Perplexity has launched Frontier Research, a lab that says it aims to build an automated scientist, an AI system that can carry out research on its own. The lab is backed by at least 10,000 GPUs and is led by researcher Andrew Gordon Wilson, according to launch coverage. It is an early research program rather than a released product, and the company has not yet shown results.

What Perplexity Announced

The lab's stated goal is to create an automated scientist that can perform insightful research and to put that capability in the hands of everyone. Perplexity says the work will produce new algorithms, models and evaluations, and that it will be shared openly. The company lists papers, blog posts, model weights and open repositories, including code, kernels and benchmarks, as its channels.

CEO Aravind Srinivas framed the effort around compute. Coverage reports that the company has committed at least 10,000 GPUs to the lab. That is a substantial commitment for a company best known for its answer engine, and it signals that Perplexity intends to train models, not just build products on top of other labs' systems.

Six Research Areas

Frontier Research is organized around six areas: multi-agent orchestration, pre-training, continual learning, reinforcement learning, systems, and frontier evaluations. Each is a distinct discipline. An automated researcher has to train capable models, keep them learning after deployment, coordinate several agents on experiments, run those experiments on efficient infrastructure and then judge whether the results are real.

The evaluation area matters most for credibility. An AI that generates research claims is only useful if its findings can be checked, and benchmarks that resist gaming are hard to build. Listing frontier evaluations alongside model training suggests the lab wants to measure its own progress, though the details of those methods have not been published yet.

Leadership and Hiring

Coverage of the launch names Andrew Gordon Wilson as the lab's leader. The lab's site lists openings in San Francisco, Palo Alto, New York and London. Applicants have a priority deadline of December 5, and one application covers all six research areas. Hiring for a lab of this scope suggests a multi-year build rather than a short experiment.

The Wider Push Toward Automated Research

Perplexity's launch lands in a period when several AI companies are investing in tools that can assist or partly automate scientific work. OpenAI's program for academic researchers is one example. Its update says the program's target is doubling to 200,000 researchers, with about 20,000 researchers in 33 countries having joined since July. Applications for 50,000 new grants are set to open on November 23. In a pilot study cited by OpenAI, 251 scientists self-reported that their work would take 1.9 times as long without AI. That figure rests on self-reporting, so it shows perceived time savings rather than measured productivity.

OpenAI has also described outside evaluation work, including first-phase agreements with Apollo Research, METR, Transluce and Goodfire, which give independent experts access to internal models, source code and training data. Such arrangements matter for a lab like Perplexity's because an automated researcher that produces claims will need external checks before anyone trusts its output.

The difference with Perplexity's approach is focus. Rather than offering research tools to academics, the Frontier Research lab is aiming to build the researcher itself, and it lists pre-training and reinforcement learning among its core areas. That puts it closer to a frontier model lab than to a research-software vendor.

What Is Not Known

No results are shown yet. The launch materials do not include a benchmark, model release or paper from the lab, and there is no public timeline for when research output will appear. "Automated researcher" is a goal rather than a demonstrated capability, and the phrase covers a wide range of claims, from literature search and code generation to running experiments and producing original findings. Readers should separate those meanings when they evaluate any future announcement.

Compute commitments also do not guarantee results. Large GPU fleets make experiments faster, but the hard problems are often about data quality, training methods and knowing which experiments are worth running.

Why It Matters

The lab matters for three reasons. First, it moves a search and answer company deeper into model research, which changes how it competes beyond consumer products. Second, its openness commitments, if followed, would give outside researchers model weights and benchmarks to test. Third, it sits in a growing group of labs that are aiming at automated research directly.

For researchers and developers, the practical question is what will actually be released and when. The lab's site does not give timelines, so the right approach is to watch its research page and any published papers, then judge the claims against independent replication. If the openness commitments hold, the community will have a concrete way to test whether an automated scientist is more than a goal.

Readers should also watch how the lab handles evaluation. Claims of automated discovery are easy to make and hard to verify, and the strongest early signal would be a result that outside groups can reproduce with the released code, data and benchmarks. Until then, Frontier Research is best understood as a serious, well-funded bet on a research agenda, not as evidence that the automated scientist exists.

Sources