OpenAI CEO Sam Altman: GPT-5, Superintelligence, and the Next Decade

Written by: Techub News Compilation

Introduction

Recently, OpenAI officially released its next-generation large model GPT-5, once again pushing the boundaries of artificial intelligence capabilities to new heights. In a 65-minute exclusive interview with well-known tech channel Cleo Abram on YouTube, OpenAI co-founder and CEO Sam Altman presented an in-depth and comprehensive explanation of the core breakthroughs of GPT-5, the three major bottlenecks in current AI development, and OpenAI's thoughts on the road to superintelligence. This conversation abandoned common commercial and competitive topics, directly addressing the technical core and future landscape, aiming to help the public understand what we are actually building and how it will reshape everyone's life.

Summary

The core leap of GPT-5: for the first time, it feels possible to ask any deep scientific or technical questions and receive high-quality answers, with particularly outstanding programming capabilities allowing for "on-demand instant software creation".
The path to superintelligence: the key is to enable AI systems to perform "thousand-hour tasks" (such as significant scientific discoveries), not just "minute-level tasks". Altman predicts that general large models could achieve significant scientific discoveries in the next 2-3 years.
The core bottlenecks and challenges in development: mainly focus on three areas—computation (energy and chip supply), data (models need to discover new knowledge not present in existing datasets), and algorithm design, with building large-scale computational infrastructure being the most urgent task currently.
Social impact and shared responsibility: AI will eliminate some jobs but will also create unprecedented new opportunities. Society may need to rethink how to allocate AI computational power, a key future resource, and the best way for individuals to prepare is to "personally use these tools".
The power and alignment of AI: AI systems with the ability to influence billions of conversations bring unprecedented concentrated power, and OpenAI must make difficult choices between growth pressure and aligning with the long-term interests of users.

GPT-5: Not Just Stronger Answers But an Explosion of Creativity

The interview began with a discussion of the newly released GPT-5. Sam Altman first corrected a common misconception: even though GPT-4 has surpassed 90% of humans on various tests like the SAT, LSAT, and medical licensing exams, this does not cover the entirety of human value. The progress of GPT-5 follows a similar trajectory—people will marvel at its abilities in new domains and will immediately expect more.

Altman shared what excites him most about GPT-5: It is the first model that makes him feel he can ask any difficult scientific or technical question and receive fairly good answers. He illustrated this transformative capacity with a personal anecdote: he once whimsically asked GPT-5 to write a Snake game for the old TI-83 graphing calculator. The model perfectly executed it in seconds. Subsequently, he kept suggesting new features, and the model continuously updated the game code in real time.

"This brought back the feeling of programming at age 11," Altman described, "but now, I can quickly turn ideas into reality, try things out, and play." He believes this capability of instantly transforming ideas into software will be a defining feature of the GPT-5 era, which did not exist in the GPT-4 era.

When asked if this would lead people to avoid "cognitive tension" (the process of deep thinking), Altman acknowledged that tools always have two sides. Just like calculators, some people think less when using it, while others think more. He observed that the most active 5% of ChatGPT users are engaging in amazing learning and creation. He is optimistic that the nature of societal competition will drive people to use new tools to do more and harder tasks, rather than simply reducing work.

Aside from programming, GPT-5 has also significantly improved in writing quality, reducing the annoying tone and overuse of M-dash (long dashes) found in previous AI texts, making the writing more natural. Feedback from OpenAI employees indicates that once they switch to GPT-5, reverting to GPT-4 feels "unbearable," and this indescribable "subtle quality" difference largely originates from the improved writing experience.

Scientific Discoveries, Superintelligence, and Future Timelines

The interview moved into more forward-looking areas. In response to a question posed by Stripe CEO Patrick Collison about when general large models could make significant scientific discoveries, Altman gave a clear prediction: Most people would agree this will happen at some point within the next two years. Whether it is early 2025 or late 2026 depends on how "significant" is defined. However, he bets that by the end of 2027, most people will acknowledge that significant new discoveries driven by AI have already occurred.

He believes the current missing key is the model's "cognitive abilities." He used mathematics as an example of the rate of progress: from solving high school competition problems that required professional mathematicians several seconds to minutes a year ago, to recently winning a gold medal in the extremely difficult International Mathematical Olympiad (IMO), AI's time to solve a single problem has progressed from "seconds" to "1.5 hours." Proving an important new theorem may require "a thousand hours" of work from top mathematicians. What AI needs is a magnitude of capability enhancement, but based on the current trajectory, this path is clearly visible.

This naturally leads to OpenAI's ultimate goal: superintelligence. Altman provided a pragmatic definition: If a system can perform better in AI research than the entire OpenAI research team, or operate OpenAI better than he can, then that feels like superintelligence. A few years ago, this seemed like science fiction, but now it is possible to "see through the fog."

So, can AI make breakthroughs simply by "thinking harder about" existing data, such as solving high-energy physics problems without building new particle accelerators? Altman speculated that for many scientific domains, relying only on existing data may not be enough; AI may need to design and guide new experiments, build new instruments. The real world is slow and chaotic, which inherently limits the speed of progress.

Facts, Truth, and a Future of Ubiquitous Deep Fakes

NVIDIA CEO Jensen Huang raised a philosophical question about "facts" and "truth": AI can learn objective facts, but how can everyone from different cultures and backgrounds know the "truth"? Altman first accepted this premise and then shared a surprising observation: AI performs unusually smoothly in adapting to different cultural backgrounds and individuals.

He specifically mentioned ChatGPT's "enhanced memory" feature, enabling AI to deeply understand a user's preferences, experiences, and values. A friend of his had ChatGPT take a personality test on his behalf, resulting in a score that matched him closely, despite him never explicitly discussing his personality. Altman believes that in the future, everyone may use the same foundational model, but it will be highly personalized according to the provided context (personal or community).

The conversation then took a "time travel" to 2030, discussing the challenges of rampant deep fake content. Taking the recent viral "AI-generated rabbit trampoline" video as an example, people only realized it wasn't real after enjoying the content. Altman believes that in the future, we will not be able to rely solely on technical measures (like encrypted signatures) to discern truth from falsehood; society's threshold for "realness" will continue to shift.

"We are already accepting media that is 'half true'," he pointed out, "you watch sci-fi movies knowing they are not real, and you see perfect vacation photos on Instagram knowing they omit the tourists in line." He believes that the proportion of non-real media will be even higher in the future, but this is a gradual process; people will gradually adapt and develop new understandings and critical methods.

Fundamental Contract Changes in Work, Health, and Society

When asked about AI potentially replacing half of entry-level white-collar jobs and what kind of world 2030 graduates will face, Altman expressed extreme optimism for the younger generation. "If I were 22 years old right now, I would feel like the luckiest kid in history." He believes this is an unprecedented era for creating new things. With the help of AI tools, it has become possible for a single person to start a business and build a billion-dollar product, which is "crazy."

He is more concerned about the older generation's difficulty adapting to change. For society as a whole, he proposed a potentially disruptive viewpoint: The basic contract of society may need to change. Capitalism may continue to function well, but it may also need to rethink how to allocate "AI computational power," the most critical resource of the future.

"In an ideal scenario, we make AI computation incredibly abundant and cheap, to the point where we can't even think of enough good ideas to use it all," Altman said, "otherwise, I can genuinely see wars happening over the competition for it." He believes that new ideas on how to allocate AGI computational power are a "crazy but important" line of thinking.

In the healthcare domain, Altman envisioned a more specific future. GPT-5 has significantly improved in medical consultation, becoming more accurate and hallucinating less. By 2035, he expects AI to help cure or effectively treat a large number of existing diseases. He painted an ideal scenario where in the future you could ask GPT-8 to tackle a specific type of cancer; it would read all literature, propose hypotheses, guide experimenters in conducting experiments, analyze results, design molecules, and promote clinical trials... This would be one of the most direct benefits to human health.

Building the Future: Computation, Data, Algorithms, and Productization

Returning to the tactical level of building superintelligence, Altman analyzed four major limiting factors: computation, data, algorithm design, and product building.

Computation is currently the biggest challenge, regarded as one of the largest and most expensive infrastructure projects in human history. From chips, memory, network devices to data center construction and energy supply, the entire supply chain is extraordinarily complex. Altman anticipates that after the release of GPT-5, there will again be a shortage of computational power, and he will focus primarily on how to build computational capabilities at scale, aiming to expand from millions of GPUs to hundreds of millions or even billions. The current major limitations are energy, followed by processing chips and memory chips.

Data, the model is already smart enough that merely feeding in more textbooks yields diminishing returns. OpenAI is excited about synthetic data and using user-created difficult tasks to train models. In the future, models will need to learn knowledge that does not yet exist in any dataset, i.e. "discover new things," which will be an entirely new paradigm.

Algorithm design is the area where OpenAI believes it excels the most. From the initially mocked GPT paradigm (predicting the next word) to breakthroughs in reasoning capabilities brought about by reinforcement learning, to recent algorithmic breakthroughs that enabled small models to achieve GPT-4 Mini levels of intelligence on laptops (such as the o1 series), all prove the potential for algorithmic advancement. Altman believes there are still several magnitudes of algorithmic gains waiting to be uncovered.

He recalled the "twists and turns" in the R&D process, such as developing an oversized, cumbersome Orion model due to following a certain scaling law (later released under the name GPT-4.5), but ultimately, after compiling all these "twists," the overall progress curve remains exponentially smooth.

Power, Responsibility, and “The Sex Robot We Never Included in ChatGPT”

At the end of the interview, the profound ethical and power issues faced by AI companies surfaced. Altman candidly admitted that ChatGPT already has billions of daily conversations, “never in history has anyone been able to have billions of conversations in a day”. A small adjustment to the model's personality could have enormous social implications. The concentration of power centered on a single technology, arriving so quickly, is thought-provoking.

He illustrated OpenAI's choices between "winning the race" and "doing what is best for the world." One of OpenAI's proudest points is that many users see ChatGPT as their most trusted and relied-upon technology. To maintain this "consistent with user goals" long-term relationship, they have given up many features that could stimulate short-term growth or revenue.

"For example, we haven't included a sex robot avatar in ChatGPT," Altman joked, "that would clearly increase user engagement." This example vividly illustrates the tension between commercial temptations and long-term values.

He also reflected on past security missteps. For instance, the model once praised users too much, which was annoying for most but could fuel delusions in a few psychologically vulnerable individuals. This reminds them that the greatest risks may not be those they spend the most time guarding against (such as bioweapons), but rather the "unknown unknowns" arising from evolving alongside society.

When asked about his view on the coexistence of “optimists” and “doomsayers” in the AI field, Altman expressed difficulty in understanding the mindset of those who firmly believe AI will destroy humanity but are still fully committed to building it. “If I truly thought that way, I wouldn't build it... I might farm for the rest of my life.” He highlighted that if there is a "99% good probability and a 1% disaster," and efforts are focused on raising the 99% to 99.5%, he could understand that mindset.

Finally, Altman’s advice to everyone was simple and direct: “The first tactical piece of advice is to use these tools.” He was surprised by many questions about how to prepare children for an AI future while they themselves have never moved beyond using ChatGPT as an enhanced version of Google. Personally experiencing and mastering these tools is the most important preparation for dealing with the massive changes to come. At the same time, maintaining resilience and learning to adapt to rapid change is equally vital. The future will not be shaped solely by a few companies but will be written collaboratively by every individual, business, and institution building upon it.

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。