Universally Accessible Intelligence with Character.ai's Noam Shazeer

Strategic Vision & Product Choice

• Character.ai aims to operate as both an AGI-focused and product-first company.
• The team selected dialogue/chatbots as the primary vehicle to force work on generalizable technologies, rather than narrowing into specific high-stakes domains like healthcare or education.
• Noam Shazeer (co-founder) cites a desire to avoid the brand risks and slow launch cycles associated with large corporations like Google, where the AI project was internally renamed "Lambda."
• The company is pursuing a vertically integrated "full-stack" approach, allowing co-design across the model, infrastructure, and application layers to accelerate innovation.

Company Metrics & User Behavior

• Since launch, the platform has processed over 20 billion human messages.
• Daily Active Users (DAUs) average two hours of engagement per day on the platform.
• Shazeer posits that the "parasocial relationship" and entertainment industry ($2 trillion annually) are driven by unmet needs for imaginary friends, a gap Character.ai fills by allowing characters to be conversational and "make stuff up" without the constraints of factual accuracy required in mission-critical applications.

Technology, Scaling, and Roadmap

• Shazeer states he has not observed any limits to current AI scaling laws, noting that model intelligence continues to improve as compute increases.
• The current model cost approximately $2 million in compute cycles to train last year; this cost is projected to drop to ~$500,000 following optimizations.
• The company plans to launch a model "tens of IQ points smarter" by the end of the year.
• Future improvements focus on multimodal capabilities (voice, face), multi-person interaction contexts, and significantly larger context windows to enable long-term memory of user interactions.
• Shazeer believes "Theory of Mind" in AI is an emergent property of scale that will improve naturally as models grow larger.

Market Outlook & Accessibility

• Character.ai views the current moment as the "Wright Brothers first airplane" phase of AGI: useful for some cases now, with massive future scaling without requiring fundamental architectural breakthroughs.
• Shazeer predicts that by the next year, NVIDIA will supply ~2 million H100s, enabling roughly 250 trillion operations per second globally, making intelligent interaction accessible to everyone.
• The goal is "universally accessible intelligence," where the cost of computation drops to $10^-18 per operation, making interaction with advanced models far cheaper than the value of human time.
• Open source advancements are welcomed as drivers of innovation, but Character.ai maintains its vertical integration strategy to leverage proprietary data loops and full-stack control.

Risk Assessment

• When asked about the potential for AGI to destroy the world, Shazeer answered "no, not yet."
• He suggested a global pause of approximately four months to ensure sufficient H100 infrastructure is online to train the next model, though he expressed strong belief in the immediate potential benefits of the technology.