Frontier Research | Perplexity

Open original

Perplexity is recruiting for Frontier Research, an initiative to build software that runs the full research loop (asking questions, gathering evidence, forming hypotheses, testing them) and make it available to anyone. The work spans multi-agent orchestration, pre-training, continual learning, reinforcement learning, systems, and new evaluations that measure discovery rather than recall.

  • The core bet is that automating scientific research is the next big advance, with agents that search widely, go deep, propose experiments, and critique each other.
  • Perplexity says it will back the effort with more than 10,000 GPUs and publish papers, models, data, and benchmarks, with results shipping into a product used by millions.
  • It argues current evals reward proxy skills rather than general ability, so it plans new benchmarks that test qualitative understanding and discovery.

Perplexity is recruiting for a group called Frontier Research, and the pitch is ambitious: automate scientific research itself. The authors frame research as a loop: pick a question, survey what's known, guess at an explanation, run an experiment, then revise. Their bet is that this loop can run inside software for anyone, not just inside a few institutions with a handful of people. The envisioned system would have specialized agents, some searching thousands of sources, some going deep on one thread, others proposing experiments or critiquing hypotheses, all coordinating as one. They say the work will be published as papers, models, data, and benchmarks, and they point to more than 10,000 GPUs supporting it. Results will also feed a product used by millions, which they call the hardest test of all. The application deadline is December 5. The technical agenda is split across six areas: multi-agent orchestration, pre-training, continual learning, reinforcement learning, systems, and frontier evaluations. Some of the questions are concrete. Can a small on-device model call a larger one only when needed, to save cost? Can we build synthetic training data that beats real data? Can a model keep learning after release without forgetting what matters? The reinforcement learning section is candid about a hard problem: when a long chain of actions ends in one sparse reward, which steps deserve credit, and can simulated environments be faithful enough to transfer to the live web? The post also mentions work on serving very large models with weight transfers in about a second, and on evaluations meant to measure discovery rather than recall. Most of the article is a reading list of Perplexity blog posts and arXiv papers, which tells you more about where the team's thinking already sits than the vision statement does. Topics include agentic search, embedding models, memory as a knowledge wiki, and forgetting in language models. For someone deciding whether this is worth tracking, the useful signal is the research trail rather than the recruiting copy. The stated goals are broad, and whether an automated scientist can actually produce insight at that scale remains an open bet, one the post itself admits.

Ask about this article