Sam Altman at TED2025: ChatGPT's Evolution, AI Agents, and the Superintelligence Question

When TED's Chris Anderson demonstrated Sora's video generation capabilities—including a simulated TED talk scene—Altman acknowledged the model's contextual understanding. The real breakthrough emerged when GPT-4o generated a conceptual diagram distinguishing intelligence from consciousness, showcasing:

• Integrated reasoning within multimodal systems
• Ability to translate abstract concepts into visual metaphors
• Emergent capabilities beyond pattern replication

Anderson confronted the elephant in the room: AI-generated Charlie Brown content without licensing. Altman outlined OpenAI's current approach:

Altman acknowledged creative professionals' polarized reactions while advocating for systems where artists could opt into style usage with revenue sharing—a complex challenge given AI's training on collective human creativity.

When questioned about open-source rivals like DeepSeek, Altman revealed strategic shifts:

• OpenAI will release "near-frontier" open-source models
• 500 million weekly active ChatGPT users (growing rapidly)
• Compute constraints remain the primary growth bottleneck

Despite competition, Altman emphasized that model intelligence is commoditizing, shifting competition toward product integration and personalized experiences like ChatGPT's evolving "Memory" feature.

The conversation pivoted to AI agents—systems that autonomously execute tasks. Anderson demonstrated "Operator" booking a restaurant, highlighting:

Altman framed safety and capability as converging priorities: "A good product is a safe product." He revealed OpenAI's preparedness framework for evaluating agent risks before deployment.

Altman challenged conventional AGI definitions, arguing current systems lack three critical capabilities:

1. Continuous self-improvement
2. Autonomous task execution ("do my job")
3. Novel scientific discovery

He proposed focusing less on an AGI milestone and more on managing the exponential capability curve: "We have to build a society to get the tremendous benefits of this and figure out how to make it safe."

Addressing safety team departures, Altman pointed to OpenAI's operational record while acknowledging evolving challenges:

• External safety testing for advanced models
• Collective governance mechanisms beyond "elite summits"
• AI-assisted public deliberation on value alignment

He defended iterative deployment as a safety learning tool: "The way we learn to build safe systems is this iterative process of deploying them... while the stakes are relatively low."

Confronted with the "moral authority" question by GPT-4o itself, Altman reflected:

He connected recent fatherhood to intensified future-focused responsibility, envisioning a world where his son perceives current limitations as archaic: "They lived such horrible lives. They were so limited. The world sucked so much. I think that's great."