OpenAI releases its first official podcast! Sam Altman reveals more details about GPT-5, Stargate, and next-generation AI hardware

OpenAI’s CEO Sam Altman shared more details about GPT-5, the Stargate project, and next-generation AI hardware on his official podcast. He explored the application of AI in parenting, education, and the potential impact of AGI on social structure, and discussed the complex relationship between AI memory, personalization, and privacy. Altman also introduced OpenAI’s hardware plans that could revolutionize human-computer interaction.

On June 19, OpenAI officially released its first podcast episode, where CEO Sam Altman systematically responded to a series of questions for the first time about the pace of GPT-5 advancement, the Stargate project, the development of next-generation AI terminal devices, the controversy over model memory capabilities, and the evolution of social structures after the arrival of AGI.

Altman talked about the real experience of using AI in parenting and education as a “new father”, and also revealed the core decision OpenAI is facing from the perspective of corporate decision-makers: how to maintain a balance between technological transitions, privacy boundaries and trust structures.

“My children will never be smarter than AI, but they will grow up to be much stronger than our generation.” Altman admitted in the show that this generation of children will grow up in a world where AI penetrates in all directions, and their dependence, understanding and interaction ability on intelligent systems will be as natural as the previous generation is used to smartphones. The new role of models such as ChatGPT in family companionship and knowledge enlightenment has opened up a new paradigm for parenting, education, work, and creativity development.

AI is becoming the next generation of growth environment

Altman mentioned that although society has not yet developed a unified definition, “every year more and more people think that we have reached an AGI system.” In his view, public demand for hardware and software is changing extremely rapidly, and current computing power is far from meeting potential demand.

When the conversation turned to Altman’s new fatherhood, he confessed that ChatGPT has been a huge help in the early days of parenting. “While many people were able to take care of their children in the days without ChatGPT, I’m not sure I can do it.” After the first few weeks of “ask everything” phase, he gradually focused on the rhythm of infant development and behavioral habits. He pointed out that such AI tools have begun to take on the role of “information intermediary” and “confidence enabler” in parenting.

Not only that, Altman is also thinking about the impact of AI on the next generation of growth paths. He bluntly said, “My children will never be smarter than AI, but they will grow up much better than our generation,” and emphasized that this generation of children will naturally grow up in an environment where AI is ubiquitous, and the dependence and interaction on AI will be as natural as smartphones have been in the past decade.

Altman shared a story circulating on social media: a father imported a character into ChatGPT’s voice mode to avoid repeating the story of “Thomas’s Train” to his child, and the child talked to him for more than an hour. This phenomenon raises Altman’s deep concern that the extension of AI in companion roles may lead to the alienation of “social-like relationships”, which in turn poses new challenges to social structures. He emphasized the need for society to reset boundaries, but also pointed out that there has always been a way to deal with the shock of new technologies in society.

How can product managers do a good job in B-end digitalization?
All walks of life have taken advantage of the ride-hail of digital transformation and achieved the rapid development of the industry. Since B-end products are products that provide services for enterprises, how should enterprises ride the digital ride?

View details >

In the field of education, Altman observes the positive potential that ChatGPT exhibits in the classroom. “ChatGPT performs very well with good teachers and good courses,” but also admits that when students use it alone for homework, it tends to degenerate into “Google-style copying.” He cited his own experience as an example, pointing out that people were worried that “he would only Google”, but in the end, he found that both children and schools were able to quickly adapt to the changes brought about by the new tools.

When asked about what ChatGPT will look like in five years, Altman said that “ChatGPT in five years will be something completely different,” and although the name may still remain, its capabilities, interactions, and positioning will change fundamentally.

AGI is dynamically defined, and Deep Research is a capability leap

When it comes to the industry buzzword “AGI,” Sam Altman gives a more dynamic explanation. He pointed out, “If you had asked me or someone else to define AGI five years ago based on the cognitive capabilities of the software at the time, the definition given at that time has now been far surpassed.” With the continuous enhancement of model intelligence, the standard of AGI has also been raised, showing a state of “dynamic transition”.

He emphasized that there are systems that can significantly improve human work efficiency and perform tasks with economic value, but what is really worth asking may be: what kind of system can be called “superintelligence”? In his view, systems with independent scientific discovery capabilities or can greatly improve the efficiency of human scientific discovery are close to this standard. “This would be an extremely beautiful thing for the world.”

This judgment has also been mapped within OpenAI. Andrew Mane recalled that when they tried out GPT-4, there was a feeling that “a decade of exploration space was opened.” In particular, the moment when the model can self-call and demonstrate its preliminary reasoning ability, it makes people realize the possibility of a new stage.

Altman agreed, further noting: “I have always believed that the core driving force for improving the quality of human life is the speed of scientific progress. “Slow scientific discovery is a fundamental limiting factor in human development, and AI’s potential at this point has not yet been fully realized. Although he admitted that he has not yet mastered the complete path of “AI automatic scientific research”, the research team’s confidence in the direction is rapidly increasing. He shared that from GPT-4.0.1 to GPT-4.0.3, the ability to come up with a new key idea every few weeks and almost always work, is an exciting pace that confirms the creed that “breakthroughs come suddenly.”

Andrew Mane added that OpenAI recently switched its default model to GPT-4.0.3, and the most significant update was the introduction of Operator mode. In his opinion, many Agentic systems in the past, despite their high promises, were not “vulnerable” enough and collapsed when they encountered anomalies. GPT-4.0.3, on the other hand, performs very differently. Altman responded, “Many people have told me that they feel the breakthrough moment of AGI, which is the Operator mode of GPT-4.0.3. Although he did not feel it particularly strongly, the feedback from external users is worth paying attention to.

The two further discussed the new capabilities brought by “deep research”. Andrew said that when he used the tool to research Marshall McLuhan, the AI was able to find, sift, organize and generate complete data packages online, which was more efficient than manual research. He also developed an app that generates audio files from questions to meet the needs of “limited memory but curiosity”.

Altman then shared another extreme use case: a “learning addict” uses Deep Research to generate complete reports on various topics of interest, sitting there all day reading, asking, iterating, and fully immersing himself in an AI-powered learning loop.

Although Altman claims to be unable to fully use these tools due to time constraints, he is willing to prioritize reading Deep Research’s generated content for a limited time.

As functions continue to be strengthened and user scenarios become increasingly diverse, the attention of the outside world to the next generation model is also rising. Andrew directly asked the question that users are most concerned about: when exactly will GPT-5 be released? Altman responded, “It might be this summer, but I’m not sure exactly when.” He revealed that the company is facing a question that has been repeatedly discussed internally: whether the new version still needs to adopt the previous “fanfare” release form, or continue to iterate under the premise of the same name like GPT-4.

He further explained that today’s model system structure is much more complex than in the past, and it is no longer a linear process of “one training, one launch”, but a dynamic system that supports continuous optimization. “We’re thinking about this question right now: if we continue to update GPT-5 after releasing it, should we call it GPT-5.1, 5.2, 5.3, or keep the name GPT-5?” Differences in user preferences also add complexity to decision-making: some users like snapshots, others want to continue to improve, but the boundaries are difficult to unify.

Andrew notes that even people with a technical background can sometimes get confused when it comes to model selection. For example, whether to use O3, O4 Mini, O4 Mini High, etc., the inconsistency of the names exacerbates the difficulty of selection.

Altman gave a background note, saying that this is actually a “by-product of a paradigm shift”. The current system is somewhat like running two model architectures at the same time, but this chaos is nearing its end. He added that although he does not rule out the possibility of a new paradigm emerging again in the future, which may lead to a “split” of the system again, “I am still looking forward to entering the stage of GPT-5 and GPT-6 as soon as possible”, when users will no longer be troubled by complex naming and model switching.

AI memory, personalization, and privacy controversies

Talking about the biggest experience change in ChatGPT recently, Sam Altman bluntly said: “The memory function is probably my favorite new feature of ChatGPT recently. He recalled that when GPT-3 was used, the conversation with the computer was already amazing, but now the model can give accurate responses based on the user’s background, and this feeling of “knowing who you are” is an unprecedented leap. Altman believes that AI is opening a new phase where it will have a deep understanding of users’ lives and provide “extremely helpful answers” based on it, as long as the user wants to.

However, functional evolution also sparks more complex discussions at the societal level. Andrew Mane mentioned a recent lawsuit filed by the New York Times against OpenAI, asking the court to force OpenAI to retain ChatGPT user data beyond the compliance period, which attracted widespread attention. Altman said: “Of course we would oppose this request. I hope and believe that we will win. He criticized the other party for claiming to value privacy while making over-the-line demands, pointing out that this just exposes the current institutional gaps in AI and privacy.

In Altman’s view, while regrettable, the lawsuit also has positive implications for “promoting serious discussions about AI and privacy in society.” He emphasized that ChatGPT has become a “private conversation partner” in the daily lives of many users, which means that platforms must establish more serious institutional safeguards to ensure that sensitive information is not misused. He bluntly said: “Privacy must be a core principle of AI use. ”

The discussion extends further to data usage and advertising possibilities. Andrew questioned whether OpenAI has access to user conversation data and whether it will be used for training or commercial purposes. In response, Altman said that users do have the option to turn off the use of training data, and OpenAI has not yet launched any advertising products. He is not completely against advertising, “Some ads are good, such as Instagram ads I have bought a lot. But he emphasized that in products like ChatGPT, “trust” is a critical cornerstone.

Altman pointed out that social media and search platforms often feel “commoditized” and that content seems to exist for ad clicks, a structural issue that is at the root of common user concerns. If the output content of AI models in the future is manipulated by advertising bidding, it will be a complete collapse of trust. “I hate it myself.”

Instead, he prefers to build a business model that is “clear, transparent and targeted”: users pay for quality services rather than being manipulated by hidden advertising. Under the premise of control, he does not rule out exploring models such as “click-to-platform commission” in the future, or displaying some practical advertisements in addition to the output content, but the premise is that it will never affect the independence and reliability of the core output of the model.

Andrew expressed similar concerns, citing Google as an example. He thinks the Gemini 1.5 model is excellent, but as an ad-driven company, Google’s underlying motivations make it difficult to fully reassure. “I’m fine with their API, but when using a chatbot, I always think: Is it really on my side?”

Altman understands this and admits that he has also been a loyal user of Google Search, “I really like Google Search. Despite the ads, it was once “the best tool on the Internet”. However, structural problems remain. He appreciates the Apple model, believing that “paying for a product for a cleaning experience” is a healthy logic, and also revealed that Apple has tried the advertising business iAd without success, and may not be enthusiastic about this kind of business model in nature.

In their view, users also need to maintain their judgment. “If one day we find that a product is suddenly ‘pushing hard’, then we have to ask one more question: what is the motivation behind this?” Andrew said. Altman added that regardless of the business model adopted in the future, OpenAI must always adhere to the principles of “extreme honesty, clarity, and transparency” to maintain the boundaries of user trust in the platform.

Stargate, building a smart energy landscape

When the conversation turns to “the evolution of AI-user relationships,” Altman begins by reviewing the structural mistakes of the social media era. He pointed out, “The deadliest problem with social platforms is the misaligned targets of recommendation algorithms – they just want to keep you longer, not really care about what you need. “The same risks can arise in AI. He warned that if the model is optimized to “only cater to user preferences,” it may weaken the consistency and principles of the system while seemingly friendly, which will be harmful in the long run.

This deviation is not the case in DALL· E 3. Andrew observed a clear monotony of style in early image generation, and Altman did not confirm its training mechanism, but acknowledged the possibility. The two agreed that the new generation of image models has significantly improved in quality and diversity.

The bigger challenge comes from the bottleneck of AI computing resources. Altman admits that the biggest problem at the moment is that “we don’t have enough computing power for everyone to use.” That’s why OpenAI launched Project Stargate. This is a global computing infrastructure financing and construction project with the goal of integrating capital, technology and operational resources to create an unprecedented scale computing platform.

“The core logic of Stargate is to lay a cost-controllable computing power base for intelligent services for the whole people.” He explained that unlike any previous generation of technology, AI will have extremely huge infrastructure requirements to truly reach billions of users. Although OpenAI does not currently have a budget of $500 billion in its account, Altman is confident in the implementation of the project and the performance of the partner, revealing that construction of its first construction site has begun, accounting for about 10% of the total investment.

He was shocked by his personal experience at the scene: “Although I know what a gigawatt-level data center is in my head, I actually see thousands of people building GPU rooms, and the complexity of the system is beyond imagination. He compared “no one person can make a pencil alone”, emphasizing the wide range of industrial mobilization behind Stargate, from mining, manufacturing, logistics to model calling, which is the ultimate embodiment of human millennium engineering cooperation.

In the face of doubts and interference from the outside world, Altman responded positively to reports of Elon Musk’s attempt to intervene in the Stargate project for the first time. He said, “I was wrong before, I thought Elon would not abuse government influence to engage in unfair competition. He regretted this and stressed that such behavior not only undermines industry trust, but is also not conducive to the overall development of the country. Fortunately, in the end, the government was not affected by it and stood firm.

He expressed his satisfaction with the current AI competition landscape. In the past, there was a general anxiety of “winner takes all”, but now more people realize that this is an ecological co-construction. “The birth of AI is much like the invention of the transistor, which will eventually form the technological foundation of the entire world, although it will start out in the hands of only a few people.” He firmly believes that countless enterprises will create great applications and businesses based on this foundation, and AI is essentially a “positive sum game”.

When it comes to the energy sources required for computing power, Altman emphasizes “all of them”. Whether it’s natural gas, solar, fission nuclear energy, or future fusion technology, OpenAI must mobilize all means to meet the needs of the hyperscale operation of AI systems. He pointed out that this is gradually breaking the geographical boundaries of traditional energy, training centers can be located anywhere in the world with resources, and intelligent results can be disseminated at low cost through the Internet.

“Traditional energy cannot be dispatched globally, but intelligence can.” In his view, this path of “transforming energy into intelligence and then output into value” is reshaping the entire human energy map.

This also extends to the field of scientific research. Andrew cited the James Webb space telescope, for example, that the vast amount of data accumulated but was difficult to process due to a lack of scientists, resulting in a large number of “untapped scientific discoveries.” In this regard, Altman envisions whether there is a smart enough AI in the future to deduce new scientific laws based only on existing data without relying on new experiments or new equipment.

He mentioned that he once joked that OpenAI should build its own giant particle accelerator, but on second thought, maybe AI can solve high-energy physics problems in a completely different way. “We have already accumulated a lot of data, and the problem is that we don’t yet understand where the limits of intelligence itself lie.”

In the field of drug discovery, such cases of “missing the known” are more frequent. Andrew mentions that drugs like orlistat were discovered in the 90s, but were shelved for decades due to perspective limitations and have not been repurposed until today. Altman believes that “there may be many of these forgotten but valuable scientific materials, and a little guidance can lead to huge breakthroughs.” ”

Altman expressed keen interest in anticipation of the next generation of models. He mentioned that Sora can understand classical physics, but whether it can advance deeper theoretical science remains to be verified. “The ‘reasoning model’ we are developing is expected to be the key to exploring this capability.”

He further explained the differences between the inference model and the existing GPT series. “From the very beginning, we found that as long as you tell the model to ‘take it one step at a time’, the quality of the answers will improve significantly. This indicates that the model has a potential inference path. The goal of the reasoning model is to systematically and structurefully enhance this ability, so that the model can perform “internal monologues” like humans.

Andrew added to Anthropic’s case of assessing model quality through “thinking time.” Altman also expressed surprise: “I thought the most annoying thing for users was waiting. But the fact is that as long as the answer is good enough, everyone is willing to wait. ”

In his view, this is the watershed moment in the evolution of AI: no longer a mechanical response in pursuit of speed, but a move closer to an agent that truly understands, reasons, and invents.

The next generation of hardware and individual potential revolution

Regarding OpenAI’s hardware plans, Andrew mentioned Sam Altman’s collaboration video with Jony Ive and directly asked if the device has entered the trial phase.

Altman admitted, “It’s still early.” He said that OpenAI has set a very high quality threshold for this product, and this is not a goal that can be achieved in a short period of time. “The computers we use now, whether hardware or software, are still inherently designed for an ‘AI-free world.'”

He pointed out that when AI can understand human context and make rational decisions on behalf of humans, the way humans interact will be completely changed. “You may want the device to be more sensitive, aware of the environment, understand the context of your life – and you may want it to be completely free from screens and keyboards.” That’s why they’re always exploring new device formats and are very excited about some of them.

Altman envisions a new paradigm of interaction—an AI that truly understands the user, grasps the context, and can participate in meetings, understand content, manage information boundaries, connect with stakeholders, and drive decision-making on their behalf. This will bring the relationship between man and device into a new state of symbiosis. “If you just say one sentence, it knows who to contact and how to act, and the way you use your computer will be completely different.”

From the perspective of evolutionary logic, he believes that our current interaction with ChatGPT is both “shaped by the device form” and “shaped by the device in turn.” The two are in a continuous dynamic co-evolution.

Andrew further noted that the popularity of mobile phones is largely due to their compatibility with “public use (screen viewing)” and “private use (voice calling)” scenarios. Therefore, the challenge of the new device is also how to achieve “both private and universal” in diverse scenarios. Altman agrees. He used listening to music as an example: using stereos at home and headphones on the street, this “public-private differentiation” is a natural occurrence. However, he also emphasized that new device forms still need to pursue stronger versatility in order to become truly viable AI terminals.

When asked when he would see the product hit the market, Altman did not give a specific time, only saying that “it will be a while longer,” but he believes that in the end “it will be worth the wait.”

The conversation naturally transitioned to Altman’s advice to young people. The obvious strategic advice, he said, is: “Learn to use AI tools.” In his view, “the world has quickly switched from ‘you should learn to code’ a few years ago to ‘you should learn to use AI’.” And this may still be just a phased transition, and he believes that new “key skills” will emerge in the future.

At a broader level, he emphasized that many abilities traditionally considered “talent” or “character” can actually be trained and learned. This includes resilience, adaptability, creativity, and even the intuition to recognize the real needs of others. “While not as easy as practicing with ChatGPT, these soft abilities can be trained in ways — and they will be extremely valuable in the future world.”

When asked if he would give similar advice to the 45-year-old, Altman responded clearly: basically the same. Learning to use AI well in your own career scenarios is a skill transfer challenge that must be addressed at any age.

Regarding the organizational structure changes after the arrival of AGI, Andrew asked a common question: “OpenAI is already so powerful, why is it hiring?” He argues that some people mistakenly believe that AGI will be a direct replacement for everything. But Altman’s answer is succinct: “We’ll have more employees in the future, but everyone will be much more productive than they were before the AGI era.” ”

He added that this is the essential goal of technological progress – not to replace humans, but to greatly enhance individual productivity. Technology is not the end, but a ladder to higher human potential. Can the return only depend on OpenAI? Silicon Valley 20-Year Dollar Fund founder warns VC model is on the verge of failure.

End of text
 0