Written by: Techub News Compilation

Recently, OpenAI CEO Sam Altman participated in an in-depth interview with DRM News. In this approximately 35-minute conversation, Altman elaborated on OpenAI's current development priorities and strategies, as well as providing profound insights into the current state of artificial intelligence (AI) technology, its future trajectory, and its far-reaching impact on the socio-economy. As a key figure leading the global AI wave, his views are crucial for understanding the direction of the industry.

AI has crossed the economic utility threshold, evolving towards "super employees"

Sam Altman opened by pointing out that the development of AI has recently crossed a critical threshold: achieving significant economic utility. He believes that previously, people needed time to figure out how to use these models, but now, the models themselves have become smarter, and the related "pipelines" (i.e., usage interfaces and tools) have matured, causing AI's capabilities to start astonishing the world.

This utility is most evident in the programming field, but it is also happening in scientific research and many knowledge work sectors. Altman described a phenomenon of "directed acceleration": people suddenly realize that things they thought would take years to achieve are happening now. Their work nature is shifting from direct technical or legal work to managing a team of AI agents.

He predicted that this trend will rapidly deepen. Currently, one might trust an AI software engineer to complete a task that takes a few hours; soon, it will handle tasks that take days, then weeks. Before long, the paradigm will shift again, and AI systems will feel seamlessly connected to your life or company, proactively thinking, working continuously, possessing all necessary contextual information, and handling matters as if they were a trusted senior employee.

Regarding whether businesses truly understand how to leverage these systems to reshape operations, Altman believes the situation is varied. The mindset of the new generation of startups is completely different: they no longer talk about how many employees to hire, but instead focus on how much compute capacity they can acquire. They care about whether they can reserve enough capacity, secure cloud service agreements, and obtain sufficient tokens. This shift of thinking is progressing slowly in large companies, but some have started.

A clear sign is that engineering and product departments are beginning to discuss plans to double or triple the amount of products delivered this year, "which has never happened before."

The definition of AGI and two key thresholds

When asked about the arrival of artificial general intelligence (AGI), Altman first noted that the term "AGI" itself has lost much meaning, as the definitions vary widely. Some believe we have already reached it, some think it is close, while others believe it may take another year.

He proposed two more interesting thresholds to measure progress:

• The first threshold: When the global cognitive capacity within data centers exceeds that of the external (i.e., humans). Altman thinks this could happen around the end of 2028 (with a large margin of error). This would be an extraordinary global transformation.
• The second threshold: When CEOs of large companies, presidents of major nations, and Nobel Prize-winning scientists will not be able to perform their jobs without heavily relying on AI. This does not mean there will be AI CEOs or AI presidents, but it signifies a fundamental change in the role of human CEOs. Humans will still need to stand behind decisions, exercising human judgment and our expectations for leaders of significant organizations, but an increasing portion of the actual work will have to rely on AI, as no one can simultaneously converse with all employees and customers of a company, attend all meetings, and master all fields. Thus, these roles will increasingly shift to overseeing a group of AIs, providing oversight, deciding how to trust outputs, and how to provide guidance. This threshold may come later, but not too much later.

Altman himself is increasingly reliant on OpenAI tools as CEO at an "incredibly quick" pace. When he has new ideas, the first thing he does is ask about the company's tools. As the tools gain the ability to access closer to the complete context of the company (including internal documents, communications, code, customer data, etc.), the quality of the responses also improves significantly.

$11 billion financing and the vision of making intelligence as cheap as water and electricity

The interview mentioned OpenAI's recent completion of an $11 billion financing round, significantly surpassing the largest public offering in history. Altman explained the purpose of this substantial capital: to meet the extremely costly and forward-committed infrastructure demands.

He stated that he has never seen any other industry like AI infrastructure. If the demand growth curve remains as steep as it is now, some "quite unusual" measures must be taken. OpenAI has done many seemingly strange things: making huge investments in infrastructure before the revenue arrives, trying new business models (like advertising), etc. All of this stems from a fundamental belief in the abundance of intelligence.

"One of the most important things for the future is that we want to make intelligence—borrowing an unfulfilled old saying from the energy sector—'too cheap to meter.' We want to flood the world with intelligence. We want people to use it for everything. We hope that future generations will not think about it, but expect it to be ubiquitous, and that everyone can access the required amount of 'genius' in any necessary field."

This guiding principle has led to many behaviors at OpenAI that differ from other companies, one of which is a strong desire to escape the situation of "always being capacity-constrained." Altman reiterated his view: compute is revenue. Fundamentally, he believes that OpenAI and all model providers' future business will be akin to selling tokens. These tokens might come from larger or smaller models, using more or less reasoning, with varying costs. The AI might be running in the background trying to assist you, or it could operate only when needed to reduce costs. They might invest tens of millions, hundreds of millions, or even tens of billions in the future to solve a genuinely valuable problem.

His envisioned future is one where intelligence becomes a utility like electricity or water, purchased and used by measure. If there is a shortage of supply, either it cannot be sold, or prices skyrocket, directing it only to the wealthy, or society makes some central planning decisions that he believes are almost always poor. Therefore, in his view, the best practice running through the history of capitalism and innovation is to flood the market.

Stargate Super Data Center and a thousandfold efficiency increase

A key part of addressing compute demand is the super data center project codenamed "Stargate." Altman described the shock of visiting these megawatt-scale facilities under construction or operation: "The scale is really hard to explain... inside it looks like a spaceship, incredibly unbelievable."

Currently, OpenAI is training at its first site in Abilene, which he believes will become "the best model in the world, hoping to be far ahead." From multiple visits to the construction process, to a day when an OpenAI researcher inputs a command triggering an unbelievable number of GPUs to start collaborating on a massive computation, this process left a deep impression on him.

Throughout the construction and expansion process, there have been both expected challenges and unknown unknowns, such as Abilene's rare extreme weather events and supply chain challenges. Any project of such scale will encounter numerous issues. A positive surprise has been how various organizations have come together in a very short time and co-operated under immense pressure like a team.

Power demand is a focal point for many. Altman remains cautiously optimistic, believing that humans will learn to build massive amounts of power generation facilities, aided by AI itself. In light of current demand growth, he is "sort of hoping for a miracle," wishing for a significant increase in efficiency per watt to buy time for infrastructure development.

He shared an astonishing efficiency improvement data point: from OpenAI’s first inference model "01" (released about 16 months ago) to the latest integrated inference model "5.4", the cost of solving the same problem has dropped by about 1000 times. This points to two things: First, we are still in the early stages of this paradigm, with much room for improvement in our understanding of how to develop, train, and efficiently operate models; second, human ingenuity and our ability to find solutions within constraints continue to yield surprises. It’s not just that models are becoming better; core engineers are helping to write more efficient cores, data center designers are finding more efficient methods, and all parties are responding to calls for increased efficiency.

In-house inference chips, global competition, and the speed of India

In addition to being a significant customer of NVIDIA and AMD, OpenAI is also developing its own chips. Altman clarified that this chip is for inference only. The underlying idea is that, in the challenges faced in the future, a chip specifically designed for inference—not necessarily the fastest, but the most efficient per watt and the lowest cost under given constraints—will be vital for all future agent needs. This is a bet with a clear point of view, a chip with limited functions, but it will be extremely important in a world with energy constraints.

He briefly explained the distinction between training and inference: training is like humans spending 22 years receiving education and learning vast amounts of knowledge; inference is about efficiently solving a problem when posed with a physical question. The high efficiency of AI models is often compared to a human's 22 years of training time against an adult's 1-second problem-solving time. If models' efficiency in solving the same physical problem is compared to that of humans, models might already be more energy-efficient.

On chip progress, Altman expressed optimism, expecting the first batch of chips to flow back in a few months and to be deployed on a large scale by the end of this year.

Regarding competition with China, Altman first offered a framework for thinking: the discovery of deep learning is closer to discovering an element or fundamental property of physics rather than a secret technology. This means that ultimately (and this "ultimately" may not be too far off), the fundamental ideas that make models so powerful will be simplified and widely known. Just as most principles in physics are understood, we will view many principles of artificial intelligence as scientific principles. The best historical analogy is the transistor: it was also a fundamental scientific breakthrough, hard to discover, requiring time to refine, but once understood, the scientific principles are clear to all. However, there is still much operational knowledge surrounding it (like the capabilities of TSMC), and there will still be many differential advantages regarding industrialization processes, workflow integration, training data, model availability, etc. But the most crucial differentiation may lie in who possesses the infrastructure and how much of it.

Specifically, he believes: the world's most powerful models (cutting-edge) are led by the United States; the cheapest inference use of models from two generations ago is dominated by China; infrastructure, the U.S. currently leads, but China is building faster; in terms of industrialization and productization, China moves faster; in closed-source areas, the U.S. leads, while in open-source areas, China leads; overall, the U.S. may still be ahead.

Altman was impressed after a recent visit to India. The extent to which Indian startups utilize AI technology amazed him. Codex's usage in India has multiplied tenfold in just a few months. When communicating with Indian startups, he sensed a stronger version than in the U.S.: people were saying the world had changed. Some talked about "one-person startups," while he, in contrast, was trying to establish "zero-person startups"—just wanting to write a prompt to create an entire startup, handling software, customer support, legal affairs, etc., and then go on vacation. Major companies in India directly inquired, "How much capacity can we purchase from you? How long can we reserve? We want to finalize this with you now and won't let you leave the room until an agreement is reached." This ambition, speed, and firm belief that AI will reshape India's business landscape left a significant impression on him. He believes that India is aligned in direction with U.S. clients but seems to be going further or moving faster.

Democratizing AI, societal challenges, and the future of GDP measurement

Altman distinguished between "autocratic AI and democratic AI." He believes that occasionally there arises a technological revolution that transforms society to such an extent that decisions about it should not be left to the few companies fortunate enough to be involved. He himself is a staunch believer in capitalism, corporate rights, and that government should not intervene excessively, but he sees AI as a period of exception, where society has legitimate interests in its impact. The internet era was similar, but it was not handled entirely well then. He hopes we can learn from those lessons and do better.

If the predictions of AI companies hold true—that AI will reshape the economy, geopolitical power, and change everyone's way of life—then how to use AI should not be imposed by companies or a particular government’s will, but rather belong to the will of the people and be determined through democratic processes. Companies like OpenAI are rapidly entering a role of "critical infrastructure," and it must be acknowledged: we create this technology, we are the experts, we should have a genuine voice in it; we have insights and understanding of its limitations and areas where it is not yet ready for use and could cause significant harm. However, rules and restrictions must be agreed upon through societal processes. With technology evolving too quickly, there is hope that democratic processes can run a bit faster.

Turning back to the global AI competition, Altman believes the U.S. is most vulnerable in three aspects:

• Dependency on global supply chains and U.S. infrastructure: he cannot over-emphasize how terrifying this is for him. If lagging behind on infrastructure and unable to catch up, or globalization decomposing in any possible manner, the U.S. cannot independently sustain the construction of AI infrastructure, which would be a huge vulnerability. He doesn't dislike, but also doesn't particularly care for, the U.S.'s current global position.
• The speed of economic adoption: it’s a competitive world. If the U.S. is not as quick as other countries in the economic application of AI, it will lose the advantages that come with being an economic powerhouse. This relates to the speed at which companies, scientists, and governments adopt it. On the positive side, this is a once-in-a-generation opportunity to genuinely improve the economy and rewrite some ineffective social rules based on this new incredible wealth fountain. The U.S. has the chance to turn this into the greatest competitive advantage, but the current trajectory is not obvious.
• The diffusion of technology to other parts of the world: Will the world primarily build on the U.S.'s AI technology stack (chips, models, applications, etc.)? Or will the U.S. Implement a series of policies that make it more difficult?

Altman believes that if properly utilized and adopted, AI could become the foundation for immense productivity prosperity. However, the measurement methods must change. He envisions a potential world: productivity prosperity is astonishing, quality of life continues to improve, and most of the things we care about get better, yet GDP measured in the current way continues to decline, leading to long-term deflation. He does not know what living in a forever deflationary world means, nor how to comprehend the relationship between GDP and quality of life when cognitive capacity within data centers outpaces that outside. There will be a lot of debate about the right measurement metrics in the coming years.

He believes society has already begun discussing these challenges, but there is no simple consensus answer. For centuries, even millennia, we have learned extensive knowledge on how to build societies to manage scarcity, but this knowledge is of little help as we rapidly learn to manage abundance. This represents a genuine shift in how capitalism works. Capitalism also relies on some degree of power balance between labor and capital. But if it becomes difficult to surpass GPUs in many of our current jobs, then that balance will change.

Altman is not a long-term pessimist about employment; he believes humans will find new things to do. Nor is he a long-term pessimist about capitalism; he is confident in it. But he says that the next few years will be a painful adjustment period where we will collectively redefine what the new system and this incredible prosperity will look like, with some very intense and uncomfortable debates along the way.

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。