Perplexity has opened applications for Frontier Research, a new initiative that aims to build an automated scientist, with a priority deadline of December 5. The company says the effort will be supported by more than 10,000 GPUs, which puts real compute behind what is otherwise a hiring page.
The pitch is about the research loop itself. Perplexity describes science as a cycle of asking questions, collecting what is known, spotting gaps, forming hypotheses, running experiments, and updating. Its argument is that this cycle has lived inside a handful of institutions and moved slowly because so few people were involved. The company wants it running in software, available to anyone.
That framing separates the project from a better chatbot. Perplexity says it cares about interpretable insight, the kind of theory that explains why something holds, over black box predictions. It cites general relativity as the example: GPS is a useful application, but the lasting value is the understanding behind it. Whether a language-model system can produce that kind of understanding is the open bet.
The plan is a multi-agent system. Some agents would search broadly across thousands of sources, some would dig into a single thread, and others would propose experiments or critique the results. Perplexity says the system should know what it does not know and remember what it learns. The company also wants models that generate their own training signal and keep learning after release, a stated requirement for what it calls a recursive self-improving loop.
Frontier Research is organized around six areas:
- Multi-agent orchestration
- Pre-training
- Continual learning
- Reinforcement learning
- Systems, covering inference, training, environments, and security
- Frontier evaluations
Three of those get top billing (pre-training, continual learning, and RL), with orchestration binding them together. For orchestration, the stated goal is output on the Pareto frontier of quality for whatever cost a user chooses to spend.
The evaluations track may be the hardest part. Perplexity says it wants tests that measure discovery rather than recall: can a system pose good questions, design experiments, and improve itself? Those tests have to tell reasoning apart from memorization. The page does not say what the benchmarks will look like or who will validate them, and a lab that grades its own scientist faces an obvious credibility problem.
On openness, Perplexity commits to publishing. Output is promised in four forms: papers, posts, weights, and repos. The company says the work will also feed its consumer product, which it calls the toughest test. The page offers no timeline for any release and no named researchers or leads.
For candidates, the practical details are on the careers page of Perplexity (research.perplexity.ai). Roles sit in San Francisco, Palo Alto, New York, and London, with competitive salary and equity. One application covers all six areas. A separate Research Fellowship has tracks before and after the doctorate for early-career researchers.
If you work on agents or evaluation, the published weights and repositories are the part to track. The first papers will show whether this is a research program or a recruiting banner, and they will land well before any automated scientist exists.
Perplexity, Frontier Research page at research.perplexity.ai, accessed November 2026.