Article reprint source: AI Trends
Original source: White Whale Out to Sea
Image source: Generated by Unbounded AI
After the weekend's crazy rumors, on Monday, the news that "most of Chaoxi Guangnian's employees will be laid off" was basically officially announced. "There will indeed be adjustments to the business direction and organization, and we will focus more on the exploration of some innovative games and related technologies," this is the official response from Chaoxi Guangnian.
ByteDance’s abandonment of its gaming business can be described as decisive.
As for the reasons, everyone has various guesses. Some people think that games cannot sustain as a cash flow business line and have no room for imagination, so ByteDance will return to AI; some people think that Tencent and ByteDance have made peace, and after ByteDance gave up its game business, Tencent games will no longer have the psychological burden of buying traffic from ByteDance for live broadcasting; there are even people who say that Zhang Yiming doesn't like games...
Anyway, we think both of them make sense. The last one is definitely not Zhang Yiming’s standard for judging business logic.
What’s interesting is that on November 27, 2023, the same day that ByteDance confirmed that it would abandon its gaming business, 36kr reported that ByteDance recently established a new AI department, Flow, with Hong Dingkun, ByteDance’s vice president of technology, as the technical director, and Zhu Wenjia, the head of ByteDance’s big model team, as the business leader. In the recruitment post released by Flow, Flow claimed to be ByteDance’s AI innovation business team, and currently has launched two products, Doubao and Cici, in China and overseas, respectively, and has multiple AI-related innovative products in incubation.
On the same day, ByteDance completed -1 & +1 in the department on various online information channels.
Cici is climbing the overseas rankings, but is it a GPT shell application?
In fact, judging from the situation of Doubao and Cici, ByteDance established Flow some time ago, or the business has been going relatively smoothly, and now a department has been officially established.
In February 2023, ByteDance began to form a large model team, focusing on two modalities: language model and image. The language model team is led by the search department, and the image model is led by the intelligent creation team.
In early June, ByteDance started testing an AI conversation project codenamed Grace. On August 17, Grace was renamed Doubao. Cici was also launched in multiple overseas markets in August. Given that China is not friendly to AI bots and ByteDance has a strong global gene, overseas markets should be the market where ByteDance will focus more.
Diandian data shows that Cici is currently online in 36 overseas countries and regions. Except for mature markets such as the United States and Europe, almost all regional markets are covered, including almost all must-go overseas markets in Japan, South Korea, the Middle East, Southeast Asia, Africa, and South America, such as Saudi Arabia, Indonesia, Brazil, etc.
From the perspective of emerging markets, App also mainly focuses on the Android side. Whether it is the time of launching the delivery, the number of delivery materials, or the continuity of the delivery intensity, the Android side is more valued.
In the past few days, Cici started to climb the charts. On the 28th, it took the first place in the overall rankings of Google Play in Argentina and Peru. In the Philippines and Indonesia, it also hovered around the top 10 of the overall rankings.
In terms of positioning, Cici focuses on "Your Helpful Friend" + "Free". ByteDance, which has its own large model, Skylark, has included the text "Powered by GPT-4" in Cici's Southeast Asian materials. So, is ByteDance's overseas business a charity business that allows users in emerging markets to use GPT-4 services for free?
To verify this, we downloaded Cici and tested it.
GPT3 + search, the product is overall qualified
Let’s answer the above question first.
When I asked Cici, "What big model did you use to train it?", Cici's answer instantly slapped the advertising material in the face, "I trained it based on OpenAI's GPT-3 model." And I asked, "Are you sure?" Cici once again gave an affirmative answer.
However, when I asked questions like "Do you know ChatGPT-4?" and "What will happen in November 2023?", Cici was also able to answer.
If Cici did not have "hallucinations", then the developers should have trained it based on GPT-3 + search engine. According to my several questions, when Cici gave search answers, it seemed to provide users with information sources.
Despite using an older model version, Cici is still capable.
Cici supports 12 languages, which can barely cover the Middle East, Southeast Asia, Japan and South Korea. ChatGPT currently mainly provides languages of some European countries, Japan, South Korea and Russian. Cici has done corresponding localization work for the market planning of the App. Users can set the language in the upper right corner of the dialog box with Cici. However, if the user directly switches the question language in the dialog box, Cici can also switch the language directly.
After the user enters the dialog box, Cici turns on voice reading of text by default. After the user sends a question, Cici will read the text again by voice when replying.
In addition, the App also provides some chatbots tailored to specific needs, such as tarot reading, language learning, copywriting, programming, etc. Users can also create their own chatbots and choose to make them public or for personal use.
In order to judge whether Cici is effective, I tested it from two dimensions: one is basic functions, including reply speed, multi-round conversation understanding ability, and sensitive content recognition ability. The other is the amount of knowledge reserve, to see whether it can solve various problems and the depth of the answers.
1. Response speed: fast.
If the big model is answering, Cici will reply almost instantly. If a search is required, Cici can also respond within a few seconds.
2. Multi-round dialogue comprehension capability: depends on the number of dialogue turns.
If the questions are asked shortly, Cici can remember the context. For example, when my first question is what is Japan's Gini coefficient, and my second question is followed by "What about Korea?", Cici can understand that I am asking about Korea's Gini coefficient.
When I asked Cici to call me Cecilia, Cici agreed. However, after a few rounds of irrelevant questions, when I asked her "What's my name?", Cici still found it difficult to answer.
ChatGPT is obviously more capable in this regard. When I asked it to answer the question I asked a few days ago again in the same dialog box, I only needed to prompt the keywords and ChatGPT could understand which question I was talking about.
3. Ability to identify sensitive content: Pass.
When I asked Cici some sensitive questions, Cici was able to identify sensitive words and refused to answer the questions directly. Below the answers, Cici would also set some related questions to try to guide users to think positively. However, compared with ChatGPT3.5, Cici's responses were more routine, while ChatGPT3.5 was more humane.
In fact, from the several default segmented chatbots provided by the developer, we can see that Cici has a wide range of knowledge, and is not as good at liberal arts content as Baidu Wenxin Yiyan when it was first launched. In terms of the depth of questions, Cici's answers on some topics are more specific.
Take language learning as an example. When I asked Cici to teach me Chinese in an English context, Cici could accurately point out grammatical errors. Although the voice function automatically skipped Chinese in the English system language, when I asked Cici to teach me how to pronounce it, it marked the pinyin in brackets. For the same question, ChatGPT's answer was much more sloppy.
Cici's answer (Part 1) ChatGPT (Part 2)
Cici, this overseas chatbot app, is not just a simple API call, but the product is still very rudimentary. What is the thinking behind ByteDance's promotion of such a standard AI conversation app to the top of the charts in the overseas market at this point in time?
ByteDance’s AI Path Conjecture
1. AI bot is a traffic portal, returning to its core business
ByteDance’s flagship products are all entertainment or content products, but ByteDance itself is a platform and does not produce content. Controlling traffic entrances is what ByteDance is better at. AI bot is at least the most obvious new traffic entrance in recent years.
Looking at ChatGPT alone, before Meta launched Threads, ChatGPT broke TikTok's record, with over 100 million users in just 2 months after its release.
Image source: Yahoo Finance
Since then, there have been countless apps in the app store that call the GPT interface to directly create chatbots, as well as apps that integrate chatbots into existing products.
2. Improve technology in toC
The consensus now is that domestic large models are still far behind GPT. Sticking to large models sounds like an option with extremely low ROI. The establishment of the Flow department is also a focus on the application layer.
At the same time, ByteDance will also have a path of success dependence. The recommendation algorithm has always been something that ByteDance is proud of, and Meta is jealous. We do not rule out that ByteDance's algorithm has a high starting point, but the subsequent improvement depends on a large amount of user behavior and data feedback.
The same applies to AI models. If the GPT-3 model is really used as Cici said, ByteDance’s current clear idea is to enter the market first to reach out to users.
3. Positioning Differences
Currently, free AI bots are being developed in emerging markets to avoid the edge of ChatGPT (but based on the premise that GPT-4 is free to use), and also to distinguish themselves from many small and medium-sized companies that develop AI bots. They mainly rely on subscription monetization and are typical apps that make money through app design and price difference. Of course, Cici seems to be one of them at present.
If ByteDance's miracle trick is demonstrated again, and the promotional version allows users in emerging markets to use GPT-4 for free as scheduled, they can avoid existing products on the market and quickly acquire a large number of users.
In fact, the current cost of calling the GPT-4 API is still quite high, especially the cost of advanced large models and some open source large models can differ by hundreds of times. After Turbo came out, the price was lowered a bit, but many developers said that the calling cost is still very expensive, and is higher than the previous 3.5.
If ByteDance can adjust the model and obtain a better experience, it can monetize through advertising based on acquiring a large number of users, and the profits will not be too low.
After today's news came out, some netizens joked, "Are they only planning AI now, waiting for competitors' products to be launched and then cutting off?" But we have to say that even so, we may tend to think that ByteDance's AI applications, at least, sound more reliable than making games...
