What Still Stops Perplexity From Making Scientists?

Aravind Srinivas wants to push past web lookups into longer research tasks. Some questions have no known answers, and he would let AIs help find them. He says some of it may involve prediction models, experiments in the world, real-time hypothesis testing, and refining a hypothesis. If Perplexity can help a person test ideas in the real world, everyone can turn into a scientist. He calls real-world testing the only bottleneck for a scientist, and he calls the result the dream.

Molly O'Shea asked about the biggest challenge to date. Aravind Srinivas said capabilities. The models are not truly long horizon, persistent, and parallel. No frontier model can do its own research end to end, autonomously. Math and code can be hill climbed because correctness can be verified. Most things people are curious about have no clear way to verify the truth. Answers take real-world deployments and rollouts, and models cannot yet handle the slow cycle of feedback. The interface is not clear either. A day before, he and a colleague imagined dropping 10 research interns onto a team, really good and working 24-7, then asking what the team would do with them. He had not thought about it. The start is Perplexity's own internal research teams, built so intuitive that a person does not need a PhD to use it. Anybody should feel like a PhD.