Written by: Techub News Compilation
Recently, OpenAI CEO Sam Altman held an internal communication meeting with a group of developers. During the nearly hour-long conversation, Altman answered dozens of questions from the audience and Twitter, covering topics ranging from the technical characteristics of GPT-5, the future of AI agents, to the profound impact of AI on the economy, security, education, and even human creativity. This dialogue rarely showcased OpenAI's direct reflections on current technological bottlenecks, future roadmaps, and social responsibilities, providing significant reference value for understanding the next stage of development in the AI industry.
A Paradigm Shift in Software Engineering: From "Writing Code" to "Commanding Computers"
At the start of the meeting, a question from Twitter set the tone: As AI makes writing code faster and cheaper, will the demand for software engineers decrease? Sam Altman's response pointed to a fundamental paradigm shift.
He believes that the definition of software engineering itself will undergo a "super change." In the future, there will be "far more than now" who can create and capture value by commanding computers. The nature of work will change dramatically—time spent typing and debugging code will be significantly reduced, while the core tasks will shift to "making computers do what people want" and "creating useful experiences for others."
Altman offered a historical analogy: every leap in efficiency in the engineering field has eventually attracted more people to participate and resulted in more software creation, while the world's demand for software "seems to have shown no signs of slowing down." He predicts a future where much software will be written for an individual or a very small group, with people constantly customizing their own software. This means that if commanding computers in new ways is viewed as software engineering, its scale will become larger and its share of global GDP will be higher.
A developer present immediately raised a practical challenge: using AI tools (like Cursor, Codex) to build products is no longer difficult, but the real bottleneck lies in "going to market" (GTM)—how to get people to discover and use what you have built. Altman acknowledged that this has always been the hardest part of entrepreneurship, and now it feels even more pronounced due to the ease of building. AI hasn't simplified other aspects of entrepreneurship, and the condition of human attention as a scarce resource hasn't changed. He painted a picture of a "highly abundant" future, where human attention may become the last scarce commodity, which means that GTM competition will remain fierce, requiring creative ideas and excellent products.
The Approach, Cost, and Evolution of AI Agents in GPT-5
Concerning the development of the model itself, developers' questions focused on specialization and generalization. Some pointed out that while GPT-4.5 performs excellently in writing, GPT-5, as an agent model, is stronger in tool use and reasoning but appears "clumsy and hard to read" in writing. Altman candidly stated, "I think we just messed up." OpenAI focused its limited bandwidth on enhancing GPT-5's intelligence, reasoning, and coding abilities, temporarily neglecting the writing dimension. However, he emphasized that the future development direction of models remains to be excellent general models because even models focused on coding need good writing abilities to generate complete applications and communicate clearly with users. He promised that future versions of GPT-5.x will catch up quickly in writing and other dimensions.
Cost and speed are another core concern for developers. When asked how to achieve “intelligence at an unmeasurable cost” and a significant drop in costs, Altman provided a specific expectation: by the end of 2027, OpenAI should be able to offer GPT-5 level high intelligence at a cost of "at least 100 times cheaper." However, he also pointed out a new dimension that is becoming prominent: speed. As model outputs become exceptionally complex, more people are starting to focus on output speed over cost. OpenAI needs to think about how to balance cost and speed, potentially even opting for a higher price in exchange for a fraction of a percent of generation time, which is crucial for many agent application scenarios.
Regarding the “durability” of AI agents—meaning the ability to autonomously run long workflows without frequent human intervention—Altman and members of the OpenAI team believe this depends on the type of task. For some clearly defined tasks, customized tools (like SDK encapsulations) can already allow agents to run continuously. The real challenge lies in open-ended questions, such as "having an agent create a startup." The solution will involve breaking down large problems into smaller tasks, allowing agents to self-validate and gradually expand their scope of work.
In response to developers' concerns about OpenAI's "platform swallowing" issue, Altman reiterated that the basic rules of business have not changed. Being able to quickly create software does not mean that other rules of entrepreneurship are invalidated. He advised developers to build businesses that "would be delighted by GPT-6 making astonishing progress," rather than relying on the current flaws of the model as "marginal patches."
Security, Biological Risks, and the Shift in Thinking from "Lockdown" to "Resilience"
The dialogue inevitably touched on the most concerning areas of AI: security, particularly biological safety. A founder of a biosafety startup asked about the potential dangers brought by AI-enabled biological design and the position of safety in OpenAI's roadmap.
Altman confessed that biological risks are one of OpenAI's most pressing concerns for 2026. The current strategy mainly involves limiting model access and using classifiers to prevent the generation of new pathogens, but he believes "this won't last long." The world needs to shift from "lockdown" thinking to "resilience" thinking. He quoted co-founder Ilya Sutskever's analogy: fire has brought benefits to society but has also burned down cities; humans ultimately did not eliminate fire but developed fire regulations, flame-retardant materials, and other resilience measures. Similarly, for risks brought by AI, such as bioterrorism and cybersecurity, society needs to work towards building resilient infrastructures, and AI itself is also part of the solution.
He warned that if there is a significant AI accident this year, the biological field "is a reasonable guess." Over time, other risks will emerge. This will not be entirely about technical solutions, and the world needs to think about these issues in ways that differ from the past.
Another underestimated risk pattern comes from the large-scale deployment of AI agents. Altman shared his own experience: he initially believed he would never give Codex unsupervised access to his computer, but the high convenience and seemingly reasonable behavior quickly changed his mind. He worries that as model capabilities steeply rise and become increasingly difficult to understand, people may "sleepwalk" into risky situations without establishing adequate safety infrastructure due to strong convenience. The lack of this "big picture safety infrastructure" is an area he believes needs to be highly prioritized and invested in.
Human-Machine Collaboration, Education, and Creativity: The Human Role in the AI Era
When AI can learn quickly and provide answers, what is the value of human collaboration and education? A student posed this fundamental question.
Altman recalled similar concerns from teachers when Google emerged and firmly believes that the current educational approach needs to change rather than resist tools. Learning to think and write remains important, but the methods of teaching and assessment must adapt to the new era. He predicts that in an AI-filled world, interpersonal connections will become more valuable, and people will cherish collaboration with others more. AI has the potential to empower group collaboration in unprecedented ways, for instance, participating as a team member in collective brainstorming or problem-solving, thereby enhancing overall productivity.
Regarding education, especially for young children, Altman expressed caution. He believes that the use of computers and AI should be minimized at the kindergarten stage, allowing children to engage more in physical interactions and outdoor activities. He is concerned about the negative impact of technology (especially social media) on teenagers and believes that the impact on young children could be even more severe but is insufficiently discussed. Without fully understanding the implications, young children should not be allowed extensive use of AI.
As for the value of university education, as a former dropout, Altman believes that in this "special period" of rapid AI development, university may not be the best use of time for "highly motivated" individuals aspiring to contribute to AI. He suggests informing parents that they can return to campus anytime in the future, but the present is a time to seize opportunities.
Creativity is another focus. Altman points out that in the field of image generation, it has been observed that consumers' appreciation and satisfaction "dropped sharply" upon learning that the work was generated by AI. People deeply care about the individuals behind the work and their stories, showing little interest in machine-generated creations. He believes this will be a profound trend for the coming decades. Art completely generated by AI (in a broad sense) may not be what the public ultimately desires, while processes guided, edited, or curated by humans will be crucial. The human factor and life stories will continue to imbue creations with value.
Personalization, Memory, and Future Skills
Developers have high expectations for the memory and personalization features of ChatGPT, hoping for more refined controls (such as distinguishing between work and personal identity memories). Altman stated that OpenAI will "super effort" to advance memory and personalization due to the significant utility it brings. His own attitude has also shifted: he is now ready to allow ChatGPT access to his entire computer and network to understand everything for the immense value a perfect digital life representative can bring. The key is that AI needs to profoundly understand the complex rules, interactions, and hierarchies in users' lives, knowing when to use which information. He believes this is what most users ultimately desire.
Finally, when asked about the most important skills in the AI era, Altman's answers were entirely "soft skills": becoming highly proactive individuals, being good at generating ideas, being resilient, and being able to adapt quickly to the rapidly changing world. He believes these skills are more important than any specific technology and can all be learned.
In his closing remarks, Altman envisioned the capabilities of future models: 100 times smarter than current models, 100 times longer context length, 100 times faster, 100 times cheaper, possessing perfect tool invocation abilities and extreme long-term consistency. He once again urged developers: "Tell us what you want us to build." This may indeed be key for OpenAI to maintain a competitive edge by keeping a close dialogue with the builder community and collaboratively shaping the future of AI.
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。
