SK Telecom's dream of AI Agents for 8 billion people; ByteDance & the TikTok Ban
January Week 1: Jan 7 - Jan 13
Hi friends, welcome back from the holidays 👋,
In this week’s edition of Coconut Capitalists:
SK Telecom unveils “Aster”, a personal AI Assistant inspired by “Her”
A deep-dive into ByteDance, the eight-headed monster & a future without TikTok
Quick-Fire Startup News from Around Asia
Let’s get into it.
SK Telecom unveils “Aster”, a personal AI Assistant inspired by “Her”
The Scoop
At the CES Conference in Las Vegas, SK Telecom - South Korea's largest telco - showcased its secretive AI agent "Aster." This personal AI assistant will soon be available to the masses in March of 2025 through an iOS and Android app.
SK Telecom's pitch is that products like ChatGPT are only capable of helping people with the planning phase of their daily tasks. For instance, take vacation planning: Aster's vision is to have a flexible AI agent that could instantly transform itself into your personalized travel agent - one that not only plans your next trip, but also navigates the internet (on its own) and books your end-to-end vacation itinerary.
Aster would start by autonomously booking your hotel rooms, flights, and even taxi rides based on historical knowledge about your travel preferences. Aster can also handle dinner reservations, coordinate daily travel excursions, and even magically take care of niche tasks like finding nearby ATMs to pick up local currency.
However, this concept of having AI agents complete end-to-end tasks isn't exactly new - many of the largest AI firms have spun up 50+ person teams over the last few years dedicated to building fully autonomous agents that can interact with the web & do all sorts of things on your behalf. For instance, Google DeepMind has a project that, funny enough, is called "Astra" - with a similar promise to SK’s "Aster." OpenAI has a project codenamed "Operator," that’s to be announced in the coming weeks, and even Anthropic recently introduced a product called "Computer Use" that allows their LLM "Claude" to control your computer screen and take actions on your behalf.
Breaking Down Aster
But getting back to Aster, the main difference between SK Telecom’s AI Agent and the rest of the pack is their vision of turning "Aster" into a consumer brand. If you take a quick peek at Aster’s website: www.asterapp.ai, you can see that the brand's UI/UX is incredibly warm, inviting, & consumer-friendly. And I give kudos to the product managers at Aster for thinking through the minor details of the experience, like credit card uploads - so the AI agents can quickly access your financial details when making online purchases - as well as timely push notifications to keep you on schedule, and personalized recommendations that are based on historical data of your past actions.
Also, it’s worth pointing out that Aster's branding reminds me a lot of www.pi.ai. For my non-American friends who might not be familiar with Pi - this was the very hyped, consumer chatbot product built by the folks at Inflection AI in Palo Alto. If you’re new to Inflection, the company’s founders are Mustafa Suleyman (creator of DeepMind) and Reid Hoffman (creator of LinkedIn). However, Pi ended up being a massive flop in the end. Inflection AI raised over $1.3 billion, couldn't find product/market fit in the face of competition from ChatGPT & Gemini, and fast-forward to now, most of the original Inflection engineering team has been acqui-hired by Microsoft.
But just because the branding of Aster & Pi are strikingly similar, I’m not saying that Aster will suffer the same fate as Pi. Aster, for one, has a much bolder vision: integrating not just the planning capabilities of LLMs, but turning them into autonomous agents that can complete the most complex of tasks, on your behalf. If Aster can actually pull this off, their total addressable market is literally billions of people - it's an app that gives everyone more of their most precious resource; time.
An elephant in the room, the internet hates is not a fan of Bots
However, for a product like Aster to work, it will need to interact with travel websites like Airbnb, Expedia, & Trip.com. But these companies hate bots. As a quick stat, over 87% of the top 1,000 travel websites employ bot-detection - essentially JavaScript code - in their website, in an attempt to block any non-human interactions. Based on this one - very blunt - stat, we can probably assume travel website’s won’t welcome our AI friends (like Aster) with open arms anytime soon….
But why is this? And is there really a difference (for example) between a human using a credit card to book a room on Airbnb versus the human’s AI assistant booking the room instead? The answer as it turns out is YES. Let’s break down why:
Consider what happens if we shift to a world where the majority of a travel service’s website traffic comes from AI agents instead of humans. In such a scenario, the very user experience moats and brand goodwill that each of these travel sites has built up over many years would completely disappear. Overnight, their products would transform from content-rich applications where hundreds of designers have put in blood, sweat, and tears (I’m being dramatic, I know) to tweak every little button color, making the experience perfect for human interaction - but for AI Agents, they interact with nothing more than an ocean of backend APIs.
This would cause the unique value proposition that differentiates sites like Airbnb, VRBO, and Hotels.com to drop near zero. To an AI agent, these are all just commodity services. Take Airbnb: while humans appreciate its beautiful user interface, stellar customer support (unless I happen to be the customer...), and key features like the interactive map showing available properties - these are all elements that currently make the product superior to its competitors - but to an AI agent, this carefully crafted user experience becomes irrelevant, reduced to simple API endpoints.
Given that there’s currently an incentive misalignment between many websites and the creators of these AI agents, certain startups see an opportunity to strike. Take www.browserbase.com which helps AI agents look more human and avoid bot detection. They’ve raised over $28 million in funding on the premise that your agents can bypass captcha restrictions, IP bans, Playwright headless browser detection, and everything in between. And even though there's no way sites like Airbnb are fans of these startups, they do represent legitimate workarounds at the moment.
To better understand the technical complexities at play, let's dig slightly deeper into the details (apologies ahead of time if you're not a Computer Science nerd like myself). Today, there’s two primary ways that AI agents interact with websites. Using Aster as our example, let's say we want to have Aster book a Grab Taxi on our behalf. Aster is (most likely) presented with two distinct approaches:
The first option is Aster navigates to Grab.com with a browser automation tool like Playwright or Puppeteer. From here, the AI agent develops a sequence of website actions (like button clicks or filling out input boxes) that eventually lead to a taxi getting booked. In this example, Aster may use a language model API like OpenAI's 01, with a vision encoder. Aster would send OpenAI’s endpoint two critical pieces of information:
Screenshots of the grab.com website (showcasing all the different parts of the website’s page that the AI agent could interact with)
The raw HTML of the website's codebase (With this, OpenAI's API could respond with the exact HTML snippet that needs to be clicked on)
The second option is Aster could reverse engineer the sequence of API calls on the Grab.com website, which leads to booking a taxi. Making these API requests versus taking the first option of multi-step DOM node interactions, saves a ton of money on compute costs (as you don't have to load entire web pages), but for sites with top-tier detection systems, this approach usually screams "I'M A BOT."
However, there might be a third, more collaborative path forward that's a little less "move fast & break things." SK Telecom is in a unique position to leverage its brand credibility to create partnerships with these travel service providers. While industry leaders like Airbnb, for example, might hesitate initially, their competitors may see an opportunity. For example, Airbnb's 2nd or 3rd largest competitor may view a partnership with SK Telecom's Aster as a chance to quickly gain market share.
This situation perfectly illustrates the classic innovator's dilemma: market leaders often hesitate to take risks on new user experiences that might completely disrupt their current business model. But if they don't take the risk, a competitor may seize the opportunity instead, and eventually surpass them as the new incumbent.
Why it Matters
While my comparison to the innovator's dilemma might suggest a slow adoption by industry leaders, the hypothesis could actually turn out to be very wrong. Maybe a travel company like Airbnb will hop on the AI Agent Train much sooner than we would expect. It’s hard to know!
We can, however, take a look at an example of an innovator's dilemma that allegedly occurred between The Wall Street Journal (a prominent news publication in the US) and a few Silicon Valley based AI startups. Supposedly, when these startups were small, they could scrape publications like The Wall Street Journal on a daily basis from a single paid subscription account—storing thousands of articles in their internal databases (I read this on Y Combinator’s Hacker News, so I can’t fully confirm). But allegedly, this allowed them to surface bits and pieces of the original articles’ content thousands of times per day—because they no longer needed to scrape the WSJ ad-hoc for each user query in real-time (which would trigger bot detection). Instead, they could simply ping their own database for a copy of the WSJ’s content.
Eventually, The Wall Street Journal discovered this practice and filed multiple lawsuits. However, smaller publications (the WSJ’s competitors) saw this as an opportunity. Places like Fortune, Time, and The Independent sought out official partnerships with these startups, granting legitimate access to their content—while The Wall Street Journal got bogged down in legal disputes. Many would argue this is a textbook example of the innovator’s dilemma.
So does this foreshadow anything specifically related to Aster? I’m not sure. But Aster will need to develop a business model that incentivizes all parties. By the way, what Aster is attempting to build should be applauded, and I’ll be rooting for them from the sidelines—화이팅!
An uncovering of ByteDance, the eight-headed monster + a future without TikTok
The Scoop
For our friends who haven’t been keeping up with America’s quest to ban force a divestiture of TikTok to a US Company, here’s a quick recap of the two main events that led to this “ban or sell” scenario:
First, in June of 2022, the House of Representatives (one of the three branches of the US government) began an investigation into TikTok & its relationship with the Chinese Communist Party. This came after multiple software engineers turned whistleblowers alerted the House Intelligence Committee that US citizen user data had allegedly been accessed by ByteDance employees in Beijing. Overseas access to a US citizen's data is not usually a national security concern, but given China's National Intelligence Law stating that "companies operating in China must cooperate with the government and provide all data upon request," the US found this to be concerning in regards to TikTok. Now, one could argue that companies like Temu (PDD), Tencent, and Alibaba are also based in China + handle large amounts of US user data - so why aren't they being targeted? Well, in the US government's report, TikTok user data has been deemed highly sensitive for a number of reasons, including the ability to 1) access real-time location data and 2) listen to a user's audio conversations at any time (allegedly).
Second, in April of 2024, after months of hearings that included TikTok’s CEO Shou Zi Chew testifying in front of the US Congress for 4+ hours as well as a few of ByteDance's US investors being subpoenaed to appear on Capitol Hill, the investigation finally concluded. In the end, congress decided to enact the "Protecting Americans from Foreign Adversary Controlled Applications Act," which required ByteDance to divest TikTok's U.S. operations by January 19, 2025, or face a nationwide ban of the app.
But without TikTok, where does this leave ByteDance?
In a very good spot actually. The $300 Billion Dollar parent company of TikTok—ByteDance—will probably end up just fine. Just like the legendary 8-headed Japanese monster known as “Yamata no Orochi” - if you chop off one of ByteDance’s heads - you still have 7 more to go.
Outside of TikTok, ByteDance has 7 other products heads you should pay attention to:
1. Douyin, the China-only version of TikTok has 748 million monthly active users where the average users spend almost 2 hours on the app per day.
2. Capcut, an AI powered video editing tool for iOS, Android, and Desktop. Think of the tool as an AI-ified version of Adobe Premiere Pro or Apple Final Cut Pro, but with a massive user base of over 320 million monthly actives
3. Toutiao, the information-sharing platform built for the Chinese population (similar to Google or Apple News) has over 292 million monthly active users
4. Doubao, the ChatGPT competitor that has seen rapid growth in China & throughout South East Asia, has over 74 million monthly active users
5. Lark, the enterprise collaboration suite (strikingly similar to Google Workspace or Microsoft M365) has over $400 million in annual recurring revenue
6. Lemon8, a social media app focused on beauty, fashion, & food that’s targeted at Southeast Asia and Japan, has over 16 million monthly active users
7. Melolo, is a newly launched mobile app that offers short movie dramas from 2 minutes to 10 minutes long, where the content is produced exclusively by ByteDance’s in-house creative studio. It's a similar concept to Netflix Originals or more-so Quibi… if you remember what that is haha).
Why it Matters
If TikTok, the viral social media app that 170 million Americans use each month, shuts down in the US, then obviously it's bad for its parent company ByteDance, but the question remains: how bad is it really? This is why highlighting ByteDance's 7 other heads is critical. Additionally, we haven’t even mentioned ByteDance’s world-class AI Research team that rivals OpenAI & DeepMind - so maybe the 8 heads do remain…
And if TikTok gets banned in the US, one could argue that it's losing only half a head, not a full one. TikTok still has over 190 million people in Indonesia on the platform, 120 million in Brazil, and another 100 million scattered across Europe.
It's worth noting, however, that the bulk of spending power by TikTok users is concentrated in the US, where TikTok brought in over $6 billion in revenue for 2023 by serving ads to American consumers. It doesn't matter that ByteDance is worth $300 billion - $6 billion disappearing overnight will sting.
🇯🇵 Japan News
Toyota has announced that Phase 1 of construction is complete for a new tech-driven city called "Woven" at the base of Mount Fuji. While founders and CEOs constructing new cities isn't an entirely new concept, Toyota's project represents Japan's first venture into this space. In the USA, entrepreneurs like Marc Andreessen (A16Z) and Patrick Collison (Stripe) are involved in the construction of a new city just north of San Francisco, while Marc Lore (Walmart) is creating Telosa in rural Nevada (Telosa is expected to have a sci-fi look, reminiscent of something you'd find in the year 4500).
Toyota’s Woven is a $10 billion smart city that will serve as a testbed for autonomous cars, drone delivery, and smart homes. It's worth noting that the city of Woven is at a significantly smaller scale than the American projects I mentioned - this first version of Woven V1 is designed to accommodate just 2,000 residents.
FujiSoft finds itself in the middle of a high-stakes takeover battle between American firms KKR & Bain Capital. Both firms are attempting to acquire the company at roughly a $4 billion USD price point. If you're not familiar with FujiSoft, it's a technology consulting firm (similar to Accenture in the US) whose job is to find innovative technology in the world and implement it into the workflows & products of their clients (large Japanese enterprises). It's important to highlight that the proposed acquisition of FujiSoft comes at a time when technology consulting firms are booming - for example, Accenture just reported $3 billion in new AI consulting revenue for the year of 2024.
To explain the knife-fight between the two firms: initially, KKR made the first offer of $4 billion USD to FujiSoft, but Bain quickly came in with a competing offer of $4.30 billion USD. In response, KKR made a counter offer of $4.31 billion USD. Now the most recent offer from Bain Capital is a "counter-counter" offer of $4.36 billion USD. However, the latest offer from Bain Capital has become highly controversial as Bain is attempting to buy out the majority of shareholders without board consent (there's nothing inherently wrong with this approach, but it's often called a hostile takeover).
🇰🇷 Korea News
Samsung is launching what they call an "AI Subscription Club". Users can pay a monthly fee to rent a number of AI devices from the company. The concept is similar to leasing a car, like a Model 3, but instead of receiving a Tesla, you get access to Galaxy phones, Galaxy laptops, plus dozens of other product options with AI at the core. There are even rumors that Samsung will include access to AI robot dogs (being built by their subsidiary Rainbow Robotics).
It's worth noting that we don't yet have a price on the AI Subscription Club, even though the planned launch is in February of 2025. There are only a handful of companies capable of offering such an AI subscription service: Maybe just Apple, Google, and Samsung that have enough of a product mix of AI software and hardware for mass appeal.
🇮🇳 India News
Accel, a global venture capital firm that's best known for its early investments in Facebook, Spotify, and Flipkart, announced a $650 million early-stage fund for founders in India.
If we zoom out, we can see that there's been an uptick over the last five years with many tier-1 venture firms including Lightspeed, Peak XV, & Accel - continuing to back Indian founders building for their home country, at the earliest of stages. Traveling west to pursue your entrepreneurial dreams is becoming less and less the “only” path for achieving startup success.
🇨🇳 China News
MiniMax, the China-based AI startup company backed by Alibaba—and most recently valued at over $3 billion—had the 4th most downloaded app in the USA for the first half of 2024 with their product Talkie AI, according to analytics from Sensor Tower. If you're not familiar with Talkie, it’s similar to the US’s Character AI (Character was founded by Noam Shazeer, one of the co-authors of the Transformer LLM Whitepaper) Both the Talkie & Character AI apps look strikingly similar - they're both packed full of virtual characters from prominent anime figures to Super Mario Bros. - where you can strike up a fictitious conversation instantaneously.
However, the app has now completely disappeared from the iOS App Store. It's currently unknown what the underlying reason was for the app being taken down. But if we put our tin-foil conspiracy hat on for a quick second, this does come during a time in which Chinese social media apps in US markets are under heavy scrutiny.
China's government (specifically Beijing's Municipal Science and Technology Commission) announced plans for a nationwide elderly care service network that will come into effect in 2029 - an elderly care plan with humanoid robots at the center. As stated in the document, these are "robots that will provide companionship, monitor health for the elderly, and help them with household services."
By 2029, approximately 20% of China's population will be over the age of 65. This creates an opportunity for robotic startups to take a break from AI dogs (cough cough Unitree) and instead build critical infrastructure for their home nation.
Zixuan Huang, an engineer from Stability AI who received his bachelor's degree at the University of Science and Technology of China, has open sourced a new AI model called SPAR3D (Stable Point Aware3D).
SPAR3D is able to generate complete 3D objects from a single image in less than a second. This is incredibly helpful to game designers who spend a significant amount of time hand-crafting in-game assets. Games like Fortnite, Skyrim, and GTA 5 have tens of thousands of these assets, each taking many hours to create manually. It's worth noting that SPAR3D is not the first AI model to generate 3D objects from single images, but it offers best-in-class quality while being open source. The model weights can be easily downloaded via Hugging Face.
Alibaba's Qwen vision LLM is now being integrated directly into AR smart glasses. The system uses a tiny version of the Qwen base model that physically sits on the hardware, while a much larger & more intelligent version runs server-side, streaming in for complex tasks.
This approach mirrors Apple Intelligence, which also uses a small model on-device, while concurrently running a multi-trillion parameter model in its data centers. It's worth noting that the company creating the AR glasses is not Alibaba, but a startup called Thunderbird Innovation. The glasses come with a Qualcomm Snapdragon chip, weigh 39 grams, have a 7-hour battery life, and a 1.3-second response time for streaming in AI content.