PLAYER-HANDBOOK.TXT
A technical manifesto. Read in any order.
Abstraction layers are gifts. Understanding the gift is a separate skill.
Frameworks hand you a working answer in five minutes. Building the same thing from scratch costs three weeks but teaches you why the answer is shaped that way — what failure modes it absorbs, what trade-offs it bakes in, what you'd choose differently if you owned the constraints.
Implementing a shell, a transformer, a renderer, or a fiber runtime forces you to confront the actual shape of the problem. The mental model you build by doing it once survives every framework migration after it.
I'm not against abstractions — I depend on them daily. I am against treating them as opaque. The point of writing PSH wasn't to replace bash; it was to never again look at a pipeline and wonder what's happening under it.
AI is most valuable where humans stay in the loop, can see why, and own their data.
Human-in-the-loop. AI suggestions belong on the critical path of a human decision, not behind it. The clinician chooses; the model argues its case.
Explainable. Every consequential prediction should carry a rationale a domain expert can argue with. SHAP, attention maps, feature attributions — these are not optional polish, they're the product.
Responsible. Subgroup performance gaps are launch gates, not appendices. If a model can't show fair behaviour across the populations it serves, it isn't ready.
Privacy-preserving. The best place to run a model is on the user's device. The second best is in a service that holds nothing it doesn't need.
Software should belong to the person using it.
Local-first. Treat the user's device as authoritative. Sync is an optimisation, not a prerequisite.
Offline-first. If the network is the difference between working and broken, the product is broken. Cache aggressively. Degrade honestly.
User ownership. Data formats should be open, exportable, and grep-able. If the user can't take their work elsewhere, the tool is a cage.
Performance is design. A 200 ms p95 is a different product than a 2 s p95 — different conversations, different trust, different uses. Performance budgets belong in the spec.
Areas I want to spend the next decade getting deep in.
Systems programming — shells, runtimes, schedulers, allocators.
Artificial intelligence — transformer internals, training stability, evaluation that survives contact with reality.
GPU computing — CUDA kernels, memory hierarchies, the gap between theoretical and achieved FLOPs.
Graphics programming — real-time rendering, physically-grounded shading, ray-marched curiosities.
Cybersecurity — packet-level intuition, threat modelling, blue-team workflows.
Distributed systems — consensus, replication, the small handful of ideas that show up everywhere.
Developer tools — the tools that make the next ten engineers faster than I am.
