In May 2023, the White House held a closed-door meeting on AI. The guest list brought together three generations of young and middle-aged people – veteran Google and Microsoft, rising star OpenAI, and start-up Anthropic, but Meta was not included.
Even after the meeting, the official response dealt another blow to Zuckerberg: “Only companies that are currently leading in the field will receive invitations.”
But at the same time, Meta was a different story in the industry.
The media, contrary to their usual practice, described Zuckerberg as “the hero AI needs”; netizens flocked to his post to leave messages, suggesting that Meta be renamed OpenAI.
Just because his open source big model, to some extent, broke the monopoly of OpenAI and Google, achieving the “Android moment” in the AI world.
To this day, it is still difficult to conclude whether “open source or closed source, who can lead AI technology better”, but what is certain is that AI has taken Meta to the sky.
It can be seen that, unlike other big model companies that are still “chasing dreams”, behind Meta’s five-fold stock reversal, every financial reporting season is supported by tangible performance growth.
As for the underlying reasons, as Zuckerberg revealed in the 2023Q1 earnings call, “AI affects every application and service under its umbrella.”
The third generation of algorithms stands on the shoulders of TikTok
“Let Ins return to Ins, I just want to see my friends’ cute photos.”
In mid-2022, Kim Kardashian and Kylie Jenner (Kim’s younger sister) complained about black and white photos, which triggered a protest against Meta.
The cause was the increase in the proportion of video content on Ins, while testing the Feed function on a small scale; Facebook was the same, pushing the content of “non-followed users”. For a time, everyone’s “friend circle” private land turned into a “Weibo + Douyin” square.
However, in the face of hundreds of thousands of “angry” petitions, Zuckerberg said, “What we are doing is very unique, and I don’t think people want to be restricted to one format.”
This is not like Facebook’s style at all. You know, in the past, if there was any slight change that caused user dissatisfaction, it would almost respond immediately.
The reason why people who “follow the wind” suddenly become “righteous” is usually not complicated, that is, there is a bigger backer behind them.
At the 2022Q1 earnings conference, Zuckerberg could not hide his excitement and said: “With the addition of artificial intelligence, we are building not only a recommendation system, but also a discovery engine.”
Specifically, in his vision, the new generation of content recommendation technology not only understands user behavior better, but also understands the content itself.
For example, if you want to post a weight loss video titled “I am reborn, reborn on the fattest day” on Bilibili or YouTube, you need to select tags such as “sports” and “fitness”, and bring topics such as “the meaning of weight loss” to be pushed to the target users.
Otherwise, it may be pushed to users who like “novels” and “funny” because of the title.
And Facebook, with AI addition, even if you don’t select a tag, the system can automatically capture the content about fitness in the video, and label it with more detailed tags such as “fitness”, “hips and legs”, “with equipment”, and push it accurately.
In addition to content, the understanding of user behavior is also further.
For example, previous algorithms focused on making recommendations based on likes, completion rates, etc., but now, compared to likes, they pay more attention to emotional expressions in user comments and other interactions – such as anger, sadness, etc.
Compared with the previous generation of algorithm star ByteDance (focusing on user behavior analysis), this algorithm means technology iteration to some extent.
Instead of focusing on the user’s content becoming “catering to their tastes”, coupled with the addictive nature of short videos themselves, user controversy gradually disappeared, replaced by more and more time spent on Facebook.
The 2023Q1 conference call mentioned: Since the application of artificial intelligence to Reels (short video function on Facebook and Instagram), the user time of Instagram has increased by more than 24%.
Facebook is the same, and the total time has been rapidly repaired since 2022H2.
In comparison, TikTok, which once dominated the world, seems to have a weak growth in user time. It can be seen that Reels already has the ability to compete with TikTok.
In addition to being restricted by the “traction” of the algorithm, users are also actively returning.
It can be seen that, unlike before the emergence of Douyin in China, Weibo was the only one in the public domain, and the overseas social field is flourishing.
In the private domain, there are Snapchat for Generation Z, WhatsApp for those born in the 1980s, and Nextdoor for the middle-aged and elderly; in the public domain, there are Facebook, Ins, Twitter, YouTube, etc., which also correspond to various age groups.
This means that it is difficult for TikTok to shake the original social chain of Facebook like Douyin impacted Weibo.
Therefore, when Facebook focuses on launching short videos and is equally addicted, everyone will tend to flow to where the sisters are – as of July 2023, YouTube Shorts has more than 2 billion monthly logged-in users; the same is true for Facebook.
Of course, not only friends, but creators also affect the flow of users.
Data shows that from July 2021 to July 2022, the number of creators with more than one million fans on TikTok dropped from about 2,000 to 1,200-1,300, with a loss of nearly 800 people.
One of the reasons for the escape is that it is not easy to “make money” on TikTok. As Hank Green, a top influencer, mentioned in an interview: The same video is posted on Facebook, Instagram, TikTok and YouTube, and TikTok has the lowest revenue.
After all, compared with Facebook and YouTube, which have mostly middle-aged users, TikTok, with nearly half of its users aged 13-24, is very young, but also “poor” – TikTok’s advertising conversion rate is 0.7%-3%; Facebook’s is 9.21%.
Under the influence of various factors, it can be seen that since the beginning of 2023, TikTok’s user growth rate has basically been at the bottom of the four.
Facebook, which is wary of TikTok, has lifted the alarm – TikTok and competition issues have rarely been mentioned in the performance meeting since Q1 2023.
However, it is not only TikTok that Zuckerberg regards as a thorn in his side, but also Apple.
Bypassing Apple’s new privacy regulations, advertising business is back on the table
In 2022, the North American advertising industry is shrouded in a gloomy cloud.
Whether it is major advertising companies or experts from advertising agencies, the opinions shared cannot avoid the “uncertainty” and “concern” about the environment.
Even at every performance exchange meeting of Meta, Zuckerberg’s reserved program is to report: the situation this season looks worse than the previous season.
However, after a year, the whole industry has collectively changed its tone.
In the 2023Q2 earnings season, Meta executives excitedly said: This year’s advertising demand has stabilized. Google executives also couldn’t hide their excitement and said: We are satisfied with the accelerated growth of search advertising revenue.
The reason for the big reversal of attitude can be summarized in 6 words: thank Chinese e-commerce.
The Wall Street Journal reported that in 2023, Temu spent nearly $2 billion on Meta advertising, contributing 10% of its advertising revenue; at the same time, it also ranked among the top five Google advertisers.
Similar “money-spending boys” include Shein, which has become the 16th largest advertiser in the United States, with advertising spending increasing by 120% year-on-year.
To put it bluntly, Chinese e-commerce going overseas has saved North American giants. However, this can only solve the problem temporarily, after all, Temu can’t keep burning this way.
Compared with the impact of the macro economy, Meta’s more core contradiction is: how to deal with the impact of Apple’s privacy regulations.
In December 2020, Facebook took out three full-page advertisements in the New York Times, the Wall Street Journal, and the Washington Post, claiming that it “would stand up for small businesses around the world against Apple’s hegemony.”
At the same time, it also wrote a “small essay” on its own platform, “yin and yang” Apple’s user privacy protection policy is just for self-interest.
The reason why they are willing to tear their faces and “throw a tantrum” is that Apple’s continued introduction of new privacy regulations has the greatest impact on Meta in the North American circle.
It can be seen that under the new privacy regulations, when an App tries to collect personal data across platforms, the phone will prompt and the user has the right to refuse.
Industry survey data show that nearly 80% of iOS users choose to prohibit App from tracking data across platforms.
This greatly weakens the accuracy of personalized recommendation ads and has dealt a heavy blow to platforms that rely on information flow ads.
As shown in the figure below, after 2021Q2, the growth rate of advertising revenue of Meta and Google has declined, but the decline of Meta is much greater than that of Google, which mainly relies on search ads.
Even among information flow platforms, Meta is the first to bear the brunt. Unlike Snapchat, which serves big brands, Meta’s customers are mainly small and medium-sized enterprises, and it is difficult to quickly adapt to this round of industry changes.
For example, with reduced accuracy, if you want to achieve the previous advertising effect, it means spending more budget.
An executive of the ophthalmology company Pearle Vision revealed that the cost of advertising and marketing on Instagram in 2021 will be 15%-30% higher than the previous year.
In the long run, big brands with deep pockets may be able to support it, but small and medium-sized businesses cannot afford it.
From this point of view, it is not difficult to understand why Facebook is the most active in criticizing Apple.
However, Apple is not TikTok. Not only is it useless to “throw a tantrum”, but it is also exported by Cook’s face, “Facebook can still track users on apps and websites as in the past.”
Meta, which just passed, tried to open up new tracks. For example, Facebook, Instagram, and WhatsApp are all put on the “little yellow car”; advertisers are moved into the metaverse, etc.
But all these take time, and distant water cannot quench immediate thirst.
When everything was at a loss, the explosion of AI brought a turning point – more accurate data analysis capabilities, more automated material delivery, more targeted crowd matching, etc., all mean that AI has technological iterations for the advertising industry.
As a result, advertising giants have flocked to evolve into AI technology.
Meta is no exception. Internal employees revealed that the money it spent on transforming AI advertising capabilities is more than what Zuckerberg spent on the Metaverse.
It can be seen that soon after GPT-4 swept the screen, Meta launched the AI advertising product Advantage+.
With this tool, advertisers can submit their audience range, overall advertising budget, expected sales targets, promotional advertising materials, etc. to the platform at one time, and AI will analyze how to deliver and what materials to deliver.
In this way, it not only simplifies the work of advertisers, but also improves the accuracy of delivery.
For example, executives of marketing agency iProspect said that through this product, every $1 spent will generate a return of $7 – almost as high as before the introduction of Apple’s privacy policy.
In the 2023Q1 earnings call, Meta was very proud: “We have made positive progress in mitigating the direct impact of Apple’s privacy policy.”
Data from August 2023 showed that the number of advertisers using Advantage+ per week has tripled compared to six months ago.
It is also worth noting that in addition to external influences, Meta’s advertising business was tripped up by its own operating rhythm.
It can be seen that after Ins introduced Reels short video ads, the growth rate of advertising revenue continued to decline, and even gave a historical low guidance in 2022Q1.
If we break down and calculate the revenue, we will find that the root cause lies in the misalignment of traffic and advertising monetization: short videos squeeze the user time of the main site, resulting in a decrease in overall advertising exposure; its own ecology has not been cultivated, the advertising loading rate is limited, and the advertising revenue is in a state of shortage.
However, the calculation model also shows that this decline in revenue growth is only a short-term and medium-term pain. In the long term (more than three years), the short video project will “accumulate and develop” and significantly increase advertising revenue.
Judging from the time, Meta has obviously reached the period of release of advertising revenue growth.
As shown in the figure below, the advertising loading rate has been rising. Citigroup analysis shows that the average advertising loading rate of Reels 2024Q1 has reached 20%.
The advertising business that was stuck in the quagmire has turned around successfully, and both the market and Meta itself have breathed a sigh of relief. What’s more, it is not only the old business that benefits.
The open source of large models allows developers to “work” for Meta
In fact, AI is to some extent an unintentional achievement for Meta.
In the summer of 2021, at the Technology and Media Tycoon Conference in Sun Valley, Idaho, Zuckerberg was confused by Google CEO Sundar Pichai’s praise for Facebooke’s technological breakthrough in the field of AI. He was full of the metaverse and didn’t know what Pichai was talking about.
After pretending to be calm and dealing with the meeting, Zuckerberg immediately returned to the headquarters and convened the artificial intelligence research group FAIR to listen to all the latest work briefs. After that, he began “a lot of self-study.”
The attention of the “top leader” has returned, and Meta has stayed at the table safely.
According to the calculation of Thundermark Capital, a venture capital firm, Meta’s “AI Research Index” ranks third among global companies in 2022, second only to Google and Microsoft.
As time enters the GPT era, Meta, with a ticket in hand, has become a “hero” again.
In May 2023, a memo from a Google engineer, “We (Google) have no moat, and OpenAI is the same”, swept the news headlines.
He believes that under the impact of Meta’s open source ecosystem (Llama2 large model), closed Google and OpenAI will lose the game.
Although not everyone agrees with this argument, the power of open source cannot be ignored.
As shown in the figure below, Llama2’s performance in reasoning tasks is similar to that of GPT-3.5, but its model training cost is much lower than that of GPT-3.5.
The reason is nothing else, it is purely the training and calling costs of open source models, which crush closed source large models.
Take the call as an example. A closed-source large model will generate a lot of call costs when used once. In layman’s terms, the pot is welded to the OpenAI stove, and you have to use OpenAI’s gas to cook. Although OpenAI will charge you, this amount of money is difficult to cover its costs.
Open source models can usually be downloaded and used freely, which is equivalent to the user taking the pot home and cooking on his own stove. Meta will not incur gas and other costs.
This can be seen on the developer side. Under the open source model, not only is the call price lower.
Data shows that the call price of Llama 270b is $4/1 million tokens; GPT-3.5 with similar performance requires $8/1 million tokens.
Developers can also change their skills at any time according to market trends to avoid being “left behind” when the underlying technology route is switched.
This time has attracted many small and medium-sized developers and researchers to flock to it. According to statistics, as of September 2023, the number of downloads of the Llama model on Hugging Face, the world’s largest open source community platform, has exceeded 30 million times.
Around Llama, various open source data sets and new models have begun to emerge.
For example, developers from domestic universities such as Tsinghua University, Jiaotong University, and Zhejiang University have gathered in the Chinese community of Llama2, not only giving birth to many Chinese open source large models based on Llama, but also contributing data for “Chinese special training”.
This further reduces the iteration cost of Meta’s large models.
To put it bluntly, the open source ecosystem will allow institutions and researchers from all over the world to participate and “work” for the iteration and evolution of large models for free.
Not only has AI become a “hotshot”, but it has also helped many of its peers to “rise to prominence”. For example, the social and advertising businesses mentioned above, as well as the much-questioned Metaverse.
It can be seen that Meta’s head-mounted displays are moving from high-end to popular: the newly launched Quest3, with a starting price of $499, is quite cheap for the same specifications in the industry; at the same time, the old models are also being reduced in price.
The reason why Quest3 is priced this way is that after the early trial period of its VR equipment, the first round of accelerated penetration seems to be just around the corner.
You know, although the iPhone opened the era of smart phones, it was OV and Xiaomi that really made it popular in China, with prices below 1,000 yuan.
And the driving force behind this reversal is still AI.
Previously, the market’s biggest concern about the Metaverse was that the amount of money spent on software development was huge, and the prospects for its landing applications were also a mystery.
The immersive experience of the Metaverse has high requirements for software and hardware: virtual scene rendering, human-computer interaction, and even the flaws in virtual character modeling will be infinitely magnified in the Metaverse.
But the reality is that the software side is a mess. For example, the quality of Horizon Worlds, the metaverse version of Facebook that Zuckerberg personally promoted, makes people “dream back to the 1990s”.
Amidst the disdain, Meta’s internal documents show that Horizon Worlds has less than 200,000 active users by Q3 2021.
Due to the “desolation after entry”, VR hardware has been in the early stages of trial and error for a long time, and sales are limited.
After the emergence of the AI big model, everything began to change. It can be seen that whether it is human-computer interaction, body positioning, or content creation, AI technology has reached a new level.
In terms of body positioning, in the past, due to the problem of field of view coverage, VR headsets were difficult to capture lower body movements, and people in Horizon Worlds only had weird upper bodies.
The use of AI algorithms can not only achieve “limb freedom”, but also more accurately predict whole body movements. When the data shows that the user is under a chair, it can be predicted that his downward posture is not necessarily squatting, but sitting on the chair.
With the leapfrog change, there are more and more fun and non-useless applications.
For example, the newly launched “PILLOW Pillow” on Quest3 is like an XR version of a drifting bottle, where you can fish and open letters on the gravity-reversed ceiling; in the game “Drop Dead: The Cabin”, zombies will even enter the player’s house and attack.
The increase in available scenarios has further boosted the user acceptance and penetration of hardware devices.
Data shows that the shipment volume of Quest in Q4 2023 exceeded 300,000 units, and even during Black Friday, Quest’s sales on Amazon exceeded AirPods.
Overall, AI overturned Meta’s old game and pulled everyone to restart a new game.
Summary
At the trial of the Oculus intellectual property dispute in 2014, Zuckerberg, who had a stern face in the dock throughout the trial, suddenly began a heartfelt speech in the eyes of everyone in surprise:
“A few months ago, my daughter Max took her first step. I recorded the whole scene in virtual reality so that I could send it to my parents and share it with the world. People can experience the situation in person, just like being in our living room.”
That was the first time the prototype of the concept of the metaverse came out, and Zuckerberg firmly believed from the beginning that this would be the “next generation of the Internet.”
However, in the eyes of Wall Street, it is extremely unwise to insist on “long-termism” in the uncertain track of technological innovation. After all, capital has always been “not seeing the rabbit, not releasing the eagle.”
For example, in this round, there was a definite signal landing, and a five-fold reversal was staged; but once there was an unclear sign, it immediately scattered like birds and beasts-Zuckerberg “spare no money to build Meta into a world-leading artificial intelligence company” just finished his voice, and the market value evaporated by 200 billion US dollars.
Musk’s fans couldn’t help but sigh after watching it: “Compared with ‘pie-painting master’ Musk, Zuckerberg’s ‘storytelling’ ability needs to be improved.”
