Customer story · Lossfunk
How Lossfunk residents run AI research on Jarvis Labs
One GPU cloud for interpretability, reinforcement learning, and world-model research, from long L4 experiment runs to A100 and H100 training.
5
research projects profiled
40K+
reasoning rollouts generated on L4s
100+
steering runs across 4 open-source models
The workload
One cloud, five research agendas.
Lossfunk is a research residency where independent researchers pursue their own agendas, from interpretability to world models. Residents run their experiments on Jarvis Labs, and each project needs a different shape of compute.
- 01
Interpretability and model steering
A100 and H100 GPUs for activation capture and patching.
Residents capture activations, build steering vectors, and patch open-source models, from NeurIPS-bound work on how expert cues drive sycophancy to reasoning studies for drug repositioning. Most runs use A100 40GB and 80GB GPUs, with H100s for the heaviest interpretability jobs.
Explore H100 GPUs - 02
Rollout generation and RL
Cheap L4s for long inference jobs, A100s only for training.
One adaptive-reasoning project generated around 40,000 vLLM rollouts across two 1.5B models on L4s, since the job is long and memory light rather than compute heavy, then moved to an A100 only for supervised fine-tuning and GRPO. Paying for the big card only while training made a real difference to how much could run.
Explore L4 GPUs - 03
Long-horizon experimentation
Sequential-task sweeps and world models on persistent instances.
Continual-learning experiments train models on sequences of up to 200 tasks across multiple seeds and baselines, and JEPA world models learn MuJoCo mazes, puzzles, and Push-T. Both run mostly on L4s, stepping up to A100 80GB GPUs for larger runs.
Explore A100 80GB GPUs
Why Jarvis Labs
Built for how researchers work.
Lossfunk residents arrived from GCP, Modal, Vast.ai, and RunPod. What keeps them on Jarvis Labs: low-latency VS Code in the browser, a persistent home folder that cuts storage cost when an instance is paused, a CLI that tracks experiments across instances, and launches that take minutes instead of a setup curve.
“The JarvisLabs service is the first which I have come across that makes the process of training and running models incredibly researcher friendly.”
— Rithin Nagaraj · Researcher, Lossfunk · Compute provider for the Lossfunk residency
Results and comparisons are reported by Lossfunk researchers and reflect their workloads and testing environments.
Build on Jarvis Labs
One cloud from first experiment to final run.
Start cheap on L4s, scale to A100s and H100s when the work demands it, and keep every experiment in one place.