The best AI is built, not bought.
The Hook
“Hi, I'm Yash from Applied Compute, and today we're introducing AC2, the training and inference cloud for open models. We're excited to bring research out of the lab and enable every AI team to train, serve, and continuously improve their LLMs. Let me show you how it works.”
Problem Agitation
“Let's say I want to train a code Q&A agent on top of Qwen 3.6 35B because it's fast and cost-efficient. In this case, I've already created a data set from GitHub repos. I also have a custom harness built on top of OpenCode that gives the model access to some special tools I wanted to use.”
Product Reveal
“To get started training with these, I'll create a new AC2 project and implement a few simple abstractions. First, I'll register my data set. Then, because I'm using my own harness, I'll wrap it in the lifecycle functions required by our bring-your-own-harness interface. This lets me directly train with the compaction logic and tools that OpenCode ships with. And finally, I'll define a reward function, which in this example uses an LLM judge to score the model's final answer.”
Feature Teaser
“Now, I can launch and monitor my run with a single command using the CLI. AC2 comes with optimized configs for all the top models, so you don't have to worry about GPU performance or tuning YAML. We also provide the observability to monitor your training runs live and see how they're progressing. Our post-training stack supports a bunch of different algorithms like GRPO, SFT, and OPSD.”
Feature Teaser
“Let's check back in on our training. Looks like I got a ping on Slack. It's from Ari, our AC2-native research agent. Ari monitors run health, reads through rollouts, and takes actions even when I'm offline. Here, the remote tool call server hosting the codebase appears to have failed. Ari saw the reward tank and decided to restart the run. I can even inspect Ari's execution trace and generate custom reports on data points or do comparisons across runs.”
Product Reveal
“Once my model has finished training, I'll use AC2 inference to one-click deploy to a fully autoscalable serving endpoint. Now that the model is being served, I can collect traces online to train further with techniques like self-distillation.”
Call to Action
“AC2 is already being used to power models that we train for companies like Microsoft, NVIDIA, Cognition, Harvey, Mercor, DoorDash, and others. If you're interested in learning more about training and serving your own models, please reach out to us.”
Synthesizable in Remotion · key components of the developer-platform-walkthrough template.
Draws dynamic SVG bezier curves from repository listings directly to individual JSON objects.
Use Remotion interpolate on strokeDashoffset over an SVG path calculated between bounding boxes.
Array of rounded frosted cards drifting smoothly in 3D space around the presenter.
Wrap items in absolute positioned containers with spring-driven translateY and opacity transforms.
Concentric dashed and solid circular lines expand, rotate, and snap to reveal the central brand logo.
Animate SVG circle stroke and scale using spring physics, fading into the company mark.
Purpose. Establish product category and core value proposition directly from the founder.
Execution. Medium talking-head shot with brand lower-third.
Purpose. Demonstrate the end-to-end technical workflow from data ingestion to reward modeling.
Execution. Terminal commands, code snippets, and dynamic data lineage graphics.
Purpose. Highlight automated failure detection and agentic recovery.
Execution. Slack alert pop-up followed by deep-dive analytics charts and report artifacts.
Purpose. Build enterprise trust and invite outreach.
Execution. Tier-1 company logo grid fading into dynamic vector outro logo.