Seven major problems in platform construction: an in-depth summary of Ant AI platform practice


Introduction: In the process of supporting the operation and development of almost all core businesses of Ant, we have made many attempts in various aspects such as platform construction, business support, platform operation, AI innovation, and overall AI operation, and gained a lot of gains and insights. Share it with everyone here.

Seven major problems in platform construction: an in-depth summary of Ant AI platform practice

In the past few years, my team and I have been responsible for the design and operation of related platform products within Ant Group.

These platform products include the A/B testing platform, machine learning platform, financial knowledge graph platform, NLP platform, intelligent copywriting platform, financial vision (CV) platform, search platform, robot platform, labeling platform, etc. of the artificial intelligence department, as well as risk control The team’s related platform products. These platform products support the operation and development of almost all core businesses of Ant.

Throughout the process, we have made many attempts in various aspects such as platform construction, business support, platform operation, AI innovation, and overall AI operation, and gained a lot of gains and insights.

Recently, I spent some time sorting it out and writing this article.

The content of the article covers seven aspects including “demand management, platform design, product verification, platform collaboration, human confrontation, cross-border thinking, challenge/growth”. Sensitive case.

The length is relatively long, about 15,000 words. If you are interested, you can bookmark it and read it later.

I hope this article has inspired you, and I look forward to being able to throw bricks and spark jade, and have in-depth discussions and exchanges with you.

1. Demand management: “role dislocation” and “selfless state”

1. Excavate needs, be alert to “misplacement of roles”, and put an end to “working behind closed doors”.

The first step in making a good product is to grasp the needs. It is necessary to figure out who the real users of each product and function are.

For C-end products, this problem is easier to solve, because designers and users often overlap. But for technology platform products and B-end products, the two are often misplaced, that is, the designer may not be the real user.

For example, the product manager of Alipay uses Alipay to pay and manage money every day in his daily life. He is a typical Alipay user, so the designer and user are the same person. In technology platforms and B-end products, product designers can use their own products, but they are basically limited to testing and verification, and the real users are other people.

Therefore, some reasoning and judgment of the designer for product requirements may be different from the real situation. Even if he uses it, there is still a difference between the use for testing purposes and the real use.

It can be seen that it is precisely because of the dislocation of this role in technology platform products that it is easy to cause problems in demand control.

Next, let’s start with a short story about our labeling platform.

One day in December last year, the relevant students of our labeling platform held a meeting to review product design.

In the meantime, regarding the details of product design on a marked page, the students in the positions of product manager, UED, front-end, and back-end sitting here expressed their opinions and debated deeply.

Suddenly, I realized a serious problem – that is, all the students in the conference room are not users of this feature.

Because the specific labeling work is done by hundreds of labelers from outsourcing companies, they are the real users of the labeling pages.

If you are not a real user and you are not in that scene, it is difficult to understand the real situation. Therefore, everyone can only judge and deduce based on their own experience and professional ability.

Making products cannot be done behind closed doors. Therefore, we immediately arranged for relevant students to go to the labeling outsourcing company to do on-site research.

At the beginning, we had a small preliminary communication with the team leaders of several annotation teams. At that time, I casually asked about the usage of the product, and they unanimously responded that “there is no problem, it is very easy to use”.

Such an answer is normal. After all, it is difficult to obtain valuable information and understand the needs of users with such a simple and direct question.

In the industry of product managers, we often say that before the invention of the car, if you directly ask the user what he wants, he can only say “I want a faster horse”.
Wuzhao, the former head of Dingding, also talked about this issue when he came to Ant to share the “Dingding Entrepreneurship Road”.
His point of view is that seeing users can’t just “discuss the facts” and only ask shallow questions related to product use. (Even if you ask such a question, you can’t ask straightforward questions such as “what do you need” that are difficult to obtain real needs).
The correct way is to leave the specific products behind first, and learn more about the customer’s background, business, status and other overall, background, ins and outs, to show “interest” in the customer, and to become the customer’s friend.
Only in this way will customers be willing to chat with you more and in depth, and only in this way can you capture valuable information. In addition, by observing the specific behaviors and operations of customers, real needs can be captured and insights can be achieved.

So, after the meeting, we asked to go upstairs to the office area where the employees were marked to see the situation in detail.

When we stood behind the labelers, carefully observed their operations, and had in-depth conversations with them, we made new discoveries.

Many unimaginable usage methods, scenarios, and details of product design have emerged in the continuous operation of labelers. The questions that everyone debated at the previous product review meeting naturally have answers.

After half a day, we recorded a total of dozens of valuable feedback and findings, and processed and followed up one by one in the follow-up work.

It can be seen that if you are not a real user and you have not observed the operation of a real user with your own eyes, you cannot predict many problems.

Everyone’s IQ is not bad. When encountering problems, we are often used to talking about methods and logic. We often fight with each other in the conference room or even slap the table and stare. In the end, no one can convince the other and no effective conclusion can be obtained.

At this time, you might as well ask yourself “Who are the real users?”, and then try the “stupid way”, go out of the office, go to the customer, ask them, chat with them, and see how they use us The product.

At that time, many problems suddenly became clear.

2. Satisfy the needs, constantly “from the shallow to the deep”, and cultivate the “state of selflessness”.

Next, let’s think a little deeper.

Now, assuming that you have identified who the users are and have grasped the general context of the needs, you should also consider the question of “whether the understanding of the needs is deep”, that is, the question of shallow needs and deep needs.

In other words, it is also a matter of means and purpose – “shallow needs” are often just means, while “deep needs” are the purpose.

For example, for the financial vision platform we are in charge of, some users feedback “I need a model report”, that is, after the model is trained, some indicators such as “accuracy rate, recall rate, AUC, etc.” are displayed in the form of charts .

If you just meet this requirement, it is not enough.

why? Because the model report that the user wants is only a “shallow demand” – he does need to look at various indicators, but what he wants most is to compare the effects of different versions of the model after the new model is trained – Not only do we need to know what the indicators are, but we also want to know the specific changes in the indicators, which ones have risen, which ones have dropped, and what the specific values ​​are.

Only in this way can the deep needs be met.

The reason is the same, and similar problems will also be encountered in C-end products.

If you pay attention, you will find that the product managers of many e-commerce websites and car shopping guide products have already touched the deep-seated needs.

For example, there is basically a “model comparison” function in the car website: not only can the configurations and parameters of different models be listed item by item in a table, but also provide “highlight different configurations, hide the same configuration” and other intimate functions. Function. This is to deeply satisfy the needs of users.

Therefore, for a requirement, ask a few more reasons, ask yourself “Is this the real purpose of the user? What does he want to do with this function?” and so on. Only in this way can it be possible to touch the deep-seated needs of users and make functions that make users feel very intimate.

To further satisfy user needs, in addition to doing superficial and in-depth analysis, the idea of ​​”divide and conquer” can also be used to layer products from modules and functions, that is, to separate “N-level rockets”, and each level of “rocket “Used to meet the needs of different types of users, or the needs of the same user at different stages.

For example, although the functions of our graph, NLP, CV, search, robot, labeling and other platform products are different, we still found a commonality, that is, we abstracted the “five levels” of demand classification and business empowerment. “Rocket” includes five levels including “function embedding, API calling, data training, model customization, and algorithm development”. The business side can choose different access methods according to specific situations.

  • The first level, function embedding: embed a certain functional module of the platform into its own system through iframe and other low-cost means.
  • The second level, API call: directly call the mature API provided by the platform, such as calling the OCR recognition API such as ID card and driver’s license.
  • The third level, data training: the model of the platform meets the requirements, but it needs to provide its own training data or dictionary data to solve the specific scene requirements.
  • The fourth level, model customization: the on-site model of the platform does not meet the requirements, so it is necessary to configure the algorithm parameters, and then train a new model that meets your needs.
  • The fifth level, algorithm development: the most advanced situation is that the business side understands the algorithm and needs to develop a new algorithm. The platform provides a series of in-depth capabilities such as “algorithm development, data management, model training, model testing and release” to improve the efficiency of algorithm research and development.

The above-mentioned “five-stage rocket” meets the needs of different types of users and the same user at different stages from the shallower to the deeper.

I remember many years ago, I participated in an advanced training course in management. The training lasted for several days, and there was a lot of content, but I forgot almost all the training content—except for a “universal four-step method” accidentally introduced by a teacher.
The so-called four-step method is the four steps of “classification-sorting-finding rules-application”. Whether you are learning new domain knowledge, taking over a new job, or coming to a new environment, you can try this universal four-step method. I believe that no matter how complicated the problem is, it can be easily solved.
User stratification and five-stage rocket are an application of “classification-sorting”.

After talking about “needs/user stratification, five-stage rocket”, does it mean that the user needs are met 360 degrees without dead ends?

The answer is no, because we have not yet achieved the “state of selflessness”.

The so-called “no-self” state means that when meeting the needs of users, we should not only consider “who I am and what I have”, but forget about myself and see what users need and what is most useful to users.

For example, although you are a product manager of an AI technology platform, you can’t only see AI, algorithms, and models in your eyes—to achieve “no self”, you must do this: if there is a non-algorithm, non-AI product strategy, If it can really help the business, then it should be done.

In the eyes of business students, it doesn’t matter whether there is an algorithm or not, and whether it is high-tech is not important-but whether there is a business effect is the key. As the saying goes, it doesn’t matter whether a cat is black or white, as long as it catches mice, it is a good cat.

For example, our intelligent copywriting platform can intelligently generate marketing copywriting for thousands of people. In the past, we have been iterating products, improving algorithm capabilities, and trying to generate more intelligent, accurate and personalized copywriting.

However, we all know that the improvement of the algorithm cannot be achieved overnight, and the effect of the algorithm is slowly polished and optimized.

In this process, product manager students cannot wait.

So, we are thinking, no matter how advanced the algorithm and how smart the platform is, what we produce is still copywriting. As for the position of copywriting, with the development of the advertising industry, it has existed for hundreds of years, so there must be a mature methodology and model.

As Internet practitioners, we advocate innovation and subversion, but we must also remain in awe of the industry.

So, our product manager classmate went to study some classic books on marketing and advertising copywriting, and summed up the so-called “18 high-quality copywriting sentence patterns/templates”. Scientific principles in advertising, psychology, and more.

After commercializing these “high-quality sentence patterns” and “copywriting rules”, with the help of algorithms and technologies, more effective copywriting can be output to the business.

We believe that machines cannot completely replace humans, and that machine intelligence and human wisdom such as industry knowledge and expert experience will definitely complement each other.

2. Platform design: platform products must also be “understandable”

After talking about the requirements, let’s talk about the design.

In the Internet industry, products for C-end users are not only abundant and abundant, but also generally free, and the acquisition cost is basically zero.

If you don’t pay, you won’t “cherish”.

Therefore, for users, the product must be easy to use, that is, it must be “understandable in seconds”. If the user can’t understand or use it for a few minutes or even tens of seconds, then he basically gives up, and the product has no chance.

For mid-stage and platform products, the same is actually true, but users can only endure the unpleasant experience, because using your product to solve his business needs is his essential job.

However, this does not mean that the product can be messed with casually, because he can have other choices. You have to know that there are often similar products in the company, not to mention external, free open source or paid solutions.

Therefore, you also need to work hard on platform design. You must be able to quickly capture users, let users quickly get started, access, and go online, and help businesses get business results.

How can we achieve “understanding in seconds”? It can be considered from several aspects such as “product framework, terminology system, help guide, product demo, unified interaction”.

1. Have a clear product framework

As soon as the user opens the page of the platform, they should clearly perceive what the platform can do, what the product framework looks like, what functional modules are included, the relationship between modules (including, sequence, etc.), what to do in the first step, what to do in the second What to do next, and so on.

This does not seem to be profound, but a common problem is that product managers do not think enough about the framework in the early stage of product design. It is often at the stage of PRD and visual review that it is discovered that the module design is unreasonable, the process is not clear, and so on. At this time, the cost of rework and modification will be high.

To make matters worse, frequent rework and changes will completely destroy the professionalism and authority of the product manager. In the future, how can we raise demand and grind resources for technology?

In order to avoid such tragic things from happening, product managers must work harder off stage.

A good habit is to reconstruct in the mind first, and then draw with the pen. Many product managers are used to drawing demos as soon as they come up. This is wrong-the cognitive and computing resources of the brain are limited, and if you focus on “this”, you will lose “the other”. When you get caught up in various details, it is impossible to fundamentally Think about the problem in terms of the above and the framework.

then what should we do? You can make full use of the tool of mind map. Specifically, you should not consider any demo diagrams, but sort out the entire product hierarchy of the platform, including navigation and modules at all levels, pages contained in each module, and core functional sections. After drawing the brain map, from the perspective of the user, sort out and simulate repeatedly until the horizontal and vertical logic and processes are no problem, and then do specific demos and PRDs.

2. There is a terminology system as the name implies

After sorting out the overall framework of the product, we should also pay attention to the “term/concept system”, that is, the naming of core concepts in the product and the design of the logical relationship between concepts.

The reason why this is important is that the concept and terminology system is the result of the accumulation of knowledge in each field, and it is also a medium for people to learn new things and communicate.

If the concept is complex, the product must be complicated; if the concept is simple, the product can be simple.

For example, the same is instructions and way of human-computer interaction, WeChat “shake” can let users “as the name implies,” and immediately the functioning sense to do so, and we pay treasure “xiu xiu”, is more difficult to understand and put them into action.

For another example, when Jobs released the iPod, he did not directly and abstractly say “the storage space is as high as 4.8G”, but said “put 1,000 songs in your pocket”.

It can be seen that the unreasonable naming of new concepts in the product, or the direct exposure of obscure underlying terms will cause great trouble to users.

For another example, in the A/B experiment platform, the initial conceptual system from top to bottom is “business domain-business line-product-experiment”.

We found that it is difficult for users to distinguish the difference between “business domain” and “business line”, and the “products” in it are not products such as “payment, borrowing, Huabei, Yu’e Bao” as everyone understands, so there are many problems .

Later, with the help of the well-known “physics laboratory, chemistry laboratory” and other things, we transformed the conceptual system into this: Darwin is an “experimental platform”, in which “xxxx laboratory” and “yyyy laboratory” can be created. In a laboratory, various “experiments” can be done. In this way, it is much easier to understand.

In addition to this, we have also made changes to the character naming in the lab.

In the previous experiment authority management, there were two common role settings of “administrator” and “member”. We also referred to the job titles of laboratory staff in real life and changed them to “laboratory director” and “researcher “.

Interestingly, “researcher” has the hierarchical meaning of “high P/organization department” in the Ali system. Such a small modification of copywriting also contains the “humanistic care” of platform designers-for those who use A/B Students who practice data-driven innovation and pursue scientific and rigorous ways of doing things through experiments, give a little warmth and honor.

Moreover, it will be easier to do future operational activities, such as appraising the “Top Ten Researchers, Top Ten Laboratories” and so on.

In short, when designing the terminology system of products, first of all, “don’t add entities if not necessary”, and secondly, try to use the concepts already in everyone’s minds, instead of directly copying technology to achieve, or creating new concepts.

3. There are appropriate help guides

Even if you work hard on conceptual design, there is no guarantee that users will not have any questions.

Therefore, it is necessary to design a “help system” for further explanation and elaboration.

Here, it is not that you are asked to write a lengthy product document. The documentation should be written, but it’s not the point, because most people don’t read the product documentation carefully before operating it-he can only look up the manual if he encounters a problem.

The “help system” mentioned here refers to the productized help system, that is, “documentation productization”. Specifically, it is to embed the main points in the help document into the product page as much as possible, so that the product can realize “self-explanation”, instead of putting it outside the product and only storing it in the help document.

“Document productization”, the specific measures include the following aspects:

There are auxiliary instructions on the page

A common situation is that our page is too clean and empty, and we are reluctant to put an explanation. When users encounter problems, they will be at a loss. Therefore, you can explain in small words under the title, and give a tip bubble reminder on the concept. For complicated situations, you can also add a “Learn more” link after the help text-directly jump to the corresponding place in the help document, instead of asking the user to search from the beginning.

New features are online, with prompts and notifications

The platform continues to iteratively improve, but it is often found that users do not know that new features are available. Therefore, you can make appropriate reminders and notifications about this: large iterations can cover pop-up windows, small changes can produce small red dots, and so on.

4. There is a simple and intuitive demo of the whole process

You can’t learn to swim just by watching teaching videos, and you can’t learn to drive in optics “subject one”.

Talk about it for a long time, it is better to play it again.

The status quo is that many technical platforms have no demo and experience capabilities at all. Then, it is difficult for users to get started.

Therefore, the platform must build a set of “full-process, body-feeling, and easy-to-use” demos for users to experience for themselves.

The whole process means that your demo should cover all the links and steps of the platform. Body sense means to have intuitive results (instead of only displaying abstract values, json code output, etc.). Ease of use means that it should be simple enough and can be completed in a few minutes (so you need to build several sets of demo corpus, graphs, data sets, etc.).

For example, in the NLP platform and financial vision platform, users can easily experience the effects of financial NER/text classification, ID card/bank card OCR online.

It can also complete the whole process of “project creation, data upload, data marking, model training, model testing” and other links.

It is worth pointing out that the demo of the platform must be as simple as possible, and do not overestimate people’s patience.

I remember that after the first version of the full-process demo of the financial vision platform was launched, when the project team members experienced it in detail, they found it was still very cumbersome and even had to give up.

To complete the demo, you still need to write a bunch of forms, such as project name/introduction, model name/introduction, dataset name/introduction, and prepare training data by yourself, so you have to search online and download dozens/hundreds picture…

Later, we greatly simplified this. If users can click the mouse, do not let users enter words, such as automatically filling in various names and profiles. In addition, the platform also has built-in test data sets for users to use and so on.

After some simplification, the user can complete the whole process and a very physical demo within a few minutes.

5. There is a standard/unified interactive experience

In addition to doing a good job in the design of each platform, it is also necessary to consider the consistency of experience on different platforms, that is, the unification of platforms.

Doing this well will not only allow users to reduce learning costs, smoothly switch between different platforms, but also reduce duplication of work for UED, product managers, and technical students.

First of all, the common framework and modules of the platform can be abstracted and unified, including Portal page, project management, authority management, data management, task management, release management, etc.

Secondly, unify the experience of details, specific to the design, naming, color, position and so on of the components.

When we precipitate a set of classic product framework and interaction standards, the product iteration speed and user experience will be greatly improved.

3. Product verification: if you don’t use “deep”, you can’t do it well

1. In-depth verification, not superficial

If a product manager wants to really make a product well, he must use it more.

The reason is very simple, but what we want to talk about here is the “depth” of use – just click and look, it is very different from the use of depth.

For example, if you are asked to design an intersection turn prompt in a navigation product, you may think that it is no problem to design it as something like “turn right at the intersection 500 meters ahead”.

You see, it not only includes the distance, but also clarifies the direction. It feels perfect. However, when you use the product in depth, when you drive yourself, you will find that this is not the case – it is difficult for you to accurately grasp whether you have reached 500 meters, and it is very likely that you will mistakenly advance to the right at an intersection at 300 meters. Sended.

Therefore, the current navigation prompt will not only say “Turn right at the Nth intersection 500 meters ahead”, but also prompt “Passing the N-1th intersection” at the intersection that should not turn right. Users will not go wrong.

For our annotation platform, the in-depth use is reflected in the number of data annotations-your perception is completely different between several times and dozens or hundreds of times.

Some design details in the annotation page are not obvious when you do the annotation once or twice. After you do it dozens or hundreds of times, even the smallest problems will be exposed and magnified.

For example, there is an image classification task where you only need to label “true” or “wrong”.

The previous design was to display a large picture on each page, and switch to the next page after answering the questions. When we personally marked dozens of sheets, we felt that the efficiency was very low.

So, we changed it to display ten or twenty pictures on one page, and the annotators only need to glance at them, check the “right” or “wrong” ones, and then submit them as a whole (it also reduces every Page refresh page, waiting time for loading pictures). Such a simple change is actually not technically difficult, but the labeling efficiency has been directly improved many times.

2. “Do business” by yourself, the result is very different

To really do a good job of a platform, not only as mentioned above, you should be more of an “annotator”, but also a “business side”. There is still a big difference between supporting business, empowering business, and doing business on your own.

Next, let’s use the case of our waste intelligent classification project “Category Treasure” as an illustration.

In July 2019, many cities across the country began to implement garbage classification.

Based on the accumulated image, NLP and map and other AI technical capabilities, our students quickly developed technologies and products for intelligent garbage classification. The project is named “Sorting Treasure”. Users can experience AI garbage sorting on Alipay applets and smart garbage recycling bin IoT devices through convenient interactive methods such as “taking photos and voice search”.

This project did not start when various business BUs gave us demands. This time, we have dual identities. We are both the platform side and the “business side” for the first time.

After starting the business side, we discovered that waste sorting seems simple, but actually involves many complicated links, from “acquisition of training data, sorting of item categories, maintenance of waste sorting standards, online return data Correction of item category weight and priority, confirmation of labeling results, coordination with various internal departments, connection with outsourced ISVs, handling of holidays and special items, etc.

After a lot of frantic tossing, I finally finished the project stumbled.

During this process, we encountered many problems that we did not know before, including product problems with unreasonable platform design, and technical problems such as too long training time.

More importantly, let us see the “vacuum zone” between different processes, different systems and different teams-this is exactly what is often said in large companies due to the division of labor and boundaries. question. And these connection problems are hidden problems that greatly affect efficiency, which need to be discovered and solved through mechanisms such as products and processes.

This practice of “doing business by yourself” made our platform students change their perspective, deeply realized the difficulty of business students, and directly promoted the iterative improvement of the platform, as well as the improvement of team cooperation and process setting.

4. Platform Collaboration: Connection, Generate Value

A lot has been said before, but most of them still focus on individuals on a certain platform.

A platform that exists in isolation may be reduced to a tool, and its value and energy will become very limited.

Therefore, to do a good job and expand the platform, we need to jump out of the platform itself and look at it from the perspective of connection, overall situation and ecology.

If different platforms are allowed to collaborate and connect, the effect of “1+1>2” will be produced. If the “control flow and data flow” closed in the platform are extended and turned into a closed loop, many innovations will burst out.

Below, several methods and cases are introduced.

Cross link, with exposure and traffic

This is the simplest method of platform collaboration. Each platform should not only fulfill its own mission, but also consider doing something for other platforms, such as bringing exposure and traffic. Therefore, we added an “AI product matrix” menu to the navigation bar of each platform product, listing the logos, names, and links of seven or eight products. The data shows that this small menu can bring considerable exposure and conversion to other platforms every day, and the ROI of making this menu is very high.

Reuse of platform capabilities to eliminate waste

In the process of continuous iterative upgrading of the platform, for a new requirement, do not do it yourself as soon as it comes up, but first check whether other platforms have ready-made capabilities that can be reused, even if it is “curve to save the country” or “expedient measure” .

For example, the knowledge update of the knowledge graph platform and the release of copywriting on the smart copywriting platform all need to go through the marking and confirmation process. We found that the labeling capabilities of the labeling platform are sufficient. Therefore, instead of redeveloping, we opened up connections between platforms and quickly solved this problem.

Feedback and closed loop to achieve common development

If a platform only has a one-way output capability, but does not get feedback from the downstream and does not form a closed loop, it is not a perfect system and platform.

For example, our labeling platform has cumulatively marked hundreds of millions of pieces of data, and these labeled data make it possible to train various models. As the saying goes, without artificial intelligence, there is no intelligence.

In this process, the labeling platform only outputs value and assists intelligence, and does not benefit from intelligence itself.

Later, we considered to form this chain into a closed loop, that is, let the model trained by marking data be fed back to the labeling platform, so as to realize “intelligent assisted labeling”.

In this way, the entire platform is transformed from “pure manual labeling” to “intelligent assisted labeling”, which greatly improves labeling efficiency and reduces labeling costs.

Precipitate data assets to create greater value

If a platform has data accumulation, then these data need to be deeply mined to generate more and greater value.

For example, each business is initially connected to the knowledge graph platform. In order to solve its own business problems, it has to build a schema and import data from scratch. But with the development of the platform, the accumulated knowledge is getting richer and richer. Then, subsequent platforms can directly benefit from the knowledge accumulated before, without having to rebuild themselves. This is the value precipitated by platform data.

For another example, the life cycle of the labeled data in the labeling platform ends after the model training is completed, and it is a pity that there is no one to manage it.

Now we plan to deposit and open up these data, so that the data can generate greater value.

First, the labeled data is open to the public. When the business is just connected to the AI ​​platform, there is a cold start stage, and the most lacking is the marking data. Therefore, it is possible to sort out and open up the massive labeling data in the labeling platform, so that the business can first search in the platform to see if there is any existing data, and if there is, it can be reused. If not, consider rebuilding the data again.

Second, the labeled data is open to the outside world. We can open up some data that does not involve privacy or our core technical capabilities to create greater value for society.

For example, in the intelligent garbage sorting “Sorting Treasure” project, hundreds of thousands of marked garbage image data have been accumulated. In addition to opening up the relevant model APIs, and opening up some of the data, we will intelligently process the garbage of the entire society and contribute to the strength of ants.

Access to an open platform to achieve strong alliances

Here, let’s talk about the specific approach of opening up. If you directly open it to the outside world, it will be more troublesome to do it, and there will be a lot of docking and maintenance. You should consider connecting your capabilities to existing, large platforms, such as the Alipay applet platform/open platform, Alibaba Cloud platform, and so on. With the help of these large platforms, many things such as customer acquisition, docking, and operation and maintenance can be covered.

Here, I would like to share another way of thinking about platform collaborative innovation, which is the “graphic method and exhaustive method”.

In the beginning, collaborative innovations on the platform happened at scattered points, and one was made when one thought of it, which was very unsystematic and systematic.
Later, in order to exhaust all the possibilities of “connection” and “synergy”, we drew a large map and matrix diagram of system collaboration, put all the platforms in it, and thought about what is there between the platforms in an all-round way. What is the possibility of collaborative innovation?

Seven major problems in platform construction: an in-depth summary of Ant AI platform practice

This method can also be used as a reference when doing other work.

5. Confrontation of human nature in the platform

It is often said that where there are people, there are rivers and lakes. A platform is also a world.

People with different roles and appeals participate in it, and the humanity is displayed.

Therefore, it is necessary to think about people’s affairs, and it is necessary to operate and govern the platform.

1. Misuse of the platform

First, incorrect usage that occurs on the platform must be corrected.

Why does this happen?

The reason is that although the product manager will try his best to prevent most mistakes when designing the product, and there are corresponding rules in the platform’s gameplay to inform users, but everyone will not “obey the rules” as you imagine , They will intentionally or unintentionally “magically use”, “misuse” or even “abuse”.

For example, when I was in charge of the A/B experiment platform last year, we conducted an in-depth analysis of all the experiments in the platform, and found many surprising phenomena.

  • There is only one version of hundreds of experiments: Normally, two or more versions are needed to conduct controlled experiments, but many experiments have only one version. One of the big “magical uses” or “misuses” is that users only The platform is used as a grayscale platform.
  • The traffic in hundreds of experiments is 0: Some users did not use the distribution capability of the platform, but did the distribution themselves, which we did not expect.
  • Hundreds of experiments run for less than 3 days or more than 30 days: Normally, experiments need to run for about a week. However, many students run the experiment for a day or two, and push the experiment to completion or take it offline as soon as they see a change in the data. This is actually unscientific. Some experiments have been running for dozens of days. The reason is that someone forgot to deal with it. Maybe the experimental scene does not exist anymore.
  • ……

It can be seen that everyone’s understanding of A/B experiments is still not enough, resulting in various “unique” usages on the platform. Then, more work needs to be done in terms of platform training and product design.

In addition to platforms such as A/B experiments, many problems have also been found on platforms such as our financial knowledge graph.

We know that in the Schema specification of the knowledge graph, the same entity can only have one type.

For example, for “company”, the most common entity type in the financial field, it is enough to globally define a type called “Company”. Different business domains can have different business scenarios, but the types should share one.

However, the reality is that business students often want to create a type for simplicity and control. As a result, repeated types like Company1 and Company2 appeared on the platform.

On the graph platform, in addition to schema duplication, data also has duplication and inconsistency, which need to be managed one by one.

However, the matter of platform governance is both science and art-neither laissez-faire nor too strict. Especially in the early stage of platform construction, if the restrictions are too rigid, it will be difficult for the business side to understand and cooperate, and even lose customers.

Therefore, we must grasp the strength.

2. “Abuse” and “Violation”

The problems of platform governance mentioned above are actually not too bad.

Next, I will introduce to you some “abuse and violation” behaviors that need to be taken seriously and dealt with seriously.

They are two real cases in the annotation platform: “task release” and “colluding with foreign workers”.

Let me talk about the first one, the abuse of the “task release” function.

Considering that there are many changes in outsourced labelers, the product manager designed a “task release” button on the label page to prevent the task from being stuck in the hands of one person.

However, the labeling team leaders later reported that they “hope to cancel this button”, saying that this button is used by many labelers to “select tasks”: when encountering difficult labeling questions, they click “task release” to skip up.

Therefore, we withdrew this function from the front-line labelers and only opened it to the team leader (this problem was also discovered during field research in outsourcing companies, and the team members did not expect it before).

The second is a violation of regulations, which refers to the collusion of personnel to “move foreign workers”.

For a while, algorithm classmates reported that the labeling speed had dropped. We analyzed the report and found that the labeling speed of multiple labelers in individual groups has decreased, including those who did it faster before.

After investigation, it was found that some black sheep not only lazy themselves, but also instigated and colluded with other people to reduce the labeling speed together to collectively “grind foreign workers”.

Of course, the most fundamental reason for the problem of “colluding with foreign workers” lies in the performance management plan of these labelers-a monthly salary system was used instead of a piecework system, and there were performance bonuses, but they were minimal.

Recently, we have established task difficulty grading standards in special projects, and are improving the overall management plan for outsourced personnel.

3. “Too smart” doesn’t work

Finally, one more very interesting thing.

We know that if a product is not considerate, smart and intelligent enough, users will definitely not like it, but conversely, if it is “too smart”, sometimes it will not work.

People are restless and anxious. If it makes him feel “too amazing to know what’s going on inside”, he dare not use it.

For example, among the products of the model service platform, some students have designed the “one-click model deployment” function, which automates the complex and tedious feature processing in the offline model deployment to the online process.

However, when everyone spent several months developing it, they found that they couldn’t find a business partner because everyone said they dare not use it. In the end, this “smart” one-click deployment function had no choice but to go offline.

(It should be noted that this is not to say that there is a problem with the product direction of “simplified model deployment”, but the above-mentioned “black box, leaving users with no idea” solution requires more consideration and more consideration from the perspective of users )

6. Cross-border, cross-border, cross-border

The so-called cross-border is the behavior of breaking through the original industry conventions and routines, and achieving innovation and breakthroughs by grafting ideas and technologies from other industries.

Charlie Munger, the world-renowned investor and Warren Buffett’s golden partner, is a man of great wisdom. He highly admires cross-border thinking. He pointed out:

  • You have to think interdisciplinaryly.
  • You must regularly use all the concepts that can be learned from freshman courses in various disciplines.
  • If you can master these basic concepts proficiently, you will not be limited in the way you solve problems.

To do a good job in the design, operation and promotion of technology platforms, you also need cross-border thinking and style of play-for example, you can combine marketing thinking with products and technologies cross-border.

The so-called marketing thinking, in simple terms, includes several key points such as “cognitive law, brand system, material carrier, and communication path”: first, we must obey people’s understanding of new things (simple and intuitive), and build a set of The system of brand recognition and memory (logo, naming), constantly planning creative activities and materials, and exposing and disseminating them in appropriate places.

Then, for the operation and promotion of technology platforms, the theories and methods in the above-mentioned marketing field can also be used across borders.

Specifically, we can start from the following aspects:

Platform products need brands

We sorted out the brand identity systems of all the platforms, referring to the practice of “Ali Zoo”, and named them Zhispider Financial Knowledge Graph Platform, Whale Language NLP Platform, Figure Eagle Financial Vision Platform, Thousand Sturgeon Search Platform, Lingxi Robot Platform , the selection of each animal reflects the characteristics of the platform products as much as possible (the names of the Picasso intelligent copywriting platform and the AlphaQ intelligent labeling platform have already been recognized to a certain extent, so they have not been modified).

In addition to the name, our awesome UED students also designed a very differentiated, memorable, and extremely exquisite logo. With the name and logo, it is much easier to communicate, disseminate and promote.

The product system needs a brand

Not only should each platform be given a degree of memory and recognition, but also how multiple platforms should be remembered and disseminated as a whole. Also considering Ali’s martial arts culture, we packaged the overall brand concept of “AI Zhongtai Tianlong Babu” to spread the eight major AI technology platform products. It was later discovered that the “Tianlong Babu” has a high internal influence, and many people use “Tianlong Babu” to refer to the AI ​​technology platform family as a whole.

Operational activities require a brand

To do operation and promotion, we also need to have a brand system. Therefore, we constructed an image of an “AI commissioner”. All articles, videos and posters we publish internally are included in this system. For example, all intranet article titles and the beginning and end of the article are in a unified format, and the name and image of “AI special commissioner” are added, which not only facilitates the formation of unified cognition, but also facilitates everyone to retrieve information in the future.

In addition, in the design of operational activities and materials, there is also brand marketing thinking. No matter how advanced the technology and platform are, interaction, creativity and fun must also be considered when disseminating.

To this end, we customized interesting Coke bottles with the platform name and slogan printed on them, and issued “Letters of Appointment” to students who marked their product experience, etc.

It can be seen that the cross-border integration of marketing, technology and products, and the design of product brand system, operation activities and materials from the perspective of users will receive better results.

7. Challenges and growth of platform product managers

After reading this, you may think that being a platform is very interesting and easy.

In fact, it is not the case, it is difficult for everyone.

For product managers of technology platforms, they will face all-round challenges of “heart, brain, and body”.

In terms of professional skills, in addition to the “requirements management, product design, project promotion” and other capabilities necessary for the product manager position, “technical understanding” is also required. To understand the R&D process, you need to understand the terminology and principles of various algorithms and models, because you not only have to talk to the development team of the platform, but also to talk to the users of the platform—most of these users are also technical classmates.

This does not require you to know more about technology than your technical classmates, and to replace technical classmates to do technical things, but requires you to understand the essence of technical points, to know what this technology can and cannot do, and how this technology differs from other technologies. What is the difference, and what is the development context of this technology.

When you work hard to figure out these problems, you will not be in a too passive situation.

However, “lack of initiative and lack of sense of accomplishment” still plagues the product manager students of the technology platform.

To solve this problem, the following aspects can be considered.

In-depth understanding of business needs, improve business sense

Ultimately, the platform serves the business. No matter how powerful the platform is, it will not be able to gain a foothold if it does not help the business. Therefore, when you have a full grasp of business needs, you can plan the direction of platform construction in a reasonable and well-founded manner, and you will feel a sense of accomplishment.

Consider what unique value you can bring to the team

The success of a project and a platform requires, in addition to professional capabilities, sufficient communication, coordination, promotion, BD, and sales capabilities. It is no exaggeration to say that to make a good product, the product manager is not only a product manager, but also has the role of a “little CEO” of the product. When you accomplish one thing through your own efforts, you will be very happy and win the recognition of the team.

Everything has room for innovation and improvement

For the labeling platform, you can follow the old path of “manual labeling”, or you can innovate in the direction of “intelligent assisted labeling”. For the intelligent copywriting platform, you can only rely on the path of algorithm improvement, or you can take the initiative to innovate and commercialize domain knowledge and industry experience to realize product manager drive. For the acquisition of user feedback and the iterative evolution of products, you can use the traditional method of “face-to-face conversation and questionnaire survey”, or try the new method of “analyzing user logs and using big data + AI”. We must believe that as long as we start from the end, start from the business, and start from the user, we will be able to find opportunities for product innovation.

Always be in awe of products and users, and do everything seriously

We once used this sentence to encourage the students in our team: We have to use the mentality of making hundreds of millions of DAU products to polish the technology platform of hundreds or thousands of DAUs. Serious people will not suffer. Every effort you make today will generate value and improve yourself. There is no way to go in vain in life, every “need” counts.

8. Conclusion

Finally, it’s the end. Here, I will make a summary of the content of the article:

Demand Management: “Role Misplacement” and “Selfless Realm”

The more basic and simpler questions are, the more difficult they are to answer, and the easier they are to be ignored intentionally or unintentionally. The first step in making a product is to answer these basic questions: find out who the user is and what the user’s real needs are. In order to deeply satisfy the needs of users, we must ask why and understand the real purpose of users. Forget about yourself and think more from the perspective of users.

Product design: platform products must also be “understood in seconds”

If a product looks messy at first glance, and you can’t figure out what’s going on, it’s basically a failure. Therefore, it is necessary to start from multiple aspects such as “product framework, concept system, help system, demo experience, and interaction unity” to achieve “understanding in seconds”.

Product verification: if you don’t use “deep”, you can’t do it well

If you want to make a good product, you must do a good job in product verification. Product managers must find ways to use their products frequently and deeply. If you have the opportunity, you have to “do a little business” yourself, and you will be amazed “ah, there are still so many problems”. In this process, you will also have many unexpected gains.

Platform Synergy: Connecting to Generate Value

The value and energy of a single platform are limited. When you break through the boundaries of the platform and create more connections and closed loops, you will create a thriving system and ecology.

Human confrontation in the platform

Where there are people, there is humanity. For a platform with multiple roles involved, operations, guidance and governance are required to ensure the stable and healthy development of the entire platform.

Crossover, crossover, crossover

Facing a complex and ever-changing environment, diverse talents and complementary skills are required, as well as cross-border integration of different industries and fields. Cross-border will produce chemical reaction, and cross-border will produce innovation.

Challenges and Growth of Platform Product Managers

In the dictionary of adults, there is no word easy. Only when there are problems and difficulties can platforms, teams and individuals improve and develop. The position of product manager is a complex, not a single skill can gain a foothold, product manager students need to constantly meet challenges and constantly cultivate themselves.

Believe in the power of the platform and the power of the product.

We’re just getting started, and we keep going.

Author: Bai Ning, a senior product expert of Ant Group, has been in charge of the products of Ant AI platform and risk control platform successively.
Original link
This article is the original content of Alibaba Cloud, and shall not be reproduced without permission