entry~2 min readUpdated 2026-07-16#orientation#vocabulary#constraints

Must Know

The hardest mistakes in WebAssembly AI are usually category errors: treating a format as a runtime, a runtime as a backend, a sandbox as a complete security policy, or one successful browser as universal support.

Core distinctions

Pair Distinction
Wasm vs WebGPU Wasm executes portable code, normally on CPU; WebGPU submits explicit GPU work.
WebNN vs WebGPU WebNN describes neural operations for an implementation-selected device; WebGPU exposes lower-level buffers, shaders, and dispatch.
WebNN vs WASI-NN WebNN is a browser API; WASI-NN is a WASI host proposal for non-browser component/module environments.
Model format vs runtime ONNX, TFLite, safetensors, and GGUF encode artifacts; a runtime must support their operators and metadata.
Runtime vs backend The runtime owns sessions and tensors; the backend or execution provider performs operations on CPU, GPU, or NPU.
Local vs offline Local inference avoids a remote compute call; offline additionally requires cached application, model, metadata, and storage behavior.
Local vs private Local execution can improve privacy, but logs, analytics, model downloads, and browser extensions can still transmit data.
Sandbox vs policy Wasm constrains guest execution; the host still decides capabilities, limits, secrets, network, storage, and updates.
Quantized vs equivalent Quantization reduces representation precision; task and numerical differences must be measured.
Fallback vs failure A fallback can preserve function, but it changes backend, performance, or quality and must be reported.

Minimum vocabulary

  • A module is validated core Wasm code; an instance is a module connected to imports, memories, tables, and state.
  • Linear memory is a byte-addressed guest memory, not a JavaScript object graph or GPU buffer.
  • An ABI defines how values cross a binary boundary; WIT and the Canonical ABI extend that contract for components.
  • A tensor combines element type, shape, strides, layout, and storage; matching dimensions alone does not prove matching semantics.
  • An operator is a graph operation; unsupported operators can block export or force fallback.
  • An execution provider, backend, or delegate maps work to a concrete implementation and device.
  • A golden output is a versioned reference result paired with input, model, preprocessing, and tolerance.
  • Cold start, first inference, and warm inference measure different product costs.

Evidence rules

  • Record browser, operating system, hardware, runtime, backend, model, quantization, and date.
  • Never claim a performance improvement without workload, warm-up, samples, distribution, and raw context.
  • Never claim numerical equivalence without an explicit comparison and tolerance.
  • Never hide which backend ran or whether a fallback occurred.
  • Mark standards, drafts, proposals, experiments, and implementation-specific APIs distinctly.
  • Treat model and Wasm binaries as versioned supply-chain artifacts with checksums.
  • Keep deterministic computation outside the model whenever correctness can be expressed in code.

Continue

Open Phase 00: Orientation to establish prerequisites and scope, then begin with WebAssembly Execution Model. Keep Always Active attached to every phase.