30d roadmap · 4 milestones
Trace a Tensor
A compact AI-systems synthesis path. Four milestones connect the work expressed by a model to the machine and serving layers that execute it, ending in a measured bottleneck diagnosis and defended optimization.
Outcome: Follow one model workload from representation and gradients to memory, kernels, scheduling, and serving economics.
How to work this roadmap
Treat each milestone as a claim that must be supported by evidence. Before reading, write what you think the mechanism does and where it will fail. After studying the linked concepts, run the drills without copying an answer, preserve the output, and revise the explanation. Move forward when you can connect the milestone goal to a working implementation, benchmark, architecture decision, or reviewable design artifact.
The 30d horizon is a sequencing aid, not a completion badge. Spend more time where your prediction and the observed behavior disagree. Keep a short decision log containing the mechanism selected, alternatives rejected, expected failure mode, measurement used, and remaining uncertainty. Review that log with FSRS prompts so the roadmap produces durable system judgment rather than a temporary tour of terminology.
At the end, explain Follow one model workload from representation and gradients to memory, kernels, scheduling, and serving economics. from first principles to a reader outside the domain. A strong explanation should survive follow-up questions about correctness, cost, latency, resource use, security, recovery, and operational visibility. If it cannot, return to the milestone that contains the missing mechanism and build a smaller falsifiable example.
Milestones
Milestone 1
Week 1 — Represent and differentiate
Explain how values become tensors and how a forward computation creates the work required by backpropagation.
Concepts
Milestone 2
Week 2 — Move and execute
Locate tensor data across the memory hierarchy and profile the runtime before choosing an optimization.
Concepts
Milestone 3
Week 3 — Change the work
Reduce data movement or numerical work with hardware-aware kernels and model representation changes while preserving quality.
Concepts
Milestone 4
Week 4 — Schedule, serve, and defend
Connect engine scheduling to latency, throughput, utilization, quality, and cost, then defend the measured system.
Concepts
Build evidence
- Trace a Tensor: Diagnose and Optimize One Workload — Trace one tensor-producing model operation from its numerical representation and computation graph through memory movement, kernel execution, engine scheduling, and request-level serving. Build or precisely model a reproducible workload, identify its dominant bottleneck, apply one justified optimization, and defend the resulting quality, latency, resource, and cost trade-offs.