At OpenAI’s 2024 DevDay, CEO Sam Altman joined Chief Product Officer Kevin Weil for a fireside chat that offered an illuminating glimpse into where AI is heading. The conversation—rich with insights on AGI, alignment, product development, and user experience—highlighted OpenAI’s ongoing dedication to pushing the boundaries of AI research while considering the real-world implications of deploying increasingly capable models.
Reframing AGI: An Evolution in Perspective
A central theme was the evolving definition of Artificial General Intelligence (AGI). Altman discussed how OpenAI’s understanding has shifted over time. Initially, the team viewed AGI as a distant, binary milestone—one day, no AGI, and the next, a sudden breakthrough. But reality has proven to be a gradual, iterative process, made murkier by the astonishing leaps in model capability. Altman introduced a conceptual framework of "levels" rather than a single AGI threshold:
According to Altman, with its new model “o1,” OpenAI believes it has reached Level 2—systems capable of surprisingly sophisticated reasoning. While this does not feel like “true AGI,” it represents a huge step forward. Altman expects Level 3—agentic AI that can take autonomous action—to arrive “in the not-distant future,” and anticipates a rapid escalation in capability over the next two years.
A Smooth Gradient, Not a Sudden Jump
Far from a sci-fi scenario where humanity awakens to AGI overnight, Altman predicts a smooth progression. As capabilities progress exponentially, there may never be a universally recognized “AGI birthday.” Instead, future historians might view the quest for AGI as an incremental curve, with today’s breakthroughs blending seamlessly into tomorrow’s.
Commitment to Research and Iterative Deployment
Both Altman and Weil emphasized that research remains at OpenAI’s core. While there have been eras when scaling compute delivered the biggest returns, recent breakthroughs—like o1—demonstrate that the next leaps require profound research. OpenAI’s greatest strength, they argued, is its culture of true research: the willingness to push beyond known paradigms, to search for new frontiers. Research informs product, and product pressures research to find new capabilities that can be applied directly to solve user problems.
This iterative loop is augmented by OpenAI’s principle of “iterative deployment.” By releasing models and features gradually, the company can learn from real-world usage and feedback at scale. Users, developers, and the broader ecosystem help reveal blind spots and highlight opportunities. This approach ensures safety and alignment remain robust as capabilities advance.
Alignment and Safety: Reality Over Theory
The conversation addressed concerns that OpenAI may be paying only “lip service” to alignment. Altman pushed back, describing how rapidly evolving model behavior forces the team to keep alignment grounded in reality. The challenges faced two years ago are nothing like those encountered today—and by applying iterative deployment and learning from actual use, OpenAI can integrate safer practices as capabilities grow.
The team’s core strategy is to continue to build systems that are both more capable and more aligned. With increasing sophistication comes new alignment techniques and better safeguards. This iterative approach, Altman contends, works better than static theoretical frameworks. The company also balances short-term alignment tasks with long-term “sci-fi” concerns—both immediate safety issues and future existential risks warrant attention.
Agents: The Next Major Transformation
Both Altman and Weil agreed that truly agentic AI will be transformative. While chat-based interfaces have driven the current wave of AI adoption, the future lies in agents that can reason deeply, take extended actions, and interact with external tools and environments. They envision systems that can accomplish in hours what once took humans months, or even do thousands of such tasks in parallel. This shift will dramatically alter workflows and productivity.
However, they cautioned that agentic AI needs to meet a high bar for reliability and trust. Before agents have the freedom to click around your computer or manage your tasks autonomously, alignment and safety must be ironclad.
Building for Tomorrow’s Frontier
Weil emphasized a unique challenge for startups building on AI platforms: the technology is evolving so fast that by the time one integrates current capabilities, the frontier has already moved. Companies need to aim for the near future, building for the capabilities that models will soon have, not just what they can do today. The reward for this foresight is significant: as models improve, products that anticipated these shifts can suddenly become far more powerful.
Voice, Manipulation, and the Human Connection
One of the newly introduced capabilities, Advanced Voice Mode, reveals how the medium of interaction can affect user perception. The more lifelike the voice, the harder it is for humans to remember it’s only an AI. This new interface taps into deep human circuitry evolved for person-to-person communication. While beneficial, it also raises subtle alignment questions around trust, manipulation, and emotional influence. For OpenAI, ensuring responsible use of voice and other advanced modalities is integral to building safe, beneficial AI.
The Road Ahead: Faster, Smarter, More Capable
The fireside chat ended on a forward-looking note. Function calling, system prompts, and structured outputs will be coming to o1 by year’s end. Beyond that, both Altman and Weil see a world where models grow rapidly more capable, context windows expand from thousands to possibly millions or even trillions of tokens, and truly universal interfaces—where you simply speak your intent, and the AI executes—become normal.
Conclusion
OpenAI’s 2024 DevDay fireside chat painted a vivid picture of AI’s trajectory. Far from having all the answers, Altman and Weil acknowledged uncertainty as part of the journey. Research-driven breakthroughs, guided by iterative deployment and widespread developer input, will shape the next wave of AI. As these systems become more agentic, more capable, and more natural in their interactions, they promise to redefine not just human-computer interaction, but also human potential itself.
