The Genesis of Bytedance's Algorithm

The following post is a translation of Zhang Yi Ming's keynote address at a tech conference 2017. This was just before Bytedance exploded into the stratosphere.

I found it interesting because this keynote explains Yi Ming's thinking around algorithms. First, he used it for the news headlines with Tou Tiao and then expanded it towards social media with Bytedance.

The original speech can be found here.

(Keith's little note: 1) In translating this speech, I tried to be true to the spirit of his words and did not translate it word for word.

The Algorithm As a Living Organism

Compared to the earlier generation of internet entrepreneurs, this generation of mobile internet entrepreneurs faces more intense competition in a compressed time period.

This is because the mainstreaming of mobile internet was first built on the widespread adoption of the internet.

Unlike our predecessors, who had limited technical talents, insufficient capital injections and lesser business experience during the rise of the internet era, we no longer face these problems. Thus, conditions have become ripe for the industry to grow.

I wanted to create a product that would bring a ton of value to as many people as possible. When I graduated, I knew I wanted to collaborate with high-calibre talent to accomplish challenging things.

Many people said that to earn my first 1 million, I must make big money through entrepreneurship. My desire for money was not particularly strong, and I was more interested in technology and products. At that time, I had only one intuition-to create interesting and challenging products with best in-class talent.

(Dang, this is the mindset we need to have!)

I didn't think too much about how to enter the market. Instead, I focused on how I could help users obtain information more effectively. This stemmed from my personal dissatisfaction with existing social network products at that time.I was not satisfied with Twitter, Fanfou, and Renren.

Here's an example: When using social networks in the past, I was often bombarded by a user who frequently posted pictures of his cats and sent Happy Farm game invites. While I had no interest in these posts, the same person occasionally shared insightful commentaries on technology that I found valuable. This raises an important question: How can I effectively filter content to focus on what interests me?

In 2009, I realised that machines could approach human judgment in certain fields through extensive learning. I saw many products related to artificial intelligence, data mining, and personalised applications, and although they were not very successful, they were quite interesting.

Developing machine systems that exceed human intelligence is very difficult, and the human brain is very complex.

Still, it is possible to achieve human judgment in a certain field.The algorithm is a living organism; you must train and tame it. This idea only began to germinate in 2011, when I was thinking about better meeting users' informational demands.

Later, as I conducted more research, I discovered nuances in the nature of users' informational needs. For example, different people have different information needs, but the same person has different information needs at different times.

Algorithms' feature selection requires continuous training and tuning, and because the features vary from user to user, they have to self-evolve. You watch it (algorithm), and it also watches you.

It watches whether you are engaged with the content and constantly monitors , stores and analyses your in-app actions.

Algorithms are also social.

Observing your actions can determine how best to distribute content to someone who shares similar interests with you. The recommendation engine is interesting because it is autonomous.

Once you have developed it, you can leave it on its own, and it will improve itself iteratively.

The algorithm is not sensitive at times, providing some content you are not interested in, but the original posters could make it insensitive. (similar to SEO hacking) Therefore, readers should provide real-time feedback.

If the content is poor, cross it out and let us know. If you find that the recommendations you have recently been given do not suit your taste and you always have to click dislike- I know there is a problem on my end, and I will improve my recommendation engine.

For example, if the system wants to promote a program to female white-collar workers in Beijing at night but later finds that many users do not like it, then we need to study why the system has incorrect results, whether it is a code error.

There is still a lot of room for improvement in recommendation search engines like Jinri Toutiao, and the difficulty is quite high.

There must be more than one opponent ahead, and blocking one person will only delay your progress. It's like a race; your goal should not be to block one person because if you block one person, others will surpass you. You should try to look forward and run forward as much as possible.

For me, the challenge of doing things well is greater than competing with others.

A Message by Lark:

I have been experimenting with Bytedance's Lark workplace solutions recently.

It is clear that Yiming's philosophy has percolated into Bytedance and Lark's DNA.

First, I admire the concept of addressing a clearly defined internal problem.Yiming and the team created Toutiao because they were dissatisfied with the current news platforms. Similarly, Bytedance created Lark because they felt existing tools did not meet their needs as they scaled globally.

Traditional Western SaaS providers expect users to have strong fluency in English so that they can use their tools effectively. While this approach worked for most Western companies, it posed significant challenges for a company rapidly scaling across diverse regions.

So, they solved this bottleneck with Lark, which embeds automated translation functions across all their work applications.

Now, a Singaporean company can seamlessly share SOPs, documents, and presentations with overseas colleagues in China, Vietnam, and Indonesia, who might have varying levels of English proficiency, without the fear of being misinterpreted. This function significantly reduces communication gaps and improves boardwide efficiency.

I must add that Bytedance continues to dogfood its own product.

This means that when it rolls out new features, you can trust that they have been robustly implemented at scale first.

Second, the idea of building a best-in-class product. Larksuite's Docs is the best word editor with an embedded translator. When I read Chinese essays, I can copy and paste the entire text, translate it within the document, and then focus on fine-tuning it.

The time I have spent translating Chinese texts, such as Yi Ming's keynote, has halved.

Other features in their workplace solutions are excellent for cross-cultural collaboration. (If I were to go on, you would have clicked off.)

If you are a business leader with a diverse team looking to enhance your productivity- you can try Lark here!