ShiftWing: an inference engine built from scratch
Why ShiftWing manages graphics memory, system memory and fast storage as one runtime instead of treating local inference as a smaller copy of server inference.
The thesis
Powerful frontier AI for regular end users, on consumer-grade hardware
ShiftWing makes frontier-scale inference possible across the memory already inside a consumer computer. LocalForge applies that capability to private code review. The research log records what worked, what failed and what changed next.
Featured
All notesWhy frontier AI has to run on the hardware people own
Frontier capability matters most when regular people can use it without surrendering their work, context or control to a remote service.
The work
AboutShiftWing
A from-scratch inference engine that treats graphics memory, system memory and fast storage as one hierarchy.
Meet ShiftWingLocalForge
Private local code review designed so source, context and inference stay on the user’s own machine.
Meet LocalForgeResearch practice
Every experiment records its hypothesis, controls, correctness checks, negative results, limitations and decision.
See the practiceLatest
All notesThe reporting rule we hold ourselves to
Every experiment report must preserve its question, controls, correctness checks, negative results, limitations and the decision that followed.
LocalForge: code review that never leaves your machine
LocalForge is a private local reviewer designed to keep repository context, inference and disagreement on the computer where the code lives.
A negative result worth publishing
A larger target missed the threshold written down before the run, so the release decision stayed negative even though parts of the result looked promising.
Pushing experts below four bits
Low-bit expert sidecars reduced storage pressure, but the quality study made clear that this is a trade to qualify rather than a free saving.
Speculative decoding that changes no output
A speculative path was admitted by confidence, verified against the ordinary path and rejected whenever it could not preserve the exact result.
Treating VRAM, RAM and NVMe as one memory hierarchy
Experiments with direct storage, learned expert locality and predictive prefetch turned three separate resources into one managed path.