What happened
The Import AI newsletter highlighted a significant development this week. Fable, a research group, demonstrated an AI system capable of writing its own GPU kernels. This points to a broader trend of AI systems starting to automate and optimise their own underlying infrastructure. The goal is to create more efficient code for specialised hardware without direct human intervention, tackling complex optimisation problems at the machine level.
How the room's reading it
The initial reaction from developers on X is cautiously optimistic. Many see this as the logical next step for AI — systems that improve their own performance over time. Infra teams are particularly interested, framing this as a potential way to escape the complexity of CUDA and other low-level optimisation frameworks. The consensus among researchers is that this isn't just about speed. It's about letting AI discover novel optimisation strategies that humans might miss, ultimately lowering the cost of inference and training for everyone building on top of these models.
Sailfish's take
We've spent enough time wrestling with CUDA to see this as a welcome shift. The real prize isn't just faster kernels — it's the abstraction. Builders shouldn't need a PhD in parallel computing to get maximum performance from their hardware. We think this trend will create a new layer of AI-managed infrastructure, where models continuously re-compile and optimise themselves for the specific hardware they're running on. This is more than a speed boost. It's a step toward hardware-agnostic development. We'd watch this space closely, but we wouldn't bet on it replacing skilled infra engineers just yet.