整理: 有新
On June 19, OpenAI officially launched its first podcast episode, where CEO Sam Altman systematically responded to a series of questions about the progress of GPT-5, the Stargate project, the development of next-generation AI terminal devices, the model's memory capability controversies, and the evolution of social structures after AGI arrives.
Altman spoke about the real usage experience of AI in parenting and education from the perspective of 'a new father,' also revealing the core choices OpenAI is facing: how to maintain a balance between technological leaps, privacy boundaries, and trust structures.
'My child will never be smarter than AI, but they will grow up to be much stronger than our generation.' Altman candidly stated in the program that this generation of children will grow up in a world fully permeated by AI; their dependence, understanding, and interaction with intelligent systems will be as natural as the previous generation's familiarity with smartphones. The new roles of models like ChatGPT in family companionship and knowledge enlightenment have opened new paradigms for parenting, education, work, and creativity development.
AI is becoming the growing environment for the next generation.
Altman mentioned that although society has not formed a unified definition, more and more people believe 'we have reached AGI systems every year.' In his view, the public's demand for hardware and software is changing extremely rapidly, while current computing power is far from meeting potential needs.
When the conversation shifted to Altman's new identity as a father, he candidly stated that ChatGPT provided great help in early parenting. 'Although many people could raise their children well without ChatGPT, I am not sure I could do it.' After going through the initial few weeks of 'asking about everything,' he gradually focused on the rhythm of infant development and behavioral habits. He pointed out that such AI tools have begun to play the roles of 'information intermediaries' and 'confidence enhancers' in parenting.
Moreover, Altman is also thinking about the impact of AI on the growth path of the next generation. He bluntly stated, 'My child will never be smarter than AI, but they will grow up to be much stronger than our generation,' emphasizing that this generation of children will naturally grow up in an environment where AI is omnipresent, and their dependence and interaction with AI will be as natural as smartphones were for the past decade.
Altman shared a story that circulated on social media: a father, to avoid repeating the plot of 'Thomas the Tank Engine' to his child, introduced the characters into ChatGPT's voice mode, resulting in the child conversing with it for over an hour. This phenomenon raised deep concerns for Altman: the extension of AI into companionship roles may trigger the alienation of 'social-like relationships,' posing new challenges to social structure. He emphasized that society needs to reset boundaries, but also noted that throughout history, society has always found ways to cope with the shocks of new technology.
In the field of education, Altman observed the positive potential demonstrated by ChatGPT in the classroom. 'Under the guidance of good teachers and good courses, ChatGPT performs very well,' but he also admitted that when students use it alone for assignments, it can easily devolve into 'Google-style copying.' He cited his own experience, pointing out that people worried back then that 'he would only Google things,' but eventually found that children and schools could quickly adapt to the changes brought by new tools.
When asked about the form of ChatGPT five years from now, Altman stated, 'ChatGPT five years from now will be something completely different,' although the name may still be retained, its capabilities, interaction methods, and positioning will undergo fundamental changes.
AGI is dynamically defined, the capability leap of Deep Research
When discussing the industry buzzword 'AGI,' Sam Altman provided a more dynamic explanation. He pointed out, 'If you had asked me or anyone else to define AGI based on the cognitive abilities of the software five years ago, the definitions given then have now been far surpassed.' As the intelligence of models continues to enhance, the standards for AGI are also being raised, presenting a state of 'dynamic shift.'
He emphasized that there are already systems that can significantly enhance human work efficiency and execute economically valuable tasks, and what is truly worth questioning may be: what kind of system can be called 'superintelligent'? In his view, systems that possess the ability for autonomous scientific discovery or can greatly enhance the efficiency of human scientific discovery are closer to this standard. 'This would be extremely wonderful for the world.'
This judgment has also been reflected internally at OpenAI. Andrew Mane recalled that when they tried out GPT-4, it felt like 'a decade of exploratory space has been opened.' Especially the moment the model could perform self-calling and display preliminary reasoning capabilities made people realize the possibilities of a new stage.
Altman agreed and further pointed out: 'I have always believed that the core driving force behind the improvement of human living standards is the speed of scientific progress.' The slow pace of scientific discovery is a fundamental limitation on human development, and the potential of AI in this regard has yet to be fully realized. He admitted he has not mastered the complete path of 'AI automatic scientific research,' but the confidence of the research team in the direction of progress is rapidly increasing. He shared that from GPT-4.0.1 to GPT-4.0.3, a new key idea can be proposed every few weeks, and almost all have been effective. This pace is exciting and confirms the belief that 'breakthroughs will come suddenly.'
Andrew Mane added that OpenAI recently switched the default model to GPT-4.0.3, and the most important update is the introduction of the Operator mode. In his view, many Agentic systems promised a lot in the past but lacked 'anti-fragility' and collapsed at the first sign of abnormality. However, the performance of GPT-4.0.3 is quite different. Altman responded, 'Many people tell me that their feeling of a breakthrough moment in AGI is precisely the Operator mode of GPT-4.0.3.' Although he himself did not have a particularly strong feeling, the feedback from external users is worth noting.
The two further explored the new capabilities brought by 'Deep Research.' Andrew stated that when he used this tool to research Marshall McLuhan, AI could search, filter, and organize materials online, generating complete material packages more efficiently than manual research. He also developed an app that turns questions into audio files to meet the needs of 'limited memory but strong curiosity.'
Altman then shared another extreme usage scenario: a 'learning addict' using Deep Research to generate complete reports on various interest topics, sitting there all day reading, questioning, and iterating, completely immersed in an AI-driven learning loop.
Although Altman claims he has not been able to fully utilize these tools due to time constraints, he is still willing to prioritize reading the generated content from Deep Research in his limited time.
As functionalities continue to strengthen and user scenarios become increasingly diverse, external attention on the next-generation models is also rising sharply. Andrew directly raised the question that users are most concerned about: When will GPT-5 be released? Altman responded, 'Perhaps this summer, but I'm not sure of the exact timing.' He revealed that internally they are facing a repeatedly discussed issue: whether the new version should adopt the previous 'grand announcement' format or, like GPT-4, continue to iterate without changing the name.
He further explained that the current model system structure is much more complex than before; it is no longer a linear process of 'one training, one launch,' but a dynamic system that supports continuous optimization. 'We are now thinking about this issue: if we release GPT-5 and continue to update it, should we call it GPT-5.1, 5.2, 5.3, or keep the name GPT-5?' The differences in user preferences also complicate the decision: some users like snapshots, while others hope for continuous improvement, but the boundaries are difficult to unify.
Andrew pointed out that even those with a technical background can sometimes feel confused when selecting models. For example, whether to use O3, O4 Mini, O4 Mini High, etc., the inconsistency in naming exacerbates the difficulty of choice.
In response, Altman provided background information, stating that this is actually a 'byproduct of paradigm shift.' The current system is somewhat like running two model architectures simultaneously, but this chaotic state is nearing its end. He added that while he does not rule out the possibility of new paradigms emerging again in the future, which could lead to another 'split' in the system, 'I am still quite looking forward to quickly entering the stages of GPT-5 and GPT-6,' at which point users will no longer feel troubled by complex names and model switches.
AI Memory, Personalization, and Privacy Controversies
When discussing the biggest experiential change in ChatGPT recently, Sam Altman said directly: 'The memory function is probably my favorite new feature of ChatGPT recently.' He recalled that when he first used GPT-3, the conversation with the computer was already astonishing, but now the model can provide precise responses based on user context, this feeling of 'knowing who you are' is an unprecedented leap. Altman believes that AI is opening a whole new stage, and as long as users are willing, it will have a deep understanding of users' lives and provide 'extremely helpful answers' based on that.
However, functional evolution has also sparked more complex discussions at the social level. Andrew Mane mentioned the recent lawsuit against OpenAI by The New York Times, demanding the court to force OpenAI to retain ChatGPT user data beyond the compliance period, which has drawn widespread attention. Altman stated: 'We will certainly oppose this request. I hope, and believe we will win.' He criticized the opposing side for claiming to value privacy while making overreach demands, pointing out that this exactly exposes the current institutional void regarding AI and privacy.
In Altman's view, although this lawsuit is regrettable, it also has the positive significance of 'promoting serious discussions in society about AI and privacy.' He emphasized that ChatGPT has become a 'private conversation partner' in the daily lives of many users, which means that the platform must establish more serious institutional safeguards to ensure sensitive information is not misused. He bluntly stated: 'Privacy must become a core principle of AI usage.'
The discussion further extended to data usage and advertising possibilities. Andrew questioned whether OpenAI can access user dialogue data and whether this data will be used for training or commercial purposes. To this, Altman responded that users can indeed choose to turn off the use of training data, and OpenAI has not launched any advertising products. He personally does not completely oppose advertising, 'Some ads are good; for example, I've bought quite a few from ads on Instagram.' However, he emphasized that in products like ChatGPT, 'trust' is an extremely critical cornerstone.
Altman pointed out that social media and search platforms often make people feel 'commodified,' where content seems to exist for advertising clicks, and this structural problem is the root of users' general concerns. If the output content of future AI models is manipulated by advertising bids, it would result in a complete collapse of trust. 'I would hate that myself.'
On the contrary, he prefers to establish a 'clear, transparent, and goal-consistent' business model: that is, users pay for quality services rather than being manipulated by hidden ads. Under controllable premises, he does not rule out exploring models like 'platform commission after clicks' in the future, or displaying some useful ads outside of output content, but the premise is that it never affects the independence and reliability of the model's core output.
Andrew expressed similar concerns, citing Google as an example. He believes the Gemini 1.5 model is excellent, but as an advertising-driven company, Google's underlying motivation makes it difficult to fully trust. 'I have no problem using their API, but when using the chatbot, I always wonder: is it really on my side?'
Altman expressed understanding and admitted he was once a loyal user of Google Search, 'I really liked Google Search.' Although it had a lot of ads, it was once 'the best tool on the internet.' However, structural problems still exist. He praised the Apple model, believing that 'paying for products in exchange for a clean experience' is a healthy logic. He also revealed that Apple attempted the advertising business iAd but did not succeed, perhaps fundamentally not being keen on such business models.
In their view, users also need to maintain their judgment. 'If one day we find that a certain product is suddenly being 'pushed very hard,' then we should ask one more question: what is the motivation behind this?' Andrew said. Altman added that regardless of what business model is adopted in the future, OpenAI must always adhere to the principles of 'extreme honesty, clarity, and transparency' to maintain the trust boundary of users towards the platform.
Stargate, Building an Intelligent Energy Landscape
When the conversation turned to 'the evolution of the relationship between AI and users,' Altman first reviewed the structural errors of the social media era. He pointed out that 'the most fatal problem of social platforms lies in the misaligned goals of recommendation algorithms—they only want you to stay longer, rather than genuinely caring about what you need.' The same risks could also arise in AI. He warned that if models are optimized to 'only cater to user preferences,' they might seem friendly but could weaken the consistency and principles of the system, ultimately being harmful in the long run.
This bias has been evident in DALL·E 3. Andrew observed that early image generation had a significant problem with a single style, and although Altman did not confirm its training mechanism, he also acknowledged that this possibility exists. Both agreed that the new generation of image models has significantly improved in quality and diversity.
The greater challenge comes from the bottleneck of AI computing resources. Altman admitted that the biggest issue currently is 'we do not have enough computing power for everyone to use.' This is precisely why OpenAI has launched Project Stargate. This is a global-level financing and construction project for computing infrastructure, aiming to integrate capital, technology, and operational resources to create an unprecedented scale of computing platform.
The core logic of Stargate is to lay a cost-controlled computational foundation that serves the intelligence of all. He explained that unlike any previous generation of technology, AI will require an extremely large infrastructure to truly cover billions of users. Although there is currently no budget of $500 billion in the OpenAI account, Altman is confident in the implementation of the project and the performance of partners, revealing that the first construction site has already begun, accounting for about 10% of the total investment.
The firsthand experience on-site shocked him: 'Although I know what a gigawatt data center is in my mind, seeing thousands of people building GPU rooms, the complexity of the system exceeded imagination.' He compared it to 'no one can make a pencil alone,' emphasizing the vast industrial mobilization behind Stargate, from mining, manufacturing, logistics to model invocation, all representing the ultimate expression of human thousand-year engineering collaboration.
In the face of external doubts and disturbances, Altman responded directly for the first time to reports about Elon Musk attempting to intervene in the Stargate project. He stated, 'I made a wrong judgment before; I thought Elon would not abuse government influence to engage in unfair competition.' He feels regret for this and emphasized that such behavior not only undermines industry trust but also is not conducive to the overall development of the country. Fortunately, the government was ultimately not influenced by him and stood firm in the legitimate position.
Regarding the current AI competitive landscape, he expressed satisfaction. In the past, everyone generally had anxiety about 'winner takes all,' but now more people realize that this is an ecological co-construction. 'The birth of AI is very similar to the invention of the transistor; although it started in the hands of a few, it will ultimately form the technological foundation of the entire world.' He firmly believes that countless companies will create great applications and businesses based on this foundation; AI is essentially a 'positive-sum game.'
When discussing the energy sources required for computing power, Altman emphasized 'we need them all.' Whether it is natural gas, solar energy, fission nuclear energy, or future fusion technology, OpenAI must mobilize all means to meet the ultra-large-scale operational needs of AI systems. He pointed out that this is gradually breaking down the geographical boundaries of traditional energy, allowing training centers to be located anywhere in the world with resources, while intelligent outcomes can be disseminated at low cost via the internet.
'Traditional energy cannot be globally dispatched, but intelligence can.' In his view, this path of 'transforming energy into intelligence and then outputting it as value' is reshaping the entire human energy landscape.
This extends to the field of scientific research. Andrew cited the example of the James Webb Space Telescope accumulating massive amounts of data, but struggling to process it due to a lack of scientists, resulting in a large number of 'unexplored scientific discoveries.' To this, Altman envisioned whether it would be possible in the future to have an intelligent AI that, without relying on new experiments or new equipment, could deduce new scientific laws solely based on existing data?
He mentioned that he once joked that OpenAI should build a giant particle accelerator, but then thought that perhaps AI could solve high-energy physics problems in completely different ways. 'We have actually accumulated a lot of data, the problem is that we do not yet understand the limits of intelligence itself.'
In the field of drug discovery, such cases of 'missing the known' are more frequent. Andrew mentioned drugs like Orlistat that were discovered in the 1990s but were shelved for decades due to limited perspectives, only to be repurposed today. Altman believes, 'There may be a lot of such forgotten but highly valuable scientific materials; with a little guidance, they could bring about significant breakthroughs.'
Regarding expectations for the next-generation models, Altman expressed strong interest. 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 become key in exploring this capability.'
He further explained the differences between reasoning models and the existing GPT series. 'From the very beginning, we found that as long as you tell the model to 'take it step by step,' the quality of answers significantly improves. This indicates that the model possesses potential reasoning pathways.' The goal of reasoning models is to systematize and structurally enhance this ability, allowing the model to perform 'internal monologues' like humans.
Andrew supplemented the case of Anthropic evaluating model quality through 'thinking time.' Altman also expressed surprise: 'I thought users hated waiting the most. But the fact is—as long as the answer is good enough, everyone is willing to wait.'
In his view, this is precisely the watershed of AI evolution: no longer pursuing speed in mechanical responses, but moving closer to truly understanding, reasoning, and inventing intelligent agents.
Next-generation hardware and individual potential revolution
Regarding OpenAI's hardware plans, Andrew mentioned a collaboration video between Sam Altman and Jony Ive and directly asked: Has the equipment entered the trial phase?
Altman admitted, 'It's still very early.' He stated that OpenAI has set extremely high quality thresholds for this product, and this is not a goal that can be achieved in a short time. 'The computers we are currently using, whether hardware or software, are essentially still designed for a 'non-AI world.'
He pointed out that when AI can understand human context and make reasonable decisions on behalf of humans, the way humans interact with machines will change fundamentally. 'You may want devices to be more sensitive, to perceive the environment, to understand your life background—you may also want them to completely break free from screens and keyboards.' For this reason, they have been exploring new types of device forms and feel very excited in some directions.
Altman envisioned a new interactive paradigm—a truly understanding user and mastering context AI can replace users in meetings, understand content, manage information boundaries, contact relevant parties, and drive decision execution. This will bring humans and devices into a new state of symbiosis. 'If you just say one sentence, it knows who to contact and how to act; your way of using a computer will be completely different.'
From an evolutionary logic perspective, he believes our current interaction with ChatGPT is both 'shaped by device forms' and 'shaping device forms in return.' The two are in a continuous dynamic mutual evolution.
Andrew further pointed out that the popularity of mobile phones largely benefited from their compatibility with 'public use (looking at the screen)' and 'private use (voice calls)' scenarios. Therefore, the challenge for new devices is: how to achieve both 'private and general' in diverse scenarios. To this, Altman expressed agreement. He used listening to music as an example: using speakers at home and headphones on the street, this 'public-private differentiation' naturally exists. But he also emphasized that new device forms still need to pursue stronger generality to become truly vibrant AI terminals.
When asked when this product would be launched, Altman did not give a specific time, only stating, 'It will take some time,' but he believes it will ultimately 'be worth the wait.'
The conversation naturally transitioned to Altman's advice for young people. He stated that the obvious strategic advice is: 'Learn to use AI tools.' In his view, 'the world has quickly shifted from a few years ago of 'you should learn to code' to 'you should learn to use AI.' This may still just be a transitional phase, and he believes there will be new 'key skills' that emerge in the future.
On a more macro level, he emphasized that many abilities traditionally seen as 'talents' or 'characteristics' can actually be trained and acquired. This includes resilience, adaptability, creativity, and even the intuition to recognize others' true needs. 'Although it is not as easy as practicing with ChatGPT, these soft skills can be trained through methods—and they will be extremely valuable in the future world.'
When asked whether similar advice would be given to a 45-year-old, Altman responded clearly: basically the same. Learning to effectively use AI in one's career context is a skill migration challenge that must be faced at any age.
Regarding the organizational structural changes after AGI arrives, Andrew raised a common question: 'OpenAI is already so powerful, why still hire?' He believes some people mistakenly think AGI will directly replace everything. But Altman's answer was concise: 'In the future, we will have more employees, but everyone's work efficiency will far exceed that before the AGI era.'
He added that this is precisely the essential goal of technological progress—not to replace humans, but to greatly enhance individual productivity. Technology is not the endpoint but a ladder leading to higher human potential.