AI agents search 1000x more than humans and have different needs (full sentence queries, variable latency, non-link outputs). This massive scale and requirement shift means search tech built for humans is inadequate, creating an entirely new market.
The amount of compute spent on web search for an AI agent should be proportional to the cost of the LLM it's feeding. Expensive models warrant more pre-processing on the search side to optimize their costly context windows, while cheaper models do not.
As AI agents become the primary consumers of web content for tasks like shopping or research, the traditional advertising model collapses because agents don't see or click ads. This necessitates a new monetization system to compensate content creators.
The AI market won't converge on one model size. Frontier models will keep growing to tackle the hardest problems. Simultaneously, every six months, much smaller models will achieve the performance of today's best, creating a wide range of useful model sizes.
The most valuable way to monetize data in the AI era won't be selling it for training. Instead, a new market will emerge for AI agents to programmatically access and pay for unique data from providers (like Pitchbook) at the moment of inference.
Instead of agents constantly polling (pulling) the web for updates, search infrastructure will evolve to a 'push' model. It will monitor for changes and notify persistent, always-on agents when pre-defined conditions are met, saving vast amounts of compute.
Former Twitter CEO Parag Agrawal says Elon Musk's core skill is compressing time with unreasonable expectations. This forces people to overcome their tendency to 'sandbag' and discover they are capable of much more than they previously thought.
Current web search API prices are too high for the coming wave of AI agents using cheap models. Spending 80-90% of a task's cost on search for a cheap model is 'entirely silly.' The market must race to the bottom on price to enable the 1000x scale increase.
To ensure the US has a leading open-source AI model, simply having one isn't enough. Parag Agrawal argues you need at least two strong domestic players competing against each other to build the best American open model, fostering innovation and preventing complacency.