Aiming to build a system that does research autonomously

Perplexity said on Nov. 16 (early Nov. 17 Japan time) through its official account on X that it has launched a research unit, "Perplexity Frontier Research." The company said the goal is to create a system that can carry out research autonomously. The work will be backed by more than 10,000 GPUs, and the unit will regularly publish papers, models, data and benchmarks (tests that measure performance).

Today we're launching Perplexity Frontier Research. The goal of Perplexity Frontier Research is to create a system that can perform autonomous research. The work will focus on multi-agent orchestration, reinforcement learning, continual learning, systems, and pre-training. We will develop algorithms, models, and evaluations, released openly, and tested in Perplexity products. Our work will be supported by more than 10,000 GPUs. The Frontier Research team will regularly publish papers, models, data, and benchmarks. We're actively hiring researchers and engineers for Perplexity Frontier Research in San Francisco, Palo Alto, New York, and London. https://research.perplexity.ai

Research spans five areas, from pre-training to coordinating multiple agents

The company listed five research areas. They are "multi-agent orchestration," in which multiple AI agents split roles and work together; "reinforcement learning," which trains models with rewards based on the results of their attempts; "continual learning," in which models keep learning new knowledge while in use; "systems," which covers the computing infrastructure; and "pre-training," which builds a model's foundation from large amounts of data.

Results to be released openly and tested in its own products

The company said it will develop algorithms, models and evaluation methods, release them openly and test them in Perplexity products. It listed papers, models, data and benchmarks as what it will publish. The computing resources backing the research amount to more than 10,000 GPUs, it said.

The post did not give a concrete timeline or frequency for releases, or licensing terms. It also did not say what kind of GPUs will be used, whether the company owns or rents them, or the unit's headcount and budget. When ORIENS tried to access the official website linked in the post, a bot-protection page was returned, and its contents could not be confirmed.

Research Lead describes running "the loop" of research inside software

Andrew Gordon Wilson, whose X profile lists the title of Research Lead at Perplexity and a professorship in machine learning at New York University, posted a thread announcing the unit on X the same day. Perplexity's official account reposted it. Wilson wrote that the unit will openly advance the frontier of models, data and algorithms.

Wilson said the unit aims to create an "automated research scientist." He said research is a loop of asking the right questions, hypothesizing, navigating ambiguity, designing experiments and updating beliefs, and that the team thinks this loop can run inside software, for anyone.

We aim to create an automated research scientist. Research is a loop that involves asking the right questions, hypothesizing, navigating ambiguity, designing experiments, and updating beliefs. We think this loop can run inside software, for anyone. 2/5

Wilson also wrote, as a personal view, that going beyond black-box predictions to build systems that provide fundamentally new knowledge is one of the most inspiring directions for AI research. Wilson listed the questions the unit will consider as multi-agent orchestration, pre-training, continual learning, reinforcement learning and systems, as well as evaluations. Wilson also called it an environment where one could propose the "next transformer" (the transformer is the architecture underlying today's major AI models), but this is Wilson's outlook, and no concrete plan was given.

On how the unit will release its work, Wilson said it will publish papers, blogs and code, and that its results end up being tested in a product used by millions of people. In a follow-up post, Wilson repeated that the effort will be backed by more than 10,000 GPUs, the same figure as the official announcement.

Hiring researchers and engineers in four cities

The company said it is hiring researchers and engineers for the unit in San Francisco, Palo Alto, New York and London. It did not disclose how many people it plans to hire or the size of each location.

In November, Perplexity released the weights of "pplx-decider-v1.3," a new version of its decision-only AI model, and also published research that tested an isolated environment for AI agents from the attacker's side. The official post did not mention any relationship between these and the new unit.