Perplexity Builds an Automated Scientist Backed by 10,000 GPUs
Perplexity launched Frontier Research , a lab aimed at building an automated scientist. Backed by more than 10,000 GPUs supporting pre-training, RL, and continual learning experiments. Six focus areas: multi-agent orchestration, pre-training, continual learning, RL, systems, and evaluations.

Takeaways
- Perplexity launched Frontier Research , a lab aimed at building an automated scientist.
- Backed by more than 10,000 GPUs supporting pre-training, RL, and continual learning experiments.
- Six focus areas: multi-agent orchestration, pre-training, continual learning, RL, systems, and evaluations.
- Papers, model weights, data, and benchmarks will be released openly on GitHub and Hugging Face.
- Hiring researchers and engineers in San Francisco, Palo Alto, New York, and London.
- Andrew Gordon Wilson, ex-NYU ML professor, leads the initiative with a December 5 priority deadline.
Perplexity has launched Frontier Research , a lab focused on building AI systems that can conduct research through repeated questioning, experimentation, and revision. The company says the effort will have access to more than 10,000 GPUs and will publish papers, models, data, and benchmarks.
A research loop that keeps turning
Perplexity defines its target system as an “automated scientist” that can identify unanswered questions, gather existing evidence, find gaps, form hypotheses, run experiments, estimate uncertainty, and update its conclusions. That scope extends the company’s existing retrieval and question-answering work into model training, persistent learning, agent coordination, and evaluation.
Andrew Gordon Wilson is listed as research lead. Wilson previously taught machine learning at New York University and has published research on Bayesian deep learning, model generalization, and uncertainty. The lab is recruiting in San Francisco, Palo Alto, New York, and London.
Six tracks connect models to experiments
The lab divides its work into six related tracks, each covering a different part of the proposed research system.
| Track | Research focus | Developer relevance |
|---|---|---|
| Multi-agent orchestration | Question decomposition, information gathering, collaboration, and escalation between models. | Routing work among models based on cost, latency, capability, and task complexity. |
| Pre-training | Architectures and training methods designed to use data and compute efficiently. | Potential improvements to model quality per training token and GPU-hour. |
| Continual learning | Models that learn after deployment and retain useful information across sessions. | Persistent memory without excessive forgetting, privacy leakage, or repeated retraining. |
| Reinforcement learning | Training models to propose hypotheses, test them, and learn from outcomes. | Post-training against retrieval systems, tools, environments, and production feedback. |
| Systems | Training, inference, agent environments, networking, and security. | Infrastructure for long-running agents and faster movement of models into production. |
| Frontier evaluations | Benchmarks intended to measure reasoning and discovery while limiting memorization. | Tests for agent reliability, novel problem-solving, and benchmark contamination. |
In the lab’s advisor-escalation design, a small model handles a task locally and calls a larger model when a router predicts that the added capability will justify the cost. The same design appears in Perplexity’s Portable Computer launch . Its practical value will depend on whether the router can preserve answer quality while reducing token use, latency, and external API calls.
The lab also says its reinforcement-learning stack operates against live retrieval and production systems. Perplexity claims its infrastructure can move trillion-parameter weights from training to inference in about one second. At 16-bit precision, one trillion parameters occupy roughly 2 terabytes before overhead, so meaningful comparison will require details about precision, network topology, storage, and the number of inference replicas receiving each update.
GPU count needs technical context
Perplexity’s stated pool of more than 10,000 GPUs could support pre-training, reinforcement-learning sweeps, and large-scale inference experiments. Actual capacity depends on the accelerator models, interconnects, cluster allocation, utilization, and access period, none of which the announcement specifies.
The research plan also covers systems security, an important constraint for agents that run code, call tools, and retain memory. Useful releases in this area would need to address sandbox isolation, credential handling, network access, audit logs, and the deletion of stored user information.
Open releases are the first test
Perplexity says it plans to publish research papers, models, datasets, and benchmarks. Its research page already lists several artifacts:
- SPACE runtime: infrastructure for long-running agents and their execution environments.
- pplx-garden repository: mixture-of-experts dispatch and combine kernels, plus Apple Silicon inference code. Mixture-of-experts models activate selected subnetworks for each input instead of running the full model.
- fabric-lib paper: research on point-to-point communication using remote direct memory access, which transfers data between machines with little CPU involvement.
- Decider v1.2: a 27-billion-parameter open-weight model listed on Hugging Face.
The listed artifacts vary in form, from source repositories to papers and model weights. Developers will need to inspect each project’s license, model card, training-data disclosure, hardware requirements, and maintenance status before adopting it. Access to weights alone does not establish permission for every commercial, research, or redistribution use.
Perplexity moves deeper into the stack
Perplexity’s current products already combine web retrieval, model routing, and generated answers. The company says its Computer product divides complex requests into subtasks and distributes work across more than 20 models. Frontier Research could give Perplexity greater control over the routers, specialized models, post-training methods, and evaluations behind that workflow while reducing dependence on outside model providers.
The automated scientist remains a research objective, so near-term evidence will come from reproducible releases and measured system performance. The most useful results would include cost and latency baselines for model routing, contamination-resistant evaluations, continual-learning tests that measure forgetting, and reinforcement-learning experiments that disclose environments, reward signals, and safety controls.
Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.
Read original report

