FB Pixel no scriptWhy the secret AI sauce behind TikTok is such a vital ingredient in luring potential buyers | KrASIA
MENU
KrASIA
Insights

Why the secret AI sauce behind TikTok is such a vital ingredient in luring potential buyers

Written by South China Morning Post Published on   5 mins read

Share
ByteDance has long been a proponent of content recommendation systems, and uses it on other products, such as the popular news aggregator Jinri Toutiao.

When ByteDance rebranded the US teen karaoke app Musical.ly it acquired as TikTok in 2018, it was just another short video app for American teens.

Today it is the most downloaded app in the world, proving so popular that it has become a flashpoint in the escalating US-China tech war, which was previously focused on heavy-duty areas such as chips and 5G networks.

Washington has demanded a forced divestment of TikTok in the US from its Chinese owner ByteDance on concerns over the safety of personal data. But updated export controls from Beijing last week – which cover two key technologies used by the short video app – have cast an added shroud of uncertainty over the sale.

Among the newly-added export restrictions – the first time China has updated such rules in over a decade – are “personalized information push technologies based on data analysis” and “artificial intelligence interactive interfaces”.

Both of these tools are used to build ByteDance’s powerful recommendation system, which feeds curated content to users based on their interests and activity. While Musical.ly may have given ByteDance a foothold in the US market, it was the secret sauce – its algorithm – that allowed it to build momentum.

Ever since its formation in 2012, ByteDance has been a proponent of content recommendation systems, and uses it on other products such as the popular news aggregator Jinri Toutiao. TikTok takes into account three elements to build a recommendation: user interaction on the app, such as liking a clip or following an account; what the content contains – in short videos, things like sounds and hashtags; and what environment the user is in, such as one’s language preference, country setting, and device type, according to information revealed by TikTok in June.

At the same time, the app also feeds in a certain amount of video content outside the user’s direct interests.

China digest

The Chinese policy change may mean that a quick deal to sell TikTok’s US operations, by September 15 at Trump’s request, “is very unlikely to take place”, the Post reported this week, citing a company source.

But why is the algorithm so important?

After ByteDance acquired Musical.ly and later merged it with TikTok, it introduced its algorithm to the lip-synching platform, boosting user time spent on the app significantly. The change “was anything but subtle”, says Eugene Wei, a product expert, on his personal blog.TikTok users on Android phones spent more than 68 billion hours in total on it last year, more than three times the figure the year before, according to market research firm App Annie. And it had nearly 92 million monthly active US users by June 2020, more than eight times the figure from January 2018, according to the legal complaint that ByteDance filed in late August against the US government.

TikTok was the most-downloaded non-game app in the world in the first half of 2020, attracting more than 596 million installs, excluding its Chinese version Douyin, according to analytics firm Sensor Tower.

Although the basics of the algorithm TikTok uses are similar to ones found in apps from other tech companies, it is the special features that each company can add that differentiates the AI engine, said Wong Kam-fai, a professor in engineering at the Chinese University of Hong Kong and one of the first batch of national experts appointed by the Chinese Association for Artificial Intelligence.

Wong, who does not believe TikTok’s AI engine is truly unique, said a new recommendation system could be built for the short video app with fresh user data, in one year‘s time or so, but losing the existing tool would have “a very big impact” on TikTok’s current valuation.

“The technology works only when the algorithm and user data are both good. Part of the reason why ByteDance’s apps have an advantage over the competition are their user data,” said Hao Peiqiang, widely known as ‘Tinyfool’ who worked as a software engineer and now runs a tech blog and advises companies.

“The regulation [on privacy] in China is too weak and the awareness of privacy protection is relatively low,” said Hao, referring to the treasure trove of user data ByteDance has been able to amass via the app.

TikTok has repeatedly said that it stores US user data outside China and that the data is not subject to Chinese law.

Wong pointed out that some users, and in turn investors, may not be prepared to wait for the time it takes to build a new algorithm.

“You can’t wait for the team to rebuild the algorithm as TikTok is already very popular,” said Wong. “It’s like your favourite TV show closing because of a technical issue … I don’t think users will accept that.”

“For bidders such as Microsoft and Walmart, they want to buy the app and make it work immediately,” said Wong. “But if they need to wait a while to make it work well, maybe they won’t want to buy it any more.”

Not all experts agree that TikTok’s AI engine is truly unique though.

Read more: ByteDance to comply with China’s latest tech export rules, complicating TikTok sale

kr asia community

“While Tiktok could not exist as-is without its recommender system, that doesn’t exactly mean the system is anything special,” said Julian McAuley, an associate professor at the University of California San Diego, who studies the field.

“The early drivers of recommender systems included e-commerce companies. For example, Amazon has used recommendation technology for almost two decades, though early systems involved simple item-to-item similarity, rather than anything machine learning-based,” said UCSD’s McAuley.

“Netflix was also a big driver of recommendation technology in the mid-2000s, culminating in things like the Netflix Prize (2006), which led to a lot of academic interest in recommendation technology,” said McAuley.

In the modern smartphone era, the technology has however been criticized for the so-called “filter bubble” problem, whereby users surround themselves with content that serves to reinforce their own biases – rejecting all information that does not conform with their own world view.

“Companies want to optimize engagement metrics and don’t want to inject diverse or more balanced content if doing so hurts their key metrics,” said McAuley, adding they have little incentive to solve the problem.

“We’re living in a time when our demand for preferred information has never been so high.”

This article was originally published by the South China Morning Post.

Share

Auto loading next article...

Loading...