A goal-only AI agent writing its own GPU kernels beat open-source libraries on 8 of 10 operators, and a same-model test found specialist kernel machinery added nothing
AspireAI community · arXiv (cs.DC)
GPU kernel development is usually sold as a specialist craft: hand-tuned code, a kernel-specific knowledge base, a search workflow, a team that knows the hardware. JustGoal asks what that scaffolding is actually worth now that general-purpose models are strong, and answers with a deliberately austere baseline - give the agent only the optimization objective and the constraints, give it no kernel-specific infrastructure or predefined workflow, and let it build its own evaluation loop.
On NVIDIA B200 GPUs, that goal-only agent outperformed open-source libraries on 8 of 10 operators, with per-operator geometric mean speedups of 1.14x to 2.34x, and beat the fastest publicly released specialized-agent kernel on 5 of 9 operators where such a comparison existed. The sharper result is the controlled one: with the same model, Fable 5.1, on single shapes, the KDA-0.5 specialist harness reaches 0.95x of JustGoal's performance, and single-mechanism ablations show no consistent benefit from domain knowledge or predefined workflows.
The authors' conclusion is a governance point rather than a benchmark score - specialized machinery should have to justify itself against a same-model, same-scope goal-only baseline, and be re-evaluated whenever the foundation model changes. The limitation is narrowness: ten operators on one GPU generation, and the same-model control restricted to single shapes, so this is a strong signal about evaluation discipline rather than a settled verdict on kernel tooling.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.