Perplexity wants AI to produce scientific understanding, not just retrieve information or make predictions. Its new Frontier Research initiative sets out a program to automate scientific research, supported, the company says, by more than 10,000 GPUs. The announcement lays out the work toward an automated scientist, rather than introducing a finished one.
The company defines research as a cycle: choose an unanswered question, examine existing knowledge, identify gaps, form a hypothesis, run an experiment and revise the explanation. Search remains part of that process, but the intended output is new knowledge that people can understand. Perplexity explicitly distinguishes that goal from a system that makes accurate predictions without explaining the underlying science.
Its proposed system divides the work among AI agents. Some would search across thousands of sources; others would investigate a single thread, suggest experiments or critique hypotheses. The ambition is for those agents to combine their strengths while retaining what they learn. Perplexity describes this as a system that recognizes gaps in its knowledge and actively seeks information to close them.
Frontier Research spans six areas: coordinating agents, initial model training, learning after deployment, reinforcement learning, computing systems and evaluations. Coordination is meant to connect these efforts while improving quality for a given cost. One example is a small model running on a device that asks a larger model for help only when that assistance is worthwhile.
- Initial training: Perplexity is pursuing efficient, specialized models with downloadable weights. Its research questions include whether generated training data can be more valuable than real data, and how lessons from small experiments transfer to larger training runs.
- Continual learning: The aim is to keep adapting after deployment. Perplexity emphasizes curiosity, memory and persistence—seeking missing information, carrying useful experience into another task, and preserving progress through pauses rather than starting over each session.
- Reinforcement learning: Training uses rewards that can be checked where possible and scored against rubrics elsewhere. Open questions include how to credit individual steps in a long task and whether simulated environments teach skills that transfer to the live web.
The systems work addresses both computing efficiency and the environments where agents act. Perplexity notes that fewer calculations do not necessarily mean faster execution when moving data consumes much of the time. It also points to SPACE, its isolated execution environments, as a way to preserve and resume agent work while keeping credentials and the host system outside an agent’s control.
We want to measure qualitative understanding, so that we can build systems for research that teach us something interpretable that we don’t already know.
Perplexity, Frontier Research announcement
Evaluation is a central unresolved problem in the agenda. Perplexity argues that tests designed as proxies for human ability can give misleading impressions of a model’s broader competence. It wants evaluations that separate reasoning from memorization and scientific understanding from predictive success. Its example is the difference between predicting a particle’s path and producing the equations or physical theory that explain it.
Perplexity promises to publish papers, models, data and benchmarks, with model weights and open repositories among the planned outputs. It also says research results will feed into a product used by millions of people. That gives the initiative two stated destinations: publicly shared scientific work and systems used outside the research setting.
Recruitment is already open. The initiative lists December 5 as its priority application deadline, although the page does not specify a year.
Reader comments
Newest comments first. Replies stay oldest first.