Written by: Techub News Coordination

Introduction

Recently, OpenAI co-founder and CEO Sam Altman returned to his alma mater, Stanford University, to guest lecture in the CS153 "Frontier Systems" course, engaging in an in-depth dialogue with the course instructor. The background of this dialogue is that Altman had previously taught the widely popular entrepreneurship course CS183 "How to Start a Startup" at Stanford in 2014. Ten years later, he returns as a leader of OpenAI to share firsthand experiences and thoughts on AI scaling, AGI development, and future societal structures. This dialogue provides a valuable perspective from a forefront practitioner for understanding the core logic, technological bottlenecks, and future direction of the current AI industry.

Summary

• Scale is Everything: Based on his experiences from Y Combinator to OpenAI, Sam Altman emphasizes that "scale" itself generates emergent properties, resulting in super-linear returns, which is a core driving force behind current AI advancements and many successful systems.
• The Accidental Birth of ChatGPT: ChatGPT was not a meticulously planned product but originated from the unexpected behavior of users "chatting" with the API key, following the YC principle of "observe what users love and put it into practice," ultimately sparking a global phenomenon.
• AI Will Become a New Utility: Altman believes AI intelligence is becoming a foundational utility like electricity, but the current challenge is explaining the value of this "intelligence" to the world, potentially requiring a specific entry point like "night lighting."
• Two Major Forks in the Future: In the coming decade, the world will face two critical forks: whether AI technology will be highly democratized or concentrated in a few companies; and how to fairly distribute computing power as a crucial resource.
• Ongoing Computing Shortage: Altman predicts that as long as AI capabilities continue to improve and costs decrease, the demand for computation will be almost limitless, and computing shortages could become a long-term norm.

Scale: A Severely Underestimated Engine of Emergence

Sam Altman directly shares a core, almost "metaphysical" observation from his career: All the most interesting and successful things are related to emergent properties brought about by scale. These returns often far exceed the consensus expectation. He acknowledges that he lacks a perfect theoretical explanation, but empirical data strongly supports this view.

He uses Y Combinator (YC) as an example. Many believed at the time that YC was too large and should return to the "golden age" of investing in only 10 startups per round. However, Altman insisted on scaling up and discovered that once the number of invested companies reached a critical scale, the network effects within the batch became a powerful emergent property, an advantage that did not exist at smaller scales. This phenomenon of "scale leading to qualitative change" is most evident in the scaling laws of AI models. When OpenAI decided to fully commit to scaling deep learning models, many top researchers in the field were skeptical, thinking it was "not real science" and "not interesting enough." But Altman's team chose to believe in the curve and continued to push for scale expansion, ultimately achieving breakthroughs.

Why do people often resist scaling? Altman points out that systems tend to "break down" in accelerated and unpredictable ways during expansion. This requires the team to have the capability to deconstruct and tackle various reasons for "why it cannot be done"—technical feasibility, capital requirements, business models, cultural resistance. This is in itself a complex system engineering challenge. For OpenAI, convincing the team to concentrate vast computing resources on one project instead of spreading them out was a typical "scaling" challenge.

At the organizational level, scaling also tests human nature. Altman believes the key lies in setting clear goals, clear paths, and providing clear answers on how to achieve those goals and how to make decisions along the way. For instance, when OpenAI decided to bet on scaling deep learning, it communicated clearly: "This is our direction. If we are wrong, we will fail. But this is what we need to do; this is the future vision we believe in." This clear conviction is crucial in combating people's natural cognitive barriers to exponential growth—people find it hard to imagine that scaling laws will lead to continuous exponential growth, that revenues will grow exponentially, and that organizations can withstand exponential complexity.

From API to ChatGPT: The Accidental Birth of a Killer App

Sam Altman reflects on the birth of two key OpenAI products—ChatGPT and Codex—revealing the interaction between planning and chance in cutting-edge exploration.

After the release of GPT-3, OpenAI faced a practical issue: it needed to make money to support subsequent computational investments amounting to billions or even tens of billions of dollars. However, while the team had a strong model, they could not find a clear product direction. After multiple unsuccessful attempts, they decided to take the most straightforward approach: open GPT-3 as an API, hoping the developer community would discover killer applications.

The GPT-3 API, launched in the summer of 2020, initially received a lukewarm response. But about a month later, things changed. Several developers accidentally created interesting applications using the API and shared them on Twitter, triggering viral spread. A large number of developers began experimenting with the API. Altman candidly stated that, according to today's standards, the GPT-3/3.5 models were "shockingly bad," yet they still sparked tremendous enthusiasm. However, developers ultimately found significant commercial success in only one area: copywriting. This was not an exciting, massive market.

The turning point came when the OpenAI team observed that many developers, while unable to build successful businesses with the API, were using their API keys to "chat" with the model. This spontaneous user behavior became a key signal. At that point, OpenAI had already developed a new model (version 3.5, between GPT-3 and GPT-4) and mastered new post-training techniques that allowed the model to better follow instructions, thus optimizing the chat experience.

Following Y Combinator's principle of "observe what users love and put it into practice," the team decided to develop a product around the chat experience, initially intended as a research demonstration to persuade more people to build chat products based on the API. This is ChatGPT. After its launch, it experienced the classic growth curve: traffic soared and then fell back, questioned as a "hype cycle," but the next peak was even higher, followed by another drop... and so on. By the fourth or fifth day, Altman, drawing on his experience at YC, determined: when a mediocre product starts to grow wildly, what you hold may very well be a future blockbuster. The team quickly entered emergency mode, fully devoted to company building and product refinement and swiftly introduced a paid plan to cover the surging computational costs. The story of ChatGPT is a typical case of "emergence": the core application scenario was not pre-designed but naturally emerged from user behavior.

As for Codex (and its represented AI programming capabilities), Altman revealed that before the ChatGPT explosion, OpenAI's original plan was to "fully bet on code." They believed code was the means by which the model controlled the digital world, just as robots control the physical world. Although the unexpected success of ChatGPT disrupted their rhythm, their focus on coding capabilities never changed, and they saw its immense value inflection point.

When asked whether the "standardized" process of current AI capability development (pre-training, mid-training, post-training, RL and supervised feedback loops) would remain stable, Altman provided a negative prediction. He believed that while this is the current mainstream pipeline, significant reconstructions will inevitably occur in the future, and the task of finding better architectures will likely be handed over to the AI itself. OpenAI has set a goal: by September 2024, to utilize computational power equivalent to 500,000 A100 GPUs as "AI research interns"; by March 2028, to achieve AI-led, end-to-end new architecture explorations.

AI as a New Utility: Finding the "Night Lighting" Metaphor

How do we explain the essence and value of AI to the world? Sam Altman believes that we are creating a new "utility," similar to electricity or the internet. Historically, the promotion of new utilities often requires clever metaphors.

He studied the history of electricity's popularity. Early power companies did not market "electricity" directly, as no one knew what it was and many felt afraid (a strange entity that could enter homes and cause death). They marketed "strong night lighting." They told people: what you're getting from us is not electricity but the light of night. As for how electricity could wash clothes or drive appliances in the future, that was secondary, and at the time seemed too far-fetched.

Altman believes that AI intelligence faces a similar challenge. Even if OpenAI's vision is entirely correct—that intelligence will become an omnipresent infrastructure that every company, individual, and government needs to access and utilize—directly selling "intelligence" or "intellect" may fail to resonate. They need to find AI's equivalent of "night lighting," a specific, intuitive, and emotionally appealing value entry point for the public. He has not yet found the perfect answer but firmly believes that this is a crucial step toward "utility-ization."

Regarding the 'utility' analogy, different versions have appeared in the course: NVIDIA founder Jensen Huang compares "computing" itself to a utility, emphasizing its infrastructural attributes; while Altman emphasizes "intelligence" or "reasoning ability" as a utility. Are these contradictory? Altman believes that from the perspective of end-users or businesses, what they care about are higher-level abstractions like "tokens" or services, such as whether they can use intelligence cheaply, extensively, and efficiently, without delving into what kind of chips or hardware are involved. It's like how people pay for mobile plans, focusing on call duration and data rather than how base stations connect to the internet. Therefore, in the future, what users perceive and pay for will be the intelligence service itself, rather than the underlying computing layer.

The Next Decade: Democratization, Computing Distribution, and the Education Crisis

When predicting the "forks" in the next decade, Sam Altman pointed out two of the most critical issues.

The first fork is the level of democratization of technology. Altman admits that there exists a powerful "attractor state," where AI technology might be concentrated in a few companies, which could accumulate a significant proportion of the world's wealth as a result. He believes this outcome would be "terrifying" and promises that OpenAI will strive to avoid it. Promoting the technology towards a "utility" model is an important way to combat this centralization. A world controlled by a few entities with powerful intelligence is not only unfair but also more vulnerable and unstable regarding alignment issues. He encourages all practitioners to promote technological democratization, even if this comes with debates around safety and stability. He personally estimates that there is an 80% probability that the world will move toward a democratized path, though this requires a collective global effort.

The second fork concerns the distribution of key resources—computing power. Altman has become less pessimistic about the "doom of work" in the short term, believing that humanity will always find new things to do. However, he sees a more fundamental challenge: computing shortages. As AI capabilities improve and costs decrease, the demand for computation may become limitless. If computing prices soar due to severe supply-demand imbalances, how to fairly distribute computing resources will become a severe social issue. This may even spark discussions around "computing rights."

When asked whether solutions to inequality (such as universal basic income or citizen wealth funds) require entirely new ideas, Altman believes new concepts may not be necessary, but he personally leans towards ownership models rather than fixed cash subsidies. He has funded UBI research and witnessed the changes brought about by startup investments, believing that ownership of capital aligns more with human psychology and brings a stronger sense of agency. His ideal model is to establish a "citizen wealth fund," similar to Norway's sovereign wealth fund, allowing everyone to own a part of capitalism as economic leverage shifts from labor to capital.

Regarding the current computing shortage, Altman confirms its severity and predicts it could be a long-term state. As long as AI capabilities continue to progress, demand will consistently outstrip supply. He likens it to how energy demand is highly correlated with prices: If AI intelligence is good enough and cheap enough, demand will be virtually unlimited. In the future, everyone may want to run 10 or even 100 personal intelligent agents, which will consume enormous reasoning power.

Finally, Altman expressed concerns about the lag of the education system. He initially thought that within one or two years after the release of ChatGPT, the education system would be comprehensively reshaped to adapt to the AI era, focusing more on cultivating students' abilities to utilize AI for project practice and deep thinking. However, three and a half years have passed, and he can hardly see any systematic significant changes. He warns that if teaching and assessment continue in the "pre-AGI era" manner, it will lead to a decline in students' critical thinking abilities. Of course, some basic skills (such as writing and programming), due to their value in training metacognition and thought processes, are still worth teaching, but many other teaching and assessment methods must be thoroughly reformed.

In the Q&A session, Altman also responded to Yann LeCun's criticism of LLMs, stating that he believes "betting that LLMs cannot scale is quite erroneous at this point," that world models are crucial for domains like robotics, but LLMs have already surpassed humans in many aspects and demonstrated their ability to discover new knowledge. He appeared calm towards challengers, believing that when a person has tightly bound their identity to a particular viewpoint, it is difficult to change even in the face of opposing evidence.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。