Sam Altman on Sora, Energy, and Building an AI Empire
Sam Altman discusses OpenAI's strategy, the path to AGI through world models like Sora, the importance of societal co-evolution with AI, and the massive infrastructure and energy requirements for future models. He covers topics from AI safety and regulation to monetization and the future of scientific discovery driven by AI.
OpenAI’s Vision: A Vertically Integrated Approach to AGI
OpenAI’s strategy is built on a vertically integrated stack comprising three core components: a research lab, mega-scale infrastructure, and a consumer technology business. The ultimate mission is to build Artificial General Intelligence (AGI) and make it widely useful, primarily through a personal AI subscription that users can access via first-party applications and third-party services.
Initially against vertical integration, Sam Altman now believes it is essential to achieving the mission. The research enables the creation of powerful models, and the massive infrastructure is required to both conduct the research and serve the products. This tight integration is necessary to deliver on the promise of AGI.
The Role of Sora and Societal Co-evolution
While products like Sora might not seem directly on the path to AGI, they play a crucial role. Altman posits that building great “world models” through video generation will be more important to AGI than many people think.
Furthermore, releasing models like ChatGPT and Sora serves a strategic purpose in the co-evolution of technology and society. By giving the world a tangible sense of what’s coming, OpenAI aims to help society adapt gradually.
“I’m a big believer that society and technology have to co-evolve. It’s you can’t just drop the thing at the end. It doesn’t work that way.”
This approach allows for a continuous dialogue and adjustment period, which is particularly important for video models, as they have a greater “emotional resonance” than text and will soon be ubiquitous.
The Future of AI: Scientific Discovery and Interfaces
The most exciting near-term development for Altman is the emergence of the “AI scientist.” He considers an AI’s ability to make novel scientific discoveries to be his personal equivalent of the Turing test. Within the next couple of years, he expects models to contribute significantly to scientific progress, accelerating discovery in fields like medicine and physics.
“To a first order scientific progress is what makes the world better over time and if we’re about to have a lot more of that that’s a big deal.”
Regarding interfaces, the current chat model is far from saturated in terms of capability. While current models are good at chit-chat, the text interface can go much further in solving complex problems. Future interfaces will evolve beyond text, potentially to real-time rendered video and new hardware devices that are “ambiently aware” of a user’s context.
Reflections on Model Progress and Capabilities
One of the biggest surprises has been the relentless pace of breakthroughs in deep learning. What once felt like a single “giant secret” with scaling laws has turned into a continuous stream of discoveries, such as the recent breakthroughs in reasoning.
There is now an immense “capability overhang,” where the abilities of the most advanced models far exceed what is widely understood or utilized by the public. Altman believes that the current LLM-based technology is powerful enough to create a system that can, in turn, figure out the next major architectural breakthrough.
Strategy: Infrastructure, Partnerships, and Monetization
OpenAI is making an “aggressive infrastructure bet” based on strong confidence in its research roadmap and the future economic value of its models. This requires partnering with a wide array of companies across the entire stack, from energy and chip manufacturing (AMD, Nvidia) to cloud infrastructure (Oracle). When resources are constrained, the company almost always prioritizes allocating GPUs to research over product support, reinforcing its core mission to build AGI.
Monetization models are still evolving based on user behavior. With Sora, for example, the observation that users are creating content far more casually and frequently than anticipated (e.g., memes for group chats) necessitates a different model than originally planned, likely involving a per-generation cost. While open to advertising, Altman is extremely cautious about breaking the high level of trust users have with ChatGPT, stating that recommending a product based on payment rather than quality would “vanish” that trust.
Safety, Regulation, and Copyright
Altman anticipates that, like all powerful technologies, AI will lead to some “really bad stuff.” The right regulatory approach, in his view, is to focus on creating careful safety testing and guardrails for only the most “extremely superhuman capable” models at the frontier. He warns against broad, premature regulation on less capable models, which could stifle innovation and put the US at a strategic disadvantage.
On copyright, his current guess is that training on public data will be considered fair use. However, a new model will emerge for generating content in the style of or with the IP of a creator, where rights holders have control. Interestingly, some rights holders are now more concerned about their characters not being included in AI generations, as it could diminish the value and relevance of their franchise.
Energy: The Fuel for the AI Revolution
Altman’s long-standing interests in AI and energy have converged, as he believes cheaper, abundant energy is the single most impactful driver of quality of life. The massive computational needs of AI will require a significant expansion of energy production. He foresees a future where the two dominant sources will be a combination of solar-plus-storage and nuclear power, including advanced SMRs and fusion. He argues that if nuclear energy becomes “crushingly economically dominant,” the political will to streamline regulation and build it quickly will follow.