"2023 is referred to as the 'AI Year.' At this current juncture, artificial intelligence has completed a technological evolution from quantitative to qualitative, initiating the idealistic journey of AI for everyone and AI for all beings. After a rapid surge, the public's excitement, anxiety, and confusion about AI intersect and gradually emerge."
On November 30, 2022, an uneventful day, on the other side of the ocean, a company named OpenAI officially released a chatbot program called ChatGPT. The small program seemed to enter the scene quietly, without a press conference, pre-launch, or roadshow. However, within just two months, its monthly active users surpassed 100 million, making it the fastest-growing consumer application in history.
Originally incubated in Google’s lab, OpenAI, led by figures like YC Altman, PayPal founder Peter Thiel, and Elon Musk, aimed initially at challenging Google’s DeepMind. In the initial years, with sky-high investments and product development within a small circle, it did not attract much external attention.
It was only when GPT-3.5 was made freely available to the public, after three iterations of high-tech products, that AI transcended simple customer service, education, creative writing, knowledge queries, etc., becoming a profound “man-machine” interaction after deep learning.
Bill Gates publicly stated that the birth of ChatGPT has significant historical significance, comparable to the birth of the Internet or personal computers. After 50 years of fluctuation in AI from logical reasoning, knowledge learning to machine learning, it finally achieved an epic transformation in this wave of technological revolution.
Battle of the Giants
Entering 2023, the ripples of ChatGPT continue to spread. Antony Aumann, a philosophy professor at the University of Northern Michigan, discovered a high-quality paper assignment in a world religions course, and upon inquiry, the student admitted that ChatGPT had completed the paper. Overnight, it seemed like people had found the right way to wield magic, bombarding ChatGPT with questions, experiencing waves of awe from the future.
The collective wisdom of the masses propelled ChatGPT into boundless transformations. The capital market also envisioned the scene of AI for everyone and all beings, a clearer and more tangible target than the metaverse. They turned their focus to the AI track, manifesting the essence of “heating up.”
In February, Microsoft and OpenAI started a subtle competition. On the day OpenAI announced the Plus subscription plan, Microsoft was rumored to be working on integrating a faster version of ChatGPT into Bing. Google hurriedly responded by launching the AI conversation product Bard. Simultaneously, the search side made efforts, integrating the new versions of Bing and Microsoft Edge, incorporating ChatGPT, with user feedback suggesting a better experience than ChatGPT.
Soon, the flames of the battle and ambitions extended to China. In late February, Fudan University released an application similar to ChatGPT called MOSS. In mid-March, Baidu took the lead in launching the AI application Wenxin Yiyuan. Regardless of the effectiveness, as the first realization of China’s internet company’s long-term investment in AI technology and large models, Wenxin Yiyuan was the result of Baidu’s accumulated efforts, waiting for this moment.
In May, AI research and development entered a fever pitch, dubbed the “Battle of the Giants.” Alibaba Cloud’s AI robot Tongyi Qianwen went online for testing; Elon Musk’s new AI dialogue robot TruthGPT started warming up; Adobe’s Adobe Firefly embedded in more of its products; Google announced an upgrade to Bard with the PaLM 2 model supporting over 100 languages.
Overall, these large models each have their advantages. Among them, the general AI models similar to ChatGPT account for a whopping 68.7% of the traffic. This is because general models do not require local deployment and do not need to join Discord channels; they can be used directly on web pages, making them the most straightforward and effective “what you see is what you get.”
Following this, Huawei introduced the Pangu large model, and Sensetime launched the SenseChat model. Subsequently, 360, ByteDance, iFlytek, JD, Tencent, Meitu, almost all of China’s internet giants successively entered the arena. According to the “China Artificial Intelligence Large Model Map Research Report,” as of the end of May 2023, at least 79 large models with parameters exceeding 1 billion have been released domestically, compared to 100 in the United States, with a global cumulative release of 202 large models.
Although the likelihood of China large models being directly comparable to ChatGPT products is small, it is noticeable that they excel in vertical fields, such as 3D models, video models, multimodal models, etc.
Generative AI Storm
If the first half of 2023 was just the beginning of the melee, the second half, with the official release of OpenAI GPT-4V, marked the “second half” of the battle of large generative AI models, entering the competition stage of “multimodal” capabilities.
“Multimodal” does not signify a one-to-many relationship but rather a geometric increase in possibilities. The use of AI capable of understanding images and videos is unimaginable, just the tip of the iceberg. Industry analysts speculate that converting videos into new training data, training smarter AI, and forming a closed loop could accelerate intelligence.
Following this, Elon Musk announced the establishment of the new company xAI and launched its first AI model, Grōk. One of its major advantages is its ability to bind with Musk’s other company, the social media platform X (formerly Twitter), fetching real-time information from X for more timely answers. The Grōk team consists of researchers from renowned technology companies such as OpenAI, Google, DeepMind, and Microsoft, making Grōk fully comparable to the top three models in terms of logic and technology.
At this point, NVIDIA jumped into the scene, releasing the new GH200 Grace Hopper platform, causing significant upheaval in the industry, touted by CEO Jensen Huang as the “world’s fastest memory.” The entry of the chip giant opened a core technical window for AI.
However, in the second half of the year, the contrast between the enthusiasm in the domestic capital market and the “calmness” on the product side became apparent. With successive venture capital investments, entrants were more concerned with the rapid commercialization combining big data and models. Baidu Wenxin Yiyuan has now reached a user base of 70 million and 4,300 scenes; Huawei Cloud is collaborating with relevant enterprises for joint innovation based on Ascend AI cloud services and the Pangu large model; Baidu, Tencent, and Alibaba’s large models have passed standard compliance testing, obtaining “commercial licenses.”
Overall, the level of domestic leading large models is roughly equivalent to OpenAI’s GPT-3.5, still needing to overcome three major obstacles to reach GPT-4: computing power, algorithms, and data. Moreover, compared to the weak general attributes of domestic large models, except for Baidu, many large models are designed for specific users in specific industries.
It seems that everyone has understood that “top-level large models are a race that only giants can participate in.” Without billions in assets, it is challenging to generate substantial business value. Some believe that, except for automatically generating advertising graphics and interesting but not entirely replacing human-written copy, large models currently do not have a better place to shine. Some even argue that 99% of large model companies will eventually perish.
“Disruption-Innovation-Application-Landing,” in this year, AI, driven by capital, becomes increasingly “real.” It’s as if people have gold in their pockets but can’t exchange it for a cup of instant coffee, leading to a mix of excitement, anxiety, and confusion. Therefore, many startups are shifting towards a “small workshop making small tools” model, seizing the window of opportunity, rapidly expanding through operational means, and earning some money first.
Searching for AI Values
In November, OpenAI staged a drama where co-founder and CEO Sam Altman was swiftly dismissed by the board. After a brief stint at Microsoft, he ultimately returned in a “palace struggle” spectacle. This has exposed the deep-seated internal disagreements and contradictions within the industry under the rapid development. Many also expressed concerns about OpenAI’s future prospects.
Under such tremendous changes, what are we to make of it? What we least want to see is that Generative AI, like the metaverse, goes astray. AI needs time but more importantly, guidance. Fortunately, the year-end success of the Pika video application and the launch of Google’s multimodal large model Gemini have brought about a turnaround in the market. This is the charm of technology.
Perhaps hidden within the vast sea of AI is a set of numbers, a line of code, scrutinizing us, expecting us to search for the next peak with rationality and care.