01
Neural prediction
Meta FAIR's TRIBE v2 predicts activation across 7 brain regions for every window of your video. Visual cortex peaks during the cuts. Language regions activate on the tagline.
AMD Developer Hackathon 2026
A.R.C Studio
AI-powered pre-testing for marketing content.
Predicts how content will land — neurally and socially — before you publish. Running on a single AMD MI300X.
- Hardware
- 1× MI300X · 192 GB
- Runtime
- ROCm 7.2 · vLLM 0.17.1
- NVIDIA in path
- 0
DEMO
Live MI300X run
Live MI300X run on the Apple 1984 ad. Full pipeline end-to-end.
PIPELINE
What it does
02
Social simulation
1000 personas grounded in real-world entities react to your content — sharing, drifting, counter-narrating. Behavioral signal at scale.
03
Iterative optimization
Qwen3.5-27B reads both signals and writes the next variant. Loops until thresholds clear.
STACK
Built on AMD
- Hardware1× AMD MI300X — 192 GB HBM3, AMD Developer Cloud
- RuntimeROCm 7.2 · vLLM 0.17.1 (rocm/vllm container)
- Variant tierQwen3.5-9B — variant generation + MiroFish agents
- Orchestrator tierQwen3.5-27B — cross-system analysis + report generation
- Neural encoderMeta FAIR TRIBE v2 — trimodal brain encoder, ROCm PyTorch
Zero NVIDIA in the inference path.
REPRODUCE
Run it yourself
-
1
Provision MI300X
AMD Developer Cloud → 1× MI300X with the vLLM Quick Start image.
-
2
Clone the branch
git clone --branch competition/amd-hackathon --recursive https://github.com/AR6420/ARC_Studio.git -
3
Follow the runbook
Ten steps, ~30–45 minutes from clean droplet to first completed campaign.