Real-time cloud inference for physical AI
The Hook
“Today we're releasing training on Dreamscale. Let me show you how you can fine-tune your own VLA in just three simple steps.”
Product Reveal
“So we go train a model, select your base model, so pi 0.5 here. You can connect your Hugging Face and see your datasets that you have. So let's say this one.”
Feature Teaser
“And Dreamscale automatically reads the dataset and configures the best fine-tuning recipe.”
Feature Teaser
“You can adjust any settings here. We automatically bake in real-time control, and also clean the dataset for any faulty frames and idle time.”
Feature Teaser
“And then you can just click start training. It will take a few minutes to a few hours, and once it's done...”
Product Reveal
“...you'll get a checkpoint that you can see the loss curve, and you're actually able to deploy and download any of the checkpoints.”
Call to Action
“So you can see the deployment here, and you can easily run this with a single CLI command on your...”
Synthesizable in Remotion · key components of the saas-walkthrough-pip template.
Circular/squircle video bubble anchored in the bottom right corner with smooth drop shadow.
Absolute positioned rounded-3xl container at bottom-6 right-6 with border border-white/20, shadow-2xl, and continuous OffthreadVideo playback.
Floating rounded white application window set over a vibrant abstract gradient backdrop.
Wrapper div with rounded-[32px], bg-[#fcfbfa], p-8, shadow-xl, layered above a high-saturation background graphic.
Custom SVG arrow cursor smoothly moving between interactive controls using spring interpolation.
Interpolate cursor x,y coordinates via Remotion spring() based on predefined timestamp waypoints.
Purpose. Capture AI/robotics engineers with instant VLA fine-tuning capability.
Execution. Dashboard view showing completed pipelines and primary action trigger.
Purpose. Prove low barrier to entry with 3-step setup and auto-configuration.
Execution. Progressive step form selection: model choice, dataset linking, and recipe generation.
Purpose. Demonstrate tangible engineering outputs: training curves and immediate CLI deployment.
Execution. Loss chart evaluation and single CLI command copy block.