Perplexity wants an AI that does research on its own, with 10,000+ GPUs
Perplexity Frontier Research targets an AI that runs the whole research loop, across six pillars, with over 10,000 GPUs and a promise to publish papers, data, models and code.

Perplexity has launched Perplexity Frontier Research, an initiative to build an “automatic research scientist”: an AI system that can conduct research on its own. The company says the work will run on more than 10,000 GPUs.
Six research pillars and open publication
Perplexity describes research as a loop: asking the right questions, forming hypotheses, designing experiments, analyzing results, handling ambiguity and updating beliefs. It says the aim is new interpretable knowledge, not black-box predictions in narrow settings.

The work is organized around six pillars: multi-agent orchestration, pre-training, continual learning, reinforcement learning, systems and evaluations. The announcement lists the kind of questions it wants to answer. Can new, more efficient architectures be proposed? Can synthesized data help models generalize more broadly? When does multi-agent collaboration work best? Can evaluations measure research ability and extrapolation rather than recall?
Perplexity says results will be shared openly as papers, blog posts, data, models and code. It also says the findings will feed a product “used by millions of curious people,” which it calls the most challenging test of all.
The initiative’s research website adds detail. According to that page, the pre-training pillar covers efficient, specialized open-weight models and synthetic data. The reinforcement learning pillar covers verifiable and rubric-based rewards. The systems pillar includes security for long-running agents. Andrew Gordon Wilson is named as research lead.
Who can take part, and what is missing
For now, the visible entry points are for researchers, not customers. Perplexity points to full-time positions and a research fellowship program. The research site says one application covers all six areas, with offices in San Francisco, Palo Alto, New York and London. It lists a priority deadline of December 5 and describes pay as a competitive salary plus equity. The fellowship has predoctoral and postdoctoral tracks for early-career researchers.
Readers who use Perplexity should not expect a new feature. The announcement names no product, release date, model or paper. It gives no budget, no headcount and no definition of what would count as an automatic scientist. The 10,000 GPU figure is the company’s own and is not independently verified. The commitment to open release is also a promise, not a track record yet: nothing has been published under the new banner.
A race toward automated researchers
Perplexity is entering a field where a larger rival has already set a public timeline. OpenAI said on September 6 that it had reached its goal of an “automated research intern”, according to coverage of that announcement by eWeek and Help Net Security. By OpenAI’s own measurements, its research organization used 3.1 agent-workdays for each human workday, based on mid-August data. The company called the figures preliminary and says its next target is an automated AI researcher by March 2028. Critics noted that OpenAI set the goal, wrote the definition and graded itself.
The two efforts differ in emphasis. OpenAI frames its work around speeding up its own AI research, with humans still setting priorities. Perplexity stresses interpretable knowledge across a broad range of applications, and says it will publish openly. That openness could matter: if papers, data and weights do appear, outsiders can check the claims, which they cannot do for internal productivity measurements.
The move also fits Perplexity’s recent product direction. Its Teammates agents, introduced on November 5 to work on their own and message coworkers in Slack, already put multi-agent orchestration in front of paying users. Perplexity has since added outside models such as Fable 5.2 to those agents, as we reported on November 9. Frontier Research suggests the company now wants its own findings, and possibly its own models, behind that layer. The announcement does not say so directly.
Perplexity has not said what an automatic research scientist would first be asked to do, or how it will test the claim. Its sixth pillar, evaluations that measure discovery rather than recall, is the part to watch: without a benchmark that tests whether a system extrapolates beyond what it has seen, the label is hard to verify from outside.
Sign Up for the Daily AI Pulse
One email a day. All the stories that matter.
By signing up, you agree to our Terms of Use and acknowledge the data practices in our Privacy Policy. You may unsubscribe at any time.
Alex CarterAI News & Big Tech CorrespondentAlex Carter writes for Aiholics, keeping readers updated on the fast-paced world of AI and Big Tech.



