A recent discussion on the Linux kernel mailing list is considering whether machine learning can assist the kernel subsystem. IBM engineer Viacheslav Dubeyko proposed creating a general-purpose infrastructure that would allow the kernel subsystem to interact with machine learning models operating entirely in user space.
The proposal centers around a lightweight “ML proxy” within the kernel. This proxy exposes structured data such as internal state and performance metrics from kernel subsystems and receives recommendations from user-space ML models. Training, model execution, and experimentation remain outside the kernel, giving the kernel full control over the application of recommendations.
And since when you think of ML, you immediately think of AI, I would like to clarify the following points. The idea does not involve embedding AI or ML code within the kernel. Instead, the kernel subsystem exposes data to user space, where machine learning models typically run, and receives optional recommendations in return.
The kernel remains fully deterministic and in control, deciding whether to apply, test, or ignore suggestions, and all training and inference is performed outside of kernel space. Existing mechanisms such as sysfs, character devices, FUSE, or eBPF have been identified as possible transport and execution layers.
Dubeyko also proposes a feedback loop in which the kernel evaluates the applied recommendations and reports efficiency metrics to user space. This information can be used to refine or retrain the model without affecting the performance or determinism of the kernel. The proposal emphasizes that human-written kernel logic remains the baseline, with ML serving only as an advisory layer.
An initial proof-of-concept implementation of the proposed ML library has been published, and an RFC patch series has been posted to the Linux kernel mailing list for discussion. Patches are explicitly marked as experimental and are intended to gather feedback rather than seek inclusion.
The work is still exploratory, so the patch has not been merged into the mainline kernel. Whether ML-assisted kernel subsystems will advance beyond experimentation remains an open question.
For more information, please see the proposal itself on the kernel mailing list.
