Anush Elangovan describes Helios as AMD's move into rack-scale AI systems and connects its memory and networking capacity to ROCm.ai, an open software platform designed for developers and coding agents. He explains how agent skills could help deploy models and optimize workloads on AMD hardware. The capacity and performance comparisons are vendor claims rather than independently demonstrated benchmarks in this interview.
Quentin Anthony explains why very sparse mixture-of-experts models can be limited by memory capacity and data movement rather than raw computation. More memory can reduce the parallelism needed to fit a model, while an open software stack lets his team change kernels and communication algorithms directly. He describes training models on an AMD-based cluster and frames larger frontier-scale models as a future ambition.
Anush Elangovan and Quentin Anthony discuss routing work across a laptop, an on-premises system and a cloud model according to complexity, cost and data sensitivity. They also explore AI-assisted kernel optimization, while arguing that human expertise still matters for low-level understanding, research decisions and diagnosing failures.
Anush Elangovan argues that long-running agents need substantial CPU resources for shell commands, file access and other work around inference. Quentin Anthony stresses the need to inspect outputs and check that research has not drifted over time. The interview closes by comparing open AI infrastructure with the internet and debating how widely its benefits could be shared.
Watch on YouTube
