AI Researchers
We set the research goals and review the findings. We’re developing AI Researchers to design and run experiments, then use the results to decide what to test next.
Recursive
self-improvement
Our goal is for each round of experiments to improve both the model and the system that builds, trains and serves it. We’re developing AI Researchers to test changes across that system, from their own research methods to inference software and GPU execution.
What can improve between cycles
AI Researchers and methods
Models and training
Inference software
GPU execution and scheduling
Evaluate each change against the research objective. Carry forward the ones that work.
Research priorities
Evaluation
We test whether an improvement holds across workloads and in later rounds of experiments.
Inference and GPU execution
We test changes to GPU kernels, numerical precision and scheduling to see where inference can run faster while preserving output quality.
Selected experiments
Optimizing Krea 3 on AMD Instinct MI355X with AI researchers
On third-party cloud GPU VMs, AI Researchers tested how to overlap data transfers with computation. Splitting the work into two chunks helped. Four chunks were slower.
