Semafor1h 1m
Former Twitter CEO Parag Agarwal on X, Elon, & Building A New Internet For AI Agents | Mixed Signals
Parag Agarwal, Elon, Ben Smith, Max Tani
Interview
- AI agents are projected to access the web 1,000 times more than humans and perform 15 to 50 times more searches during complex tasks, driven by their ability to query deeper, non-human information without time constraints, necessitating multiple generations of technology beyond current systems.
- Parallel Search is expected to optimize agent workflows by reducing token consumption and latency, enabling overnight processing that exceeds the 500-millisecond limits of human-oriented interfaces, with the business targeting 5x to 10x year-on-year growth to reach substantial payment volumes within two years.
- Traditional media revenue models face existential risks as current subscription and ad structures rely on human traffic and attention, which will vanish as agents do not possess wallets or require clicks, potentially leading to a "death of the content ecosystem" if value capture drops 5x to 10x annually despite renewal deal values rising 10x to 50x.
- Media companies may resort to blocking agents to preserve value capture, though this is predicted to be a value-destructive move where blocked sources fail to influence agent decisions, leaving only open sources to shape outcomes like trip planning.
- New pricing mechanisms may emerge to allow differential pricing based on data value, such as charging high-willingness-to-pay users 10,000 times more for consequential decisions, while subscription models are deemed inefficient compared to ad engines for scaling differentials without friction.
- Social platforms, particularly those offering human-to-human connection and experiential content, are expected to survive the agent era, with specific predictions that Twitter (X) will remain durable and influential on influencers despite having a materially smaller market share than short-form video or text apps.
- A failure to establish proper incentive structures for data provision could result in information becoming inaccessible to general AI, remaining available only to large labs and big players, thereby stalling the business and preventing transactions required to build information systems.
- Significant gaps exist between current revenue models and the future agent economy, characterized by anxieties similar to the first digital media bubble, with fears that scrolling and traditional reading habits may be viewed as obsolete retro activities within five to ten years.
- The competitive landscape for model pricing is volatile, with frontier-grade models currently costing 50x to 100x more than models like Luna or Haiku, suggesting that cost efficiencies and model commoditization will drive future agent adoption and economic viability.
- The core survival strategy relies on creating a sustainable open web ecosystem where high-quality data is rewarded proportionally to the GPU resources used, ensuring that the "pie" grows enough for all participants to thrive rather than leaving content owners feeling their information is being stolen.