VOICES OFAI Research
A survey of Lambda Research Grant recipients on the state of their fields, the compute bottleneck, and what comes next.
Introduction
The ideas that define AI start in labs, with a researcher, a hypothesis, a dataset, and a cluster. The papers published at ICML, NeurIPS, and CVPR today become the foundation for what frontier labs ship next.
This report is built from the responses of more than 200 researchers in the Lambda Research Grant Program. We asked them what they're working on, what progress matters most, what stands in their way, and where they think their fields are headed.
Lambda's Research Grant Program exists to remove the bottleneck between research and compute. Lambda has distributed over $1.3M in cloud credits through the program, and that compute has powered over 150 published papers. We provide compute to researchers so they can spend their time on experiments.
The work
We asked researchers about their current projects, what they set out to do, what came of it, and how they used the compute.
Most projects run on short timelines and small datasets. That fits what researchers told us about compute: long, large-scale runs need more than most labs can get.
02/Dataset tier for a typical project
“I focus on developing novel machine-learning architectures and algorithms to support robot learning and human-robot collaboration in the diverse and unstructured environments that will be encountered by these agents in the real world.”
Rohan Paleja, Purdue University
“Principled approaches to LLM post-training and evaluation.”
Jikai Jin, Stanford University
What gets in the way
The researchers were asked to stack-rank a handful of common problems: data quality, researcher headcount, meeting publication deadlines, and more. Across all areas of research, one answer topped all others.
The biggest problem for researchers is their lack of compute. It keeps novel AI research ideas from becoming findings and validated claims. Labs try to close the gap through funding cycles, but funding reviews take months, while architectures and baselines shift in weeks. By the time compute is approved, the proposed experiment is outdated. So researchers spend their time writing proposals instead of running the experiments they've trained for years to run.
That gap, between the work researchers want to do and the infrastructure they can reach, is shaping where the field goes next.
Showing data for Total
03/What slows you down the most in your work?
- 1Compute access or scale
- 2Funding
- 3Evaluation & benchmarks
- 4Data (quantity, quality, or access)
- 5Competing priorities
- 6Model architecture / training methods
- 7Researcher headcount or talent
- 8Engineering support
- 9Reviewing & publication cycles
04/What is the % of how you used your compute?
62%
/ ranked compute as their #1 blocker
“My claims live or die on breadth…Compute is what converts an anecdote into a finding.”
Alexander Xiong, UC Berkeley
“Compute is the bottleneck between having a research idea and being able to test it at the scale the data demands.”
Davide Romelli, Trinity College Dublin
The data backs that up: seven in 10 researchers can reliably reach only two accelerators or fewer. Fifty-one percent of researchers said they need more compute to run more experiments, faster, and with more rigor. The data and the expertise are already there. Compute is what stands between them and their next result.
05/Compute access tier that you can reliably reach for a typical project
06/Funding sources (multiple answers allowed)
The mismatch between grant timelines and research timelines shows up in how researchers describe the process itself. When we asked what frustrates them most about funding applications, compute and long, complex processes came up again and again. Every week spent on paperwork is a week not spent on experiments.
“Compute lets me scale dataset size, run many parallel simulation environments for faster-than-real-time rollout and MPC-guided fine-tuning, and sweep conditioning and consistency designs in days instead of weeks, which is the difference between testing one idea and finding the one that works.”
Deemo Chen, Caltech
“Compute converts almost linearly into statistical power and into the ability to validate claims at a scale where they actually transfer.”
Tianyi Zhang, University of Washington
What’s the most challenging or frustrating part of the research funding application process across any funder?
About the researchers
The Lambda Research Grant Program spans a wide range of researchers and institutions, from large university labs to independent researchers.
07/Primary area of research
08/Primary role
09/Primary affiliation
10/Team or lab size
Run your next experiment on Lambda
Every insight in this report came from a researcher running experiments on Lambda compute. The Lambda Research Grant Program puts compute in the hands of the people doing the work. Grant recipients have published at ICML, NeurIPS, CVPR, and other top-tier venues, and the program spans hundreds of institutions, from Stanford, MIT, Berkeley, and Carnegie Mellon to emerging labs worldwide.
The Lambda Research Grant Program accepts applicants on a rolling basis. If your lab is doing frontier research in finance, robotics, healthcare, AI agents, or anything in between, Lambda wants to hear from you.
Shape the future of AI
Submit your next research project to the Lambda Research Grant Program.