Written by | Wu Xianzhi Edited by | Wang Pan
On September 10, the founder of a newly listed embodied AI company, riding on the frustration of the company's stock plunging right after its IPO, posted a named complaint on WeChat Moments about the existence of deal-packaging companies in the embodied AI industry.
He mentioned the prevailing trend in the industry of generating fake and unsustainable revenue, lacking Product-Market Fit (PMF), merely releasing various sensational news for hype, and aiming for an IPO just two or three years after establishment.
Previously, this founder had publicly expressed concerns about companies generating revenue through data collection center channels in the face of data scarcity in the embodied AI industry. For instance, local governments invest in building data collection centers and purchase equipment from embodied AI companies, which then buy back the data to create book revenue.
When an industry is booming with massive opportunities, there will always be flashy ways to survive. The reason is that when an industry just takes off, very few companies can achieve PMF. During the vacuum period before delivery, few companies have the patience to stay grounded and do solid work.
In the embodied AI industry, this manifests as data scarcity and urgent demand, necessitating unconventional tricks.
On the same day in Hangzhou, Amap held a press conference to showcase another approach to the data business. Over the past year, 880 million users have browsed lists, reviews, and ratings on Amap. Users have cumulatively navigated 36.6 billion kilometers to listed merchants. 15.28 million people signed up for the Scout Creator Program, and 8.38 million completed Alipay credit authorization, leaving nearly 50 million genuine reviews.
Amap's Street Scouting List has established a low-cost data collection channel without the need to pay data licensing fees. The process from collection to feedback is completed during user engagement.
Embodied AI companies spend heavily on data, leading to financial issues. The cost for Amap to accumulate 36.6 billion kilometers of data is much lower.
Building a Data Collection Network in One Year
On September 10 last year, Amap launched the Street Scouting List, with the external messaging focused on resolving fake ratings and reviews.
CEO Guo Ning reiterated the logic at the press conference, "Writing a review is much easier than navigating 30 kilometers to visit a store." Faking a review nowadays requires no cost, but faking behavioral data is much harder. Therefore, replacing text data with behavioral data became the starting point of the entire Street Scouting List story. After all, cheating in behavioral collection has thresholds and difficulties; making the data itself valuable is the prerequisite for this data collection pipeline to work.
Over the past year, the Street Scouting List has expanded into various segmented lists around different scenarios, which looks more like precise data annotation requirements.
Within 100 days of launch, the user base exceeded 400 million; a year later, 880 million users discover new places to eat, drink, and have fun on the lists, providing Amap with a continuous stream of data from the real world.
2.6 million merchants have voluntarily joined, and Amap produces flying street views for them for free. Some merchants claim a 400% increase in user scale after being listed. Holding the report card at the press conference, Guo Ning implicitly mocked other lists for modifying their rating rules to benchmark against Amap's Street Scouting List, holding merchant conferences, clearing fake reviews, and using real evaluation data as the ranking basis. After gaining the right to speak on data standards, Amap has stood tall.
Amap's Street Scouting List acts as a free "lever" to unlock data, constructing three layers of data dimensions.
Trip data, consisting of navigation mileage, proportion of dedicated visits, returning customers, and the proportion of locals, forms the first layer of data collection dimensions. For Crab God Miemie in Wuhan, the cumulative mileage of customers making dedicated visits exceeds 470,000 kilometers, with a dedicated visit proportion of 68%; for Huajia Yiyuan in Beijing, the proportion of locals is 78%. Similar data formats mostly carry causal relationships, which did not exist in the industry before.
High-quality data after cleaning forms Amap's second layer of data sources. Specifically, 8.38 million Zhima Credit authorized users completed the preliminary preparation for data collection. The nearly 50 million reviews they contributed are each tied to a real identity. In addition to users, KOLs (Key Opinion Leaders) provide Amap with more credible data sources.
For example, @ChengduDetective49 has visited over 700 coffee shops in three years, leaving reviews for 79 brands, covering everything from bean varieties and processing methods to the supply chain. @JiangzheShenpoAichi wrote 146 seafood reviews in a year, breaking them down into 13 types of crabs. These processes, in the form of reviews, perform the work of data annotation. Real human experiences bring taste, and Amap doesn't need to pay for it; honorary titles can unlock the data flywheel.
Scenario data is a newly added data source after Amap's AI transformation, mainly including the flying street views, business status, and customer flow tides of the "2.6 million merchants" claimed by Amap, all provided and maintained by the merchants themselves. Amap has built an effective resource exchange pool. The group invests a Token budget for Amap to generate better visual displays for merchants; merchants gain traffic from the uploaded scenario displays.
Although Amap currently lacks the user mindset for redemption, it at least has traffic; in the worst case, users can go to Dianping for redemption.
Comparing costs reveals Amap's advantages. Purchasing a data collection device costs hundreds of thousands of CNY, and investing in a data collection center costs over 100 million CNY to achieve a daily output of thousands of data assets. Amap's data collection terminals are 880 million mobile phones, with collection coverage spanning hundreds of city stores. The cost structure of data collection is basically negative, and the collected parties are even grateful.
While the embodied AI industry is still exploring effective loops for data collection, Amap uses list products to drive users and merchants to participate.
The End of Data Collection Lies in Models
The Street Scouting List has established a real-human collection system, and the four products at this press conference demonstrate the capabilities that the collected data can drive.
The new features released by the Street Scouting List, from a data dimension perspective, actually continue to refine the granularity of the collected data. For example, the Top Scholar List adds three new categories: coffee, bars, and entertainment. The coverage of the Local Flavor Small Shops List expands from 133 cities to 269 cities. The lists are presented based on time, location, and person, and even a Street Scouting Guide is added to supplement the story and visiting tips for each store.
Li Gang, head of Amap's Content Ecology Center, mentioned at the press conference that the algorithm has evolved from counting how many people have visited a store to simultaneously understanding both people and scenarios. This is actually completing data accumulation through real-human annotation, and then turning statistical numbers into elements for understanding scenarios. "Real Scout" is like a productized data collection method. Food Real Scouts turn field visits into content supply, expanding the context of list evaluation sources.
Amap released the ABot Earth world model. With nearly 100 million real-human collected data points, Guo Ning no longer needs Liang Jingru's courage to benchmark against Google. According to Amap, it has created the world's first fully multimodal, predictable 3D native city world model, organizing space, time, and real-world information into an enterable, explorable, and deducible 3D space-time.
At the conference, Amap mentioned 2D network topology and 3D spatial structure, and introduced time-series and real-world scenario changes. These elements together constitute Amap's model of the real world.
The first two layers are the map data accumulated by Amap over the past 20 years and the offline scenario assets of 2.6 million merchants. The latter two layers are based on congestion prediction (simulation data) and 36.6 billion kilometers of flow data. The Street Scouting List continuously provides a steady stream of data supply for the world model, including theater seat selection, finding stores in malls, and looking at steps in Qianhu Miao Village. The three scenarios demonstrated at the press conference all come from real collection.
Combining the launch of the Street Scouting List a year ago to establish a data loop, and the promotion of context capabilities at the press conference a year later, along with the space-time knowledge graph and real-time collaborative simulation of 15 types of dynamic variables. The actions of the Street Scouting List over the past year aim to reconstruct the review system from a data perspective.
This time, Amap presented its own practical results. Three five-day travel guides for Guizhou—two from general large models and one highly detailed guide with over 10,000 likes across the network—all failed when put into Amap's space-time deduction. Sun Chong, head of Amap's navigation products, explained that Amap can calculate the time sequence corresponding to the day of the guide, as well as the rules by which the physical world operates.
Upgrading TOP 100 to the annual BEST 100 list is just a change of name, which is not important. Another noteworthy move for Amap right now is to further seek more and newer physical world interfaces based on the proven effectiveness of the Street Scouting List. For example, the conference mentioned cooperation with Mercedes-Benz to enter in-car systems, as well as cooperative ecosystems with smart glasses, watches, electric vehicles, and robot dogs.
Terminals distribute services for Amap while also providing another dimension of data collection touchpoints. Regardless of the manufacturer or the form factor of the product, the data will ultimately converge at Amap on the application side after list convergence.
Navigation Live is the assembly of touchpoints. According to Sun Chong's definition, navigation has evolved from guiding a single route to accompanying users throughout the day, being reconstructed into a long-horizon agent with spatial perception and action capabilities. To this end, Amap even used the "Yu Chengdong skill," marketing it with the "three fulls": full-duplex interaction, full-modality perception, and full-scenario relay.
This is actually using a free method to acquire a continuous, real-time, first-person data stream of the real world.
The Economic Rent of Data
Embodied AI has not yet jumped out of the cost fence of data. Amap's way of accumulating data assets is completely different: users provide devices, time, and labor; KOLs and merchants provide content; relevant departments step forward to endorse. The collection hardly costs Amap a dime.
The embodied AI industry pays real money for the hard-to-get real data, while internet companies can leverage 800 million data points just with tool products. Maps are the most important mass-level entry point in the real world. Behavioral data, with slight constraints, can prevent fake volume. Overlaying these two conditions, low-cost, high-value data can be continuously transmitted back.
Mr. Ma's behind-the-scenes window guidance shows some profound meaning more or less. Large models have already consumed all internet corpus and need to find the next data goldmine in the physical world.
Language models understand language, and spatial intelligence understands the world. When Guo Ning said this at the press conference, it was hard not to feel that he was delineating the boundaries of two data assets within the Alibaba ecosystem. Alibaba needs a probe extending into the real world to collect flow data of space, time, and people. Amap wants to be the only physical world probe within the system.
This probe took one year to find the necessity of continuing to provide financial support to the group, because for acquiring similar data, the cost of Amap's Street Scouting List method is lower and more seamless. Lists, merchant services, and in-car system cooperation are businesses in themselves, and the data collection cost is amortized into a profitable business.
At this time, Amap holds many options in its hands. One map can support local life services; if progress is not smooth, it can also become a data service provider. The only thing to observe is how it will deliver commercial value to the group after having data and capabilities.
Taking in-store services as an example, besides Amap, there are currently at least two other teams in the Alibaba ecosystem expanding the in-store business format. The logic of Taobao Flash Purchase is from home delivery to in-store, and the logic of Alipay is to find a landing scenario for online payments.
Returning to the war of words among embodied AI companies, on the day Shao Tianlan posted on WeChat Moments, 800 million people were still collecting data for Amap, and no one even felt they were being collected. Internet data collection might provide some reference for embodied AI; data collection should be seamless, imperceptible, and cheap.
WeChat ID | TMTweb
Official Account | Photon Planet