Cool Reflection Amid Sora's Hype: The Brilliance of Generative AI Lies Beyond the Bubble
-
"The sky has not left traces of wings, but I have flown." Tagore's timeless verse aptly describes the phenomenal impact of OpenAI's video generation model Sora (pronounced as "sky" in Japanese) in the court of public opinion.
Amid the hype surrounding Sora, Nvidia, hailed as the "hardware overlord of generative AI," has seen its market capitalization soar past the $2 trillion mark, with its founder Jensen Huang emerging as a top evangelist for human technological advancement. Yet, just over a year ago, Nvidia, battered by the collapse of the cryptocurrency bubble, saw its stock price plummet by 60%. The Economist even took a jab, questioning, "As he peers through his glasses at flashy new models he thinks will change the face of AI, and at vaguer concepts like the metaverse, is he in danger of underestimating the brutal reality of the here and now?"
The current landscape is dazzling, with a flurry of technical predictions about generative AI emerging out of thin air, and investors growing increasingly excited about its prospects. In fact, as early as a quarter century ago, Ray Kurzweil, a preacher of the technological singularity theory, envisioned the development of 'massively parallel neural network computers' in his masterpiece The Age of Spiritual Machines. He predicted that around 2020, with the support of this computational foundation, artificial intelligence technology would achieve several milestones:
- Most commercial transaction scenarios would involve a virtual person;
- The majority of roads would be equipped with autonomous driving systems;
- People would begin to form connections with robots, treating them as companions, teachers, caregivers, or even lovers;
- Virtual artists would emerge in various artistic fields; Media widely reported that computers have passed the Turing test, although these tests do not yet meet the standards recognized by experts...
After more than two decades of twists and turns marked by excitement and disappointment, the rise of OpenAI seems to have brought us to a milestone of progress as envisioned by the future scientific community.
Regarding the technological significance of generative AI, even the typically reserved deep learning pioneer Geoffrey Hinton did not hesitate to praise: "AI will change the world more than anything in the history of humanity. In scale, it can be compared to the Industrial Revolution, or the invention of the wheel and electricity." ("AI is going to change the world more than anything in the history of humanity.") Indeed, even without quoting golden sayings from big names like Hinton or Jensen Huang, the general public can easily awaken a simple yet strong intuition from the viral spread of ChatGPT and Sora, realizing that a significant transformation is happening right now. If the 2016 AlphaGo vs. human match completed the popularization of AI's 'usefulness,' then today's increasingly popular large AI models can be seen as a clear demonstration of 'ease of use.' The two prerequisites for technological diffusion are now in place, and the 'long summer' of artificial intelligence is foreseeable.
It is claimed that Sora represents OpenAI's underlying model's ability to perceive and understand the real world, that AI is capable of generating its own open world, interacting and evolving within it, and that the path to Artificial General Intelligence (AGI) is now open.
However, filtering out the 'DIY-style' players like Li Yizhou, interpretations around generative AI are abundant, but when it comes to the essential questions of 'What is it really useful for? How useful is it?' there remains no clear and definitive answer to date. In fact, within the intricately polished story of technological and business evolution, many critical milestones did not originally exist in the mind of any technical genius. They either emerged as products of mutual enlightenment and deepened understanding during engineering practices, gradually forming a consensus within the research community, or were purely inherent "emergent" properties of neural network models themselves.
"People don't know what they want until you show it to them." This famous quote by Steve Jobs applies equally to both the recipients of innovation and its creators. Take OpenAI as an example. The "emergence" of GPT model's performance was actually an "accident" resulting from engineering exploration as model parameters increased. As for the inter-frame coherence and object consistency demonstrated by Sora, project developer Tim Brooks also admitted these were unplanned capabilities. From the engineering principles of so-called Diffusion Transformer, Sora can hardly be called a "world model". According to Yann LeCun's description of world models, the essential intuitive "common sense" about the real physical world is clearly at odds with traditional neural networks' approach of approximating implicit probability distributions. The stunning video effects may only prove that Sora has learned the probability distribution of physical laws, not the physical laws themselves.
From a technical perspective, Sora still hasn't proven or disproven an exceptionally important question: Facing the "black box" of neural networks, is the Scaling Law's brute-force increase in complexity a viable path to AGI, or just a sweet illusion after all low-hanging fruits have been picked and satiety achieved? If the answer is the former, then there is no doubt that the United States has firmly grasped all the key chips leading to AGI. From infrastructure suppliers represented by NVIDIA to large model developers like OpenAI and Google, their advantages over overseas competitors are astonishing. The successive suppression of China, the main competitor, demonstrates the Americans' determination to actively defend this advantage. However, at this highlight moment when the U.S. AI industry is "winning overwhelmingly," it may be necessary to also remember a sobering law: the gifts of fate often come with a price.
According to a study by Accenture on the impact of generative AI on human jobs, banking, insurance, and software are the top three industries with the highest risk exposure. As we all know, these are currently the high-end pillars of the U.S. economy. Once the technological maturity of generative AI crosses a certain balance point, its accelerated popularization may make the U.S. itself the first and deepest to feel the pain of transformation, and the socio-economic consequences are still hard to predict.
If the answer is the latter, then the judgment during the first major trough in the history of artificial intelligence can also be seamlessly applied today: "The first person to climb a tree can claim it as a significant step toward flying to the moon." Amid the diminishing returns of Scaling Law, can large language models overcome intermittent hallucinations and catastrophic forgetting to avoid outputting absurd statements like the recent joke about 'bullet train temperatures reaching 1538℃'?
Taking Sora as an example, can its application prospects truly lead to the so-called 'one-sentence movie generation'? Based on current speculation, if the model cannot achieve continuous prompt correction and relies solely on trial-and-error like rolling dice with prompts, its application in image production scenarios may remain illusory. Even for short advertising videos, can the niche market size support the current generative AI concept stock market valuation of over $10 trillion? In any case, there is a fact worth emphasizing: the public's fervent anticipation of Sora today has occurred many times since the dawn of the Industrial Revolution. Each time, people believed that the new era of human society brought by automation was just around the corner. Consider the discussions by Norbert Wiener, the father of cybernetics, in his 1950 work The Human Use of Human Beings, on the possibility of machines replacing humans and its implications—how strikingly similar they are to today's public discourse: "From this stage on, all work can be done by machines. This mechanization applies equally to the vast majority of tasks in industrial enterprise libraries and archives. In other words, machines show no preference for manual labor or clerical work. Thus, the new industrial revolution will penetrate a very wide range of fields, including all labor that requires little thought... The new industrial revolution is a double-edged sword; it can be used for the benefit of humanity or for its destruction. If we do not use it wisely, it may quickly reach that point."
Of course, today's ChatGPT, Sora, and even the earlier milestone AlphaGO, despite delivering clear and profound perceptual shocks to the public, any ordinary person switching to a producer's perspective can immediately recognize the deep chasm between their capabilities and scenarios and the requirements of productivity tools. Stirring public curiosity is merely the first step in the long march from technological possibility to commercial transformation. Letting time provide the answer might be the wisest attitude.
Undoubtedly, the current AI frenzy is comparable to the dot-com bubble of the millennium. Back then, enthusiastic investors and entrepreneurs were equally eager to bet everything on imagined transformative visions despite the lack of clear application scenarios. Shortly after the bubble burst dramatically, Amazon turned a profit during the 2001 Christmas shopping season, marking the moment the internet economy found its direction.
Time and again, people set out upon seeing the summit, only to hesitate when searching for the path—until, in the valleys and at the edges, major breakthroughs in engineering and applied innovation ignite from the bottom up. The thread of history is always this concise and profound. Looking back at history, the industrialization path of generative AI is likely to remain the same—moving forward in uncertainty, yet clear in hindsight. Once the AI computing power and model arms race among major platform giants concludes, the capital bubbles of OpenAI and even Nvidia may face inevitable liquidation. Seizing the moment to realize value is perhaps the intention behind OpenAI's meticulously packaged Sora PR campaign. However, the true excitement of the industry might only emerge after the bubble bursts.
Kurzweil's another prophecy could serve as a fitting conclusion and aspiration for this article: "Considering all these factors, it is reasonable to estimate that by around 2020, a $1,000 personal computer will match the human brain in terms of computing speed and capacity, especially in neural connection operations (the primary computing method of the human brain)."