Recently, the hottest topic in the tech circle is none other than Sora.
At the beginning of the Year of the Dragon, OpenAI launched a remarkable model called Sora, which can generate a one-minute video based on input. The effects are indeed stunning, putting considerable pressure on both domestic and international peers.
The emergence of a popular technology always attracts followers, but some comments are overly exaggerated. One moment it’s said to disrupt industries like film, advertising, and gaming, and the next it’s dubbed the “GPT moment.” We’ve experienced so many “GPT moments” in the past year that the term has become inflated.
Our readers are likely familiar with the aforementioned statements: these so-called “disruptive” and “revolutionary” words are merely tricks that AI in drug development has long played.
Now, Sora can generate a 60-second video. Isn’t it similar to AI helping to filter a drug molecule into clinical trials? It’s very cool and impressive, but so what? Will a multi-trillion dollar industry be disrupted because of this?
After so many years of development, AI pharmaceutical companies still haven’t seen the backs of giants like Pfizer, Johnson & Johnson, and Merck. Why do people think Sora will lead to the downfall of Pixar and Hollywood?
To truly disrupt an industry, one must at least have an understanding of that industry. Otherwise, making these exaggerated claims will only distort expectations and create anxiety, with no positive effect on the industry.
Two Core Issues
Sora and AI drug development may seem like two unrelated fields, but they face similar issues.
1. Cost Reduction Remains Uncertain
From the perspective of the pharmaceutical industry, the actual cost reduction that AI can achieve is at most 2% of the entire industry, far from any significant transformation.
Although Sora has not disclosed technical details and is currently limited to internal testing, we do not know the actual cost behind the videos it generates. However, based on some competitors’ performance (Runway charges about $0.2 for a 4-second video and nearly $1 for an 18-second video), Sora’s requirement for computing power to generate a higher quality 60-second video would be much greater. Charging you several dozen dollars per minute wouldn’t be unreasonable, right? This is already a price that most individual media cannot bear.
Note that this is just the cost of using Sora; the cost of hiring more professional personnel has not been included.
Despite some claims about “one-click video generation,” it is clear that the final product still requires the involvement of professionals, which further raises the requirements for practitioners.
Moreover, industry professionals who master AI technology undoubtedly demand higher prices. Although the number of required employees may decrease, the unit price has increased, making it uncertain whether personnel costs have decreased or by how much.
This brings to mind the current situation of some AI+CRO companies: no matter how much AI is touted to replace human labor, in actual projects, a large amount of human cost is still indispensable. The salaries of some AI technicians are often several times that of traditional biochemistry experts, leading to related costs even exceeding those of traditional CROs.
2. Limited Breakthroughs in a Long Industry Chain
Many human errors stem from “assuming.” Overemphasizing the importance of a single breakthrough in a long industry chain is a mistake, whether it’s Sora or AI drug development.
For example, in the film industry, even if an AI model like Sora is powerful, what it replaces is merely a part of the upstream production process, and the efficiency improvements have a minimal impact on the final result.
Whether a film can achieve significant commercial value relies not only on production but also on critical aspects like preparation, research, marketing, and distribution, which Sora cannot accomplish.
For instance, in the film industry, IP is an unavoidable topic. From the world-building and story setup to stars and cartoon characters, everything requires intentional construction, integrating the creator’s thoughts. Major companies are very familiar with this operation process, which is their core competitiveness, something Sora lacks.
Additionally, in today’s information explosion world, “good wine fears no deep alley,” platforms often influence a film’s market performance. This is why many excellent studios need to maintain cooperation with streaming giants like Netflix. The creation of a work follows one logic, while commercialization follows another; Sora provides virtually no assistance for the latter.
This point serves as a cautionary tale for AI drug development.
Many once believed it would change drug development models, even leading to “one-click drug generation,” but it has been proven to be wishful thinking. Currently, the problems AI can actually solve are still limited to the preclinical stages, while its role in more critical stages (such as clinical trials) and more important decisions (like selecting targets, indications, and patient populations) remains quite limited.
This also results in AI drug molecules still facing high risks, long cycles, and enormous costs when entering clinical trials, with some first batches of AI drug molecules ending up in the “no survivors” category.
Whether in film, advertising, gaming, or pharmaceuticals, all are industries with decades or even centuries of history, where every link and department has its significance and value, and numerous professionals depend on them for their livelihoods; they cannot be easily disrupted with mere words.
Understanding “Understanding the World”
A core controversy regarding Sora is whether this technology can understand the physical world.
Supporters argue that Sora is a “data-driven physical engine,” a “world model” capable of simulating both real and virtual worlds.
Opponents contend that the process of generating videos is entirely different from causal prediction based on a world model; the correlation of probability statistics cannot accurately express the causality of physical laws.
If we look at the technical documents released by OpenAI, we can find that their own statements are quite ambiguous: “This exploration (referring to Sora) indicates that developing a general simulator capable of simulating the dynamics of the physical world is a promising avenue.”
Simulating the physical world and understanding the physical world are two different concepts: the former only requires that the final presented results conform to our common sense and logic, while the latter is more complex; how to define “understanding” is a major question, whether it is according to human understanding or AI’s own understanding.
OpenAI emphasizes the former while deliberately avoiding the potential debates triggered by the latter. Previously, Sam Altman expressed a similar viewpoint at the Davos Forum: “I cannot delve into your brain to observe trillions of synapses to understand what is happening, but I can ask you to demonstrate your reasoning process to determine if it is reasonable. I believe our AI systems can do the same.”
Altman’s viewpoint is essentially a form of outcome determinism; as long as the outcome is good, how AI understands the physical world or how it understands it is secondary.
In fact, even among supporters, industry experts express restraint in claiming that Sora “understands the world,” while those outside the field tend to make such claims more assertively.
Onlookers often enjoy making a big deal out of it, as is their habit.
A similar situation occurred in AI drug development. When Alphafold2 was first released, there were many voices claiming it would open the so-called “black box of life” and unveil the mysteries of life.
But nearly three years later, reality tells us that AI is indeed powerful, but more so as a function fitting computation process; whether it understands biological mechanisms is unknown, but it certainly does not do so in a human way, and explainability remains a major issue.
Conclusion
Since the dawn of humanity, most progress has been made by optimizing and improving upon existing foundations, with truly “disruptive” events being exceedingly rare.
When the AI drug development craze subsides, the founder of DeepMind solemnly remarked: “AI truly reconstructing drug development requires at least six AlphaFold-level breakthroughs.” How many Sora-level breakthroughs are needed to reconstruct the video industry?
As for why people are so enamored with new technologies today, I believe that, besides the influence of some marketing masters (including Altman and Musk), it is more a reflection of societal mentality. The frenzy behind it is an overflowing anxiety.
Source | Intelligent Pharmacy
