Interview
AI Agents Could Drive 10,000x More Web Activity | Former Twitter CEO
- Web agent usage is projected to grow from current levels of 10–20% of human volume to a thousand times greater, with the trajectory suggesting at least 5,000x to 10,000x growth remaining and consumer agent expansion potentially outpacing prior benchmarks like ChatGPT.
- Agents are expected to evolve from reactive to hyper-proactive systems by learning implicit and explicit user needs, while models are predicted to become simultaneously larger, more capable, longer-running, and significantly cheaper, disrupting all computer-based work categories.
- A shift from a "pull" to a "push" web model is anticipated where agents provide high-precision alerts based on deep user understanding, driven by the speaker's invariant belief that one to two units of web search spend will occur for every 10 GPUs spent on model inference.
- Parallel Web Systems plans to grow alongside inference with a short-term benchmark of 10x growth, aiming to create an auto-differential pricing system that compensates content owners and keeps the web open for agent access.
- The web is forecast to transition from human-centric navigation to agent-centric consumption, necessitating a re-architecture of infrastructure and business models to replace attention-based advertising, while voice agents driving 150–200 millisecond latency expectations will require improvements across all underlying infra components.
- Security risks are expected to escalate with cyber events occurring every few weeks, bad actors leveraging margins where humans cannot distinguish bots from agents, and the critical need for defense in many layers of the tech stack to prevent adversarial attacks.
- Business viability for agents depends on solving business model alignment to prevent widespread blocking and enabling agents to inherit human privileges, with web search identified as a potential primary bottleneck for application capability despite strong underlying models.
- Market inefficiencies in the GPU sector will necessitate urgent acquisition strategies to secure the right scale, location, and timing for hardware, although the overall scaling of models and inference costs remains consistent with linear predictions made three years ago.