo, Liberia and Zimbabwe and those in the UK is clearly shrinking.
His most vivid customer is in Palestine: the merchant buys daily necessities from Yiwu, loads them into 6 mini vans, posts photos on Facebook and runs Lexi ads. Residents in Gaza circle the products in images and place orders through messages, after which he sorts and delivers them every day.
The customer was “having an especially good time” running the business during wartime, focused only on serving customers and doing the work well. 郭振宇 uses this to put his own entrepreneurial pressure in perspective: “My business is much easier and safer than his.” It has become a practical anchor for staying low-anxiety and optimistic over the long term.

125: Using AI to Recreate a “Me”: A Conversation With Mindverse’s 陶芳波 on Identity Models DHH: Future of Programming, AI, Ruby on Rails, Productivity & Parenting | Lex Fridman Podcast #474